Quasi-Likelihood And Its Application: A General Approach to Optimal Parameter Estimation Christopher C. Heyde Springer Preface This book is concerned with the general theory of optimal estimation of parameters in systems subject to random effects and with the application of this theory. The focus is on choice of families of estimating functions, rather than the estimators derived therefrom, and on optimization within these families. Only assumptions about means and covariances are required for an initial discussion. Nevertheless, the theory that is developed mimics that of maximum likelihood, at least to the first order of asymptotics. The term quasi-likelihood has often had a narrow interpretation, associated with its application to generalized linear model type contexts, while that of optimal estimating functions has embraced a broader concept. There is, however, no essential distinction between the underlying ideas and the term quasi-likelihood has herein been adopted as the general label. This emphasizes its role in extension of likelihood based theory. The idea throughout involves finding quasi-scores from families of estimating functions. Then, the quasilikelihood estimator is derived from the quasi-score by equating to zero and solving, just as the maximum likelihood estimator is derived from the likelihood score. This book had its origins in a set of lectures given in September 1991 at the 7th Summer School on Probability and Mathematical Statistics held in Varna, Bulgaria, the notes of which were published as Heyde (1993). Subsets of the material were also covered in advanced graduate courses at Columbia University in the Fall Semesters of 1992 and 1996. The work originally had a quite strong emphasis on inference for stochastic processes but the focus gradually broadened over time. Discussions with V.P. Godambe and with R. Morton have been particularly influential in helping to form my views. The subject of estimating functions has evolved quite rapidly over the period during which the book was written and important developments have been emerging so fast as to preclude any attempt at exhaustive coverage. Among the topics omitted is that of quasi- likelihood in survey sampling, which has generated quite an extensive literature (see the edited volume Godambe (1991), Part 4 and references therein) and also the emergent linkage with Bayesian statistics (e.g., Godambe (1994)). It became quite evident at the Conference on Estimating Functions held at the University of Georgia in March 1996 that a book in the area was much needed as many known ideas were being rediscovered. This realization provided the impetus to round off the project rather vi PREFACE earlier than would otherwise have been the case. The emphasis in the monograph is on concepts rather than on mathematical theory. Indeed, formalities have been suppressed to avoid obscuring "typical" results with the phalanx of regularity conditions and qualifiers necessary to avoid the usual uninformative types of counterexamples which detract from most statistical paradigms. In discussing theory which holds to the first order of asymptotics the treatment is especially informal, as befits the context. Sufficient conditions which ensure the behaviour described are not difficult to furnish but are fundamentally uninlightening. A collection of complements and exercises has been included to make the material more useful in a teaching environment and the book should be suitable for advanced courses and seminars. Prerequisites are sound basic courses in measure theoretic probability and in statistical inference. Comments and advice from students and other colleagues has also contributed much to the final form of the book. In addition to V.P. Godambe and R. Morton mentioned above, grateful thanks are due in particular to Y.-X. Lin, A. Thavaneswaran, I.V. Basawa, E. Saavendra and T. Zajic for suggesting corrections and other improvements and to my wife Beth for her encouragement. C.C. Heyde Canberra, Australia February 1997 Contents Preface v 1 Introduction 1 1.1 The Brief . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.3 The Gauss-Markov Theorem . . . . . . . . . . . . . . . . . . . 3 1.4 Relationship with the Score Function . . . . . . . . . . . . . . . 6 1.5 The Road Ahead . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.6 The Message of the Book . . . . . . . . . . . . . . . . . . . . . 10 1.7 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2 The General Framework 11 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.2 Fixed Sample Criteria . . . . . . . . . . . . . . . . . . . . . . . 11 2.3 Scalar Equivalences and Associated Results . . . . . . . . . . . 19 2.4 Wedderburn's Quasi-Likelihood . . . . . . . . . . . . . . . . . . 21 2.4.1 The Framework . . . . . . . . . . . . . . . . . . . . . . . 21 2.4.2 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.4.3 Generalized Estimating Equations . . . . . . . . . . . . 25 2.5 Asymptotic Criteria . . . . . . . . . . . . . . . . . . . . . . . . 26 2.6 A Semimartingale Model for Applications . . . . . . . . . . . . 30 2.7 Some Problem Cases for the Methodology . . . . . . . . . . . . 35 2.8 Complements and Exercises . . . . . . . . . . . . . . . . . . . . 38 3 An Alternative Approach: E-Sufficiency 43 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 3.2 Definitions and Notation . . . . . . . . . . . . . . . . . . . . . . 43 3.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.4 Complement and Exercise . . . . . . . . . . . . . . . . . . . . . 51 4 Asymptotic Confidence Zones of Minimum Size 53 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 4.2 The Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . 54 4.3 Confidence Zones: Theory . . . . . . . . . . . . . . . . . . . . . 56 vii viii CONTENTS 4.4 Confidence Zones: Practice . . . . . . . . . . . . . . . . . . . . 60 4.5 On Best Asymptotic Confidence Intervals . . . . . . . . . . . . 62 4.5.1 Introduction and Results . . . . . . . . . . . . . . . . . 62 4.5.2 Proof of Theorem 4.1 . . . . . . . . . . . . . . . . . . . 64 4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67 5 Asymptotic Quasi-Likelihood 69 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 5.2 The Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . 71 5.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 5.3.1 Generalized Linear Model . . . . . . . . . . . . . . . . . 79 5.3.2 Heteroscedastic Autoregressive Model . . . . . . . . . . 79 5.3.3 Whittle Estimation Procedure . . . . . . . . . . . . . . 82 5.3.4 Addendum to the Example of Section 5.1 . . . . . . . . 87 5.4 Bibliographic Notes . . . . . . . . . . . . . . . . . . . . . . . . . 88 5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88 6 Combining Estimating Functions 91 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 6.2 Composite Quasi-Likelihoods . . . . . . . . . . . . . . . . . . . 92 6.3 Combining Martingale Estimating Functions . . . . . . . . . . 93 6.3.1 An Example . . . . . . . . . . . . . . . . . . . . . . . . 98 6.4 Application. Nested Strata of Variation . . . . . . . . . . . . . 99 6.5 State-Estimation in Time Series . . . . . . . . . . . . . . . . . . 103 6.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104 7 Projected Quasi-Likelihood 107 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 7.2 Constrained Parameter Estimation . . . . . . . . . . . . . . . . 107 7.2.1 Main Results . . . . . . . . . . . . . . . . . . . . . . . . 109 7.2.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . 111 7.2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . 112 7.3 Nuisance Parameters . . . . . . . . . . . . . . . . . . . . . . . . 113 7.4 Generalizing the E-M Algorithm: The P-S Method . . . . . . . 116 7.4.1 From Log-Likelihood to Score Function . . . . . . . . . 117 7.4.2 From Score to Quasi-Score . . . . . . . . . . . . . . . . 118 7.4.3 Key Applications . . . . . . . . . . . . . . . . . . . . . . 121 7.4.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . 122 7.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 8 Bypassing the Likelihood 129 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129 8.2 The REML Estimating Equations . . . . . . . . . . . . . . . . . 129 8.3 Parameters in Diffusion Type Processes . . . . . . . . . . . . . 131 8.4 Estimation in Hidden Markov Random Fields . . . . . . . . . . 136 8.5 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 CONTENTS ix 9 Hypothesis Testing 141 9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 9.2 The Details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 9.3 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 10 Infinite Dimensional Problems 147 10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 10.2 Sieves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 10.3 Semimartingale Models . . . . . . . . . . . . . . . . . . . . . . 148 11 Miscellaneous Applications 153 11.1 Estimating the Mean of a Stationary Process . . . . . . . . . . 153 11.2 Estimation for a Heteroscedastic Regression . . . . . . . . . . . 159 11.3 Estimating the Infection Rate in an Epidemic . . . . . . . . . . 162 11.4 Estimating Population Size . . . . . . . . . . . . . . . . . . . . 164 11.5 Robust Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 169 11.5.1 Optimal Robust Estimating Functions . . . . . . . . . . 170 11.5.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . 173 11.6 Recursive Estimation . . . . . . . . . . . . . . . . . . . . . . . . 176 12 Consistency and Asymptotic Normality for Estimating Functions 179 12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179 12.2 Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180 12.3 The SLLN for Martingales . . . . . . . . . . . . . . . . . . . . . 186 12.4 The CLT for Martingales . . . . . . . . . . . . . . . . . . . . . 190 12.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 13 Complements and Strategies for Application 199 13.1 Some Useful Families of Estimating Functions . . . . . . . . . . 199 13.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 199 13.1.2 Transform Martingale Families . . . . . . . . . . . . . . 199 13.1.3 Use of the Infinitesimal Generator of a Markov Process 200 13.2 Solution of Estimating Equations . . . . . . . . . . . . . . . . . 201 13.3 Multiple Roots . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 13.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 202 13.3.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . 204 13.3.3 Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 208 13.4 Resampling Methods . . . . . . . . . . . . . . . . . . . . . . . . 210 References 211 Index 227 Chapter 1 Introduction 1.1 The Brief This monograph is primarily concerned with parameter estimation for a random process {Xt} taking values in r-dimensional Euclidean space. The distribution of Xt depends on a characteristic taking values in a open subset of p-dimensional Euclidean space. The framework may be parametric or semiparametric; may be, for example, the mean of a stationary process. The object will be the "efficient" estimation of based on a sample {Xt, t T}. 1.2 Preliminaries Historically there are two principal themes in statistical parameter estimation theory: least squares (LS) - introduced by Gauss and Legendre and founded on finite sample considerations (minimum distance interpretation) maximum likelihood (ML) - introduced by Fisher and with a justification that is primarily asymptotic (minimum size asymptotic confidence intervals, ideas of which date back to Laplace) It is now possible to unify these approaches under the general description of quasi-likelihood and to develop the theory of parameter estimation in a very general setting. The fixed sample optimality ideas that underly quasi-likelihood date back to Godambe (1960) and Durbin (1960) and were put into a stochastic process setting in Godambe (1985). The asymptotic justification is due to Heyde (1986). The ideas were combined in Godambe and Heyde (1987). It turns out that the theory needs to be developed in terms of estimating functions (functions of both the data and the parameter) rather than the estimators themselves. Thus, our focus will be on functions that have the value of the parameter as a root rather than the parameter itself. The use of estimating functions dates back at least to K. Pearson's introduction of the method of moments (1894) although the term "estimating function" may have been coined by Kimball (1946). Furthermore, all the standard methods of estimation, such as maximum likelihood, least-squares, conditional least-squares, minimum chi-squared, and M-estimation, are included under minor regularity conditions. The subject has now developed to the stage where books are being devoted to it, e.g., Godambe (1991), McLeish and Small (1988). 1 2 CHAPTER 1. INTRODUCTION The rationale for the use of the estimating function rather than the estimator derived therefrom lies in its more fundamental character. The following dot points illustrate the principle. * Estimating functions have the property of invariance under one-to-one transformations of the parameter . * Under minor regularity conditions the score function (derivative of the log-likelihood with respect to the parameter), which is an estimating function, provides a minimal sufficient partitioning of the sample space. However, there is often no single sufficient statistic. For example, suppose that {Zt} is a Galton-Watson process with offspring mean E(Z1 | Z0 = 1) = . Suppose that the offspring distribution belongs to the power series family (which is the discrete exponential family). Then, the score function is UT () = c T t=1 (Zt - Zt-1), where c is a constant and the maximum likelihood estimator ^T = T t=1 Zt T t=1 Zt-1 is not a sufficient statistic. Details are given in Chapter 2. * Fisher's information is an estimating function property (namely, the variance of the score function) rather than that of the maximum likelihood estimator (MLE). * The Cram´er-Rao inequality is an estimating function property rather than a property of estimators. It gives the variance of the score function as a bound on the variances of standardized estimating functions. * The asymptotic properties of an estimator are almost invariably obtained, as in the case of the MLE, via the asymptotics of the estimating function and then transferred to the parameter space via local linearity. * Separate estimating functions, each with information to offer about an unknown parameter, can be combined much more readily than the estimators therefrom. We shall begin our discussion by examining the minimum variance ideas that underly least squares and then see how optimality is conveniently phrased in terms of estimating functions. Subsequently, we shall show how the score function and maximum likelihood ideas mesh with this. The approach is along the general lines of the brief overviews that appear in Godambe and Heyde (1987), Heyde (1989b), Desmond (1991), Godambe and Kale (1991). An earlier version 1.3. THE GAUSS-MARKOV THEOREM 3 appeared in the lecture notes Heyde (1993). Another approach to the subject of optimal estimation, which also uses estimating functions but is based on extension of the idea of sufficiency, appears in McLeish and Small (1988); the theories do substantially overlap, although this is not immediately transparent. Details are provided in Chapter 3. 1.3 Estimating Functions and the Gauss-Markov Theorem To indicate the basic LS ideas that we wish to incorporate, we consider the simplest case of independent random variables (rv's) and a one-dimensional parameter . Suppose that X1, . . . , XT are independent rv's with EXt = , var Xt = 2 . In this context the Gauss-Markov theorem has the following form. GM Theorem: Let the estimator ST = T t=1 at Xt be unbiased for , the at being constants. Then, the variance, var ST , is minimized for at = 1/T, t = 1, . . . , T. That is, the sample mean X = T-1 T t=1 Xt is the linear unbiased minimum variance estimator of . The proof is very simple; we have to minimize var ST = 2 T t=1 a2 t subject to T t=1 at = 1 and var ST = 2 T t=1 a2 t - 2at T + 1 T2 + 2 T = 2 T t=1 at - 1 T 2 + 2 T 2 T . Now we can restate the GM theorem in terms of estimating functions. Consider the set G0 of unbiased estimating functions G = G(X1, . . . , XT , ) of the form G = T t=1 bt(Xt - ), the bt's being constants with T t=1 bt = 0. Note that the estimating functions kG, k constant and G produce the same estimator, namely T t=1 bt Xt/ T t=1 bt, so some standardization is necessary if variances are to be compared. One possible standardization is to define the standardized version of G as G(s) = T t=1 bt T t=1 bt (Xt - ) 2 T t=1 b2 t -1 . The estimator of is unchanged and, of course, kG and G have the same standardized form. Let us now motivate this standardization. (1) In order to be used as an estimating equation, the estimating function G 4 CHAPTER 1. INTRODUCTION needs to be as close to zero as possible when is the true value. Thus we want var G = 2 T t=1 b2 t to be as small as possible. On the other hand, we want G( + ), > 0, to differ as much as possible from G() when is the true value. That is, we want (E ˙G())2 = T t=1 bt 2 , the dot denoting derivative with respect to , to be as large as possible. These requirements can be combined by maximizing var G(s) = (E ˙G)2 /EG2 . (2) Also, if max1tT bt/ T t=1 bt 0 as T , then T t=1 bt (Xt - ) 2 T t=1 b2 t 1/2 d - N(0, 1) using the Lindeberg-Feller central limit theorem. Thus, noting that our estimator for is ^T = T t=1 bt Xt T t=1 bt, we have var G (s) T 1/2 (^T - ) d - N(0, 1), i.e., ^T - d N 0, var G (s) T -1 . We would wish to choose the best asymptotic confidence intervals for and hence to maximize var G (s) T . (3) For the standardized version G(s) of G we have var G(s) = T t=1 bt 2 2 T t=1 b2 t = -E ˙G(s) , i.e., G(s) possesses the standard likelihood score property. Having introduced standardization we can say that G G0 is an optimal estimating function within G0 if var G(s) var G(s) , G G0. This leads to the following result. GM Reformation The estimating function G = T t=1(Xt -) is an optimal estimating function within G0. The estimating equation G = 0 provides the sample mean as an optimal estimator of . 1.3. THE GAUSS-MARKOV THEOREM 5 The proof follows immediately from the Cauchy-Schwarz inequality. For G G0 we have var G(s) = T t=1 bt 2 2 T t=1 b2 t T/2 = var G(s) and the argument holds even if the bt's are functions of . Now the formulation that we adapted can be extended to estimating functions G in general by defining the standardized version of G as G(s) = -(E ˙G) (EG2 )-1 G. Optimality based on maximization of var G(s) leads us to define G to be optimal within a class H if var G(s) var G(s) , G H. That this concept does differ from least squares in some important respects is illustrated in the following example. We now suppose that Xt, t = 1, 2, . . . , T are independent rv's with EXt = t(), var Xt = 2 t (), the t's, 2 t 's being specified differentiable functions. Then, for the class of estimating functions H = H : H = T t=1 bt() (Xt - t()) , we have var H(s) = T t=1 bt() ˙t() 2 T t=1 b2 t () 2 t (), which is maximized (again using the Cauchy-Schwarz inequality) if bt() = k() ˙t() -2 t (), t = 1, 2, . . . , T, k() being an undetermined multiplier. Thus, an optimal estimating function is H = T t=1 ˙t() -2 t () (Xt - t()). Note that this result is not what one gets from least squares (LS). If we applied LS, we would minimize T t=1 (Xt - t())2 -2 t (), which leads to the estimating equation T t=1 ˙t() -2 t () (Xt - t()) + T t=1 (Xt - t())2 -3 t () ˙t() = 0. 6 CHAPTER 1. INTRODUCTION This estimating equation will generally not be unbiased, and it may behave very badly depending on the t's. It will not in general provide a consistent estimator. 1.4 Relationship with the Score Function Now suppose that {Xt, t = 1, 2, . . . , T} has likelihood function L = T t=1 ft(Xt; ). The score function in this case is a sum of independent rv's with zero means, U = log L = T t=1 log ft(Xt; ) and, when H = T t=1 bt() (Xt - t()), we have E(U H) = T t=1 bt() E log ft(Xt; ) (Xt - t()) . If the ft's are such that integration and differentiation can be interchanged, E log ft(Xt; ) Xt = EXt = ˙t(), so that E(U H) = T t=1 bt() ˙t() = -E ˙H. Also, using corr to denote correlation, corr2 (U, H) = (E(U H))2 /(EU2 )(EH2 ) = (var H(s) )/EU2 , which is maximized if var H(s) is maximized. That is, the choice of an optimal estimating function H H is giving an element of H that has maximum correlation with the generally unknown score function. Next, for the score function U and H H we find that E(H(s) - U(s) )2 = var H(s) + var U(s) - 2E(H(s) U(s) ) = EU2 - var H(s) , since U(s) = U 1.5. THE ROAD AHEAD 7 and EH(s) U(s) = var H(s) when differentiation and integration can be interchanged. Thus E(H(s) -U(s) )2 is minimized when an optimal estimating function H H is chosen. This gives an optimal estimating function the interpretation of having minimum expected distance from the score function. Note also that var H(s) EU2 , which is the Cram´er-Rao inequality. Of course, if the score function U H, the methodology picks out U as optimal. In the case in question U H if and only if U is of the form U = T t=1 bt() (Xt - t()), that is, log f(Xt; ) = bt() (Xt - t()), so that the Xt's are from an exponential family in linear form. Classical quasi-likelihood was introduced in the setting discussed above by Wedderburn (1974). It was noted by Bradley (1973) and Wedderburn (1974) that if the Xt's have exponential family distributions in which the canonical statistics are linear in the data, then the score function depends on the parameters only through the means and variances. They also noted that the score function could be written as a weighted least squares estimating function. Wedderburn suggested using the exponential family score function even when the underlying distribution was unspecified. In such a case the estimating function was called a quasi-score estimating function and the estimator derived therefore a quasi-likelihood estimator. The concept of optimal estimating functions discussed above conveniently subsumes that of quasi-score estimating functions in the Wedderburn sense, as we shall discuss in vector form in Chapter 2. We shall, however, in our general theory, take the names quasi-score and optimal for estimating functions to be essentially synonymous. 1.5 The Road Ahead In the above discussion we have concentrated on the simplest case of independent random variables and a scalar parameter, but the basis of a general formulation of the quasi-likelihood methodology is already evident. In Chapter 2, quasi-likelihood is developed in its general framework of a (finite dimensional) vector valued parameter to be estimated from vector valued data. Quasi-likelihood estimators are derived from quasi-score estimating functions whose selection involves maximization of a matrix valued information criterion in the partial order of non-negative definite matrices. Both fixed 8 CHAPTER 1. INTRODUCTION sample and asymptotic formulations are considered and the conditions under which they hold are shown to be substantially overlapping. Also, since matrix valued criteria are not always easy to work with, some scalar equivalences are formulated. Here there is a strong link with the theory of optimal experimental design. The original Wedderburn formulation of quasi-likelihood in an exponential family setting is then described together with the limitations of its direct extension. Also treated is the closely related methodology of generalized estimating equations, developed for longitudinal data sets and typically using approximate covariance matrices in the quasi-score estimating function. The basic formulation having been provided, it is now shown how a semimartingale model leads to a convenient class of estimating functions of wide applicability. Various illustrations are provided showing how to use these ideas in practice, and some discussion of problem cases is also given. Chapter 3 outlines an alternative approach to optimal estimation using estimating functions via the concepts of E-sufficiency and E-ancillarity. Here E refers to expectation. This approach, due to McLeish and Small, produces results that overlap substantially with those of quasi-likelihood, although this is not immediately apparent. The view is taken in this book that quasi-likelihood methodology is more transparent and easier to apply. Chapter 4 is concerned with asymptotic confidence zones. Under the usual sort of regularity conditions, quasi-likelihood estimators are associated with minimum size asymptotic confidence intervals within their prespecified spaces of estimating functions. Attention is given to the subtle question of whether to normalize with random variables or constants in order to obtain the smallest intervals. Random normings have some important advantages. Ordinary quasi-likelihood theory is concerned with the case where the maximum information criterion holds exactly for fixed T or for each T as T . Chapter 5 deals with the case where optimality holds only in a certain asymptotic sense. This may happen, for example, when a nuisance parameter is replaced by a consistent estimator thereof. The discussion focuses on situations where the properties of regular quasi-likelihood of consistency and possession of minimum size asymptotic confidence zones are preserved for the estimator. Estimating functions from different sources can conveniently be added, and the issue of their optimal combination is addressed in Chapter 6. Various applications are given, including dealing with combinations of estimating functions where there are nested strata of variation and providing methods of filtering and smoothing in time series estimation. The well-known Kalman filter is a special case. Chapter 7 deals with projection methods that are useful in situations where a standard application of quasi-likelihood is precluded. Quasi-likelihood approaches are provided for constrained parameter estimation, for estimation in the presence of nuisance parameters, and for generalizing the E-M algorithm for estimation where there are missing data. In Chapter 8 the focus is on deriving the score function, or more generally quasi-score estimating function, without use of the likelihood, which may be 1.5. THE ROAD AHEAD 9 difficult to deal with, or fail to exist, under minor perturbations of standard conditions. Simple quasi-likelihood derivations of the score functions are provided for estimating the parameters in the covariance matrix, where the distribution is multivariate normal (REML estimation), in diffusion type models, and in hidden Markov random fields. In each case these remain valid as quasi-score estimating functions under significantly broadened assumptions over those of a likelihood based approach. Chapter 9 deals briefly with issues of hypothesis testing. Generalizations of the classical efficient scores statistic and Wald test statistic are treated. These are shown to usually be asymptotically 2 distributed under the null hypothesis and to have asymptotically, noncentral 2 distributions, with maximum noncentrality parameter, under the alternative hypothesis, when the quasi-score estimating function is used. Chapter 10 provides a brief discussion of infinite dimensional parameter (function) estimation. A sketch is given of the method of sieves, in which the dimension of the parameter is increased as the sample size increases. An informal treatment of estimation in linear semimartingale models, such as occur for counting processes and estimation of the cumulative hazard function, is also provided. A diverse collection of applications is given in Chapter 11. Estimation is discussed for the mean of a stationary process, a heteroscedastic regression, the infection rate of an epidemic, and a population size via a multiple recapture experiment. Also treated are estimation via robustified estimating functions (possibly with components that are bounded functions of the data) and recursive estimation (for example, for on-line signal processing). Chapter 12 treats the issues of consistency and asymptotic normality of estimators. Throughout the book it is usually expected that these will ordinarily hold under appropriate regularity conditions. The focus here is on martingale based methods, and general forms of martingale strong law and central limit theorems are provided for use in particular cases. The view is taken that it is mostly preferable directly to check cases individually rather than to rely on general theory with its multiplicity of regularity conditions. Finally, in Chapter 13 a number of complementary issues involved in the use of quasi-likelihood methods are discussed. The chapter begins with a collection of methods for generating useful families of estimating functions. Integral transform families and the use of the infinitesimal generator of a Markov process are treated. Then, the numerical solution of estimating equations is considered, and methods are examined for dealing with multiple roots when a scalar objective function may not be available. The final section is concerned with resampling methods for the provision of confidence intervals, in particular the jackknife and bootstrap. 10 CHAPTER 1. INTRODUCTION 1.6 The Message of the Book For estimation of parameters, in stochastic systems of any kind, it has become increasingly clear that it is possible to replace likelihood based techniques by quasi-likelihood alternatives, in which only assumptions about means and variances are made, in order to obtain estimators. There is often little, if any, loss in efficiency, and all the advantages of weighted least squares methods are also incorporated. Additional assumptions are, of course, required to ensure consistency of estimators and to provide confidence intervals. If it is available, the likelihood approach does provide a basis for benchmarking of estimating functions but not more than that. It is conjectured that everything that can be done via likelihoods has a corresponding quasi-likelihood generalization. 1.7 Exercise 1. Suppose {Xi, i = 1, 2, . . .} is a sequence of independent rv's, Xi having a Bernoulli distribution with P(Xi = 1) = pi = 1 2 + ai, P(Xi = 0) = 1 - pi, and 0 < ai 0 as i . Show that there is a consistent estimator of if and only if i=1 a2 i = . (Adaped from Dion and Ferland (1995).) Chapter 2 The General Framework 2.1 Introduction Let {Xt, t T} be a sample of discrete or continuous data that is randomly generated and takes values in r-dimensional Euclidean space. The distribution of Xt depends on a "parameter" taking values in an open subset of pdimensional Euclidean space and the object of the exercise is the estimation of . We assume that the possible probability measures for Xt are {P} a union (possibly uncountable) of families of parametric models, each family being indexed by and that each (, F, P) is a complete probability space. We shall focus attention on the class G of zero mean, square integrable estimating functions GT = GT ({Xt, t T}, ), which are vectors of dimension p for which EGT () = 0 for each P and for which the p-dimensional matrices E ˙GT = (E GT,i()/j) and EGT GT are nonsingular, the prime denoting transpose. The expectations are always with respect to P. Note that ˙G is the transpose of the usual derivative of G with respect to . In many cases P is absolutely continuous with respect to some -finite measure T giving a density pT (). Then we write UT () = p-1 T () ˙pT () for the score function, which we suppose to be almost surely differentiable with respect to the components of . In addition we will also suppose that differentiation and integration can be interchanged in E(GT UT ) and E(UT GT ) for GT G. The score function UT provides, modulo minor regularity conditions, a minimal sufficient partitioning of the sample space and hence should be used for estimation if it is available. However, it is often unknown, or in semiparametric cases, does not exist. The framework here allows a focus on models in which the error distribution has only its first and second moment properties specified, at least initially. 2.2 Fixed Sample Criteria In practice we always work with specified subsets of G. Take H G as such a set. As motivated in the previous chapter, optimality within H is achieved by maximizing the covariance matrix of the standardized estimating functions G (s) T = -(E ˙GT ) (EGT GT )-1 GT , GT H. Alternatively, if UT exists, an optimal estimating function within H is one with minimum dispersion distance from UT . These ideas are formalized in the following definition and equivalence, which we shall call criteria for OF -optimality (fixed sample optimality). Later 11 12 CHAPTER 2. THE GENERAL FRAMEWORK we shall introduce similar criteria for optimality to hold for all (sufficiently large) sample sizes. Estimating functions that are optimal in either sense will be referred to as quasi-score estimating functions and the estimators that come from equating these to zero and solving as quasi-likelihood estimators. OF -optimality involves choice of the estimating function GT to maximize, in the partial order of nonnegative definite (nnd) matrices (sometimes known as the Loewner ordering), the information criterion E(GT ) = E(G (s) T G (s) T ) = (E ˙GT ) (EGT GT )-1 (E ˙GT ), which is a natural generalization of Fisher information. Indeed, if the score function UT exists, E(UT ) = (E ˙UT ) (EUT UT )-1 (E ˙UT ) = EUT UT is the Fisher information. Definition 2.1 G T H is an OF -optimal estimating function within H if E(G T ) - E(GT ) (2.1) is nonnegative definite for all GT H, and P. The term Loewner optimality is used for this concept in the theory of optimal experimental designs (e.g., Pukelsheim (1993, Chapter 4)). In the case where the score function exists there is the following equivalent form to Definition 2.1 phrased in terms of minimizing dispersion distance. Definition 2.2 G T H is an OF -optimal estimating function within H if E U (s) T - G (s) T U (s) T - G (s) T - E U (s) T - G (s) T U (s) T - G (s) T (2.2) is nonnegative definite for all GT H, and P. Proof of Equivalence We drop the subscript T for convenience. Note that E G(s) U(s) = -(E ˙G) (EGG )-1 EGU = E G(s) G(s) since EGU = -E ˙G G H EGU = G log L L = G L = - G L 2.2. FIXED SAMPLE CRITERIA 13 and similarly E U(s) G(s) = E G(s) G(s) . These results lead immediately to the equality of the expressions (2.1) and (2.2) and hence the equivalence of Definition 2.1 and Definition 2.2. A further useful interpretation of quasi-likelihood can be given in a Hilbert space setting. Let H be a closed subspace of L2 = L2 (, F, P0) of (equivalence classes) of random vectors with finite second moment. Then, for X, Y L2 , taking inner product (X, Y ) = E(X Y ) and norm X = (X, X)1/2 the space L2 is a Hilbert space. We say that X is orthogonal to Y , written XY , if (X, Y ) = 0 and that subsets L2 1 and L2 2 of L2 are orthogonal, which holds if XY for every X L2 1, Y L2 2 (written L2 1L2 2). For X L2 , let (X | H) denote the element of H such that X - (X | H) 2 = inf Y H X - Y 2 , that is, (X | H) is the orthogonal projection of X onto H. Now suppose that the score function UT G. Then, dropping the subscript T and using Definition 2.2, the standardized quasi-score estimating function H(s) H is given by inf H(s)H E U - H(s) U - H(s) , and since tr E(U - H(s) )(U - H(s) ) = U - H(s) 2 , tr denoting trace, the quasi-score is (U | H), the orthogonal projection of the score function onto the chosen space H of estimating functions. For further discusion of the Hilbert space approach see Small and McLeish (1994) and Merkouris (1992). Next, the vector correlation that measures the association between GT = (GT,1, . . . , GT,p) and UT = (UT,1, . . . , UT,p) , defined, for example, by Hotelling (1936), is 2 = (det(EGT UT ))2 det(EGT GT ) det(EUT UT ) , where det denotes determinant. However, under the regularity conditions that have been imposed, E ˙GT = -E(GT UT ), so a maximum correlation requirement is to maximize (det(E ˙GT ))2 /det(EGT GT ), which can be achieved by maximizing E(GT ) in the partial order of nonnegative definite matrices. This corresponds to the criterion of Definition 2.1. Neither Definition 2.1 nor Definition 2.2 is of direct practical value for applications. There is, however, an essentially equivalent form (Heyde (1988a)), 14 CHAPTER 2. THE GENERAL FRAMEWORK that is very easy to use in practice. Theorem 2.1 G T H is an OF -optimal estimating function within H if E G (s) T G (s) T = E G (s) T G (s) T = E G (s) T G (s) T (2.3) or equivalently E ˙GT -1 EGT G T is a constant matrix for all GT H. Conversely, if H is convex and G T H is an OF -optimal estimating function, then (2.3) holds. Proof. Again we drop the subscript T for convenience. When (2.3) holds, E G(s) - G(s) G(s) - G(s) = E G(s) G(s) - E G(s) G(s) is nonnegative definite, G H, since the left-hand side is a covariance function. This gives optimality via Definition 2.1. Now suppose that H is convex and G is an OF -optimal estimating function. Then, if H = G + G , we have that E G(s) G(s) - E H(s) H(s) is nonnegative definite, and after inverting and some algebra this gives that 2 EGG - E ˙G E ˙G -1 EG G E ˙G -1 E ˙G - -EGG + E ˙G E ˙G -1 EG G - -EG G + EG G E ˙G -1 E ˙G is nonnegative definite. This is of the form 2 A - B, where A and B are symmetric and A is nonnegative definite by Definition 2.1. Let u be an arbitrary nonzero vector of dimension p. We have u Au 0 and u Au -1 u Bu for all , which forces u Bu = 0 and hence B = 0. Now B = 0 can be rewritten as EGG E ˙G -1 C + C E ˙G -1 EGG = 0, where C = EG(s) G(s) - EG(s) G(s) E ˙G -1 EG G 2.2. FIXED SAMPLE CRITERIA 15 and, as this holds for all G H, it is possible to replace G by DG, where D = diag (1, . . . , p) is an arbitrary constant matrix. Then, in obvious notation i (EGG ) (E ˙G) -1 C j + C (E ˙G)-1 (EGG ) i j = 0 for each i, j, which forces C = 0 and hence (2.3) holds. This completes the proof. In general, Theorem 2.1 provides a straightforward way to check whether an OF -optimal estimating function exists for a particular family H. It should be noted that existence is by no means guaranteed. Theorem 2.1 is especially easy to use when the elements G H have orthogonal differences and indeed this is often the case in applications. Suppose, for example, that H = H : H = T t=1 at() ht() , with at() constants to be chosen, ht's fixed and random with zero means and Ehs()ht() = 0, s = t. Then EHH = T t=1 atEhthta t E ˙H = T t=1 atE ˙ht and (E ˙H)-1 EHH is constant for all H H if a t = E ˙ht (Ehtht)-1 . An OF -optimal estimating function is thus T t=1 E ˙ht() Eht()ht() -1 ht(). As an illustration consider the estimation of the mean of the offspring distribution in a Galton-Watson process {Zt}, = E(Z1|Z0 = 1). Here the data are {Z0, . . . , ZT }. Let Fn = (Z0, . . . , Zn). We seek a basic martingale (MG) from the {Zi}. This is simple since Zi - E (Zi | Fi-1) = Zi - Zi-1 16 CHAPTER 2. THE GENERAL FRAMEWORK are MG differences (and hence orthogonal). Let H = h : hT = T t=1 at()(Zt - Zt-1), at() is Ft-1 measurable . We find that the OF -optimal choice for at() is a t () = -1/2 , where 2 = var(Z1|Z0 = 1). The OF -optimal estimator of is (Z1 + . . . + ZT )/(Z0 + . . . + ZT -1). We would call this a quasi-likelihood estimator from the family H. It is actually the MLE for the power series family of offspring distributions P(Z1 = j|Z0 = 1) = A(j) (a())j F() , j = 0, 1, 2, . . . where F() = j=0 A(j)(a())j . These form the discrete exponential family for this context. To obtain the MLE result for the power series family note that P(Z0 = z0, . . . , ZT = zT ) = P(Z0 = z0) T k=1 P Zk = zk Zk-1 = zk-1 = P(Z0 = z0) T k=1 j1+...+jzk-1 =zk A(j1) . . . A(jzk-1 ) (a())zk (F())zk-1 = P(Z0 = z0) (a())z1+...+zT (F())z0+...+zT -1 × term not involving and hence if L = L(Z0, . . . , ZT ; ) is the likelihood, d log L d = (Z1 + . . . + ZT ) ˙a() a() - (Z0 + . . . + ZT -1) ˙F() F() . However = j=1 j A(j) (a())j F() , 1 = j=0 A(j) (a())j F() 2.2. FIXED SAMPLE CRITERIA 17 and differentiating with respect to in the latter result, 0 = j=1 j A(j)˙a() (a())j-1 F() - ˙F() F2() j=0 A(j)(a())j so that ˙a() a() = ˙F() F() and the score function is UT () = ˙a() a() [(Z1 + . . . + ZT ) - (Z0 + . . . + ZT -1)] . The same estimator can also be obtained quite generally as a nonparametric MLE (Feigin (1977)). This example illustrates one important general strategy for finding optimal estimating functions. This strategy is to compute the score function for some plausible underlying distribution (such as a convenient member of the appropriate exponential family) log L0/, say, then use the differences in this martingale to form ht's. Finally, choose the optimal estimating function within the class H = { T t=1 atht} by suitably specifying the weights at. The previous discussion and examples have concentrated on the case of discrete time. However, the theory operates in entirely similar fashion in the case of continuous time. For example, consider the diffusion process dXt = Xt dt + dWt, where Wt is standard Brownian motion. We wish to estimate on the basis of the data {Xt, 0 t T}. Here the natural martingale to use is Wt, and we seek an OF -optimal estimating function from the set H = { T 0 bs dWs, bs predictable}. Note that, for convenience, and to emphasize the methodology, we shall often write estimating functions in a form that emphasizes the noise component of the model and suppresses the dependence on the observations. Here T 0 bs dWs is to be interpreted as T 0 bs(dXs - Xs ds). We have, writing HT = T 0 bs dWs, H T = T 0 b s dWs, HT = T 0 bs dWs = T 0 bs dXs - T 0 bs Xs ds, so that E ˙HT = -E T 0 bs Xs ds, EHT H T = E T 0 bs b s ds 18 CHAPTER 2. THE GENERAL FRAMEWORK and (E ˙HT )-1 EHT H T is constant for all H H if b s = Xs. Then, H T = T 0 Xs dXs - T 0 X2 s ds and the QLE is ^T = T 0 Xs dXs T 0 X2 s ds. This is also the MLE. As another example, we consider the multivariate counting process Xt = (Xt,1, . . . , Xt,p) , each Xt,i being of the form Xt,i = i t 0 Ji(s) ds + Mt,i with multiplicative intensity i(t) = i Ji(t), Ji(t) > 0 a.s. being predictable and Mt,i a square integrable martingale. This is a special case of the framework considered by Aalen (1978), see also Andersen et.al. (1982) and it covers a variety of contexts for processes such as those of birth and death type. The case p = 1 has been discussed by Thavaneswaran and Thompson (1986). The data are {Xt, 0 t T} and we write XT = T 0 J(s) ds + MT , where J(s) = diag(J1(s), . . . , Jp(s)), = (1, . . . , p) , MT = (MT,1, . . . , MT,p) and we note that for counting processes MT,i and MT,j are orthogonal for i = j. Then, we seek an OF -optimal estimating function from the set H = T 0 bs dMs, bs predictable . Let HT = T 0 bs dMs, H T = T 0 b s dMs. We have E ˙HT = -E T 0 bs J(s) ds, EHT H T = E T 0 bs J(s) b s ds, and (E ˙HT )-1 EHT H T is constant for all H H if b s I. Thus H T = MT is an OF -optimal estimating function and the corresponding QLE is ^T = T 0 Js ds -1 XT . 2.3. SCALAR EQUIVALENCES AND ASSOCIATED RESULTS 19 That this ^ is also the MLE under rather general conditions follows from § 3.3 of Aalen (1978). The simplest particular case is where each Xt,i is a Poisson process with parameter i. General comment: The art, as distinct from the science, in using quasilikelihood methods is in a good choice of the family H of estimating functions with which to work. The ability to choose H is a considerable strength as the family can be tailor made to the requirements of the context, and regularity conditions can be built in. However, it is also a source of weakness, since it is by no means always clear what competing families of estimating functions might exist with better properties and efficiencies. Of course, the quasi-likelihood framework does provide a convenient basis for comparison of families via the information criterion. For examples of estimators for autoregressive processes with positive or bounded innovations that are much better than the naive QLE chosen from the natural family of estimating functions for standard autoregressions see Davis and McCormick (1989). 2.3 Scalar Equivalences and Associated Results Comparison of information matrices in the partial order of nonnegative definite matrices may be difficult in practice, especially if the information matrices are based on quite different families of estimating functions. In the case where an OF -optimal estimating function exists, however, we may replace the matrix comparison by simpler scalar ones. The following result is essentially due to Chandrasekar and Kale (1984). Theorem 2.2 Suppose that H is a space of estimating functions for which an OF -optimal estimating function exists. The condition that G T is OF -optimal in H, i.e., that E(G T )-E(GT ) is nonnegative definite for all GT H, is equivalent to either of the two alternative conditions: for all GT H, (i) (trace (T) criterion) tr E(G T ) tr E(GT ); (ii) (determinant (D) criterion) det(E(G T )) det(E(GT )). Remark There are further alternative equivalent conditions in addition to (i) and (ii), in particular (iii) (smallest eigenvalue (E) criterion) min (E (G T )) min (E (GT )) and (iv) (average variance (A) criterion) p-1 tr (E (G T )) -1 -1 p-1 tr (E (GT )) -1 -1 . 20 CHAPTER 2. THE GENERAL FRAMEWORK Conditions (i) ­ (iv) have been widely used in the theory of optimal experimental design where they respectively correspond to T, D, E and A-optimality. See Pukelsheim (1993), e.g., Chapter 9, and references therein. For experimental designs it often happens that a Loewner optimal design ( i.e., OF -optimal estimating function) does not exist, but an A, D, E or T optimal design (i.e., estimating function optimal in the sense of the A, D, E or T criterion) can be found. See Pukelsheim (1993, p. 104) for a discussion of nonexistence of Loewner optimal designs. Proof of Theorem 2.2 We shall herein drop the subscript T for conve- nience. (i) The condition that E(G ) - E(G) is nnd immediately gives tr (E(G ) - E(G)) = tr E(G ) - tr E(G) 0. Conversely, suppose H satisfies tr E(H) tr E(G) for all G H. If there is an OF -optimal G , then tr E(H) tr E(G ). But from the definition of OF optimality we also have tr E(G ) tr E(H) and hence tr E(G ) = tr E(H). Thus, we have that A = E(G ) - E(H) is nnd and tr A = 0. But A being symmetric and nnd implies that all its eigenvalues are positive, while tr A = 0 implies that the sum of all the eigenvalues of A is zero. This forces all the eigenvalues of A to be zero, which can only happen if A 0, since the sum of squares of the elements of A is the sum of squares of its eigenvalues. Thus, E(G ) = E(H) and we have an OF -optimal estimating function. (ii) Here we apply the Simultaneous Reduction Lemma (e.g., Rao (1973, p. 41)) which states that if A and B are symmetric matrices and B is positive definite (pd), then there is a nonsingular matrix R such that A = (R-1 ) R-1 and B = (R-1 ) R-1 , where is diagonal. In the nontrivial case we first suppose that E(G ) is pd. Then using the Simultaneous Reduction Lemma we may suppose that there exists a nonsingular matrix R such that for fixed G E(G ) = (R-1 ) R-1 , E(G) = (R-1 ) GR-1 , where G is diagonal. Then the condition that E(G ) - E(G) = (R-1 ) (I - G)R-1 is nnd forces det(E(G ) - E(G)) = (det(R-1 ))2 det(I - G) 0. This means that detG 1 and hence det(E(G )) = det(R-1 )2 det(E(G)) = det(R-1 )2 det(G). Conversely, suppose that H satisfies det(E(H)) det(E(G)) for all G H. As with the proof of (i) we readily find that det(E(H)) = det(E(G )) when G 2.4. WEDDERBURN'S QUASI-LIKELIHOOD 21 is OF -optimal. An application of the Simultaneous Reduction Lemma to the pair E(G ), E(H), the former taken as pd, leads immediately to E(G ) = E(H) and an OF -optimal solution. Remark It must be emphasized that the existence of an OF -optimal estimating function within H is a crucial assumption in the theorem. For example, if G satisfies the trace criterion (i), it is not ensured that G is an OF -optimal estimating function within H; there may not be one. 2.4 Wedderburn's Quasi-Likelihood 2.4.1 The Framework Historically there have been two distinct approaches to parameter inference developed from both classical least squares and maximum likelihood methods. One is the optimal estimation approach introduced by Godambe (1960), and others, from the viewpoint of estimating functions. The other, introduced by Wedderburn (1974) as a basis for analyzing generalized linear regressions, was termed quasi-likelihood from the outset. Both approaches have seen considerable development in their own right. For those based on the Wedderburn approach see, for example, Liang and Zeger (1986), Morton (1987), and Mc Cullagh and Nelder (1989). In this book our emphasis is on the optimal estimating functions approach and, as we shall show in this section, the Wedderburn approach can be regarded as a particular case of the optimal estimating function approach where we restrict the space of estimating functions to a special class. Wedderburn observed that, from a computational point of view, the only assumptions on a generalized linear model necessary to fit the model were a specification of the mean (in terms of the regression parameters) and the relationship between the mean and the variance, not necessarily a fully specified likelihood. Therefore, he replaced the assumptions on the probability distribution by defining a function based solely on the mean-variance relationship, which had algebraic and frequency properties similarly to those of log-likelihoods. For example, for a regression model Y = () + e with Ee = 0, we suppose that a function q can be defined by the differential equation q = Q() = ˙ V -1 (Y - ()), for matrices ˙ = (i/j) and V = Eee . Then E{Q()} = 0. E (Q()) = - ˙ V -1 ˙. 22 CHAPTER 2. THE GENERAL FRAMEWORK cov {Q()} = ˙ V -1 ˙. Thus Q() behaves like the derivative of a log-likelihood (a score function) and is termed a quasi-score or quasi-score estimating function from the viewpoint of estimating functions, while q itself is called a quasi-(log)likelihood. A common approach to get the quasi-score estimating function has been to first write down the general weighted sum of squares of residuals, (Y - ()) V -1 (Y - ()), and then differentiate it with respect to assuming V is independent of . We now put this approach into an estimating function setting. Consider a model Y = () + e, (2.4) where Y is an n × 1 data vector, Ee = 0 and (), which now may be random but for which E(e e | ˙) = V , involves an unknown parameter of dimension p. We consider the estimating function space H = {A(Y - ())}, for p × p matrices A not depending on which are ˙-measurable and satisfy the conditions that EA ˙ and E(A e e A ) are nonsingular. Then we have the following theorem. Theorem 2.3 The estimating function G = ˙ V -1 (Y - ) (2.5) is a quasi-score estimating function within H. This result follows immediately from Theorem 2.1. If G = A(Y - ), we have EG G = E(A e e V -1 ˙) = E(AE(e e | ˙)V -1 ˙) = E(A ˙) = -E ˙G as required. The estimating function (2.5) has been widely used in practice, in particular through the family of generalized linear models (e.g, McCullagh and Nelder (1989, Chapter 10)). A particular issue has been to deal with dispersion , which is modeled by using V () in place of the V in (2.5) (e.g., Nelder and 2.4. WEDDERBURN'S QUASI-LIKELIHOOD 23 Lee (1992)). Another method of introducing extra variation into the model has been investigated by Morton (1989). The Wedderburn estimating function (2.5) is also appropriate well beyond the usual setting of the generalized linear model. For example, estimation for all standard ARMA time series models are covered in Theorem 2.3. In practice, we obtain the familiar Yule-Walker equations for the quasi-likelihood estimation of the parameters of an autoregressive process under assumptions on only the first and second moments of the underlying distribution. 2.4.2 Limitations It is not generally the case, however, that the Wedderburn estimating function (2.5) remains as a quasi-score estimating function if the class H of estimating functions is enlarged. Superior estimating functions may be found as we shall illustrate below. It should be noted that, when () is nonrandom, H confines attention to nonrandom weighting matrices A. Allowing for random weights may improve precision in estimation. As a simple example, suppose that Y = (x1, . . . , xn) , e = (e1, . . . , en) and the model (2.4) is of the form xi = + ei, i = 1, 2, . . . , n with E ei Fi-1 = 0, E e2 i Fi-1 = 2 x2 i-1, Fi being the -field generated by x1, . . . , xi. Then, it is easily checked using Theorem 2.1 that the quasi-score estimating function from the estimating function space H = n i=1 ai(xi - ), ai is Fi-1 measurable is G 1 = -2 n i=1 x-2 i-1(xi - ) in contrast to the Wedderburn estimating function (2.5), i.e., G = -2 n i=1 (Ex2 i-1)-1 (xi - ), which is the quasi-score estimating function from H in which all i are assumed constant. If 1 and are, respectively, the solutions of G 1( 1) = 0 and G ( ) = 0, then (E ˙G 1)2 (E(G 1)2 )-1 = -2 n i=1 E(x-2 i-1) 24 CHAPTER 2. THE GENERAL FRAMEWORK -2 n i=1 (Ex2 i-1)-1 = (E ˙G )2 (E(G )2 )-1 since (EZ)(EZ-1 ) 1 via the Cauchy-Schwarz inequality. Thus G 1 is superior to G . Now it can also happen that linear forms of the kind that are used in H are substantially inferior to nonlinear functions of the data. The motivation for H comes from exponential family considerations, and distributions that are far from this type may of course arise. These will fit within different families of estimating functions. Suppose, for example, that Y = (x1, . . . , xn) , e = (e1, . . . , en) where xi = + ei, i = 1, . . . , n with the ei being i.i.d. with density function f(x) = 21/4 e-x4 /(1/4), - < x < . Then, we find that the true score function is U() = 4 n i=1 (xi - )3 . This is much superior to the Wedderburn estimating function (2.5) for estimation of , i.e., G = -2 n i=1 (xi - ) where 2 = var(x1) = Ee2 1. Indeed, after some calculation we obtain (E ˙U)-2 E(U)2 = Ee6 1 (9(Ee2 1)2)n = -1/2 (1/4) 12 n (3/4) , which is approximately 0.729477 of (E ˙G )-2 E(G )2 = 1 n E e2 1 = -1/2 (3/4) n (1/4) . This means that the length of an asymptotic confidence interval for derived from G will be 1/0.729477 1.1708 times the corresponding one derived from U. The true score estimating function here could, for example, be regarded as an estimating function from the family H2 = n i=1 (ai(xi - ) + bi(xi - )3 ), ai's, bi's constants , 2.4. WEDDERBURN'S QUASI-LIKELIHOOD 25 third moments being assumed zero, in contrast to G derived from H = n i=1 ai(xi - ), ai's constant . From the above discussion, we see the flexibility of being able to choose an appropriate space of estimating functions in the optimal estimating functions approach. 2.4.3 Generalized Estimating Equations Closely associated with Wedderburn's quasi-likelihood is the moment based generalized estimating equation (GEE) method developed by Liang and Zeger (1986), and Prentice (1988). The GEE approach was formulated to deal with problems of longitudiual data analysis where one typically has a series of repeated measurements of a response variable, together with a set of covariates, on each unit or individual observed chronologically over time. The response variables will usually be positively correlated. Furthermore, in the commonly occurring situation of data with discrete responses there is no comprehensive likelihood based approach analogous to that which comes from multivariate Gaussian assumptions. Consequently, there has been considerable interest in an approach that does not require full specification of the joint distribution of the repeated responses . For a recent survey of the area see Fitzmaurice, Laird and Rotnitzky (1993). We shall follow Desmond (1996) in describing the formulation. This deals with a longitudinal data set consisting of responses Yit, t = 1, 2, . . . , ni, i = 1, 2, . . . , k, say, where i indexes the individuals and t the repeated observations per individual. Observations on different individuals would be expected to be independent, while those on the same individual are correlated over time. Then, the vector of observations Y = (Y11, . . . , Y1n1 , . . . , Yk1, . . . , Yknk ) will have a covariance matrix V with block-diagonal structure V = diag (V 1, V 2, . . . , V k). Suppose also that V i = V i(i, i), i = 1, 2, . . . , k where i = (i1 , . . . , ini ) is the vector of means for the ith individual and i is a parameter including variance and correlation components. Finally, the means it depend on covariates and a p × 1 regression parameter , i.e., it = it(), t = 1, 2, . . . , ni, i = 1, 2, . . . , k. For example, in the case of binary response variables and ni = T for each i, we may suppose that P Yit = 1 xit, = it, log it (1 - it) = xit , where xit represents a covariate vector associated with individual i and time t. This is the logit link function. 26 CHAPTER 2. THE GENERAL FRAMEWORK The basic model is then Y = + , say, where E = 0, E = V and, assuming the i, i = 1, 2, . . . , k known, the quasi-score estimating function from the family H = {A (Y - )} for k i=1 ni × k i=1 ni matrices A satisfying appropriate conditions is Q() = ˙ V -1 (Y - ) as in Theorem 2.3. Equivalently, upon making use of the block-diagonal structure of V , this may be written as Q() = k i=1 ˙i V -1 i (Y i - i), where ˙i = i/, Y i = (Yi1, . . . , Yini ) , i = 1, 2, . . . , k. The GEE is based on the estimating equation Q() = 0. Now the particular feature of the GEE methodology is the use of "working" or approximate covariance matrices in place of the generally unknown V i. The idea is that the estimator thereby obtained will ordinarily be consistent regardless of the true correlation between the responses. Of course any replacement of the true V i by other covariance matrices renders the GEE suboptimal. It is no longer based on a quasi-score estimating function although it may be asymptotically equivalent to one. See Chapter 5 for a discussion of asymptotic quasi-likelihood and, in particular, Chapter 5.5, Exercise 3. Various common specifications of possible time dependence and the associated methods of estimation that have been proposed are detailed in Fitzmaurice, Laird and Rotnitzky (1993). 2.5 Asymptotic Criteria Here we deal with the widely applicable situation in which families of estimating functions that are martingales are being considered. Since the score function, if it exists, is usually a martingale, it is quite natural to approximate it using families of martingale estimating functions. Furthermore, the availability of comprehensive strong law and central limit results for martingales allows for straightforward discussion of issues of consistency, asymptotic normality, and efficiency in this setting. For n × 1 vector valued martingales MT and NT , the n × n process M, N T is the mutual quadratic characteristic, a predictable increasing process such that MT NT - M, N T is an n × n martingale. We shall write M T for M, M T , the quadratic characteristic of MT . A convenient sketch of these concepts is given by Shiryaev (1981); see also Rogers and Williams (1987, IV. 26 and VI. 34). 2.5. ASYMPTOTIC CRITERIA 27 Let M1 denote the subset of G that are square integrable martingales. For {GT } M1 there is, under quite broad conditions, a multivariate central limit result G - 1 2 T GT M V N(0, Ip) (2.6) in distribution, as T ; see Chapter 12 for details. For the general theory there are significant advantages for the random normalization using G T rather than constant normalization using the covariance EGT GT . There are many cases in which normings by constants are unable to produce asymptotic normality, but instead lead to asymptotic mixed normality. These are cases in which operational time involves an intrinsic rescaling of ordinary time. Estimation of the mean of the offspring distribution in a Galton-Watson branching process described in Section 2.1 illustrates the point. The (martingale) quasi-score estimating function is QT = T i=1(Zi - Zi-1), the quadratic characteristic is Q T = 2 T i=1 Zi-1 and, on the set {W = limn -n Zn > 0}, Q - 1 2 T QT d - N(0, 1), (EQ2 T )- 1 2 QT d - W 1 2 N(0, 1), the product in this last limit being a mixture of independent W 1 2 and N(0, 1) random variables. Later, in Chapter 4.5.1, it is shown that the normal limit form of central limit theorem has advantages, for provision of asymptotic confidence intervals, over the mixed normal form. Let M2 M1 be the subclass for which (2.6) obtains. Next, with GT M2, let be a solution of GT () = 0 and use Taylor's expansion to obtain 0 = GT ( ) = GT () + ˙GT ( )( - ), (2.7) where - - , the norm denoting sum of squares of elements. Then, if ˙GT () is nonsingular for in a suitable neighborhood and ( ˙GT ( ))-1 ˙GT () Ip in probability as T , expressions (2.6) and (2.7) lead to G() - 1 2 T ˙GT ()( - ) M V N(0, Ip) in distribution. Now, for {FT } the filtration corresponding to {GT } define the predictable process Gt() = t 0 E d ˙Gs() Fs- , Fs- being the -field generated by r 0. It is evident that maximizing the martingale information IG() = GT () G() -1 T GT () leads to (asymptotic) confidence regions centered on of minimum size (see for example Rao (1973, § 4b. 2)). Note that IG() can be replaced by IG( ) in (2.8). The relation between OF and OA optimality, both restricted to the same family of estimating functions, is very close and it should be noted that E GT () = E ˙GT (), E G() T = EGT () GT (). Below we shall consider the widely applicable class of martingale estimating function A = GT () = T 0 s() dMs(), s predictable (2.9) and it will be noted that OF and OA optimality coincide. Applications of this class of estimating functions will be discussed in some detail in Section 2.6. As with Definition 2.1 the criterion of Definition 2.3 is hard to apply directly, but as an analogue of Theorem 2.1 we have the following result, which 2.5. ASYMPTOTIC CRITERIA 29 is easy to use in practice. Theorem 2.4 Suppose that M M1. Then, G T () M is an OA-optimal estimating function within M if GT () -1 G(), G () T = G T () -1 G () T (2.10) for all GT M, , P and T > 0. Conversely, if M is convex and G T M is an OA-optimal estimating function, then (2.10) holds. Proof. This follows much the same lines as that of Theorem 2.1 and we shall just sketch the necessary modifications. Write G = (G1, . . . , Gp) and G = (G 1, . . . , G p) , where the subscript T has been deleted for convenience. To obtain the first part of the theorem we write the 2p × 2p quadratic characteristic matrix of the set of variables Z = G1, . . . , Gp, G 1, . . . , G p in partitioned matrix form as C = G G, G G, G G . Now C is a.s. nonnegative definite since for u, an arbitrary 2p × 1 vector, u Cu = u Z, Z u = u Z, Z u = u Z 0 and then the method of Rao (1973, p. 327) gives the a.s. nonnegative definitness of G - G, G ( G ) -1 G, G . (2.11) But, condition (2.10) gives G, G = ( G)( G )-1 G and using this in (2.11) gives OA-optimality via Definition 2.3. The converse part of the proof carries through as with Theorem 2.1 using the fact that H = G + G M for arbitrary scalar . With the aid of Theorem 2.4 we are now able to show that OA-optimality implies OF -optimality in an important set of cases. The reverse implication does not ordinary hold. Theorem 2.5 Suppose that G T is OA-optimal within the convex class of martingale estimating functions M. If ( G T )-1 G T is nonrandom for T > 0, then G T is also OF -optimal within M. 30 CHAPTER 2. THE GENERAL FRAMEWORK Proof. For each T > 0, G T = G T T , (2.12) say, where T is a nonrandom p × p matrix. Then, from Theorem 2.4 we have G, G T = GT T , (2.13) and taking expectations in (2.12) and (2.13) leads to EGT G T = (E ˙GT )T = (E ˙GT )(E ˙G T )-1 EG T G T . The required result then follows from Theorem 2.1. For estimating functions of the form GT () = T 0 s() dMs(), s predictable, of the class A (see (2.9) above) it is easily checked that G T () = T 0 (d Ms) (d M s)- dMs is OA-optimal within A. Here Adenotes generalized inverse of a matrix A, which satisfies AAA = A, A- AA= A. It is often convenient to use A+ , the Moore-Penrose generalized inverse of a matrix A, namely the unique matrix A+ possessing the properties AA+ A = A, A+ AA+ = A+ , A+ A = AA+ . Note that G T = T 0 (d Ms) ( dM s)d Ms = G T and Theorem 2.5 applies to give OF -optimality also. 2.6 A Semimartingale Model for Applications Under ordinary circumstances the process of interest can, perhaps after suitable transformation, be modeled in terms of a signal plus noise relationship, process = signal + noise. The signal incorporates the predictable trend part of the model and the noise is the stochastic disturbance left behind after the signal part of the model is fitted. Parameters of interest are involved in the signal term but may also be present in the noise. More specifically, there is usually a special semimartingale representation. This framework is one in which there is a filtration {Ft} and the process of interest {Xt} is (uniquely) representable in the form Xt = X0 + At() + Mt(), 2.6. A SEMIMARTINGALE MODEL FOR APPLICATIONS 31 where At() is a predictable finite variation process which is of locally integrable variation, Mt() is a local martingale and A0 = M0 = 0 (e.g., Rogers and Williams (1987, Chapter VI, 40)). The local martingale has a natural role to play in inference as it represents the residual stochastic noise after fitting of the signal which is encapsulated in the finite variation process. The semimartingale restriction itself is not severe. All discrete time processes and most respectable continuous time processes are semimartingales. Indeed, most statistical models fit very naturally into a semimartingale frame- work. Example 1. {Xt} is an autoregression of order k. This is a model of the form Xt = + 1Xt-1 + . . . + kXt-k + t consisting of a trend term +1Xt-1+. . .+kXt-k and a martingale difference disturbance t. Example 2. {Xt} is a Poisson process with intensity and Xt = t + (Xt - t), which consists of a predictable trend t and a martingale disturbance Xt - t. Example 3. {Xt} is a 1-type Galton-Watson branching process; E(X1 | X0 = 1) = . Here Xt = X (1) 1,t-1 + . . . + X (Xt-1) 1,t-1 = Xt-1 + [(X (1) 1,t-1 - ) + . . . + (X (Xt-1) 1,t-1 - )] = Xt-1 + t, say, where X (i) 1,t-1 is the number of offspring of individual i in generation t, i = 1, 2, . . . , Xt-1. Here Xt-1 represents the predictable trend and t a martingale difference disturbance. Example 4. In mathematical finance the price {St} of a risky asset is commonly modeled by St = S0 exp - 1 2 2 t + Wt , where Wt is standard Brownian motion (e.g. the Black-Scholes formulation; see e.g. Duffie (1992)). Then we can use the semimartingale representation Xt = log St = log S0 + - 1 2 2 t + Wt. (2.14) It is worth noting, however, that the semimartingale property is lost if the Brownian motion noise is replaced by fractional Brownian motion (e.g. Cutland, Kopp and Willinger (1995)). 32 CHAPTER 2. THE GENERAL FRAMEWORK For various applications the modeling is done most naturally through a stochastic differential equation (sde) representation, which we may think of informally as d (process) = d (signal) + d (noise). In the case of the model (2.14) the corresponding sde is, using the Ito formula, dSt = St ( dt + dWt) and it should be noted that the parameter is no longer explicitly present in the signal part of the representation. Of course the basic choice of the family of estimating functions from which the quasi-score estimating function is to be chosen always poses a fundamental question. There is, fortunately, a useful particular solution, which we shall call the Hutton-Nelson solution. This involves confining attention to the local M G estimating functions belonging to the class M = Gt K : Gt() = t 0 s() dMs(), s predictable , where the local M G Mt comes from the semimartingale representation of the process under consideration, or in the discrete case M = Gt K : Gt() = t s=1 s() ms(), s Fs-1 measurable , where Mt = t s=1 ms. As usual, elements of M are interpreted, via the semimartingale representation, as functions of the data and parameter of interest. Within this family M the quasi-score estimating function is easy to write down, for example, using Theorem 2.1. It is G t () = t s=1 E ˙ms Fs-1 E ms ms Fs-1 - ms (discrete time) = t 0 (d Ms) (d M s)- dMs (continuous time) where d Ms = E d ˙Ms Fs- , M s being the quadratic characteristic and - referring to generalized inverse. The Hutton-Nelson solution is very widely usable, so widely usable that it may engender a false sense of security. Certainly it cannot be applied thoughtlessly and we shall illustrate the need for care in a number of ways. 2.6. A SEMIMARTINGALE MODEL FOR APPLICATIONS 33 First, however, we give a straightforward application to the estimation of parameters in a neurophysiological model. It is known that under certain circumstances the membrane potential V (t) across a neuron is well described by a stochastic differential equation dV (t) = (- V (t) + ) dt + dM(t) (e.g. Kallianpur (1983)), where M(t) is a martingale with discontinuous sample paths and a (centered) generalized Poisson distribution. Here M t = 2 t for some > 0. The Hutton-Nelson quasi-score estimating function for = ( , ) on the basis of a single realization {V (s), 0 s T} gives G T = T 0 (-V (t) 1) {dV (t) - (- V (t) + ) dt }. The estimators ^ and ^ are then obtained from the estimating equations T 0 V (t) dV (t) = T 0 -^V (t) + ^ V (t) dt, V (T) - V (0) = T 0 -^V (t) + ^ dt, and it should be noted that these do not involve detailed properties of the stochastic disturbance M(t), only a knowledge of M t. In particular, they remain the same if M(t) is replaced (as holds in a certain limiting sense; see Kallianpur (1983)) by 2 W(t), where W(t) is standard Brownian motion. In this latter case ^ and ^ are actually the respective maximum likelihood estimators. There are, of course, some circumstances in which a parameter in the noise component of the semimartingale model needs to be estimated, such as the scale parameter associated with the noise in the above membrane potential model. This can ordinarily be handled by transforming the original semimartingale model into a new one in which the parameter of interest is no longer in the noise component. For the membrane potential model, for example, we can consider a new semimartingale {[M]t - M t}, {[M]t} being the quadratic variation process. Then, since T 0 (dV (t))2 = T 0 (dM(t))2 = [M]T and M T = T 0 d M t = 2 T, it is clear that 2 can be estimated by T-1 T 0 (dV (t))2 . Now for stochastic processes it is important to focus on the interplay between the data that is observed and the sources of variation in the model for which estimators are sought. 34 CHAPTER 2. THE GENERAL FRAMEWORK Stochastic process estimation raises a rich diversity of confounding and nonidentifiability problems that are often less transparent than those which occur in a nonevolutionary statistical environment. Furthermore, there are more often limitations on the data that can be collected and consequent limitations on the parameters that can be estimated. It is worth illustrating with some examples in which one must be careful to focus individually on the different sources of variation in the model. The first example (Srensen (1990)) concerns the process dXt = Xt dt + dWt + dNt, t 0, X0 = x0, where Wt is a standard Brownian motion and Nt is a Poisson process with intensity . This process may be written in semimartingale form as dXt = Xt dt + dt + dMt, where Mt = Wt + Nt - t and the Hutton-Nelson quasi-score estimating function based on Mt is G T = - 1 1 + T 0 Xs- 1 dMs. Then, the estimating equations are T 0 Xs- dXs = ^ T 0 X2 s ds + ^ T 0 Xs ds XT = ^ T 0 Xs ds + ^ T. These, however, are the maximum likelihood estimating equations for the model dXt = ( Xt + ) dt + dWt, that is, where Nt has been replaced by its compensator t. The entire stochastic fluctuation is described by the Brownian motion and this would only be realistic if 1. However, if we rewrite the original model by separating the discrete and continuous parts and treating them individually, we get true MLE's. Put dXc t = Xt dt + dWt Xc t = Xt - Nt (2.15) dXt - dXc t = dNt = dt + dNt - dt. (2.16) The Hutton-Nelson quasi-score estimating functions based on (2.15) and (2.16) individually are, respectively, T 0 Xs- dXc s = ~ T 0 X2 s ds, NT = ~ T. 2.7. SOME PROBLEM CASES FOR THE METHODOLOGY 35 The improvement over the earlier approach can easily be assessed. For example, E(~ - )2 /T while E(^ - )2 (1 + )/T as T . In the next example (Srensen (1990)) {Xt} behaves according to the compound Poisson process Xt = Nt i=1 Yi, where {Nt} is a Poisson process with intensity and the {Yi} are i.i.d. with mean (and independent of {Nt}). Here the semimartingale representation is Xt = t + Nt i=1 (Yi - ) + (Nt - t) = t + Mt, say, {Mt} being a martingale. Using {Mt}, the estimating equation based on the Hutton-Nelson quasi-score is XT = ( ) T so that only the product is estimable. However, if the data provide both {Xt, 0 t T} and {Nt, 0 t T}, then we can separate out the two martingales Nt 1 (Yi - ) , {Nt - t} and obtain the Hutton-Nelson quasi-score estimating functions based on each. The QL estimators are ^ = XT /NT , ^ = NT /T and these are MLE's when the Yt's are exponentially distributed. General conclusion: Identify relevant martingales that focus on the sources of variation and combine them. Relevant martingales can be constructed in many ways. The simplest single approach is, in the discrete time setting, to take the increments of the observation process, substract their conditional expectations and sum. Ordinary likelihoods and their variants such as marginal, conditional or partial likelihoods can also be used to generate martingales (see, e.g., Barndorff-Nielsen and Cox (1994, Sections 8.6 and 3.3)). General methods for optimal combination of estimating functions are discussed in Chapter 6. 2.7 Some Problem Cases for the Methodology It is by no means assured that a quasi-score estimating function is practicable or computable in terms of the available data even if it is based on the semimartingale representation as described in Section 2.6. To provide one example we take the population process in a random envi- ronment Xt = ( + t) Xt-1 + t, 36 CHAPTER 2. THE GENERAL FRAMEWORK where ( + t) is multiplicative noise coming from environmental stochasticity and t is additive noise coming from demographic stochasticity. Let {Ft} denote the past history -fields. We shall take E t Ft-1 = 0 a.s., E t Ft-1 = 0 a.s., E 2 t Ft-1 = 2 Xt-1, the last result being by analogy with the Galton-Watson branching process. The problem is to estimate on the basis of data {X0, . . . , XT }. Here the semimartingale representation is Xt = Xt-1 + ut, where the ut = t Xt-1+ t are M G differences. The Hutton-Nelson quasi-score estimating function based on {ut} is Q T () = T 1 Xt-1 E u2 t Ft-1 (Xt - Xt-1) and E u2 t Ft-1 = X2 t-1 E 2 t Ft-1 + 2Xt-1 E t t Ft-1 + E 2 t Ft-1 may not be explicitly computable in terms of the data; indeed a tractable form requires independent environments. If we attempt to restrict the family of estimating functions to ones that are explicit functions of the data H : HT = T t=1 at(Xt-1, . . . , X0; ) (Xt - Xt-1) , then there is not in general a solution to the optimality problem of minimizing var HS T . Suppose now that the environments in the previous example are independent and 2 = E(2 t | Ft-1). Then, E u2 t Ft-1 = 2 X2 t-1 + 2 Xt-1, so that Q T = T 1 Xt-1 2X2 t-1 + 2Xt-1 (Xt - Xt-1). The QLE is ^T = T 1 Xt Xt-1 2X2 t-1 + 2Xt-1 T 1 X2 t-1 2X2 t-1 + 2Xt-1 , 2.7. SOME PROBLEM CASES FOR THE METHODOLOGY 37 which contains the nuisance parameters 2 , 2 . However, on the set {Xt }, ^T will have the same asymptotic behavior as T-1 T t=1 Xt Xt-1 . This kind of consideration suggests the need to formulate a precise concept of asymptotic quasi-likelihood and this is done in Chapter 5. Another awkward example concerns the estimation of the growth parameter in a logistic map system whose observation is subject to additive error (e.g., Berliner (1991), Lele (1994)). Here we have observations {Yt, t = 1, 2, . . . , T} where Yt = Xt + et (2.17) with the {Xt} process given by the deterministic logistic map Xt+1 = Xt(1 - Xt), t = 0, 1, 2, . . . , while the e's are i.i.d. with zero mean and variance 2 (assumed known). Based on the representation (2.17) one might consider the family of estimating functions H = T t=1 ct(Yt - Xt), ct s constants , (2.18) although some adjustments are clearly necessary to avoid nuisance parameters, the X's not being observable. Nevertheless, proceeding formally, the quasiscore estimating function from H is QT () = T t=1 dXt d (Yt - Xt-1(1 - Xt-1)). The weight functions dXt/d are, however, not practicable to work with. Note that the system {Xt} is uniquely determined by X0 and and that Xt may be written (unhelpfully) as a polynomial of degree 2t - 1 in and 2t in X0. It may be concluded that, despite its apparent simplicity, H is a poor choice as a family of estimating functions and that one based more directly on the available data may be more useful. With this in mind, and noting that Yt(1 - Yt) + 2 - Xt(1 - Xt) = et(1 - 2Xt) + (2 - e2 t ), so that Yt(1 - Yt) + 2 is an unbiased estimator of Xt(1 - Xt), we can consider the family K = T t=1 ct(Yt - (Yt-1(1 - Yt-1) + 2 )), ct s constants (2.19) 38 CHAPTER 2. THE GENERAL FRAMEWORK of estimating functions. The element of this family for which ct = 1 for all t has been shown by Lele (1994) to provide an estimator that is strongly consistent for and asymptotically normally distributed. Quasi-likelihood methods do offer the prospect of some improvement in efficiency at the price of considerable increase in complexity (see Exercise 4, Section 2.8). 2.8 Complements and Exercises Exercise 1. Show that an equivalent form to Criterion 2.2 when the score function UT exists is E UT - G (s) T G (s) T = 0 = EG (s) T UT - G (s) T . (Note that UT = U (s) T .) (Godambe and Heyde (1987)). Exercise 2. Extend Theorem 2.2 to include the smallest eigenvalue criterion and average variance criterion in the remark immediately following the statement of the theorem as additional equivalent conditions (e.g., Pukelsheim (1993)). Exercise 3. (Geometric waiting times) If time is measured in discrete periods, a model that is often used for the time X to failure of an item is P(X = k) = k-1 (1 - ), k = 1, 2, . . . . Suppose that we have a set of i.i.d. observations X1, . . . , Xn from X. (i) Show that this is an exponential family model and find the MLE ^ of . (ii) Show that EX = (1 - )-1 and that the QLE for from the family H = { n i ci(Xi - EXi), ci constants} coincides with the MLE. Now suppose that we have censoring of the data. Suppose that we only record the time of failure if failure occurs on or before time r and otherwise we just note that the item has survived at least time r + 1. Thus we observe Y1, . . . , Yn, which have the distribution P(Yi = k) = k-1 (1 - ), k = 1, 2, . . . , r P(Yi = r + 1) = 1 - P(X r + 1) = r . Let M be the number of indices i such that Yi = r + 1. (iii) Show that the MLE of based on Y1, . . . , Yn is ^(Y1, . . . , Yn) = n i=1 Yi - n n i=1 Yi - M . 2.8. COMPLEMENTS AND EXERCISES 39 (iv) Show that the QLE for based on the family H = n 1 ci(Yi - EYi), ci constants does not coincide with the MLE ^(Y1, . . . , Yn). (v) Now subdivide the data into those indices i such that Yi = r + 1 and those indices i for which Yi r and consider these sets separately. To make this clear, define Zi = Yi I(Yi = r + 1), Wi = Yi I(1 Yi r), i = 1, 2, . . . , n, I denoting the indicator function. Next consider the families of estimating functions H1 = n 1 ci(Zi - EZ1), ci constants , H2 = n 1 di(Wi - EW1), di constants , and show that these lead to QSEF's (r + 1)(M - n r ) (2.20) and n 1 Yi - (r + 1) M - n 1 - r 1 - - r r , (2.21) respectively. Note that the first of these suggests the use of M to estimate nr and substituting this into n 1 Yi - (r + 1) M - n 1 - ^r 1 - ^ - r ^r = 0 with M = n^r leads to the ^(Y1, . . . , Yn) defined above in (iii). (vi) Show that the estimating functions (2.20) and (2.21) can also be added to produce ^(Y1, . . . , Yn). That is, there are multipliers and such that (r + 1)(M - nr ) + n 1 Yi - (r + 1)M - n 1 - r 1 - - rr = 0 has solution = ^(Y1, . . . , Yn). Find / . 40 CHAPTER 2. THE GENERAL FRAMEWORK (vii) Show that the solution in (vi) is the QS estimating function for the family H = n 1 (Zi - EZ1) + n 1 (Wi - EW1), , constants . Exercise 4. For the logistic map observed subject to noise investigate the quasi-score estimating function from the family (2.19) together with practicable variants. If t = Yt - Yt-1 (1 - Yt-1) + 2 , note that s and t are not independent unless |t-s| > 1. Obtain strong consistency and asymptotic normality results for estimators if possible and compare asymptotic variances with that obtained for the estimator based on the case ct = 1, all t. Infinite moments The quasi-likelihood theory developed herein is based on estimating functions with zero means and finite variances. When the data are derived from distributions that do not possess finite second moments, linear functions of the data cannot be used as a basis for constructing estimating functions and transformation of the data is required. Exercise 5 provides a simple illustation of the possibilities. More information is provided in Chapter 13. Exercise 5. Let Xi, i = 1, 2, . . . , n be i.i.d.r.v. having a Cauchy distribution with density b/((b2 + x2 )), - < x < . Show that n-1 n i=1 cos Xi can be obtained as a QLE for e-b . Extending this reasoning, show that, for any real t, n-1 n i=1 cos t Xi is a QLE for e-tb and investigate the question of a suitable choice for t. Non-regular cases In assessing the appropriateness of likelihood-based methods for parameter inference in a particular context, it is necessary to check the following basic set of requirements: (1) Is the likelihood differentiable with respect to the parameter (i.e., does the score function exist)? (2) Does Fisher information exist? (3) Does the Fisher information of the sample increase unboundedly as the sample size increases? (4) Can the parameter of interest take a value on the boundary of the parameter set? (5) Is the likelihood zero outside a domain that depends on the unknown parameter? 2.8. COMPLEMENTS AND EXERCISES 41 Failure of any of these regularity conditions is typically a warning signal that non-standard behavior may be expected. Similar considerations apply also to quasi-likelihood methods in respect to (2)-(4) and the information E. The following exercises illustrate consequences of warning signals (4), (5). Exercise 6. Let X1, . . . , Xn be i.i.d. uniform U(0, ) random variables. (i) Show that the maximum likelihood estimator X(n) = maxkn Xk does not satisfy the likelihood equation (obtained from equating the score function to zero). (ii) Show that n( - X(n)) converges in distribution to an exponential law. (iii) Let Sn = [n + 1/n] X(n). Show that n is preferable asymptotically to X(n) by verifying that E n(X(n) - )2 22 , E n(n - )2 2 as n . Exercise 7. Let Y1, Y2, . . . , Yn be i.i.d. such that Yj = |Xj|, where Xj has a N(, 1) distribution. The situation arises, for example, if Xj is the difference between a matched pair of random variables whose control and treatment labels are lost. The sign of the parameter is unidentifiable so we may work with = ||, with parameter space = [0, ). Show that the score function is n j=1 [Yj + tanh (Yj) - ] and observe that for > 0 the problem is a regular one and n (^n - ) d - N(0, i()) with i() = 1 - 2 2 e- 1 2 2 0 y2 e- 1 2 y2 sech (y) dy. For = 0 all this breaks down and the score function is identically zero. However, show that a first-order expansion of the score function gives ^2 n = 3 n 1 Y 2 j - n n 1 Y 4 j provided n 1 Y 2 j > n. Show that ^2 n has asymptotically a N(0, 2) distribution truncated at zero (Cox and Hinkley (1974, pp. 303-304)). Chapter 3 An Alternative Approach: E-Sufficiency 3.1 Introduction In classical statistics, the concept of "sufficiency" and its dual concept of "ancillarity" play a key role and sufficient statistics provide the essential approach to parameter inference in cases where they are available. To adapt these basic concepts to the context of estimating functions, Small and McLeish developed the expectation based concepts of E-sufficiency, E-ancillarity, local E-sufficiency and local E-ancillarity. The basic idea is that an ancillary statistic is one whose distribution is insentive to changes in the parameter and that, in an estimating function context, changes to the parameter are first evident through changes to the expectation of the estimating function. For a space of estimating functions the E-ancillary subset, A (say), are defined on the basis of expectation insensitivity to parameter changes and the E-sufficient estimating functions belong to the orthogonal complement of A with respect to . It is argued that an optimum estimating function from should be chosen from the E-sufficient or locally E-sufficient subset of , if this exists. Detailed expositions are given in McLeish and Small (1988), Small and McLeish (1994, Chapter 4). It turns out that our framework of optimal or quasi-score estimating functions in this book corresponds most closely to the notation of locally E-sufficient estimating functions and this chapter is concerned with their relationship. Our study shows that whether a quasi-score estimating function is locally Esufficient depends strongly on the estimating function space chosen, and also that, under certain regularity and other conditions, the quasi-score estimating function will be locally E-sufficient. In many cases the two approaches both lead to the same estimating equation. The treatment here follows Lin (1994a). 3.2 Definitions and Notation Let X be a sample space and P be a class of probability measures P on X. For each P, let vP denote the p-dimensional vector space of vector-valued functions {f} which are defined on the sample space X and satisfy EP f 2 < , where f 2 = f (x) f(x). Let be a real-valued function on the class of probability measures P and = {(P), P P} IRp be the parameter space. Now we consider a space in which every estimating function (, x) = () is a mapping from the parameter space and the sample space into IRp . 43 44 CHAPTER 3. AN ALTERNATIVE APPROACH: E-SUFFICIENCY Each element of is unbiased and square integrable, i.e., EP [((P))] = 0, and 2 (P ) EP [ ((P)) ((P))] < , (3.1) for all P and = (P). Assume that the space has constant covariance structure, i.e., the inner product EP [ ((P)) ((P))] depends on P only through (P), for all , . As usual, we also require that be a Hilbert space of estimating functions and weak square closed, i.e., for any sequence of functions n , if limn, m n - m = 0, then there is a function such that limn n - = 0. The choice of will, in general, depend on ; see McLeish and Small (1988), Small and McLeish (1994). Definition 3.1 An unbiased estimating function () is locally E-ancillary if it is a weak square limit of functions {} that satisfy EP [()] = o(|(P) - |), for (P) . (3.2) We note that if the order of the differentiation and integration can be interchanged, then equation (3.2) can be rewritten as E i j p×p = 0. Therefore, under suitable regularity conditions, Definition 3.1 can be replaced by Definition 3.2 below. Definition 3.2 An unbiased estimating function () is locally E-ancillary if it is a weak square limit of functions {} that satisfy E i j p×p = 0. The locally E-ancillary estimating functions form a subset of the estimating function space, which will be denoted by Aloc. Following the definition of local E-ancillarity, when Aloc, i.e., (1, . . . , p) Aloc, then all of the estimating functions ai,j = (0, . . . , ai, . . . , 0), where ai = j and i, j = 1, . . . , p, belong to Aloc. Definition 3.3 A subset Sloc of the estimating function space is complete locally E-sufficient when Sloc if and only if E( () ()) = 0 for all P, = (P) and Aloc. 3.2. DEFINITIONS AND NOTATION 45 Because of the special property of Aloc, when Sloc, for each element of Aloc, say = (1, . . . , p) , we have E( aij) = (0, . . . , ai, . . . , 0) , ai = j, and i, j = 1, . . . , p, i.e, E(ij) = 0 (i, j = 1, . . . , p). So E( ) = 0, which in turn implies that E( ) = 0. Therefore, the definition of Sloc has the following equivalent formulation. Definition 3.3 A subset Sloc of the estimating function space is complete locally E-sufficient when Sloc if and only if E(() () ) = 0 for all P, = (P) and Aloc. It follows from the definition of Sloc that Sloc = A loc, and by the assumptions on , we have = Sloc Aloc. However the decomposition is associated with . A more detailed discussion of the space can be found in Small and McLeish (1991). In general there is no guarantee of the existence of a complete locally Esufficient subspace. However, if it does exist it is a closed linear subspace of . Completeness is dispensed with in the following definition. Definition 3.4 A subset L of is locally E-sufficient if for , E( () ()) = 0 for all L and implies that is locally E-ancillary. As a simple illustration we consider the square integrable scalar process {Xt}, with Xt adapted to the increasing sequence of a fields Ft, and such that E Xt Ft-1 = t, var Xt Ft-1 = t() for an unknown parameter and Ft-1-measurable t, t() whose forms are specified. Let be the space of square integrable estimating functions of the form () = n i=1 Ai()(Xi - i), Ai's are Fi-1 measurable . (3.3) Then, if the function () = n i=1 i-1 i (Xi - i) is in the space , it generates the locally E-sufficient subspace. 46 CHAPTER 3. AN ALTERNATIVE APPROACH: E-SUFFICIENCY To check this, we note that of () is of the form (3.3), then E( () ()) = E n i=1 i -1 i iAi() = E n i=1 i Ai() = d d E n i=1 ( - )i Ai() = = d d E () = and if this is zero then is locally E-ancillary. 3.3 Results In the following, assume that satisfies the usual regularity conditions, so that Definition 3.2 can be used to define the concept of local E-ancillarity. Let G1 be a subset of . Each element G of G1 satisfies the conditions that E ˙G() = (E(Gi/j))p×p and EG() G () are nonsingular. The score function is as usual denoted by U(). The fact that = Sloc Aloc leads us to seek conditions under which a quasi-score estimating function is in Sloc or Aloc. In the following discussion, a quasi-score estimating function in G1 is always denoted by G and, for an estimating function G, if G = 0 a.s., we write G = 0. Theorem 3.1 If G1 is a convex set and the quasi-score estimating function G G1 Aloc (resp. G G1 Sloc), then G1 Sloc = (resp. G1 Aloc = ). Proof. We give only the first part of the proof. The proof of the second part is analogous. Applying Theorem 2.1, if G G1 Aloc, but G1 Sloc = , then there is a G G1 Sloc such that E ˙G -1 EG G = E ˙G -1 EG G . Following the definition of Sloc, we have EG G = 0. So the left hand side of the above equation is equal to zero. This implies EG G = 0, and G has to be a zero vector. This is contrary to the definition of a quasiscore estimating function. Therefore G1 Sloc = . 3.3. RESULTS 47 From Theorem 3.1 we note that, if there are no other restrictions on G1, it is possible that G is outside Sloc. However in general, under regularity conditions, G1 Aloc = . This implies that G can be only in the complement of Aloc. But this conclusion has not answered the questions of whether G Gloc and what the necessary conditions are for G to belong to Sloc. The conditions that force G Sloc are of interest and are provided in the following theorems. Theorem 3.2 Suppose that G1 is a convex set and G G1 - G1 Aloc. Let G ps be the projection of G on Sloc. If G ps G1 and G ps satisfies the condition that G ps + G G1 for any G with E ˙G = 0, then G Sloc. Proof. Assume that G = G ps + G pa, where G pa Aloc and G ps Sloc. Since G pa Aloc, there is a sequence {Gpa,n} in Aloc with E ˙Gpa,n = 0 (n = 1, 2, . . .), such that E((G pa - Gpa,n) (G pa - Gpa,n)) 0, as n . Let Gn = G ps + Gpa,n. By the assumption of the theorem, we have Gn G1. So by Theorem 2.1, E ˙Gn -1 EGnG = E ˙G -1 EG G , i.e., E ˙G ps -1 EG psG ps + EGpa,nG pa = E ˙G ps + E ˙G pa -1 EG psG ps + EG paG pa . (3.4) Since E Gpa,nG pa - G paG pa = E Gpa,n - G pa G pa E Gpa,n - G pa E G pa 0 as n , letting n in (3.4) gives E ˙G ps -1 EG psG ps + EG paG pa = E ˙G ps + E ˙G pa -1 EG psG ps + EG paG pa , which implies that E ˙G pa = 0. Applying Theorem 2.1 again, we have E ˙G ps -1 EG psG = E ˙G -1 EG psG ps + EG paG pa = E ˙G ps -1 EG psG ps + EG paG pa , 48 CHAPTER 3. AN ALTERNATIVE APPROACH: E-SUFFICIENCY i.e., E ˙G ps -1 EG psG ps = E ˙G ps -1 EG psG ps + EG paG pa . Therefore EG paG pa = 0 and hence G pa = 0. Thus, G G1 Sloc. Theorem 3.3 Suppose that G1 is a convex set. Let G = G ps + G pa, where G ps Sloc G1 and G pa Aloc. If E ˙G = E ˙G ps for all , then G Sloc. The proof of Theorem 3.3 is analogous to that of Theorem 3.2 and is omit- ted. Before another important property for the quasi-score estimating function is mentioned, we introduce some new notation and definitions. Let ~G1 = {G, E ˙G is nonsingular and G }, so that ~G1 G1. Definition 3.5 An estimating function G T is a sub-quasi-score estimating function in ~G1, if E ˙GT -1 EGT GT E ˙GT -1 - E ˙G T -1 EG T G T E ˙G T -1 is nonnegative-definite for all GT ~G1 and all . Remark. The relationship between a sub-quasi-score estimating function and a quasi-score estimating function is as follows. If a sub-quasi-score estimating function G T is an element of G1, then G T is a quasi-score estimating function in G1. However, if G T is a quasi-score estimating function in G1, G T need not to be a sub-quasi-score estimating function in ~G1 because the space ~G1 is larger than G1 Obviously a sub-quasi-score estimating function is not a quasi-score estimating function in general. However, if we restrict our consideration to the model Xt = t 0 fs ds + ms, and consider the estimating function space T = { T 0 s() dms, s() Fs-} with G1 ~G1 T , where {ms} is a martingale with respect to some -fields {Ft} and {s} is an increasing function, then following the proof in Godambe and Heyde (1987, p. 238), we can show that the quasi-score estimating function G is a sub-quasi-score estimating function in ~G1. Since the conditions required for the sub-quasi-score estimating function are weaker than those for the quasi-score estimating function, it should be easy 3.3. RESULTS 49 to find the conditions under which the sub-quasi-score estimating function is an E-sufficient estimating function. We now restrict by requiring that the subset Aloc of satisfy the following condition. Condition 3.1 G Aloc if and only if E ˙G = 0 for all . Theorem 3.4 Let G be a sub-quasi-score estimating function on ~G1 . Then, under Condition 3.1, G Sloc. Proof. Express G as G = G ps + G pa, where G ps Sloc and G pa Aloc. Since G pa Aloc and G ~G1, it follows that E ˙G ps = E ˙G is nonsingular for all . Following Definition 3.5 we have that E ˙G ps -1 EG psG ps E ˙G ps -1 - E ˙G -1 EG G E ˙G is nonnegative-definite. That is, E ˙G ps -1 EG psG ps E ˙G ps -1 - E ˙G ps -1 EG psG ps + EG paG pa E ˙G ps -1 is nonnegative-definite. Consequently, -EG paG pa is nonnegative-definite, which gives EG paG pa = 0 and G Sloc. Theorem 3.4 shows that, if Condition 3.1 is satisfied and if a quasi-score estimating function is a sub-quasi-score estimating function, then it will belong to Sloc. Therefore, another way to check whether G Sloc is by verifying Condition 3.1 and showing that G is a sub-quasi-score estimating function. As an example, consider a stochastic process {Xt} that satisfies a stochastic diferential equation of the form dXt = at() d M() t + dMt() (t T), (3.5) where {Mt} is a square integrable martingale with predictable variation process M t observable up to the value of the parameter . Also assume that the predictable process {at()} is observable up to the parameter value and there exists a real-valued predictable process f(t; , ) such that s 0 at() d M() t = s 0 f(t; , ) d M() t, for all 0 < s < T and all sufficiently close to . Noting that f(t; , ) = at() (see McLeish and Small (1988, p. 103)) and restricting our consideration to a space of estimating functions, in which each element has the form () = T 0 Hs() dMs(), where H is a predictable process, we obtain the following 50 CHAPTER 3. AN ALTERNATIVE APPROACH: E-SUFFICIENCY proposition. Proposition 3.1 Under the conditions of the model (3.5), Condition 3.1 is satisfied, i.e., Aloc = { : E ˙ = 0, all }. Proof. We note that, for any estimating function in the estimating function space and parameters and in the parameter space, E() = E T 0 Ht()[dMt() + at() d M() t - at() d M() t] = E T 0 Ht()[f(t; , ) - f(t; , )] d M() t . Dividing by - in the above equation and then letting , we obtain E ˙() = E () = = E T 0 Ht() f(t; , ) d M() t. For any 0() = T 0 Ht,0() dMt() in Aloc, we can show that E 0 = 0. From the definition of Aloc, there is a sequence of estimating functions {n = T 0 Ht,n dMt()} with E ˙n = E T 0 Ht,n() f(t; , ) d M() t = 0, for all , which weakly square converges to 0. Therefore, applying the KunitaWatanabe inequality (see e.g. Elliott (1982, Corollary 10.12, p. 102)), we obtain E ˙0() = E ˙0() - ˙n() = E T 0 (Ht,0() - Ht,n()) f(t; , ) d M() t E T 0 (Ht,0() - Ht,n())2 d M() t 1 2 E T 0 f(t; , ) 2 d M() t 1 2 = 0 - n 2 E T 0 f(t; , ) 2 d M() t 1 2 0 as n , 3.4. COMPLEMENT AND EXERCISE 51 which yields E ˙0 = 0 for all . Therfore, Aloc = { : E ˙() = 0, all , }. Applying Proposition 3.1 and following the remark under Definition 3.5 for models satisfying (3.5), all quasi-score estimating functions from the space G1 = { T 0 Hs() dMs(); H a predictable process} lie in Sloc. As another example consider the widely used model dXt = ft() d + dMt, where {ft()} is a predictable process and M is a square integrable martingale with d M() t = at() d. Since this model can be rewritten in the form (3.5) and also satisfies the assumption required for (3.5), the quasi-score estimating function in the space = T 0 Ht() [dXt - ft() d], {Ht} a predictable process is locally E-sufficient. Condition 3.1 plays an important role in our discussion. However, checking this condition in general is still an issue. Of course Proposition 3.1 says that Condition 3.1 is always true for the model 3.5, which covers many problems and a similar analysis can be undertaken in other particular cases. 3.4 Complement and Exercise The theory in this book has been formulated on the assumption that the estimating functions G() under consideration are differentiable with respect to . It is possible, however, to extend this theory by replacing the -E ˙G() used herein by G() = EG() = . This feature has been incorporated in the theory of this chapter. The following exercise indicates the kind of result that is thereby included. 1. Estimation of the parameter in the density function 1 2 e-|x-| , - < x < , is required on the basis of a random sample {X1, . . . , XT }. Consider the space of estimating functions that contains just the two ele- ments 1 = T t=1 |Xt - | - T, 2 = T t=1 sgn (Xt - ). Show that 1 is locally E-ancillary and 2 is locally E-sufficient for estimation of . Note that 2 leads to the sample median as an estimator of . Chapter 4 Asymptotic Confidence Zones of Minimum Size 4.1 Introduction Thus far in the book we have concentrated on the optimal properties of estimating functions that are derived from maximum information, minimum distance, or, in the case of Chapter 3, sufficiency ideas, which all involve fixed finite samples. Now we shall proceed to asymptotic considerations about confidence zones that generalize classical results holding for maximum likelihood. However, some comments about maximum likelihood itself are appropriate as a prelude since, although it is one of the most widely used statistical techniques, it does suffer from various problems. The principal justification of maximum likelihood is asymptotic. Under certain regularity conditions the maximum likelihood estimator is asymptotically unbiased, consistent, and asymptotically normally distributed with minimum size asymptotic confidence zones for the unknown parameter. Unfortunately the regularity conditions are rather stringent, and there is no shortage of striking examples of things that can go wrong, even in such seemingly innocent contexts as for exponential families. Le Cam (1990b) has given an entertaining account of problems for i.i.d. variables. However for all its faults, maximum likelihood methods do provide a benchmark against which others are judged and the exceptions just serve to underline the principle of reviewing each application individually for compliance. Not surprisingly, whatever problems can occur for likelihood based methods can also occur for quasi-likelihood. However, the quasi-likelihood framework has one big advantage and that is the opportunity to choose the family of estimating functions within which to work. Appropriate regularity can be demanded of these estimating functions whereas there is ordinarily little control over the choice of model and consequent form of the likelihood. Despite the foregoing comments we shall ordinarily take counterparts of the standard regularity conditions for maximum likelihood for granted in developing the corresponding quasi-likelihood theory. Exceptions typically need rather special treatment. The reader is reminded that the standard regularity conditions for maximum likelihood involve a finite dimensional parameter space , say, which is an open subset of the corresponding Euclidean space and a true parameter which is an interior point of . A likelihood which is at least twice differentiable with respect to the parameter is postulated and differentiation under the integral sign is permitted. Various other technical conditions are typically 53 54 CONFIDENCE ZONES OF MINIMUM SIZE imposed which are sufficient to ensure consistency and asymptotic normality of the estimators (e.g., Cox and Hinkley (1974, Chapter 9), Basawa and Prakasa Rao (1980, particularly Chapter 7)). The most common exceptions involve unknown endpoint problems (i.e., one endpoint of the range of the distribution in question is the unknown parameter), parameters on the boundary of the parameter space and unbounded likelihoods. For discussion of nonregular cases for the MLE see, e.g., Smith (1985), (1989), Cheng and Traylor (1995). Whenever something can be done for the MLE, it may be expected that corresponding quasi-likelihood analogues can be developed. Suppose that we have experiments indexed by t [0, ) and {Ft} denotes a standard filtration generated from these experiments. Both discrete and continuous time are covered. With little loss in generality we confine attention to the class G of zero mean, square integrable, Ft-measurable, semimartingale estimating functions Gt() which are vectors of dimension p such that the p dimensional matrices ˙Gt() = (Gt,i()/j), E ˙Gt(), EGt() Gt(), and [G()]t are (a.s.) nonsingular for each t > 0, [G()]t denoting the quadratic variation process. Here [G()]t = [Gcm ()]t + 0 0 a.s.) we have, as T , (EGT () GT ())-1/2 E ˙GT () ( T - ) d - Z, (4.3) say, not depending on the choice of GT . The size of confidence zones for is then governed by the scaling "informa- tion" E(Gt()) = E ˙Gt() EGt() Gt() -1 E ˙Gt() = EG (s) t () G (s) t () and we prefer estimating function G1,t to G2,t if E(G1,t) E(G2,t) for each t t0 in the Loewner ordering (partial order of nonnegative definite matrices). This is precisely the requirement of OF -optimality, as discussed in Chapter 2, but for all sufficiently large samples. We shall henceforth suppose that a quasi-score or asymptotic quasi-score estimating function Qt() has been chosen for which (4.3) holds and our considerations regarding confidence zones will be based on Qt. Meta Theorem: Under sensible regularity conditions a quasi-likelihood estimator within some H is strongly consistent and, with suitable norming, asymptotically normally distributed. It can be used to construct minimum size asymptotic confidence zones for estimators within H. This is a result for which a satisfactory formal statement and proof are elusive to the extent that any readily formulated set of sufficient conditions has obvious exceptions for which the desired results continue to hold. Furthermore, it is usually preferable, and more economical, to check directly in a particular case to see whether consistency and asymptotic normality hold than to try to verify the sufficient conditions of a general theorem that might not quite apply. The results in this area mostly make use of martingale strong laws and central limit theorems and we shall provide powerful versions of these in 56 CONFIDENCE ZONES OF MINIMUM SIZE Chapter 12. They can be used in particular cases. For specific results along the lines of the meta theorem we refer the reader to Hutton and Nelson (1986), Theorem 3.1 (strong consistency) and Theorem 4.1 (asymptotic normality), Hutton, Ogunyemi and Nelson (1991), Theorem 12.1 (strong consistency) and Theorem 12.2 (asymptotic normality) and Greenwood and Wefelmeyer (1991), Proposition 10.1. See also Section 3 of Barndorff-Nielsen and Srensen (1994). Of course the issue of "sensible regularity conditions" is crucial. For some examples where asymptotic normality of the quasi-likelihood estimators does not hold, see Chan and Wei (1988). These involve unstable autoregressive processes. Much of the complication of regularity conditions can be drawn off into a single stochastic equicontinuity condition (Pollard (1984, Chapter 7)) in the case where a stochastic maximization or minimization is involved, or similarly, into stochastic differentiability conditions (Hoffmann-Jrgensen (1994, Chapter 14)). 4.3 Confidence Zones: Theory At the outset it is important to emphasize that confidence intervals based on (4.3) are mostly difficult to formulate unless Z is normally distributed and it is desirable, wherever possible, to renormalize to obtain asymptotic normality. Indeed, as we shall see in Section 4.4, there is a specific sense in which confidence intervals based on asymptotic normality are preferable, on average, to those based on an alternative limiting distribution, at least in the scalar case. For Qt() G, the result [Q] -1/2 t Qt d - MV N(0, Ip) (4.4) seems to encapsulate the most general form of asymptotic normality result. Norming by a random process such as [Q]t is essential in what is termed the nonergodic case (for which (E[Q]t)-1 [Q]t p - constant as t ). A simple example of this is furnished by the pure birth process Nt with intensity Nt-, where we take for Qt the score function and Qt = -1 (Nt - 1) - t 0 Ns- ds, [Q]t = -2 (Nt - 1), while [Q] -1/2 t Qt d - N(0, 1), (EQ2 t )-1/2 Qt d - W1/2 N(0, 1) and (E[Q]t)-1 [Q]t a.s. - W, where W has a gamma distribution with the form parameter N0 and shape parameter N0 and W1/2 N(0, 1) is distributed as the product of independent 4.3. CONFIDENCE ZONES: THEORY 57 W1/2 and N(0, 1) variables. Here E(Qt()) = E[Q]t = EQ2 t , the MLE ^t satisfies ^t = (Nt - 1) t 0 Ns- ds and 1 [Q] 1 2 t (^t - ) d - N(0, 1), 1 E(Qt()) 1 2 (^t - ) d - W- 1 2 N(0, 1). On average, confidence intervals for from the former are shorter than those from the latter (Section 4.4). To obtain confidence zones for from (4.4) we may use the Taylor expansion (4.2) and then under appropriate continuity conditions for ˙Qt, ˙QT (1,T ) ˙QT () -1 p - Ip and, when (4.4) holds [Q()] -1/2 T ˙QT () ( T - ) d - MV N(0, Ip) (4.5) as T . For the construction of confidence intervals we actually need this convergence to be uniform in compact intervals of ; we shall use u.d. - to denote such convergence and henceforth suppose that (4.5) holds in this mode. We shall also write E(Qt()) = - ˙Qt() [Q()]-1 t - ˙Qt() . (4.6) If C is a column vector of dimension p, convergence in (4.5) is mixing in the sense of R´enyi (see Hall and Heyde (1980, p. 64)) and E(QT ()) behaves asymptotically like a constant matrix. Then, replacing E(QT ()) by the estimated E(QT ( T )), we obtain P C C T z/2 C E(QT ( T )) -1 C 1 - where (z) = 1 - , denoting the standard normal distribution function. In particular, this provides asymptotic confidence results for the individual elements of and also any nuisance parameters can conveniently be deleted. Other confidence statements of broader scope may also be derived. Noting that ( T - ) E (QT ( T )) ( T - ) u.d. - 2 p 58 CONFIDENCE ZONES OF MINIMUM SIZE and that ( T - ) E (QT ( T )) ( T - ) = max c (C ( T - ))2 C ( E (QT ( T )))-1C , we have that P max c |C ( T - )| (C ( E (QT ( T )))-1C)1/2 p, 1 - , where 2 p, is the upper point of the chi-squared distribution with p degrees of freedom, so that P C 0 C T p, C ( E (QT ( T )))-1 C 1/2 for all C 1 - . Also, if we let c be the set of all possible satisfying the inequality ( T - ) E (QT ( T )) ( T - ) 2 p,, the so-called ellipsoids of Wald in the classical case (e.g., Le Cam (1990a)), then simultaneous confidence intervals for general functions gi(), i = 1, 2, . . . , p with asymptotic confidence possibly greater than (1 - ) are given by min c gi(), max c gi() , i = 1, 2, . . . , p. This allows us to deal conveniently with confidence intervals for nonlinear functions of the components . Other methods are available in particular cases and, it often happens, in cases where likelihood-based theory is available and tractable, that the ML ratio produces better-behaved confidence zones than the ML estimator (e.g. Cox and Hinkley (1974, p. 343)). In such a case, if = ( , ) and 0 = {0, }, then the set 0 : sup Lt() - sup 0 LT () 1 2 2 d, (4.7) based on the likelihood function L, and with d = dim , may give an asymptotic confidence region for of size . It should be noted that (4.7) is invariant under transformation of the parameter of interest. It is usually the case that the quasi-score or asymptotic quasi-score estimating function Qt() is a martingale and subject to suitable scaling, - ˙Qt() - [Q()]t, (4.8) - ˙Qt() - Q() t, (4.9) and [Q()]t - Q() t (4.10) 4.3. CONFIDENCE ZONES: THEORY 59 are zero mean martingales, Q() t being the quadratic characteristic of Qt. The quantity - ˙Qt() is a generalized version of what is known as the observed information, while EQt() Qt() = E[Q()]t = E Q() t = -E ˙Qt() is a generalized Fisher information. The martingale relationships (4.8) ­ (4.10) and that fact that each of the quantities - ˙Qt(), [Q()]t and Q() t goes a.s. to infinity as t increases usually implies their asymptotic equivalence. Which of the asymptotically equivalent forms should be used for asymptotic confidence zones based on (4.5) is unclear in general despite the fact that many special investigations have been conducted. Let S be the set of (possibly random) normalizing sequences {At} that are positive definite and such that A-1 t () [Q()] -1/2 t ˙Qt() p - Ip, (4.11) as t . Then, each element of S determines an asymptotic confidence zone for , E(Qt()) being replaced by At() At(). Of course (4.11) ensures that zones will be similar for large T with high probability. In the ergodic case when likelihoods rather than quasi-likelihoods are being considered, there is evidence to support the use of the observed information - ˙Qt() rather than the expected information EQt() Qt() (e.g., Efron and Hinkley (1978)). In the general case, Barndorff-Nielsen and Srensen (1994) have used a number of examples to argue the advantages of (4.6), which, as we have seen, comes naturally from (4.4). On the other hand, in Chapter 2.5 we have used the form - Qt() Q() -1 t - Qt() , where Qt() is a matrix of predictable processes such that ˙Qt() - Qt() is a martingale. This has the advantage of a concrete interpretation as an empirical information and is a direct extension of the classical Fisher information and is closely related to E(Qt()). The general situation is further complicated by the fact that E(Qt()) ordinarily involves the unknown and consequently has to be replaced by E(Qt( t )) in confidence statements. The variety of possibilities is such that each case should be examined on an individual basis. Finally, some special comments on the nonergodic case are appropriate. Suppose that (E[Q]t)-1 [Q]t p - W (> 0 a.s.) as t . As always, a sequence of normalizing matrices {At} is sought so that AT ( T - ) d - MV N(0, Ip) as T . However, it has been argued that one should condition on the limit random variable W and then treat the unobserved value w of W as a nuisance parameter to be estimated. This approach has the attraction of reducing a 60 CONFIDENCE ZONES OF MINIMUM SIZE nonergodic model to an ergodic one but at the price of introducing asymptotics which, although plausible, may be very difficult to formalize in practice (e.g., Basawa and Brockwell (1984)). A related approach based on conditioning on an asymptotic ancillary statistic has been discussed by Sweeting (1986). This also poses considerable difficulties in formalization but may be useful in some cases. The problem of precise choice of normalization, of course, usually remains. In conclusion, we also need to comment on the point estimate itself. If confidence statements are based on (4.5) and T is an estimator such that [Q()] -1/2 T ˙QT ()(T - T ) p - 0 as T , then [Q()] -1/2 T ˙QT ()(T - ) d - MV N(0, Ip) clearly offers an alternative to (4.5). There may sometimes be particular reasons to favor a certain choice of estimator, such as for the avoidance of a nuisance parameter. 4.4 Confidence Zones: Practice In the foregoing theoretical discussion we have chosen to focus on the asymptotics of the estimator rather than the quasi-score estimating function Q() from which it is derived. However, if Q is asymptotically normal or mixed normal itself, there is significant empirical evidence in favor of basing confidence statements on Q directly. Suppose that (EQT () QT ())- 1 2 QT () d - MV N(0, Ip) resp. Q() - 1 2 T QT () d - MV N(0, Ip) . Then, we have the confidence regions : QT () EQT () QT () -1 QT () 2 p, (4.12) resp. : QT () Q() -1 T QT () 2 p, . The form of such regions will depend heavily on the context but it can be argued that they will usually be preferable to regions constructed on the basis of the asymptotic distribution of the estimator . It may be expected that our primary information concerns the distributions of Q() directly and that the Taylor expansion used to obtain from Q() introduces an unnecessary component of approximation. A simple example concerns nonlinear regression where Xi are independent random vectors with the distribution MV N(fi(), 2 I), i = 1, 2, . . . , T. Then, 4.4. CONFIDENCE ZONES: PRACTICE 61 the Hutton-Nelson quasi-score estimating function based on the martingale differences {mi() = Xi - fi()} is -2 ( ˙f()) m() where ˙f() = ˙f1() ..... . . . ..... ˙fT () with ˙fr() = (fr(i)/j), and m() = m1() ..... . . . ..... mT () which has an MV N(0, 2 ( ˙f()) ˙f()) distribution. Consequently, if 2 is known, the confidence zone corresponding to (4.12), namely, : -2 m () ˙f()(( ˙f()) ˙f())-1 ( ˙f()) m() 2 p, , is exact. On the other hand, the distribution of the quasi-likelihood estimator may be seriously skewed, leading to confidence regions based on ( T ) E(QT ( ))( T -), say, which are unreliable and sometimes very poor. This is despite the satisfactory asymptotics. The separate issue of possibly needing to replace 2 by an estimated value does not alter the general preference for confidence zones based on Q(). The problem referred to above can be alleviated by bias reduction methods and these have been widely studied. The usual approach is to seek a first order asymptotic expression for the bias E T = + bT () + rT (), say, where rT () = o( bT () ) and bT () 0 as T for fixed and then to replace T by the bias corrected estimator (BC) T = T - bT ( T ). In the regular classical setting of i.i.d. random variables, bT () is of the form b()/T. For a discussion of the basic methods of bias correction see Cox and Hinkley (1974, Sections 8.4 and 9.2). A more general approach to this problem involves modification of the score (or quasi-score) rather than the estimator itself. This has been developed by Firth (1993) for the case of maximum likelihood estimation but the ideas are readily adapted to the quasi-likelihood context. We replace the quasi-score estimating function QT () by Q (BC) T () = QT () + AT (), where AT () is chosen to be of the form ˙QT () bT (). This form is used since Taylor expansion gives 0 = QT ( ) QT () + ˙QT () ( T - ) 62 CONFIDENCE ZONES OF MINIMUM SIZE so that we have Q(BC) () - ˙QT () ( - - bT ()), which has been bias corrected to first order. The theory underlying such corrections can usefully be interpreted in differential geometry terms related to correcting for curvature (e.g., Barndorff-Nielsen and Cox (1994, Chapter 6) and references therein). For a recent further adaption to provide higher order corrections to estimating functions see Li (1996b). The idea is to remove first order bias while minimizing the mean squared error. This allows for situations in which, for example, skewness problems preclude the use of the Edgeworth expansion and the kurtosis is unknown. For another approach to adjusting score and quasi-score estimating functions, based on martingale methods, see Mykland (1995). 4.5 On Best Asymptotic Confidence Intervals 4.5.1 Introduction and Results Much asymptotic inference for stochastic processes is based on use of some obvious consistent estimator but involves a choice between competing normalizations, one being of constants and the other of random variables. In this section we study the common situation where either asymptotic normality or asymptotic mixed normality is achievable through suitable normalization. It should be remarked that asymptotic mixed nomality is also obtained under a variety of non-regular conditions (e.g., Kutoyants and Vostrikova (1995)) as well as regular ones. It is shown that there is a certain sense in which, on average, confidence intervals based on the asymptotic normality result are preferable to those based on asymptotic mixed normality. The result here is from Heyde (1992a). Theorem 4.1 Let {^t} be a sequence of estimators that is consistent for and {ct}, {dt} be norming sequences, possibly random, such that ct(^t - ) d - W, (4.13) dt(^t - ) d - -1 W (4.14) as t , where W is standard normal and > 0 a.s. is random, independent of W and such that ct d-1 t d - as t . Suppose that L (i) t (), i = 1, 2, is the minimum length of an 100(1 - )% confidence interval for based on an exact, approximate or assumed distribution of the pivot in (4.13) and (4.14), respectively, for which the 4.5. ON BEST ASYMPTOTIC CONFIDENCE INTERVALS 63 specified convergence result holds. Then lim inf t E L (2) t ()/L (1) t () 1. Remark 1. The theorem gives a sense in which, on average, confidence intervals based on the pivot in (4.13) are to be preferred to those based on the pivot in (4.14) whether or not a random norming is required. Remark 2. A principal application of the theorem is in the context of inference for the nonergodic models that are common in stochastic process estimation (e.g., Basawa and Scott (1983), Hall and Heyde (1980, Chapter 6)). In this context we have {ct} as random and {dt} as constants so that L (1) t () is a random variable and L (2) t () is a constant. One of the most important cases is that of locally asymptotic mixed normal (LAMN) families. This is the situation in which the log-likelihood ratio has asymptotically a mixed normal distribution and results of type (4.14) hold with {dt} as constants and (4.13) with {ct} as random variables, while Ect = dt for each n. For LAMN families one has, under modest regularity conditions, a striking result known as Haj´ek's convolution theorem (e.g., Theorem 2, page 47, of Basawa and Scott (1983)) to the effect that if {Tt} is any other sequence of consistent estimators of such that dt(Tt - ) d - U (4.15) for some nondegenerate U, then U d = V + -1 W, where V is independent of and W. It is readily shown that confidence intervals based on the pivot in (4.15) are wider than those for the corresponding result (4.14). That is, using (4.13) is better on average than (4.14), which is better than (4.15). If is the standard normal distribution function, we have for any real a, , with > , (( - a) y) - (( - a) y) 1 2 ( - ) y - - 1 2 ( - ) y (4.16) and hence, integrating with respect to dP( y), P( < -1 W + a < ) P - 1 2 ( - ) < -1 W < 1 2 ( - ) , (4.17) so that P( < -1 W + V < ) = P( < -1 W + v < ) dP(V v) P - 1 2 ( - ) < -1 W < 1 2 ( - ) . 64 CONFIDENCE ZONES OF MINIMUM SIZE Remark 3. A particular case of the result of the theorem has been obtained by Glynn and Iglehart (1990) who studied the problem of finding minimum size asymptotic confidence intervals for steady state parameters of the simulation output process from a single simulation run. They contrasted the approach of consistently estimating the variance constant in the relevant central limit theorem with the standardized time series approach which avoids estimation of the variance in a manner reminiscent of the t-statistic and suggested that the former approach is preferable on average. 4.5.2 Proof of Theorem 4.1 Suppose that t and t are exact, approximate or assumed distribution functions of the pivots ct(^t - ) and dt(^t - ), respectively. We are given the complete convergence results t c - , t c - as t , where (x) = P(W x) and (x) = P(-1 W x) = 0 (xy) dG(y) with G(x) = P( x). Now let at, bt be any numbers for which t(bt) - t(at) 1 - . This gives a confidence interval for of length l (1) t () = (bt - at)/ct. By passing to a subsequence, we may suppose that at a and bt b, where - a b . Then, by complete convergence, t(at) - t(bt) (b) - (a), and, in view of (4.16), b - a 2z, where 2(z) = 2 - . Therefore, noting that ctl (1) t () is not random, we have lim inf t ct L (1) t () 2z. (4.18) Next, let zt be the smallest z for which t(z) - t(-z) 1 - . As above, we may suppose that zt z and, because is continuous, t(zt) - t(-zt) = 1 - + o(1) 4.5. ON BEST ASYMPTOTIC CONFIDENCE INTERVALS 65 as t . By complete convergence and using the result (-x) = 1 - (x), x > 0, it follows that 2(z) = 2- and hence that z = z. Then, L (1) t () 2zt/ct, so that lim inf t ct L (1) t () 2z, (4.19) and (4.18) and (4.19) imply lim t ct L (1) t () = 2z. (4.20) Similar reasoning to that which led to (4.20) also applies to show that lim t dt L (2) t () = 2, (4.21) where 2() = 2 - . We merely replace (4.16) and (4.17) to show that the symmetric confidence interval is best, while (-x) = 1 - (x), x > 0, is evident from the definition and the corresponding result for . From (4.18) and (4.21), it then follows that, as t , dt L (2) t ()/ct L (1) t () /z. Since ct/dt d - , Slutsky's Theorem gives L (2) t ()/L (1) t () d - /z. and then, using Billingsley ((1968), Theorem 5.3, page 32), lim inf t E L (2) t ()/L (1) t () (E) /z. Thus, in order to complete the proof, it remains to show that (E) z. (4.22) Now note that = (; ) solves the equation ((; )) = P(-1 W (; )) = 1 - /2. Then, taking b > 0, we have 1 - 2 = P(W (; )) = P W 1 b (; ) b = P(W (b ; ) b ) so that continuity and strict monotonicity of imply that (b ; ) = 1 b (; ) and hence, if () = (E) (; ), 66 CONFIDENCE ZONES OF MINIMUM SIZE we have (b ) = (). (4.23) Now we see from (4.22) that it is required to show () -1 (1 - /2) and using (4.23) we may, without loss of generality, scale so that (; ) = 1. (4.24) But (4.24) implies (1) = 1 - /2, or equivalently, 0 (y) G(y) = E() = 1 - /2. Thus we have to show that () = E -1 (1 - /2) subject to E() = 1 - /2, and since is monotone, this holds if (E) E(). (4.25) However, since 1 - (x) is convex for x > 0, as is easily checked by differentiation, we have from Jensen's inequality that 1 - (E) E(1 - ()) and (4.25) follows. This completes the proof. Final Remarks It is worth noting that (4.22), which relates the average confidence intervals when (4.13) and (4.14) hold exactly rather than as limits, is a paraphrase of the fact that, on average, extraneous randomization does not improve confidence intervals for a normal mean. That is, if X is distributed as N(, 1), the usual symmetric confidence interval for based on X - is not improved, on average, if one instead uses the pivot (x - )/, where is independent of . In particular, when Xt, s2 t are the mean and variance of a random sample of size n from the N(, 2 ) distribution where is known, using the t-intervals for based on ( Xn - )/sn produces confidence intervals which are longer on average than z-intervals. 4.6. EXERCISES 67 4.6 Exercises 1. Suppose that X1, . . . , XT are i.i.d. random variables with mean 1/ and variance 2 . Show that a quasi-score estimating function for the estimation of is QT () = T(-1 - X), where X = T-1 T t=1 Xt and that a suitable choice for a bias corrected version is Q (BC) T () = QT () - 2 . In the case where the Xt have a Poisson distribution develop alternative forms of confidence intervals using (a) bias correction and (b) the quasi-score function directly. (Adapted from Firth (1993).) 2. Let X1, . . . , XT be a random sample from a distribution with mean and variance C2 , the coefficient of variation C being known. Taking C = 1 for convenience, consider the estimating function GT () = -2 T t=1 (Xt - ) + -3 T t=1 (Xt - )2 - 2 , which is the score function when the Xt are normally distributed. Show that a suitable choice for a bias corrected version of GT is G (BC) T () = GT () + 2 3 -1 . (Adapted from Firth (1993).) Chapter 5 Asymptotic Quasi-Likelihood 5.1 Introduction Discussion in earlier chapters on optimality of estimating functions and quasilikelihood has been concerned with exact results, where a specific criterion holds for either fixed T or for each T as T . Here we address the situation where the criteria for optimality are not satisfied exactly but hold in a certain asymptotic sense to be made precise below. These considerations give rise to an equivalence class of asymptotic quasi-likelihood estimator, which enjoy the same kind of properties as ordinary quasi-likelihood estimators, such as having asymptotic confidence zones of minimum size, within a specified family, for the "parameter" in question. One particular difficulty with the exact theory is that a quasi-likelihood estimator may contain an unknown parameter or parameters. The asymptotic quasi-likelihood theory herein allows one to focus on issues such as whether there is loss of information when an estimator is replaced by a consistent estimator thereof or, under some circumstances, is asymptotically irrelevant. As a simple illustration of the ideas take the case where {Xt} is a subcritical Galton-Watson branching process with immigration and = (m, ) is to be estimated on the basis of data {Xt, t = 0, 1, . . . , T}, where m and are, respectively, the means of the offspring and immigration distributions. This model has been widely used in practice, for example, for particle counts in colloidal solutions, and an account of various applications is given in Heyde and Seneta (1972) and Winnicki (1988). Noting that for the model in question the (t + 1)th generation is obtained from the independent reproduction of each of the individuals in the tth generation, each with the basic offspring distribution, plus an independent immigration input with the immigration distribution, we obtain E (Xt | Ft-1) = m Xt-1 + so that there is a semimartingale representation T t=1 Xt = m T t=1 Xt-1 + T + T t=1 mt, where mt = Xt - E (Xt | Ft-1) are martingale differences. 69 70 CHAPTER 5. ASYMPTOTIC QUASI-LIKELIHOOD Now suppose that X0 has a distribution for which E X2 0 < and that the variances of the offspring and immigration distributions are 2 (< ) and 2 (< ), respectively. Then using Theorem 2.1, the quasi-score estimating function based on the martingale { t s=1 ms} is QT = T t=1 Xt-1 (2 Xt-1 + 2 )-1 (Xt - m Xt-1 - ) T t=1 (2 Xt-1 + 2 )-1 (Xt - m Xt-1 - ) , (5.1) which in general involves the nuisance parameters 2 , 2 . If no immigration is present, so that = 2 = 0, the nuisance parameter 2 disappears in the estimating equation Qt = 0. Furthermore, if we have a parametric model based on m, alone, such as in the case where the offspring and immigration distributions are both Poisson, there is no nuisance parameter problem. Indeed, in the Poisson case it is straightforward to check that (5.1) is a constant multiple of the true score estimating function, which leads to the maximum likelihood estimators. In general, nuisance parameters must be estimated or avoided. In this case estimation is possible using strongly consistent estimators of 2 and 2 suggested by Yanev and Tchoukova-Dantcheva (1980). We can take these estimators as ^2 ( mT , T ) = T T t=1Xt-1 ^U2 t - T t=1Xt-1 T t=1 ^U2 t T T t=1 X2 t-1 - T t=1 Xt-1 2 , ^2 T ( mT , T ) = T t=1 ^U2 t T t=1 X2 t-1 - T t=1Xt-1 T t=1Xt-1 ^U2 t T T t=1 X2 t-1 - T t=1 Xt-1 2 where ^U2 t = Xt- mT Xt-1-T , mT and T being strongly consistent estimators of m and , respectively. Wei and Winnicki (1989) studied an estimating function closely related to (5.1) in which the term 2 Xt-1 + 2 is replaced by Xt-1 + 1. Details of the corresponding limit theory are given in Theorem 3.C of Winnicki (1988). Furthermore, for ^mT , ^T based on (5.1) with estimated 2 and 2 as suggested above (using ^mT , ^T ), a similar analysis to that undertaken by Wei and Winnicki shows that T 1 2 ^mT - m ^T - d - N(0, W -1 ) as T where, if X has the stationary limit distribution of the {Xt} process, W = E X2 (2 X + 2 )-1 E X (2 X + 2 )-1 E X (2 X + 2 )-1 E (2 X + 2 )-1 , the same result as one obtains if 2 and 2 are known. The quasi-likelihood framework ensures that this estimation procedure is optimal from the point of view of asymptotic efficiency. 5.2. THE FORMULATION 71 The substitution of estimated values for the nuisance parameters amounts to replacing the quasi-likelihood estimator by an asymptotic quasi-likelihood estimator that has the same asymptotic confidence zones for the unknown parameter. Exact solutions of the optimal estimation problem leading to ordinary quasilikelihood estimators require certain orthogonality properties, which are often only approximately satisfied in practice. For example, using Theorem 2.1 and omitting the explicit for convenience, if QT H G is OF -optimal within H, then for the standardized estimating functions G (s) T , Q (s) T H we need to have E G (s) T - Q (s) T G (s) T = E G (s) T G (s) T - Q (s) T = 0. (5.2) Here we shall typically be dealing with circumstances under which we have sequences of estimating functions and the quantities in (5.2) tend to zero as T . 5.2 The Formulation We confine our attention to the space G of sequences of zero mean and finite variance estimating functions {GT () = GT ({Xt, 1 t T}; )}, which are vectors of dimension p and are a.s. differentiable with respect to the component of and such that E ˙GT () and EGT ()GT () are nonsingular for each T, the prime denoting transpose. We shall write G (n) T = (E ˙GT )-1 GT for the normalized estimating function and drop the for convenience. It is convenient in this context to use a different standardization of estimating functions from that of Chapters 1 and 2 that we have referred to as G (s) T . The asymptotic optimality of a sequence of estimating functions will be defined as the maximization of E(GT ) in the partial order of nonnegative definite matrices in a certain asymptotic sense. Before the definition is spelt out, however, we need to discuss certain concepts concerning matrices. For a matrix A = (ai,j) we shall denote by A the Frobenius (Euclidean) norm given by A = i j a2 i,j 1 2 . A sequence of symmetric matrices {An} is said to be asymptotically nonnegative definite if there exists a sequence of matrices {Dn} such that An -Dn is nonnegative definite and Dn 0 as n . In the sequel we need square roots of positive definite matrices. Let A 1 2 (A T 2 ) be a left (resp. right) square root of the positive definite matrix A, i.e., A 1 2 A T 2 = A. In addition, let A- 1 2 = (A 1 2 )-1 , A- T 2 = (A T 2 )-1 . The left Cholesky square root matrix is defined as the unique lower triangular matrix satisfying the square root condition, and it can be calculated easily without solving any eigenvalue problems. Though we do not require any specific form 72 CHAPTER 5. ASYMPTOTIC QUASI-LIKELIHOOD of the square root matrix for our results to hold, Cholesky's square root matrix is preferred for ease of application. Definition 5.1 Suppose that {QT } H G, and EQ (n) T Q (n) T - 1 2 EG (n) T G (n) T EQ (n) T Q (n) T - T 2 - I is asymptotically nonnegative definite for all {GT } H, where I is the p × p identity matrix. Then we say that {QT } is an asymptotic quasi-score (AQS) sequence of estimating functions within H. The solution T of QT () = 0 will be called an asymptotic quasi-likelihood estimator from H. Remark 5.1 The choice of the subset H of estimating functions is open and it can be tailored to suit the particular problem. For example, features such as being linear or bounded functions of the data could be incorporated. Remark 5.2 If GT is OF -optimal within H for each T (see Heyde (1988a)), then {GT } is an asymptotic quasi-score sequence of estimating functions within H. Remark 5.3 The asymptotic quasi-score formulation provides a natural framework for extension of Rao's concept of asymptotic first order efficiency (e.g., Rao (1973, pp. 348-349)). In the case of a scalar parameter and estimator ZT , Rao's definition takes the form EU2 T 1 2 ZT - - () UT EU2 T P - 0 (5.3) as T for some () not depending on T or the observations. Here UT is the score function and the idea is that ZT - has basically the same asymptotic properties as the score function. For example, as indicated by Rao, the Fisher information in ZT is asymptotic to the Fisher information in UT under minor additional conditions. In this chapter we generalize this framework in a variety of ways. First, it is possible to confine attention to estimation within a subset of estimation functions, H say, in which case the benchmark role of an estimating function containing maximum information is a quasi-score estimating function, QT , say. In addition, we can replace the linear (or linearized) estimating function by a general normalized estimating function G (n) T = (E ˙GT )-1 GT . Now, since under usual regularity conditions for quasi-score estimating functions E ˙QT = -EQ2 T , we have Q (n) T = -(EQ2 T )-1 QT . 5.2. THE FORMULATION 73 Thus, we can think of EQ2 T 1 2 G (n) T - ()Q (n) T P - 0 (5.4) as a natural generalization of (5.3). If (5.4) holds, we could say that GT is asymptotically first order efficient for within H. Now, if {GT } is an asymptotic quasi-score sequence of estimating functions within H, we have E G (n) T 2 E Q (n) T 2 = EQ2 T -1 (5.5) as T . Also, under the standard regularity conditions EG (n) T Q (n) T = -(EQ2 T )-1 . Thus, in this case we have EQ2 T E G (n) T 2 - 2EG (n) T Q (n) T + E Q (n) T 2 2 0 and (5.4) holds with () = 1 because L2 convergence holds. It is thus clear that the asymptotic quasi-score condition is a sensible (sufficient) condition for Rao's asymptotic first order efficiency. It has the advantages of natural extensions to deal with the quasi-likelihood setting, the case where Tn is not a linear function of , and also the vector case. There are several different but equivalent versions of Definition 5.1 that illuminate the concept. These are given in Proposition 5.4. As pointed out in Chapter 2, for the ordinary quasi-score case, a definition such as that given above is of limited practical use. However, the following propositions provide a rather easy route to checking whether a sequence of estimating functions has the asymptotic quasi-score property. Proposition 5.1 The sequence of estimating functions {QT } H is AQS if there exists a p × p matrix kT such that lim T kT EG (n) T Q (n) T kT = K = lim T kT EQ (n) T Q (n) T kT (5.6) for all {GT } H, where K is some nondegenerate p × p matrix. In particular, if EQ (n) T Q (n) T - 1 2 EG (n) T Q (n) T EQ (n) T Q (n) T - T 2 I (5.7) as T for all {GT } H, or if there exists constant T such that lim T -1 T EG (n) T Q (n) T = K for all {GT } H and some nondegenerate matrix K, then {QT } is AQS. 74 CHAPTER 5. ASYMPTOTIC QUASI-LIKELIHOOD Proof. Suppose that (5.6) holds for some {kT }. It is easy to see that kT is nondegenerate when T is large enough. Given any {GT } H, let AT kT EG (n) T G (n) T - EG (n) T Q (n) T EQ (n) T Q (n) T -1 EG (n) T Q (n) T kT ; then AT is nonnegative definite using a standard argument based on properties of covariance matrices (e.g. Heyde and Gay (1989, p. 227)). We may write AT = kT EG (n) T G (n) T - EQ (n) T Q (n) T kT + kT EQ (n) T Q (n) T kT - kT EG (n) T Q (n) T kT kT EQ (n) T Q (n) T kT -1 kT EG (n) T Q (n) T kT = kT EG (n) T G (n) T - EQ (n) T Q (n) T kT + DT , say. It follows from (5.6) that DT 0 as T . Thus kT EG (n) T G (n) T - EQ (n) T Q (n) T kT is asymptotically nonnegative definite. This is equivalent, as shown in Proposition 5.4 below, to saying that {QT } is AQS within H. The two particular cases are accomplished by taking kT = (EQ (n) T Q (n) T )- 1 2 and kT = - 1 2 T , respectively. This completes the proof. Remark 5.4 (5.6) and (5.7) are actually equivalent. On the grounds that one seeks to maximize E(GT ), i.e., to minimize the estimating function variance EG (n) T G (n) T , for GT H, it would be sensible to avoid those estimating functions for which asymptotic relative efficiency comparisons are arbitrarily poor. With this in mind we give the following definition. Definition 5.2 The sequence of estimating function {GT } H is unacceptable if there exists a sequence {QT } H such that EQ (n) T Q (n) T - 1 2 EG (n) T G (n) T EQ (n) T Q (n) T - T 2 - I is unbounded as T . If we denote by HU H the subset of all unacceptable estimating functions, the complementary set H - HU consists of all the acceptable estimating func- tions. 5.2. THE FORMULATION 75 Proposition 5.2 A sequence of estimating functions {GT } H is unacceptable if and only if there exists a sequence of estimating functions {QT } H and vectors {cT } such that lim T cT EG (n) T G (n) T cT cT EQ (n) T Q (n) T cT = . The proof is fairly straightforward from the definition of unacceptability and we omit the details. Proposition 5.3 Suppose {QT } H is AQS where H is a linear space, then (5.7) holds for all {GT } H-HU . On the set H-HU , (5.6) is both necessary and sufficient for the AQS property. Proof. For T an arbitrary p×p matrix, consider the sequence of estimating functions {(E ˙GT ) T Q (n) T + GT } that belongs to H since H is a linear space. Because {QT } is AQS within H, by the definition of AQS and the fact that EQ (n) T = I, we have, upon writing L = E Q (n) T Q (n) T , limT kT L 1 2 E ˙GT T + E ˙GT -1 E (E ˙GT )T Q (n) T + GT Q (n) T T E ˙GT + GT T E ˙GT + E ˙GT -1 L- T 2 kT - kT kT 0, (5.8) for arbitrary sequence of unit vectors {kT }. For ease of notation, set M = EG (n) T Q (n) T and N = EG (n) T G (n) T . The quantity on the left side of (5.8) for fixed T can be written as kT L- 1 2 (T + I)-1 (E ˙GT )-1 E (E ˙GT )T Q (n) T + GT Q (n) T T E ˙GT + GT - (E ˙GT )(T + I)L(T + I)E ˙GT (E ˙GT )-1 (T + I)-1 L- T 2 kT = kT L- 1 2 (T + I)-1 T M + M T - T L - L T + (N - L) (T + I)-1 L- T 2 kT = kT L- 1 2 (T + I)-1 L 1 2 L- 1 2 T L 1 2 L- 1 2 M L- T 2 - I 76 CHAPTER 5. ASYMPTOTIC QUASI-LIKELIHOOD + (L- 1 2 M L- T 2 - I)(L- 1 2 T L 1 2 ) + L- 1 2 N L- T 2 - I L T 2 (T + I)-1 L- T 2 kT = kT CT (C-1 T - I)(L- 1 2 M L- T 2 - I) +(L- 1 2 M L- T 2 - I) (C-1 T - I) + L- 1 2 N L- T 2 - I CT kT ( where CT = L- 1 2 (T + I)-1 L 1 2 ) = kT (I - CT ) L- 1 2 M L- T 2 - I CT kT (5.9) + kT CT L- 1 2 M L- T 2 - I (I - CT ) kT + kT CT L- 1 2 N L- T 2 - I CT kT . Since {GT } H-HU and L- 1 2 N L- T 2 -I is bounded, there exists a constant A large enough such that x L- 1 2 N L- T 2 - I x A x 2 (5.10) for any p dimensional vector x. Suppose (5.7) does not hold. Then there exist two subsequences of vectors, denoted by {xT } and {yT } for simplicity in notation, such that yT = (L- 1 2 M L- T 2 - I)xT , yT = 1 and xT A. Because kT is an arbitrary unit vector, we can choose kT = yT . Moreover, since T is an arbitrary matrix, so is CT . Hence we may choose CT so that CT yT = xT , where = - 1 2A+A2 . Using (5.10), the quantity in (5.9) is less than or equal to 2 - 22 +A 2 xT 2 2 + 22 A + 2 A2 = < 0, which contradicts (5.8). This completes the proof of (5.7) and the remaining part of the proof follows immediately from Proposition 5.1. The next proposition gives five equivalent versions of the definition of asymptotic quasi-score. Proposition 5.4 The following conditions are equivalent. 1) EQ (n) T Q (n) T - 1 2 EG (n) T G (n) T EQ (n) T Q (n) T - T 2 - I is asymptotic nonnegative definite for all {GT } H. 5.2. THE FORMULATION 77 2) I - EG (n) T G (n) T - 1 2 EQ (n) T Q (n) T EG (n) T G (n) T - T 2 is asymptotic nonnegative definite for all {GT } H. 3) For any nonzero p-dimensional vector {cT }, limT cT EG (n) T G (n) T cT cT EQ (n) T Q (n) T cT 1 for all {GT } H. 4) There exists a sequence of p×p matrices {kT } such that kT EQ (n) T Q (n) T kT K nondegenerate, and kT EG (n) T G (n) T kT - kT EQ (n) T Q (n) T kT is asymptotic nonnegative definite for all {GT } H. 5) There exists a sequence of p × p matrices {kT } such that kT EQ (n) T Q (n) T -1 kT K nondegenerate, and kT EQ (n) T Q (n) T -1 kT kT EG (n) T G (n) T -1 kT is asymptotic nonnegative definite. Proof. 1) 3) Let xT be an arbitrary unit vector. Then, limT xT EQ (n) T Q (n) T - 1 2 EG (n) T G (n) T EQ (n) T Q (n) T - T 2 xT 1. Let cT = a EQ (n) T Q (n) T - T 2 xT ; then limT cT EG (n) T G (n) T cT cT EQ (n) T Q (n) T cT 1. Here cT is an arbitrary vector because xT is an arbitrary unit vector and a is an arbitrary nonzero constant. 3) 1). Let xT = EQ(n) T Q(n) T T 2 cT EQ(n) T Q(n) T T 2 cT . Then, limT xT EQ (n) T Q (n) T - 1 2 EG (n) T G (n) T EQ (n) T Q (n) T - T 2 xT = limT cT EG (n) T G (n) T cT cT EQ (n) T Q (n) T cT 1. Here xT can be regarded as an arbitrary unit vector because cT is an arbitrary vector. 78 CHAPTER 5. ASYMPTOTIC QUASI-LIKELIHOOD 2) 3) can be shown in a similar fashion. 1) 4) and 2) 5) are obvious. 4) 3) For an arbitrary vector cT , limT cT EG (n) T G (n) T - EQ (n) T Q (n) T cT cT k-1 T k-1 T cT 0, so that limT cT EG (n) T G (n) T cT cT EQ (n) T Q (n) T cT - 1 cT EQ (n) T Q (n) T cT cT k-1 T k-1 T cT 0, and since limT cT EQ (n) T Q (n) T cT cT k-1 T k-1 T cT min(K) > 0, because K is nondegenerate and thus is positive definite, we have limT cT EG (n) T G (n) T cT cT EQ (n) T Q (n) T cT 1. 5) 3) can be proved similarly. The final proposition indicates that two sequences of asymptotic quasi-score estimating functions are asymptotically close under appropriate norms. Proposition 5.5 Suppose { ~GT } H is AQS, H is a linear space, and there exists a sequence of matrices {kT } such that lim T kT E ~G (n) T ~G (n) T kT = K where K is some nondegenerate matrix. Then the following statements are equivalent: (i) {QT } H is AQS, (ii) limT kT EQ (n) T Q (n) T kT = K, (iii) EQ (n) T Q (n) T - 1 2 E ~G (n) T ~G (n) T EQ (n) T Q (n) T - T 2 I as T , (iv) trace EQ (n) T Q (n) T - 1 2 E ~G (n) T ~G (n) T EQ (n) T Q (n) T - T 2 p as T , (v) det E ~ G (n) T ~ G (n) T det E Q(n) T Q(n) T 1 as T , 5.3. EXAMPLES 79 (vi) cT E ~ G (n) T ~ G (n) T cT cT EQ(n) T Q(n) T cT 1 as T for all p-dimensional vectors cT = 0, (vii) kT E ~G (n) T - Q (n) T ~G (n) T - Q (n) T kT 0 as T . The proof of this proposition is straightforward and we omit the details. It is worthwhile to point out that under a small perturbation within H, an asymptotic quasi-score is still asymptotic quasi-score. (This can be seen from (vii) in Proposition 5.5.) One can see that the relative difference of the information about the parameter contained in two sequences of asymptotic quasi-score estimating functions is arbitrarily small under suitable normalization. 5.3 Examples 5.3.1 Generalized Linear Model A simple application involves the generalized linear model where the parameter is a p-vector . The distribution of the response yn is assumed to belong to a natural exponential family, which may be written as f(yn| n) = c(yn) exp(nyn - b(n)) and n and Zn are related by a link function, Zn being a p-vector of covariates. We assume yn and n are one dimensional for simplicity. Suppose the link is such that n = u(Zn) for some differentiable function u. The score function is ~GT = T i=1 Zi ˙u(Zi)(yi - ˙b(u(Zi))), which is the optimal estimating function within HT = { T i=1 Ci(yi-˙b(u(Zi)))} for all p-vectors Ci, i = 1, 2, .... Now suppose the assumption of the exponential family distribution of yn and the link function u is slightly relaxed to only assuming that yn has mean ˙b(u(Zn)) and finite variance 2 n. This guarantees that { ~GT } is within H. Now ~GT is no longer necessarily a score or quasi-score estimating function. However, under the assumption E ~G (n) T ~G (n) T 0 and ¨b(u(ZT )) 2 T 1 as T , one can still show that { ~GT } remains an AQS sequence within T =1HT . 5.3.2 Heteroscedastic Autoregressive Model Our second example concerns a d-dimensional heteroscedastic autoregressive model. Let {Xk, k = 1, 2, ..., T} be a sample from a d-dimensional process satisfying Xk = Xk-1 + k, 80 CHAPTER 5. ASYMPTOTIC QUASI-LIKELIHOOD where is a d×d matrix to be estimated and { k; k = 1, 2, ...} is a d-dimensional martingale difference sequence such that E k = 0 and E k k = k. One can conveniently write this model as Xk = (Xk-1 Id) + k, where = vec () = (11, ..., d1, 21, ..., d2, ..., d1, ...dd) , the vector obtained from by stacking its columns one on top of the other and for matrices A = (Aij) and B the Kronecker product, A B denotes the matrix whose (i, j), the block element, is the matrix AijB. The quasi-score estimating function based on the martingale difference sequence { k} is QT () = T k=1 (Xk-1 Id) -1 k (Xk - (Xk-1 Id)), which contains the generally unknown k. Let GT () = T k=1 (Xk-1 Id) (Xk - (Xk-1 Id)). We shall show that {GT } is a sequence of asymptotic quasi-score estimating functions under the conditions that k , a nondegenerate matrix. That is, GT is a quasi-score for the model when k is constant in k but unknown, and {GT } remains as AQS when k is not constant but converges to a constant. Now, writing V k for E(Xk-1Xk-1), we have EQ (n) T Q (n) T = T k=1 V k -1 k -1 , EG (n) T G (n) T = T k=1 V k Id -1 T k=1 V k k T k=1 V k Id -1 , it being easily checked that E ˙QT = - T k=1 V k -1 k = -EQT QT , E ˙GT = - T k=1 V k Id, EGT GT = T k=1 V k k, since for a d × d matrix M, E Xk-1 Id M Xk-1 Id = V k M. 5.3. EXAMPLES 81 Since -1 k -1 , a positive definite matrix, for any d-dimensional vector sequence {bk}, bk-1 k bk bk-1 bk 1. (5.11) Since V k is nonnegative definite, we can write V k = T T for some matrix T = (tij)d×d. For any d2 -dimensional vector sequence {ck}, let ck = (ck1, ck2, ..., ckd) where cki is a d-dimensional vector for 1 i d. Now we can write ck(V k -1 k ) ck = d l=1 d i=1 til cki -1 k d i=1 til cki . Replacing -1 k by -1 , the above formula still holds. Therefore it follows from (5.11) that ck(V k -1 k ) ck ck(V k -1 ) ck 1. (5.12) Also, min(E(QT )) gives cT ( T k=1 V k -1 k ) cT (5.13) as T . Combining (5.12) and (5.13), it follows that cT ( T k=1 V k -1 k ) cT cT ( T k=1 V k -1 ) cT 1 as T . But since {cT } can be an arbitrary unit vector sequence, so we have det( T k=1 V k -1 k ) det( T k=1 V k -1 ) 1 as T , and lim T det(EQ (n) T Q (n) T ) det(EG (n) T G (n) T ) = lim T det(( T k=1 V k) -1 )-1 (det( T k=1 V k))-ddet(( T k=1 V k) )(det( T k=1 V k))-d = lim T (det( T k=1 V k))-d (det())d (det( T k=1 V k))-d(det( T k=1 V k))d(det())d(det( T k=1 V k))-d = 1. Because QT is a quasi-score, by Proposition 5.4 Condition (v), we know {GT } is an AQS sequence of estimating functions. 82 CHAPTER 5. ASYMPTOTIC QUASI-LIKELIHOOD 5.3.3 Whittle Estimation Procedure As a final illustration of the AQL methodology we shall show that the widely used estimation procedure based on the smoothed periodogram, which dates back to remarkable early work of Whittle (1951), provides an AQL-estimator under a broad range of conditions. These encompass many random processes or random fields with either short or long range dependence. Here we shall deal with the one-dimensional random process case. For the multivariate random process and random fields cases see Heyde and Gay (1989), (1993). Suppose that the stationary random process {Xt}, with E Xt = 0, is observed at the times t = 1, 2, . . . , T and let IT () = 1 2 T T i=1 Xj e-ij 2 be the corresponding periodogram. The spectral density of the random process is f(; , 2 ), where the one-step prediction variance is 2 = 2 exp (2)-1 log f(; , 2 ) d , (5.14) and we shall write g(; ) = 2 -2 f(; , 2 ). We shall assume that 2 is specified while , of dimension p, is to be estimated. It should be noted that (5.14) gives log g(; ) d = 0, which is equivalent to the fact that -1 Xt has prediction variances independent of (e.g., Rosenblatt (1985, Chapter III, Section 3)). We shall initially assume further that f is a continuous function of , including at = 0 (short range dependence), and is continuously differentiable in as a function of (, ). We consider the class of estimating functions GT = A() [IT () - EIT ()] d constructed from smoothed periodograms where the smoothing function A() (a vector of dimension p) is square integrable and symmetric about zero, that is, A () A() d < , A() = A(-). We shall show that under a considerable diversity of conditions, an asymptotic quasi-score sequence of estimating functions for is given by G T = (g(; ))-2 g(; ) [IT () - EIT ()] d 5.3. EXAMPLES 83 = (g(; ))-2 g(; ) [IT () - f(; , 2 )] d + o(1) as T . The corresponding AQL estimator T obtained from the estimating equation G T () = 0 is then asymptotically equivalent to the Whittle estimator obtained by choosing to minimize log f(; , 2 ) + IT ()(f(; , 2 ))-1 d, (5.15) i.e., to solve (g(; ))-2 g(; ) [IT () - f(; , 2 )] d = 0. The idea to minimize (5.15) comes from the observation that if the random process Xt has a Gaussian distribution, it is plausible that T-1 times the log likelihood of the data could be approximated by - 1 2 log 2 2 - (2 2 )-1 IT ()(g(; ))-1 d = - log 2 - 1 2(2) log f(; , 2 ) + IT ()(f(; , 2 ))-1 d. It is, in fact, now a commonly used strategy in large sample estimation problems to make the Gaussian assumption, maximize the corresponding likelihood, and then show that the estimator still makes good sense without the Gaussian assumption (e.g., Hannan (1970, Chapter 6, Section 6; 1973)). Our first specific illustration concerns the linear process Xt = u1=- guet-u, with Eet = 0, Eeset = 2 if s = t, 0 otherwise, and u1=- g2 u < . There are various somewhat different regularity conditions under which we could proceed and since the discussion is intended to be illustrative rather 84 CHAPTER 5. ASYMPTOTIC QUASI-LIKELIHOOD than to present minimal conditions, we shall, for convenience, use results from Section 6, Chapter IV of Rosenblatt (1985). The same conclusions, can be deduced under significantly different conditions using results of various other authors (e.g., via Theorem 4 of Parzen (1957) or Theorem 6.2.4, p. 427 of Priestley (1981)). We suppose that {Xt} is a strongly mixing stationary random process with EX8 t < and cumulants up to eighth order absolutely summable. Then, from the discussion leading up to Theorem 7, p. 118 of Rosenblatt (1985), we find that if GT = A()[IT () - EIT ()] d, G T = - A ()[IT () - EIT ()] d, then TEGT G T (2) 2 - A()(A ()) f2 () d + f4(, -, )A()(A ()) d d (5.16) as T , with f() now being written instead of f(; , 2 ). Here f4(, , ) is the fourth order cumulant spectral density f4(, , ) = (2)-3 a,b,d ca,b,d e-i(a+b +d ) with ca,b,d = cum(Xt, Xt+a, Xt+b, Xt+d) = ( - 3)4 u gugu+agu+bgu+d, ( - 3)4 being the fourth cumulant of Xt (which is zero in the case where the process is Gaussian). But, analogously to the discussion of Chapter II, Section 4 of Rosenblatt (1985, pp. 46, 47), we find that f4(, -, ) = (2)-1 ( - 3)4 f()f(), and hence the second term on the right-hand side of (5.16) is (2)-1 ( - 3) 4 A()f() d - A ()f() d , which is zero if the random field is Gaussian or if the class of smoothing functions {A()} is chosen so that A()f() d = 0. 5.3. EXAMPLES 85 We shall henceforth suppose that either (or both) of these conditions hold and then T EGT G T 2(2) - A()(A ()) f2 () d (5.17) as T . Also, E ˙GT = - - A() EIT () d - A()( ˙f()) d, (5.18) provided ˙f() is square integrable over [-, ], since EIT () ˙f() in the L2 norm by standard Fourier methods (e.g. Hannan (1970, p. 508)). Then, from (5.17) and (5.18) we see that (5.6) holds with kT = T if A () = ˙f()(f())-2 , which gives the required AQL property. Under the same conditions and a number of variants thereof it can be further deduced that if T is the estimator obtained from the estimating equation GT () = 0, then (see e.g. the discussion of the proof of Theorem 8, p. 119 of Rosenblatt (1985)) T1/2 (T - 0) d - Np 0, (A B-1 A)-1 , where A = - A()( ˙f()) d, B = - A()A ()f2 () d and that the inverse covariance matrix A B-1 A is maximized in the partial order of nonnegative definite matrices for the AQL case where A() takes the value A () = ˙f()(f())-2 . These considerations place the optimality results for random processes of Kulperger (1985, Theorem 2.1 and Corollary) and Kabaila (1980, Theorem 3.1) into more general perspective. The matrix maximization result mentioned above appears in the latter paper. 86 CHAPTER 5. ASYMPTOTIC QUASI-LIKELIHOOD The next specific example on the Whittle procedure concerns a random process exhibiting long range dependence. We shall consider the case where the spectral density f(x; , 2 ) 2 |x|-() L(x) as x 0, where 0 < () < 1 and L(x) varies slowly at zero. A typical example is provided by the fractional Brownian process. This is a stationary Gaussian process with zero mean and covariance E(XnXn+k) = 1 2 c |k + 1|2H - 2|k|2H + |k - 1|2H c H(2H - 1)k2H-2 as k , where H is a parameter satisfying 1 2 < H < 1 and c > 0, while f(x; H) 1 2 cF(H)|x|1-2H as x 0 with f(H) = (1 - cos x)|x|-1-2H dx -1 . The Whittle estimation procedure (5.15) for in this context and subject to the assumption of a Gaussian distribution has been studied in some detail by Fox and Taqqu (1986), (1987) and Dahlhaus (1988). In particular, it follows from Theorem 1 of Fox and Taqqu (1987) that if GT = A(x)[IT (x) - EIT (x)] dx with A(x) = O(|x|)-- as x 0 for some < 1 and each > 0 and + < 1 2 , then TEGT GT 4 - f2 (x)A(x)A (x) dx (5.19) as T ; see also Proposition 1 of Fox and Taqqu (1986). Furthermore, provided that f j La , 1 j p, 1 a < -1 , then j EIT (x) f j in the La norm (e.g., Hannan (1970, p. 508)) and from H¨older's inequality we find that E ˙GT = - - A(x) EIT (x) dx - A(x)( ˙f(x)) dx. (5.20) 5.3. EXAMPLES 87 Then, from (5.19) and (5.20), 4T-1 (E ˙GT ) (EGT GT )-1 (E ˙GT ) A(x)( ˙f(x)) dx (5.21) - f2 (x)A(x)A (x) dx -1 A(x) ˙f(x)) dx and the right-hand side of (5.21) is maximized in the partial order of nonnegative definite matrices when A(x) takes the value A (x) = ˙f(x)(f(x))-2 (Kabaila (1980, Theorem 3.1)). This gives the AQL property for the corresponding {G T } by direct application of Definition 5.1. 5.3.4 Addendum to the Example of Section 5.1 Earlier approaches to this estimation problem have used conditional least squares or ad hoc methods producing essentially similar results. This amounts to using the estimating function (5.1) with the terms 2 Xt-1 + 2 removed. For references and details of the approach see Hall and Heyde (1980, Chapter 6.3). The essential point is that quasi-likelihood or asymptotic quasi-likelihood will offer advantages in asymptotic efficiency and these may be substantial. The character of the limiting covariance matrices associated with different estimators precludes direct general comparison other than via inequalities. We therefore give a numerical example as a concrete illustration. This concerns the Smoluchowski model for particles in a fluid in which the offspring distribution is assumed to be Poisson and the immigration distribution to be Bernoulli; for a discussion see Heyde and Seneta (1972). The data we use is a set of 505 observations from F¨urth (1918, Tabelle 1). We shall compare the asymptotic covariance matrices which come from the methods (i) conditional least squares (e.g., Hall and Heyde (1980, Chapter 6.3); Winnicki (1988, Theorem 3.B)), (ii) Wei and Winnicki's weighted conditional least squares (e.g., Winnicki (1988, Theorem 3.C)), (iii) asymptotic quasi-likelihood. All quantities are estimated from the data, and since the covariance matrices are all functions m and , common point estimates of m and (those which come from (i)) are used as a basis for the comparison. Methods (ii), (iii) produce slightly different point estimates. We find that ^m = 0.68, ^ = 0.51 and that the estimated asymptotic covariance matrices for (505) 1 2 ( ^m - m, ^ - ) in the cases (i), (ii), (iii) are, respectively, 0.81 -1.04 -1.04 2.05 , 0.76 -0.74 -0.74 1.57 , 0.75 -0.73 -0.73 1.56 . The reduction in variance estimates obtained by using (iii) rather than (i) or (ii) in this particular case is not of practical significance. Approximate 95% 88 CHAPTER 5. ASYMPTOTIC QUASI-LIKELIHOOD confidence intervals for m are 0.68 0.08 in each case, and for , (i) gives 0.51 0.12 while (ii) and (iii) improve this to 0.51 0.11. Although quasi-likelihood does not perform better than the simpler methods here, it nevertheless remains as the benchmark for minimum width confidence intervals. 5.4 Bibliographic Notes The bulk of this chapter follows Chen and Heyde (1995), a sequel to Heyde and Gay (1989). The branching process discussion is from Heyde and Lin (1992). If {GT } is an asymptotic quasi-score sequence under Definition 5.2, then it is also under the definition given in Heyde and Gay (1989), which is relatively weaker. In particular, the two definitions are equivalent if the parameter is one dimensional or if max EQ (n) T Q (n) T /min EQ (n) T Q (n) T is bounded for all T where max() (min()) denotes the largest (smallest) eigenvalue. A corresponding theory to that of this chapter, based on conditional covariances rather than ordinary ones, should be able to be developed straightforwardly. This would be easier to use in some stochastic process problems. 5.5 Exercises 1. Suppose that {X1, X2, . . . , XT } is a sample from a normal distribution N(, 2 ), where 2 is to be estimated and is a nuisance parameter. Recall that T t=1(Xt - )2 has a 2 2 T distribution and T t=1(Xt - X)2 a 2 2 T -1 distribution, X being the sample mean. Show that T t=1 [(Xt - X)2 - 2 ] can be interpreted as an AQS estimating function for 2 . 2. Let {X1, X2, . . . , XT } be a sample from the process Xt = Xt-1 + ut, where > 1 is to be estimated and if Ft = (Xt, . . . , X1) are the pasthistory -fields, E(ut | Ft-1) = 0, E(u2 t | Ft-1) = 2 X2 t-1 + 2 Xt-1, 2 and 2 being nuisance parameters. Consider the family of estimating functions H = { T t=1 t(Zt - Xt-1)} where t's are Ft-1-measurable and show that, on the set {Xt }, T t=1 X-1 t-1(Xt - Xt-1) is an AQS estimating function. 5.5. EXERCISES 89 3. For a regression model of the form Y = () + e with Ee = 0 and V = Ee e (such as is studied in Section 2.4), suppose that the quasi-score estimating function from the space H = {A(Y - ())} for p×p matrices A not depending on for which E(A ˙) and E(A e e A ) are nonsingular and E(e e | ˙) = V is ˙ V -1 (Y -) (see, e.g., Theorem 2.3). Suppose that V is not known and is replaced by an approximate covariance matrix W . Show that ˙ W -1 (Y -) is an asymptotic quasiscore estimating function iff det(E ˙ W -1 V W -1 ˙)/det(E ˙ V -1 ˙) 1 as the sample size T . This gives a necessary and sufficient condition for a GEE (see Section 2.4.3) to be fully efficient asymptotically. Chapter 6 Combining Estimating Functions 6.1 Introduction It is often possible to collect a number of estimating functions, each with information to offer about an unknown parameter. These can be combined much more readily than the estimators therefrom and, indeed, simple addition is often appropriate. This chapter addresses the issue of how to perform the combination optimally. Pragmatic considerations led to the combination of likelihood scores. Likelihood based inference encounters a range of possible problems including computational difficulty, information on only part of the model, questionable distributional assumptions, etc. As a compromise, the method of composite likelihoods, essentially the summing of any available likelihood scores of relevance was suggested (e.g., Besag (1975), Lindsay (1988)). A good example concerns observations on a lattice of points in the plane where the conditional distributions of the form f(yi | y[i]) are specified, the y[i] denoting a prescribed set of variables, for example the nearest neighbors of i. Let y be the vector of the yi's in dictionary order and write W = (wij) for the matrix of 0's and 1's such that wii = 0, wij = wji = 1 if i and j are neighbors, while wi denotes the ith column of W . Then, if yi | y[i] has a N( wi y, 2 ) distribution for parameters , 2 , there is a consistent joint distribution of the form y d = N(0, 2 (I - W )-1 ), 2 = 2 (, ). Taking 2 as known and setting it equal to 1 for convenience, it is easily seen that the score function associated with y is y W y + d d log det (I - W ). There are fundamental computational difficulties in estimating from this. On the other hand, the composite score obtained from summing the conditional scores of the yi's is y W y - y W 2 y, and this is simple to use. This kind of approach is easy to formalize into a framework of composite quasi-likelihood, and this will be the topic of the next section. The setting that we shall consider is a general one in which maximum likelihood may be difficult or impossible to use; the context may be a semiparametric 91 92 CHAPTER 6. COMBINING ESTIMATING FUNCTIONS one, for example. However, we suppose that we are able to piece together a set of quasi-likelihood estimating functions, possibly based on conditional or marginal information. The problem is how to combine these most efficiently for estimation purposes. 6.2 Composite Quasi-Likelihoods Let {Xt, t T} be a sample of discrete or continuous data that is randomly generated with values in r-dimensional Euclidean space whose distibution involves a "parameter" taking values in an open subset of p-dimensional Euclidean space. The true value of the "parameter" is 0 and this is to be estimated. Suppose that the possible probability measures for {Xt} are {P} and that each (, F, P) is a complete probability space. We shall as usual confine attention to the class G of zero mean square integrable estimating functions GT = GT ({Xt, t T}, ) for which EGT () = 0 and E ˙GT () and EGT () GT () are nonsingular for each P. Here GT is a vector of dimension p. We consider a class of K G of estimating functions GT that are a.s. differentiable with respect to the components of and such that E ˙GT = EGT,i j and EGT GT are nonsingular, the prime denoting transpose. We shall suppose that the setting is such that k distinct estimating functions Gi,T (), 1 i k, are available. Our first approach to the composition problem is to seek a quasi-score estimating function from within the set K = k i=1 i,T () Gi,T () where the weighting matrices i,T are p × p constants. The discussion here follows Heyde (1989a). Suppressing the dependence on T and for convenience, we write H = k i=1 i Gi = G, H = k i=1 i Gi = G, where = 1 ..... . . . ..... k , = 1 ..... . . . ..... k , G = G1 ..... . . . ..... Gk , 6.3. COMBINING MARTINGALE ESTIMATING FUNCTIONS 93 so that E ˙H = k i=1 i E ˙Gi = E ˙G, EH H = k i=1 k j=1 i EGi Gj j = EG G . Then, from Theorem 2.1 we see that H is a quasi-score estimating function within K if = (EG G )-1 E ˙G. This provides a general solution, although it may be less than straightforward to calculate in practice. Note in particular that since each individual Gi K, the formulation ensures that E(H ) - E(Gi) is nnd, 1 i k, so that composition is generally advantageous. Various simplifications are sometimes possible. For example, if the Gi are standardized quasi-score (or true score) estimating functions, then we can suppose that each Gi satisfies -E ˙Gi = EGi Gi, 1 i k. Circumstances under which equal weighting is obtained are especially important. Clearly this holds in the particular case when EGi Gj = 0, i = j, that is, the estimating functions are orthogonal. For an application to estimation in hidden-Markov random fields see Section 8.4. If the estimating functions are not orthogonal, then we can adjust them using Gram-Schmidt orthogonalization to get a new set Ki, 1 i k of mutually orthogonal estimating functions. For example, in the case k = 2 we can replace G1 and G2 by H1 = G1 and H2 = G2 - (EG2 G1)(EG1 G1)-1 G1 and H1, H2 are orthogonal. There is also another approach to the problem of composition that is closely related but quite distinct, and this is treated in the next section. It is applicable only to the case of martingale estimating functions but it is nevertheless of use in a wide variety of contexts. As has been noted earlier, many models admit a semimartingale description, which naturally leads to martingale estimating functions and, furthermore, true score functions are martingales under mild conditions. It is principally spatial processes that are excluded. 6.3 Combining Martingale Estimating Functions Section 6.2 deals with the general question of finding a quasi-score by combination of functions. However, as noted earlier in Section 2.5, it is often useful 94 CHAPTER 6. COMBINING ESTIMATING FUNCTIONS to adopt a martingale setting. For example, in many cases there is actually an underlying likelihood score U, say, which is typically unknown. Then U will be a square integrable martingale under minor regularity conditions. This suggests the use of a martingale family of estimating functions as the most appropriate basis for approximation of U. Indeed, if the family is large enough to contain U, then the quasi-score estimating function on that space will be U. Again we let (Xt, 0 t T) be a sample from a process taking values in r-dimensional Euclidean space whose distribution depends on a parameter belonging to an open subset of p-dimensional Euclidean space. The chosen setting is one of continuous time but the theory also applies to the discrete time case. This is dealt with by replacing a discrete time process (Xn, n 0) by a continuous version (Xc t, t 0) defined through Xc t = Xn, n t < n+1. The discussion here follows Heyde (1987). Suppose that the possible probability measures for (Xt) are (P) and that each (, F, P) is a complete probability space. The past-history -fields (Ft, t 0) are assumed to be a standard filtration. That is, Fs Ft F for s t, F0 being augmented by sets of measure zero of F and Ft = Ft+, where Ft+ = s>t Fs. Let M denote the class of square integrable estimating functions (GT , FT ) with GT = GT {(Xt, 0 t T), } that are martingales for each P and whose elements are almost surely differentiable with respect to the components of . Here GT is a vector of dimension d not necessarily equal p. Now for each martingale estimating function GT M there is an associated family of martingales T 0 as() dGs() where as() is a predictable matrix and the quasi-score estimating function from this family is T 0 (d Gs) (d G s)- dGs. As usual the prime denotes transpose and the minus generalized inverse. For n × 1 vector valued martingales MT and NT , the n × n process M, N T , called the mutual quadratic characteristic, is the predictable increasing process such that MT NT - M, N T is an n × n martingale. Also, we write M T for M, M T , the quadratic characteristic of MT . Finally, ˙MT is the n × p matrix obtained by differentiating the elements of MT with respect to those of and d Mt = E(d ˙Mt | Ft-). If we have two basic martingales HT and KT , which belong to M and are such that one is not absolutely continuous with respect to the other, then each gives rise to a quasi-score estimating function, namely, T 0 (d Ht) (d H t)- dHt and T 0 (d Kt) (d K t)- dKt, 6.3. COMBINING MARTINGALE ESTIMATING FUNCTIONS 95 respectively, and these may be regarded as competitors. If all the relevant quantities are known, a better estimating function may be obtained by combining them. We shall discuss the best procedure for combining the estimating functions and the gains that may be expected. In the particular case where the Xi are independent random variables with finite mean i() and variance 2 i (), natural martingales to consider for the estimation of are n i=1 {Xi - i()} and n i=1 (Xi - i())2 - 2 i () and the optimal linear combination of these has been considered by Crowder (1987) and Firth (1987). Here we treat the problem of combination generally. Given HT and KT we combine them into a new 2p × 1 vector martingale JT = (HT , KT ) and the quasi-score estimating function for this is T 0 (dJt) (d J t)dJt. (6.1) Now J t = H t H, K t H, K t K t (6.2) and if A and D are symmetric matrices and A is nonsingular, M- = A B B D - = A-1 + F EF -F E- -EF E- (6.3) where E = (D - B A-1 B), F = A-1 B, and M is nonsingular if A and E are nonsingular. Also, if E is nonsingular (A-1 + F E-1 F )-1 = A - B D-1 B . (6.4) The result (6.3) and its supplement follows from Exercise 2.4, p. 32 and a minor modification of Exercise 2.7, p. 33 of Rao (1973) while (6.4) comes from Exercise 2.9, p. 33 of the same reference. Suppose that d H t and d K t are nonsingular. Then, using (6.2) ­ (6.4) and after some algebra, we find that (6.1) can be expressed as T 0 (d Ht) (d H t)-1 (I - RtSt)-1 - (d Kt) (d K t)-1 (I - StRt)-1 St dHt + T 0 (d Kt) (d K t)-1 (I - StRt)-1 - (d Ht) (d H t)-1 (I - RtSt)-1 Rt dKt (6.5) 96 CHAPTER 6. COMBINING ESTIMATING FUNCTIONS where Rt = (d H, K t)(d K t)-1 , St = (d H, K t) (d H t)-1 = (d K, H t)(d H t)-1 . Note that if H, K t 0, the quasi-score estimating function (6.5) becomes T 0 (d Ht) (d H t)-1 dHt + T 0 (d Kt) (d K t)-1 dKt, the sum of the quasi-score estimating functions for H and K. The quasi-score estimating function (6.5) can be conceived as arising from the martingales H and K in the following way. Based on H, K we can consider the class of estimating functions T 0 at dHt + T 0 bt dKt where at, bt are predictable matrices. Then, the corresponding quasi-score estimating function is T 0 (at d Ht + bt d Kt) (d a H + b K t)-1 (at dHt + bt dKt) and the best of these, in the sense of minimum size asymptotic confidence intervals, is obtained by choosing for at, bt the values a t , b t , where a t , b t are such that T 0 (at d Ht + bt d Kt) (d a H + b K t)-1 (at d Ht + bt d Kt) is maximized. This result is given by (6.5). The formula (6.5) can also be easily obtained through Gram-Schmidt orthogonalization. Note that for the martingales G1T = HT and G2T = KT - T 0 d H, K t(d H t)-1 dHt we have G1, G2 0 so that, upon standardizing the estimating functions, we have as a combined quasi-score estimating function T 0 (d Ht) (d H t)-1 dHt + T 0 (d G2t) (d G2 t)-1 dG2t, which, after some algebra, reduces to (6.5). In order to compare martingale estimating functions we shall use the martingale information introduced in Section 2.5. For G M we define the martingale information IG as IG = GT ( G T )-1 GT (6.6) 6.3. COMBINING MARTINGALE ESTIMATING FUNCTIONS 97 when G T is nonsingular. Note that in the one-dimensional case where GT = UT , the score function of (Xt, 0 t T), IG is the conditional Fisher information. Also, the quantity IG occurs as a scale variable in the asymptotic distribution of the estimator obtained from the estimating equation GT ( ) = 0. The quantities IG for competing martingales can be compared in the partial order of nonnegative definite matrices. Note that the quasi-score estimating function based on G is Q S(G) = T 0 (d Gt) (d G t)-1 dGt and IQS(G) = T 0 (d Gt) (d G t)-1 d Gt. (6.7) In many cases there is one "natural" martingale suggested by the context such as through a semimartingale model representation as is described in Section 2.6. For example, this is certainly the case if (Xt) is representable in the form Xt = t 0 fs() ds + mt() (6.8) where (t) is a real, monotone increasing, right continuous process with 0 = 0, (mt(), Ft) M and {ft()} is predictable. This framework has been extensively discussed in Hutton and Nelson (1986) and Godambe and Heyde (1987). Various other martingales can easily be constructed from a basic martingale (mt(), Ft) to use in conjuction with it. The simplest general ones are, for d = 1, t 0 (dms())2 - m() t (6.9) (discrete time) and m2 t () - m() t = 2 t 0 ms-() dms() , (6.10) the last result following from Ito's formula. Generally if Hn(x, y) is the HermiteChebyshev polynomial in x and y defined by the generating function exp t x - 1 2 t2 y = n=0 tn n! Hn(x, y), then, for each n, Hn(mt, m t) is a martingale. See for example, Chung and Williams (1983, Theorem 6.4, p. 114). 98 CHAPTER 6. COMBINING ESTIMATING FUNCTIONS 6.3.1 An Example We consider the first order autoregressive process Xi = Xi-1 + i, where the i = i() are independent and identically distributed with Ei = 0, E2 i = 2 , E4 i < . Write -3 E3 i = and -4 E4 - 3 = for the skewness and kurtosis, respectively, of the distribution of . We wish to estimate on the basis of sample (Xi, 0 i T). Now the natural martingale for the process (Xi), which of course has a representation of the form (6.8), is Hj = j i=1 (Xi - Xi-1), j = 1, 2, . . . and the associated martingale given by (6.9) is Kj = j i=1 (Xi - Xi-1)2 - j2 , j = 1, 2, . . . . Put Hj = j i=1 hi, Kj = j i=1 ki. The combined quasi-score estimating function given by (6.5) is then 1 3( + 2) 1 - 2 + 2 -1 T i=1 {(2 ˙ - ( + 2)Xi-1)hi + (Xi-1 - 2 ˙)ki} , (6.11) where ˙ = d/d. Note that the quasi-score estimating functions based separately on H and K are - 1 2 T i=1 Xi-1hi, - 2 ˙ ( + 2)3 T i=1 ki, respectively. Thus, if ˙ = 0, the K martingale will contribute to the estimation of in combination with H if = 0 even though it is useless in its own right. If the i are normally distributed, then = 0, = 0 and (6.11) reduces to minus twice the score function. That is, quasi-likelihood and maximum likelihood estimation are the same. To compare the various estimators we calculate their corresponding martingale informations. We find, after some algebra, that IQS(H) = -2 T i=1 X2 i-1, IQS(K) = 4T( ˙)2 ( + 2)2 6.4. APPLICATION. NESTED STRATA OF VARIATION 99 and, writing Q S(H, K) for the combined quasi-score estimating function (6.11), IQS(H,K) = 1 2( + 2) 1 - 2 + 2 -1 T i=1 (( + 2)Xi-1 - 2)Xi-1 + 2 ˙(2 ˙ - Xi-1) = 1 - 2 + 2 -1 IQS(H) + IQS(K) - 4 ˙ 2( + 2) T i=1 Xi-1 . In the case where (Xi) is stationary (|| < 1) we have IQS(H) T-2 EX2 1 = T(1 - 2 )-1 a.s. and IQS(H,K) 1 - 2 + 2 -1 IQS(H) + IQS(K) a.s. as T . If, on the other hand, || 1, then IQS(K) = o(IQS(H)) a.s. and IQS(H,K) 1 - 2 + 2 -1 IQS(H) a.s. as T . Note that even in this latter case combining K with H is advantageous if = 0. 6.4 Application. Nested Strata of Variation Suppose we have a nested model in which e(R) represents the error term for stratum R. Let e denote all the e(R) written in lexicographic order as a column vector. Then, the obvious approach to quasi-likelihood estimation based on the family of estimating function {C e, C constant matrix} is not practicable. Because of the nesting, the e's are correlated across strata making Ee e difficult to invert. The problem can be avoided, however, by working with appropriate residuals. Define, within each stratum R, residuals r(R) = (I - W (R)) e(R) (6.12) where W (R) is idempotent so that W (R)r(R) = 0. Then, if W (R) is also chosen so that V (R) W (R) = W (R) V (R) (6.13) 100 CHAPTER 6. COMBINING ESTIMATING FUNCTIONS where V (R) = Ee(R) e(R), we have that a quasi-score estimating function (QSEF) based on the set of estimating functions {C r(R), C constant matrix} is E ˙e(R) Ee(R) e(R) -1 r(R). (6.14) The result (6.14) follows because (dropping the suffix (R) for convenience) the QSEF is, from Theorem 2.1, (E ˙r) (Er r )r = (E ˙e) (I - W ) ((I - W ) V (I - W ) )r, (6.15) the minus denoting generalized inverse. Further, in view of (6.13), (I - W ) V (I - W ) = V - W V - V W + W V W = (I - W ) V since W is idempotent and it is easily checked that ((I - W ) V )= V -1 (I - W ). (6.16) Also from (6.13), (I - W ) V -1 = V -1 (I - W ), (6.17) so that using (6.16), (6.17) and (I - W ) r = r, the right hand side of (6.15) reduces to (E ˙e) V -1 r as required. Now it is necessary to combine the QSEF from each of the strata and for this it is desirable that the residuals be orthogonal across strata. That is, we need Er(R) r(S) = 0 R, S. These requirements are not difficult to achieve in practice and then the combined QSEF is the sum over strata R (Ee(R) e(R))-1 (E ˙e(R)) r(R). (6.18) As a concrete example we shall consider a model introduced by Morton (1987) to deal with several nested strata of extra Poisson variation in a generalized linear model where multiplicative errors are associated with each stratum. The motivation for this modeling came from a consideration of insect trap catches, details of which are given in the cited paper. Suppose that there are three nested strata labeled 2, 1, 0 with respective subscripts i, j, k having arbitrary ranges. Stratum 0, the bottom stratum, has scaled Poisson errors; stratum 1 introduces errors Zij and the top stratum 2 has further errors Zi. It is assumed that the Z's are mutually independent with EZij = 1, EZi = 1, var Zij = 2 1, var Zi = 2 2 6.4. APPLICATION. NESTED STRATA OF VARIATION 101 and the model can be written as Xijk = ijk Zij Zi + eijk with the eijk uncorrelated and E eijk Zij, Zi = 0, var eijk Zij, Zi = ijk Zij Zi where is an unknown scale parameter. From these assumptions it is straightforward to calculate the covariance matrix V for the data vector x = (Xijk) , the elements being written in a row in lexicographic order. Define 1 = 2 1(2 2 + 1)/ and 2 = 2 2/. Then, the elements of V are var Xijk = ijk{1 + (1 + 2) ijk}, cov (Xijk, Xijn) = (1 + 2) ijk ijn (k = n), cov (Xijk, Ximn) = 2 ijk imn (j = m), cov (Xijk, Xlmn) = 0 (i = l). If = (ijk) depends on a vector of unknown parameters, then the quasiscore estimating function can be written as (/) V -1 (x - ). However, this is typically not a tractable form to use because of its dimension and the complexity of V . We shall use the theory developed above to obtain a useful version of the quasi-score estimating function by focusing on the sources of variation in each of the strata separately and optimally combining the resulting estimating functions. This contrasts with the approach of Morton (1987) where ideas based on partitioning of a sum of squares are used (in the spirit of the Wedderburn approach to quasi-likelihood). We write for the error terms in the three strata, eijk = Xijk - ijk Zij Zi, eij = Xij - ij Zi, ei = Xi - i, where the weighted totals are Xij = k Xijk, ij = k ijk, eij = k eijk, Xi = j wij Xij, i = j wij ij, ei = j wij eij, ei = j wij eij, 102 CHAPTER 6. COMBINING ESTIMATING FUNCTIONS with wij = ij /var ei = 1/(1 + 1 ij), and the residuals are rijk = Xijk - ijk ij Xij = eijk - ijk ij eij, rij = Xij - ij i Xi = eij - ij i ei, ri = Xi - i = ei. We find that var ei = i (1 + 2 i), var eij = ij (1 + 1 ij) and it is easily checked that the conditions (6.12) and (6.13) are satisfied for the three strata. To check orthogonality across strata we note that cov (rijk, Xij) = cov (Xijk, Xij) - ijk ij var Xij = ijk ij (1 + 2) + ijk - ijk ij {2 ij(1 + 2) + ij} = 0 and hence cov (rijk, rij) = 0, cov (rijk, ri) = 0. Also, cov (rij, Xi) = cov (Xij, Xi) - ij i var Xi = 2 ij i + ij - ij i ( 2 2 i + i} = 0 and hence cov (rij, ri) = 0. All other cross strata covariances of residuals are easily seen to be zero. The combined QSEF is then given by (6.18) as ijk ijk rijk ijk + ij ij rij ij (1 + 1 ij) + i i ri i (1 + 2 i) . For another example of a linear model with multiplicative random effects on which quasi-likelihood methods have been used, see Firth and Harris (1991). Their approach is similar to that of Morton (1987). 6.5. STATE-ESTIMATION IN TIME SERIES 103 6.5 State-Estimation in Time Series An important class of time series problems involve state space models of which a simple example is the type F1t(Yt, Yt-1, Xt) = t, F2t(Xt, Xt-1) = t, (6.19) where the Y 's are observed but not the X's, the F's are known functional forms and the { t} and {t} are each uncorrelated sequences of random variables that are mutually orthogonal. In filtering we require the estimation of XT given the observations (Y0, . . . , YT ) and in smoothing the estimation of (X0, . . . , XT ) given (Y0, . . . , YT ). Other parameters in the system are generally assumed known in the context of filtering or smoothing. The joint estimation of the parameters and the X's is known as system identification. In this section we shall indicate the estimating function approach to filtering and smoothing and in particular show how the celebrated Kalman filter follows from optimal combination of estimating functions. The material follows the ideas of Naik-Nimbalkar and Rajarshi (1995). A closely related approach, based on an extension of the E-M algorithm for missing data, is given in Chapter 7.4.4. To focus the discussion we shall specialize (6.19) to the linear case Yt = t Xt + t, Xt - = t(Xt-1 - ) + t, (6.20) where , 's, 's are known constants and if Gt and Ht are the -fields generated by Xt, . . . , X0, Yt, . . . , Y0 and Xt, . . . , X0, Yt-1, . . . , Y0, respectively, then almost surely, E t Ht = 0, E t Gt-1 = 0, E s t Ht = 0 = E s t Ht , s = t. Now the model (6.20) suggests estimating functions G1T = YT - T XT , G2T = XT - XT |T -1, where XT |T -1 is the minimum variance unbiased estimator of XT based on YT -1, . . . , Y0. Further, G1T and G2T are uncorrelated since EG1T G2T = E G2T E T HT = 0, G2T being HT -measurable. Then, according to Section 6.2 we obtain a composite quasi-score by standardizing the estimating functions G1T , G2T and adding. This leads to the estimating equation YT - T XT E 2 T (-T ) + XT - XT |T -1 E(XT - XT |T -1)2 = 0 (6.21) for XT . The solution of (6.21) gives Zehnwirth's (1988) extended Kalman filter, the classical filter corresponding to the case where E( 2 T | HT ) is nonrandom. 104 CHAPTER 6. COMBINING ESTIMATING FUNCTIONS We now consider the case of smoothing. Here we can conveniently use the estimating functions H1T = T t=1 ct(Yt - t Xt) H2T = T t=1 dt(Xt - - t(Xt-1 - )) where ct is Ht-1-measurable and dt is Gt-1-measurable. Then, using the ideas of Section 6.3 and noting that H1T , H2T are orthogonal, the quasi-score estimating equations for X0, . . . , XT are Yt - t Xt E 2 t Ht (-t) + Xt - - t(Xt-1 - ) E 2 t Gt-1 + Xt+1 - - t+1(Xt - ) E 2 t+1 Gt (-t+1) = 0, t = 1, 2, . . . , T - 1, YT - T XT E 2 T HT (-T ) + XT - - T (XT -1 - ) E 2 T GT -1 = 0. Further examples and references can be found in Naik-Nimbalkar and Rajarshi (1995). 6.6 Exercises 1. (Generalized linear model). Take X = (X1, . . . , Xn) as a sample of independent random variables, {F} a class of distributions and = (1, . . . , p) a vector parameter such that EF Xi = i((F)), var Xi = (F) Vi((F)) for all F F where i, Vi, i = 1, 2, . . . , n are specified real functions of the indicated variables. Suppose that the skewness () and kurtosis () are known: = EF (Xi - i)3 /(Vi)3/2 , = EF (Xi - i)4 /(Vi)2 - 3 for all F F and i = 1, 2, . . . , n. Let h1i = Xi - , h2i = (Xi - i)2 - Vi - (Vi)1/2 (Xi - i) and write hr = (hr1, . . . , hrn) , r = 1, 2. Show that the quasi-score estimating function for the family G = h1 + h2; , nonrandom n × p matrices 6.6. EXERCISES 105 is h1 + h2 where = -(Vi)-1 (i/j) , = (Vi)1/2 (i/j) + (Vi/j) (Vi)2( + 2 - 2) (Crowder (1987), Firth (1987), Godambe and Thompson (1989)). 2. Consider the random effects model Xi,j = + ai + eij (i = 1, 2, . . . , n, j = 1, 2, . . . , ni), where ai and eij are normally distributed with zero means and variances 2 a and 2 i , respectively. Take = 0 for convenience. Use the space of estimating functions H = n i=1 Ai Y i where the Ai are (n + 1) × (ni + ni(ni - 1)/2) nonrandom matrices and the Y i are the ni + ni(ni - 1)/2 vectors Y i = (X2 i1 - EX2 i1, . . . , X2 ini - EX2 ini , Xi1Xi2 - EXi1Xi2, . . . , Xi1Xin1 - EXi1Xin1 , . . . , Xini-1Xini - E(Xini-1Xini )) to find a quasi-score estimating function for estimation of 2 a and 2 i , i = 1, 2, . . . , n (Lin (1996)). Chapter 7 Projected Quasi-Likelihood 7.1 Introduction There is a considerable emphasis in theoretical statistics on conditional inference. Principles such as that inference about a parameter should be conditioned on an ancillary statistic have received much attention. This discussion requires full specification of distributions and is consequently not directly available in quasi-likelihood context. However, conducting inference conditionally on a statistic T is equivalent to projecting onto the subspace orthogonal to that generated by T and such projections can conveniently be carried out for spaces of estimating functions. Indeed, projection based methods seem to be the natural vehicle for dealing with conditioning in an estimating function context, and it has recently become clear that quasi-likelihood extensions of much conditional inference can be obtained via projection. It is worthwhile to trace the development of this subject from conditional likelihoods to conditional score functions to projected quasi-likelihoods. The principles of conditioning date back to Fisher but with extensive subsequent developments; see Cox and Hinkley (1974, Chapter 2) for a discussion and references. The first step to an estimating function environment was taken by Godambe (1976) who showed that the conditional score function is an optimal estimating function for a parameter of interest when the conditioning statistic is complete and sufficient for the nuisance parameters. This work was generalized by Lindsay (1982) to deal with partial likelihood factorizations and cases where the conditioning statistic may depend on the parameter of interest. The next steps in the evolution, from a likelihood to a quasi-likelihood environment, are now emerging and elements of this development are the subject of this chapter. Conditioning problems may involve either the parameter to be estimated or the data that are observed. We shall illustrate by dealing with parameters subject to constraint, nuisance parameters, and missing data (for which we provide a P-S (Project-Solve) generalization of the E-M (Expectation-Maximization) algorithm). 7.2 Constrained Parameter Estimation Suppose that the parameter is subject to the linear constraint F = d, where F is a p × q matrix not depending on the data or . For nonlinear or inequality constraints the extension of the theory is indicated briefly in Section 7.2.3. The overall discussion follows Heyde and Morton (1993). 107 108 CHAPTER 7. PROJECTED QUASI-LIKELIHOOD For an arbitrary positive-definite symmetric matrix V , we define the projection matrix P V = F (F V -1 F )F V -1 , where Adenotes a generalized inverse of A. Necessarily (Rao (1973, p. 26)) P V F = F , and (Rao (1973, p. 47)) P V is unique. In what follows, V is independent of the data but may depend on , although such dependence will be suppressed in the notation. An important property of these projections is that, for any V and W , (I - P W )(I - P V ) = I - P W . (7.1) We shall consider three methods of estimation of subject to constraint. For Methods 2 and 3 we use standardization of the type employed in Chapter 2 which ensures that the likelihood score property EG G = -E ˙G holds. Method 1: Projection of free estimator. Given any estimating function G G, let ^ solve G(^) = 0. For arbitrary V the projected estimator is ~ = (I - P V ) ^ + V -1 F (F V -1 F )d. (7.2) Provided that d is consistent in the sense that F V -1 F (F V -1 F )d = d, then the constraint F ~ = d is ensured. Method 2: Projection of standardized functions. For any standardized estimating function G G let ~ solve the equations (I - P V ) G(~) = 0, F ~ = d. (7.3) Note that from (7.1) the choice of V is immaterial, for if we multiply (7.3) by I -P W , we get (I -P W ) G(~) = 0. In practice we may like to choose V = I. Method 3: Analogue of Lagrange multipliers. If we had a quasilikelihood q, then by the method of Lagrange multipliers we would maximize q + (F - d), where is determined by the constraint. This suggests the following method. Let G G be any standardized estimating function. We solve for ~ and the equations G(~) + F = 0, F ~ = d. (7.4) This method is available even when G is not the derivative of a function q or even a quasi-score. Define V 0 = E(G) and P 0 to be projector with V = V 0. The information criterion is modified to EF (G) = (I - P 0)V 0(I - P 0) . 7.2. CONSTRAINED PARAMETER ESTIMATION 109 Its symmetric generalized inverse is EF (G)= (I - P 0) V -1 0 (I - P 0), (7.5) which is shown below to be the asymptotic variance for the projected estimator with the optimum choice of V = V 0. Our optimality criterion will be to minimize (7.5). With this choice, it is seen that the three methods are asymptotically equivalent and that optimality occurs within H when G is a quasi-score estimating function, standardized for the purpose of (7.3) and (7.4). 7.2.1 Main Results Theorem 7.1 Let ~ be the projected estimator (7.2). Suppose that ^ has asymptotic variance E(G)-1 . Then ~ has asymptotic variance (I - P V ) E(G)-1 (I - P V ) EF (G)- ; equality holds when V = E(G). Proof. The formula for the asymptotic variance ~ is obvious from the definition of ~. To prove the inequality, we use (7.1) and the fact that V -1 0 P 0 = P 0V -1 0 and P 2 0 = P 0. The difference is (I - P V ) V -1 0 (I - P V ) - (I - P 0) V -1 0 (I - P 0) = (I - P V ) {V -1 0 - (I - P 0) V -1 0 (I - P 0)}(I - P V ) = (I - P V ) {V -1 0 - V -1 0 (I - P 0)}(I - P V ) = (I - P V ) V -1 0 P 0(I - P V ) = (I - P V ) P 0 V -1 0 P 0(I - P V ) = (P 0 - P V ) V -1 0 (P 0 - P V ) 0, since V -1 0 is positive definite. Theorem 7.2 Let G G be a standardized estimating function. Assume that with probability tending to 1 all methods (7.2)­(7.4) possess unique solutions for ~ in some neighborhood of . Then in this neighborhood, when solutions are unique: (a) under mild regularity conditions the projected estimator (7.2) with V = V 0 and the projected function estimator (7.3) agree to first order; (b) the projected function estimator (7.3) and the Lagrange analogue estimator (7.4) have identical solutions. 110 CHAPTER 7. PROJECTED QUASI-LIKELIHOOD Proof. (a) Expand G(^) about to first order and assume that ˙G approximates -V 0. Then ^ - - ˙G()-1 G() V -1 0 G(). Hence ~ - (I - P 0) V -1 0 G() and G(~) G() + ˙G()(~ - ) G() - V 0(~ - ) {I - V 0(I - P 0) V -1 0 } G() = P 0 G(). Thus (I - P 0) G(~) 0, so (7.3) is satisfied to first order. By uniqueness, the two estimators must agree to first order. (b) Given that a solution to (7.3) exists, it also solves (7.4) with = -(F V -1 F )F V -1 G. By uniqueness the two solutions must be identical. Corollary 7.1 Assume that G() is asymptotically normally distributed under variance based norming. Then, under mild regularity conditions, and for large samples, ~ is approximately normally distributed with mean and singular variance matrix EF (G), whichever method is used. Also, (I -P 0) G() is approximately normally distributed with mean zero and singular variance matrix EF (G). This Corollary may be used to construct approximate confidence regions for (see Chapter 4). We would prefer to base them on the projected estimating function, since the asymptotic normality of G rather than ^ is usually the origin of the theory, so that the further approximation due to the expansion of G is avoided. Then the confidence region for is of the form G EF (G)G c, F = d, where c is a scalar obtained from the appropriate chi-squared or F distribution. Theorem 7.3 Let Q be a quasi-score estimating function within H. Then Q is optimal in that EF (G)- EF (Q)for all G H. Proof. Since E(Q) E(G), E(G)-1 E(Q)-1 . Hence (I - P V ) {E(G)-1 - E(Q)-1 } (I - P V ) 0. Upon inserting V = E(G), we get that EF (G)- (I - P V ) E(Q)-1 (I - P V ) EF (Q)- 7.2. CONSTRAINED PARAMETER ESTIMATION 111 by Theorem 7.1. Naturally, the asymptotic properties of the corollary hold for Q. 7.2.2 Examples Example 1: Linear regression. Consider the usual regression model Y = X + , E( ) = 0, E( ) = 2 I, where X is the design matrix with full rank and F = d. In this case the quasi-score estimating function Q = X (Y - X ) gives the usual normal equation Q = 0. The optimal choice for V is V 0 = 2 X X, so P 0 = F {F (X X)-1 F }F (X X)-1 . Standardization of Q gives Q(s) = -2 Q, which would be the likelihood score if were normally distributed. The free estimator is ^ = (X X)-1 X Y , so its projection is ~ = ^ - (X X)-1 F {F (X X)-1 F }(F ^ - d), which is identical to that obtained by projecting Q(s) or using the Lagrange multiplier method. The result has been long known: see Judge and Takayama (1966), who used Lagrange multipliers with least squares. Example 2: Two Poisson processes, constrained total. Let Ni = {Ni(t), 0 t T} be independent Poisson processes with constant rates i (i = 1, 2), which are constrained so that 1 + 2 = 1. We have that H is the class of estimating functions of the form T 0 b1(s){dN1(s) - 1 ds}, T 0 b2(s){dN2(s) - 2 ds} , with b1, b2 being predictable. An application of Theorem 2.1 can be used to find the quasi-score estimating function Q = -1 1 N1(T) - T, -1 2 N2(T) - T . This is already standardized and in fact is the likelihood score. Here F = (1, 1) , V 0 = diag (-1 1 , -1 2 ), so that P 0 = (1 + 2)-1 1 2 1 2 . 112 CHAPTER 7. PROJECTED QUASI-LIKELIHOOD Using (7.3) we obtain the estimator ~i = Ni(T)/{N1(T) + N2(T)} (i = 1, 2), which is identical to the solution of (7.2) and (7.4). For an example in this form a likelihood and Lagrange multiplier approach is available, but this could easily have been precluded by the use of counting processes, which were not specified to be Poisson and for which only the first and second moment properties were assumed. Example 3: Linear combination of estimating functions. Suppose that we have n > p estimating functions H = {hi(X, )} such that E(H) = 0 and E(H H ) = V H exist. We wish to reduce the dimension of H to p by choosing optimal linear combinations. That is, H consists of estimating functions of the form G = C H, where C is a p × n matrix depending only on . The standardized quasi-score estimating function is Q(s) = -E( ˙H) V -1 H H. (7.6) Suppose now that we wish to reduce the dimensions of H in the presence of linear constraints F = d. By (7.4), the theory is modified to solve for ~, from E( ˙H) V -1 H H - F = 0, F ~ = d. The theory extends to the case where C may contain incidential parameters (Morton (1981a)). An important application concerns functional and structural relationships where the elements of H are the contrasts that are free of the incidential parameters (Morton (1981b)). 7.2.3 Discussion We have presented the theory assuming that the constraints are linear in order to avoid obscuring the simple arguments by messy details. If we had nonlinear constraints, say () = c, then the Lagrange multiplier analogue replaces F by F () = / , which we assume exists and has full rank. If the projection methods are used, then this substitution is also made and in (7.2) we must also approximate d() = F (0) iteratively. In (7.3) and (7.4), F = d is replaced by () = c. Inequality constraints could in principle be handled by Lagrange multipliers. If the jth constraint is (F )j dj, the corresponding multiplier j would be set equal to zero if strict inequality holds and would be free if equality holds. In many problems it would be natural to transform the parameter to ( , ) , where has r = rank(F ) elements, which define the constraints, and is free. We then need to reduce G to dimensions p - r. Picking any p - r elements of G would not be optimal; the best linear reduction would be achieved in the manner of (7.6), which would be equivalent to projecting G as in (7.3). Algebraic elimination of from G would be equivalent to the Lagrange 7.3. NUISANCE PARAMETERS 113 multiplier method. For nonlinear problems this could lead to an estimating function outside the class G. In the case where G is the class of functions that are linear in , all three methods agree with the parameter transformation method. In the linear regression model (Example 1), suppose that = F is not fixed but is a nuisance parameter. It can be shown that, in the terminology of Chapter 3 that Q1 = (I -P 0) Q is locally E-ancillary for and if Q2 = P 0 Q, then EQ1 Q2 = 0 and hence Q2 is locally E-sufficient for . In the general situation, assuming that linear approximation is adequate, the decomposition Q = P 0 Q + (I - P 0) Q has the first term locally E-sufficient for and the second term locally Eancillary for . Thus, the method of projection of standardized functions (7.3) discards the first order sufficient information about and replaces it by the known constraint = d. 7.3 Quasi-Likelihood in the Presence of Nuisance Parameters In this section we use methods analogous to those of the last section to treat the case where the basic parameter contains a (vector) nuisance component. Suppose that the parameter is partitioned with = ( , ), being the component of interest and a nuisance parameter and that G is a standardized estimating function chosen for the estimation of . Standardization means that the likelihood score property E ˙G = E(G/) = -EG G holds. We write G = G G and partition the information matrix V G = EG G = E(G) according to V G = V V V V where V = EG G, V = EG G, V = V = E(G G). 114 CHAPTER 7. PROJECTED QUASI-LIKELIHOOD Write also F = V V = -E(G/), F = V V = -E(G/). It turns out that the projection P = F (F V -1 G F )-1 F V -1 G identifies the information about for given and the estimating equation (I - P ) G = 0 (7.7) is optimal for estimation of in the presence of the nuisance parameter . In (7.7) the sensitive dependence of G on has been removed in the sense that E {(I - P ) G} = 0 since P F = F and G has zero mean. This is, of course, a first order approach and (I -P )G will not in general be free of . There is a convenient interpretation in terms of E-ancillary and E-sufficient estimating functions as discussed in Chapter 3. It can be seen that (I - P ) G is locally E-ancillary for and P G is locally E-sufficient for . To appreciate (7.7) in more detail, first note that, in partitioned matrix form, F V -1 G = 0 I , so that P = F V -1 0 I = 0 V V -1 0 I and (I - P ) G = G - V V -1 G 0 . The "information" concerning associated with the estimation procedure is E G - V V -1 G G - V V -1 G = V - V V -1 V , (7.8) or alternatively we can think of it in partitioned matrix form as E(G) = (I - P ) V G (I - P ) = V - V V -1 V 0 . 7.3. NUISANCE PARAMETERS 115 Of course there is no loss in efficiency if G, G are orthogonal, for then V = V = 0. The formula (7.8) is a direct generalization of the familiar result obtained from partitioning the likelihood score (e.g., Bhaphkar (1989)) and the asymptotic variance of the estimator of based on G is then V - V V -1 V -1 = V , (7.9) where V -1 G = V V V V . Optimality of estimation for thus focuses on maximizing V in the partial order of nnd matrices and we have the following theorem. Theorem 7.4 Let Q be a quasi-score estimating function within H. Then Q is optimal for estimation of in the sense that V G V Q in the partial order for all G H. Proof. The result follows immediately from the fact that E(Q) = V Q E(G) = V G for all G H and then V -1 G V -1 Q . As a simple example of the methodology we shall discuss the linear model Z = X + , say, where Z is a vector of dimension T, X = (X1, X2) is T × p matrix with X1 of dimension T × r and X2 of dimension T × (p - r). Also, = ( , ) where and are, respectively, vectors of dimension r and p - r and is a vector of dimension T with zero mean and covariance matrix . In this case the standardized estimating function is G = X -1 (Z - X1 - X2 ) = X1 -1 (Z - X1 - X2 ) X2 -1 (Z - X1 - X2 ) = G G . Also, V G = EG G = X -1 X = X1 -1 X1 X1 -1 X2 X2 -1 X1 X2 -1 X2 = V V V V . 116 CHAPTER 7. PROJECTED QUASI-LIKELIHOOD Then, the estimating function G - V V -1 G = X1 -1 - (X1 -1 X2)(X2 -1 X2)-1 X2 -1 (Z - X1 ) does not involve the nuisance parameter and the estimator for is obtained from equating this to zero. This solution can, of course, be obtained directly from solving the ordinary estimating equation G G = 0 0 and then eliminating from the resultant equations. Indeed, many nuisance parameter problems are amenable to such a direct approach. The focus in this section has been on projection but there is no need for this geometric interpretation to be emphasized. The first order theory seeks the best combination G - c(, )G, which is easy to calculate analytically. Also, it is not difficult to go beyond the first order theory. Second order theory allows, incorporation, for example, of the result of Godambe and Thompson (1974) who dealt with the case in which the likelihood score is U = (U, U) , and being scalars, and showed that an estimating function of the form U - c(, )(U2 - EU2 ), if c(, ) can be chosen to make it free of , is the best choice for estimating . An example where this holds is in estimation of using a random sample of observations from the N(, ) distribution. 7.4 Generalizing the E-M Algorithm: The P-S Method The E-M algorithm (e.g., Dempster et al. (1977)) is a widely used method for dealing with maximum likelihood calculations where there are missing or otherwise incomplete data. It involves the taking of the expectation of the complete-data likelihood with respect to the available data (the E-step) and then maximizing this over possible distributions (the M-step) and the procedure suggests a simple iterative computing algorithm. It has not been available in contexts where a likelihood is unknown or unavailable. In this section we extend the E-M algorithm method to deal with estimation via estimating functions, in particular the quasi-score. The transitions to estimating functions is made since there are situations where no quasi-loglikelihood exists. The discussion here follows Heyde and Morton (1995). In our approach the E-step is replaced by a step which projects the quasiscore rather than taking expectations of a log-likelihood. In many examples the 7.4. GENERALIZING THE E-M ALGORITHM: THE P-S METHOD 117 projection is equivalent to predicting the missing data or terms. The predictor will not in general be a conditional expectation. The M-step is replaced by solving the projected quasi-score set equal to zero. The approach can reasonably be described as the projection-solution (P-S) method. When the likelihood is available and the score function is included in the class of estimating functions permitted, the standard E-M procedure is recovered from the P-S method. More broadly, however, we seek to make the point that there is no formal difference between QL estimation for incomplete data and QL estimation for complete data. 7.4.1 From Log-Likelihood to Score Function As a prelude to generalization of the E-M algorithm formalism from log-likelihoods to estimating functions, we first show how attention can be transferred from operations on the log-likelihood to corresponding ones on its derivative, the likelihood score. We denote the full data by x and the observed data by y. The parameter of interest is , a p ( 1) dimensional vector. The likelihood of based on data z is denoted by L(; z). First note that, from the definition of conditional distributions, log L(; y) = log L(; x) - log L(; x | y) . (7.10) The second term on the right hand side of (7.10) has zero expectation conditional on y under the usual regularity conditions and so log L(; y) = E log L(; x) y . (7.11) That is, the likelihood score based on data y is obtained by taking the expectation, conditional on y fixed, of the likelihood score based on the full data x. Now the M-step of the E-M algorithm is based on maximization of E0 (log L(; x) | y), and it is clear that, under the usual regularity conditions, log L(; y) = E0 log L(; x) y 0= . (7.12) The iterative computing algorithm involving (p+1) as the value of that maximizes E(p) (log L(; x) y) then has as its obvious analogue the choice of (p+1) as the value of which solves E(p) log L(; x) y = 0 (7.13) 118 CHAPTER 7. PROJECTED QUASI-LIKELIHOOD or, equivalently, to solve log L(; y)/ = 0 via E(p) log L(; x) y =(p+1) = 0, (7.14) using (7.11). 7.4.2 From Score to Quasi-Score The Framework Now suppose that likelihoods are unsuitable or unavailable and that, based on data x, a family of zero mean, square integrable estimating functions Hx has been chosen within which the quasi-score estimating function is Q(; x). The function Q may be a likelihood score but this will not be the case in general, nor will Q in general be expressable as the derivative with respect to of a scalar function. As usual Q is chosen as the estimating function G Hx for which the "information" E(G) = (E ˙G) (EG G )-1 E ˙G (7.15) is maximized in the Loewner ordering (partial order of nonnegative definite matrices). It should be recalled that estimating functions G and M G, M being any nonsingular matrix of the dimension of G, give rise to the same estimators of . For each estimating function G we choose M = -(E ˙G) (EG G )-1 and denote G(s) = M G, the standardized version of G. This satisfies the likelihood score property EG(s) G(s) = -E ˙G (s) = E(G(s) ) (= E(G)). All estimating functions in the subsequent discussion of this section will be assumed to be in standardized form and we shall drop the superscripts for convenience. Note that the likelihood score is automatically in standardized form. We shall subsequently stipulate that all families G of estimating functions under consideration, such as Hx, are convex. Then, Q is a quasi-score estimating function within G iff EG Q = -E ˙G (7.16) for all G G (Theorem 2.1). In the case where the likelihood score U exists, an equivalent formulation of quasi-score is obtained by defining Q Hx to be a quasi-score estimating function if E{(Q - U)(Q - U) } = inf GHx E{(G - U)(G - U) }, (7.17) 7.4. GENERALIZING THE E-M ALGORITHM: THE P-S METHOD 119 the infimum being taken with respect to the partial order of nonnegative definite matrices. This follows easily from the formulation based on (7.15) upon observing that E{(G - U)(G - U) } = EG G - EU U (Criterion 2.2 and Theorem 2.1). Also, we see in conjunction with (7.17) that E{(G - U)(G - U) } = E{(Q - U)(Q - U) } (7.18) + E{(Q - G)(Q - G) }. The general situation with which we are concerned is when y = y(x) is observed rather than x. We seek to adapt a quasi-score estimating function Q(; x) to obtain Q (; y) where Q is a quasi-score in a class Hy, which is typically a suitable linear subspace of Hx. Then, E{(Q - U)(Q - U) } = inf GHy E{(G - U)(G - U) }, when U exists, and subtracting E{(Q-U)(Q-U) } from both sides, we have that E{(Q - Q)(Q - Q) } = inf GHy E{(G - Q)(G - Q) }, (7.19) a formula whose use we advocate even when U does not exist. Thus, Q is the element of Hy with minimum dispersion distance from Q Hx. We note that Q = E(Q | y) provided this belongs to Hy as in the E-M algorithm. However, in general Q is just given as the least squares predictor (7.19). That is, the "expectation" step is replaced by a "projection" step in the E-M algorithm procedure. We can regard this as a generalization of the E-M technique. The Algorithm In general the algorithm proceeds as follows. Write a 2 = a a for vector a. Then, we define H( | 0; y) such that E0 H( | 0; y) - Q(; x) 2 = inf GHy E0 G(; y) - Q(; x) 2 and solve iteratively for (p+1) from H((p+1) (p) ; y) = 0 starting from an initial guess (0) . A sequence {(p) } is generated and, provided (p) , we have in the limit E H( ; y) - Q(; x) 2 = inf GHy E G((; y) - Q(; x) 2 , (7.20) 120 CHAPTER 7. PROJECTED QUASI-LIKELIHOOD and Q (; y) = H( ; y) satisfies (7.19) provided a solution does exist. This last result is a straightforward consequence of the following lemma. Lemma 7.1 If a and b are random vectors such that E(a a - b b ) is nonnegative definite and E a 2 E b 2 , then Ea a = Eb b . For our application, suppose that K = K(; y) Hy satisfies (7.20) and that a solution Q = Q (; y) to (7.19) exists. We take a = K - Q and b = Q - Q in the lemma. Since (7.19) gives nonnegativity of E(a a - b b ) while (7.20) gives E a 2 E b 2 we obtain Ea a = Eb b . However, for Hy Hx we have Ea b = 0, so that E a - b 2 = E a 2 - E b 2 = 0, which gives K = Q a.s. Finally, to prove the lemma we note that, upon taking traces, 0 tr {E(a a - b b )} = tr {E a 2 - E b 2 } 0, so that A = E(a a -b b ) is a symmetric and nonnegative definite matrix with tr A = 0. This forces A 0 since the sum of squares of the elements of A is the sum of squares of its eigenvalues all of which must be zero as they are nonnegative and sum to zero. Alternatives It may be expected, by analogy with the arguments of Osborne (1992) for the score function, that the algorithm developed above will give a first order rate of convergence compared with the second order rate typical of Fisher's method of scoring. Thus, the latter would be preferred for solving Q (; y) = 0 if the necessary derivative calculation could be straightforwardly accomplished. The E-M generalization, of course, has the virtue of avoiding any such requirement. Sufficient conditions for convergence of either method can, in principle, be formulated along the lines of Section 3 of Osborne (1992), but they involve technical assumptions that ensure smoothness and the applicability of a law of large numbers, and are not transparent, so they will not be discussed herein. It should be noted that there is no formal requirement for Q in (7.19), (7.20) to be a quasi-score estimating function. It can, in principle, be replaced by any estimating function H Hx. However, the optimality properties associated with the quasi-score estimating function will then be lost. In practice Q can usually be calculated with the aid of Theorem 2.1, only first and second moment properties being required. 7.4. GENERALIZING THE E-M ALGORITHM: THE P-S METHOD 121 7.4.3 Key Applications Estimating Functions Linear in the Data Since explicit expressions are not available in general we shall focus on the most common area of use of quasi-likelihood methods, namely, the situation of estimating functions which are linear in the data. Let = Ex, = E y, V x = E(x - )(x - ) , V y = E(y - )(y - ) . If Hx and Hy consist, respectively, of functions that are linear in x and y, the quasi-score estimating functions may be taken as Q(; x) = V -1 x (x - ), Q (; y) = V -1 y (y - ) (7.21) (e.g. using Theorem 2.1). Also, if y belongs to a linear subspace of x, y = C x say, then we have = C , V y = C V x C and the best linear predictor of x given y (in the weighted least squares sense) is ^x(y) = + V x C V -1 y (y - ), so that Q (; y) = Q(; ^x(y)). (7.22) It should be noted that even in this particular case Q (; y) = E(Q(; x) y) (7.23) in general, the right-hand side being what is suggested by the E-M prescription. Equality in (7.23) requires that E(x | y) be linear in y as holds, for example, for Normal or Poisson variables. Another context where equality in (7.23) may not hold occurs when Hy is not a subset of Hx. Then it can also happen that Q Hx. For details involving exponential variables see the example concerned with totals from data with constant coefficient of variation below. Generalization to Polynomials The linear description developed above can be exploited to deal with quadratic, and indeed higher polynomial, functions of x. To illustrate this, take Hx to include quadratic functions of x and again y = C x. We can regard Hx as linear in vectors containing all 1 2 p(p + 3) linear and quadratic elements x = (x1, . . . , xp, x2 1, x1x2, . . . , x2 p) 122 CHAPTER 7. PROJECTED QUASI-LIKELIHOOD and Hy as linear, in a similarly augmented vector y with y = C x where C is constructed so as to generate the quadratic functions of y. 7.4.4 Examples Screening Tests Our first example deals with extension of a standard setting for the E-M algorithm concerned with developing screening tests for a disease given by Laird (1985). Our purpose is to compare the approach via the score or quasi-score estimating function to that of the familiar one via the log-likelihood which is given by Laird. The problem here involves observed data consisting of a random sample of patients measured on screening tests, each of which gives a dichotomous result. The true disease status is unknown and for Test i the sensitivity has a distribution with mean Si, i = 1, 2. It is assumed that the test results are conditionally independent given disease status and that false positives are not possible. We wish to estimate S1, S2 and the disease prevalence . Retaining Laird's notation, the observed data can be put into an array of the form Test 2 + + y11 y12 Test 1 - y21 y22 The complete data are x11, x12, x21, x221, x222 where yij = xij unless (i, j) = (2, 2) and y22 = x221 +x222, with x221 being the number of diseased individuals with negative outcomes on both tests and x222 the number of nondiseased patients. We also write x1+ = x11 + x12, x+1 = x11 + x21, N for the sample size ijyij and ND for the number of diseased patients (N - x222). Laird has supposed that the test sensitivities are random and that, conditional on ND fixed, x1+ and x+1 have, respectively, the independent binomial distribution b(ND, S1) and b(ND, S2) while x222 has the binomial distribution b(N, 1-). Then the quasi-score estimating function based on linear functions of the essential data x1+, x+1 and ND is Q(; x) = x1+ - E x1+ ND var x1+ ND E x1+ ND + x+1 - E x+1 ND var x+1 ND E x+1 ND 7.4. GENERALIZING THE E-M ALGORITHM: THE P-S METHOD 123 + ND - END var ND END , where = (S1, S2, ) . We have, under Laird's assumptions, E x1+ ND = ND S1, var x1+ ND = ND S1 (1 - S1), E x+1 ND = ND S2, var x+1 ND = ND S2 (1 - S2), END = N , var ND = N (1 - ) so that Q(; x) = x1+ - ND S1 S1(1 - S1) , x+1 - ND S2 S2(1 - S2) , ND - N (1 - ) = x1+ S1 ND - x1+ 1 - S1 , x+1 S2 ND - x+1 1 - S2 , ND - x222 1 - , which is also the likelihood score estimating function as is easily seen from Laird's discussion. Then, the obvious choice for Hy yields the quasi-score E(Q(; x) y) = y1+ S1 E(ND | y) - y1+ 1 - S1 , y+1 S2 E(ND | y) - y+1 1 - S2 , E(ND | y) N - E(ND | y) 1 - and a simple conditional probability argument gives E(ND | y) = N - y22(1 - ) (1 - S1)(1 - S2) + (1 - ) . (7.24) If an algorithmic solution is sought, we set N (p+1) D = E(ND | y, (p) ) and we have H((p+1) | (p) , y) = y1+ S (p+1) 1 - N (p+1) D - y1+ 1 - S (p+1) 1 , y+1 S (p+1) 2 - N (p+1) D - y+1 1 - S (p+1) 2 , N (p+1) D (p+1) N - N (p+1) D 1 - (p+1) leading to (p+1) = N (p+1) D N , S (p+1) 1 = y1+ N (p+1) D , S (p+1) 2 = y+1 N (p+1) D as obtained by Laird for the E-M algorithm solution. The iterative solution converges quite quickly in this case but an explicit solution to the estimating equation E(Q(; x y) = 0 124 CHAPTER 7. PROJECTED QUASI-LIKELIHOOD is also available. After some algebra we find that ^S1 = y11 y+1 , ^S2 = y11 y1+ , ^ = 1 - y22 N 1 - (1 - y11 y+1 )(1 - y11 y1+ ) = y+1 y1+ N y11 , which gives the maximum likelihood estimator for based on y. Note that the full distributional assumptions made by Laird are not needed in the above analysis. All that is required is the first and second (conditional) moment behavior of x1+, x+1 and ND. Suppose now that the data are not homogeneous and perhaps not really from a random sample. It is, nevertheless, still reasonable to take E(x1+ ND) = ND S1, E(x+1 ND) = ND S2 and to suppose that E x221 y22 E x222 y22 = (1 - S1)(1 - S2) 1 - , which, together with E(ND y) = N - E x222 y22 still gives (7.24). The expressions for the conditional variances of x+1 | ND and x1+ | ND may need to be changed to incorporate a measure of dispersion. For example, if the individuals counted in x+1 (resp. x1+) displayed sensitivity to Test 1 (resp. 2) with a distribution of beta type with mean S1 (resp. S2), then formulas var (x1+ | ND) = 1 ND S1(1 - S1), var (x+1 | ND) = 2 ND S2(1 - S2) would be obtained for certain dispersion parameters 1, 2. If it were reasonable to assume that the dispersion parameters could be taken as the same, the original analysis would remain unchanged. If not, obvious adjustments could be made. Totals from Data with Constant Coefficient of Variation Here we consider a situation where the full data set is x = (xij, 1, 2, . . . , n; j = 1, 2) and it is the totals yi = xi1 + xi2 that are observed, so that y = (yi, i = 1, 2, . . . , n). The likelihood is not known but the xij are assumed to be independent with moments E(xij) = ij, var xij = 2 ij where ij = ij() and is the coefficient of variation, assumed to be constant. The aim is to estimate , taken as scalar for convenience, and for simplicity we assume that = 1 is known. 7.4. GENERALIZING THE E-M ALGORITHM: THE P-S METHOD 125 The class Hx consists of the linear functions of x and the usual quasi-score is (e.g., McCullagh and Nelder (1989)) Q(; x) = ij (xij - ij) -2 ij ij . Since we observe only the totals y we consider Hy, the class of linear functions of y. Then the corresponding quasi-score is Q (; y) = i (yi - (i1 + i2))(2 i1 + 2 i2)-2 (i1 + i2). (7.25) The linear predictor ^x of x given y is ^xij - ij = 2 ij 2 i1 + 2 i2 (yi - (i1 + i2)) and we see that Q (; y) = Q(; ^x). It should be noted that Q is not the expectation of Q, which is general is nonlinear in y. For example, suppose that xij has the exponential density f(xij; ) = -1 ij exp(-xij/ij), if 0 < xij < , 0, otherwise. Then it is easily checked that E(xij yi; ) = i yi exp(-yi/i) 1 - exp(-yi/i) , where -1 i = -1 i1 - -1 i2 . In this case Q = U is the likelihood score for the full data but U = E(U y; ) differs from Q , which is the projection of Q to be used when the exponential likelihood specification is doubtful or inappropriate. There is some evidence that there is little loss in efficiency in using Q instead of U even when the exponential assumption is actually correct. To evaluate the asymptotic relative efficiency we must compare E(U ) = E(U )2 (7.26) = i E(E(xi1 yi) - i1) -1 ij 2 + (E(xi2 yi) - i2) -1 ij 2 , 126 CHAPTER 7. PROJECTED QUASI-LIKELIHOOD and E(Q ) = E(Q )2 = i 2 i1 + 2 i2 -1 i1 + i2 2 . (7.27) The information measures (7.26) and (7.27) satisfy the inequality E(U )2 E(Q )2 but a concrete comparison requires more specific values for the ij. For example, in the particular case where i2 = i1 with fixed and taking = - 1 > 0 without loss of generality we find that E(Q )2 = (1 + )2 1 + 2 i i2 2 (7.28) while E(U )2 is given below. Indeed, temporarily dropping subscript i again, we find that f(y; 1, 2) = 1 + 2 e-y/2 (1 - e-y/2 ), y > 0, and, after some algebra, E - 2 2 log f(y; 1, 2) = 2 2 1 + (1 + ) 0 z2 e-(1+)z 1 - e-z dz = 2 2 1 + 2 (1 + ) r=1 1 (1 + r )3 = 2 2 1 + 2 1 + 2 3, 1 , where (3, 1 ) = r=1 (r + 1 )-3 is the generalized Riemann zeta-function. Then, E(E(U y))2 = 1 + 2 1 + 2 3, 1 i i2 2 (7.29) and the asymptotic relative efficiency of Q is ARE(Q )() = (1 + )2 1 + 2 1 1 + 2 ( - 1)-2 (3, ( - 1)-1) using (7.27) and (7.28). It is easily seen that ARE(Q )() approaches unity as 1 or . A plot of ARE(Q )() against (1 < < 50) has been obtained with the aid of the package MAPLE and appears in Heyde and Morton (1995). The ARE always exceeds 0.975. 7.5. EXERCISES 127 Time Series Smoothing with Missing Observations The E-M algorithm itself is mainly useful in situations where the y-likelihood is intractable but the x-likelihood is simple (so that the M-step is easily achieved). The algorithm outlined in this section is similarly useful in situations where quasi-score Q is simple but Q is intractable. Many such examples can be generated in the context of time series observed subject to noise, and with missing data as well, when Gaussian distributional assumptions are questionable or inappropriate. The example we discuss here comes from Shumway and Stoffer (1982) whose discussion has been succinctly summarized by Little and Rubin (1987, pp. 165- 168). It concerns data on expeditures for physician services and is treated using the model: xij = zi + ij, j = 1, 2, (7.30) zi = zi-1 + i, (7.31) where the ij are independent with E ij = 0, var ij = Bj, the i are independent with Ei = 0, var i = Q and is an inflation factor. The xij's are the observations but in practice only a subset yij have been observed yij = xij for i, j A, say. The object is to estimate the zi's, which are not assumed to be stationary. Now Shumway and Stoffer have assumed, without comment, that the ij and i are normally distributed and they have used the E-M algorithm to calculate the zi, the corresponding likelihood being intractable. However, if normality is not assumed it is still reasonable to use linear estimating equations for this problem. Then, the quasi-scores are identical to those based on the normal likelihood and the E-M algorithm results are preserved in our approach. The inferences remain appropriate for nonnormal data having unchanged first and second moments. 7.5 Exercises 1. (Mixture densities) Suppose that 0 1 and let X1, . . . , Xn be independent and identically distributed with mixture density h(x; , ) = f(x; ) + (1 - ) g(x; ). Here is to be estimated. If the functions h, f, g are known, so that the score function U = (U, U) can be used as a basis for the calculation, show that the optimal estimating equation for is U - [E(U U)/EU2 ] U = 0 (Small and McLeish (1989)). 128 CHAPTER 7. PROJECTED QUASI-LIKELIHOOD 2. (A form of Rao-Blackwellization) Let Q be a quasi-score estimating function. If K is an estimating function show that E(K | Q) is preferable to K in the usual sense that E(E(K | Q)) E(K) in the partial order of nonnegative definite matrices (Small and McLeish (1994, p. 85)). Chapter 8 Bypassing the Likelihood 8.1 Introduction Estimation based on the application of maximum likelihood methods can involve quite formidable difficulty in calculation of the joint distribution of the observations and sometimes even in differentiating the likelihood to obtain the corresponding estimating equations. Notable examples are the derivation of the restricted (or residual) (REML) estimating functions for dispersion parameters associated with linear models and the derivation of the maximum likelihood estimators of parameters in diffusion type models. In the former case both the derivation of the likelihood and its differentiation are less than straightforward while for the latter the Radon-Nikodym derivative calculations are a significant obstacle. Quasi-likelihood methods, however, allow such estimators (or estimating functions) to be obtained quite painlessly and under more general conditions. In this chapter we shall illustrate the power and simplicity of the approach through three quite different examples. 8.2 The REML Estimating Equations Suppose that the n × 1 vector y has the multivariate normal distribution MV N(X, V ()) with mean vector X and covariance matrix V (). For simplicity we take the rank of X as r, the dimension of the vector . Let A be any matrix with n rows and rank n - r satisfying A X = 0. Then, the distribution of A y is MV N(0, A V A) which may be singular and it can be shown, with some difficulty, that the likelihood function of based on A y has the form of a constant multiplied by n-r i=1 i 1/2 exp - 1 2 y V -1 Q y , (8.1) where Q = I - P V R(X), P V R(X) denoting the projector X (X V -1 X)X V -1 onto the subspace R(X), the range space of the matrix X that is the orthogonal projection onto the orthogonal complement of R(X) with respect to the inner product a, b = a V -1 b, while 1, . . . , n-r are the nonzero eigenvalues of V -1 Q. (See Rao (1973, Eq. (8a.4.11) p. 528) for the case where A V A has full rank.) The striking thing here is that the result (8.1) does not depend on A. That is, all rank n - r matrices that define error contrasts (i.e., for which A X = 0) generate the same likelihood function in . This follows from the 129 130 CHAPTER 8. BYPASSING THE LIKELIHOOD result A(A V A)A = V -1 Q (8.2) for all A and any g-inverse (A V A). The basic reference on this topic is Harville (1977), although detailed derivations are not included, the reader being referred to an unpublished technical report. For a more recent derivation see Verbyla (1990). The next step is to differentiate (8.1) with respect to . Again this is less than straightforward, involving vector differentiation of matrices and determinants, but there finally emerge the REML estimating equations tr V -1 Q V i = y V -1 Q V i V -1 Q y, (8.3) where tr denotes trace. We shall now show how the result (8.3) can be derived painlessly, and under more general conditions, by quasi-likelihood methods. We shall no longer suppose that y has the multivariate normal distribution but only that its mean vector is X, its covariance matrix is V (), that Ey2 i y2 j = Ey2 i Ey2 j and Ey3 i yj = 0 for i = j and that each yi has kurtosis 3. For fixed A, let z = A y and, in the expectation that it is quadratic functions of the data that should be used to estimate covariances, consider the class of quadratic form estimating functions H = {G : G = (G(Si), . . . , G(Sp)) , G(Si) = z Siz - E(z Siz), Si symmetric matrix i = 1, 2, . . . , p}, being of dimension p. Write W = A V A. Theorem 8.1 G = (G(S i ), . . . , G(S p)) H is a quasi-score estimating function in H if S i = W - W i W , i = 1, 2, . . . , p, (8.4) for any g-inverse W = (A V A). Furthermore, the G(S i ) do not depend on A and the corresponding estimating equation G(S i ) = 0 is Equation (8.3). Proof. Using standard covariance results for quadratic forms in multivariate normal variables (see, e.g., Karson (1982, p. 62)) that involve only the moment assumptions noted above, we have cov (G(Si), G(S j )) = 2 tr (W SiW S j ), (8.5) while Ez Siz = tr (W Si). Also, it is easily checked that the (i, j) element of E ˙G, (E ˙G)ij = - tr W j Si . 8.3. PARAMETERS IN DIFFUSION TYPE PROCESSES 131 Then, when S i is given by (8.4), we see that cov (G(Si), G(S j )) = - 2(E ˙G(S))ij, since tr BC = tr CB and A W W = A, W W A = A for any choice of generalized inverse (e.g., Lemma 2.2.6, p. 22 of Rao and Mitra (1971)). Consequently, the result of the theorem concerning the form of S i follows from Theorem 2.1. The REML estimating equations (8.5) then follow immediately using z = A y and the representation (8.2). The theorem is new even in the case where the mean of y is known to be zero (i.e., the case r = 0). The results could, in principle, be adjusted to deal with yi's having different moments from those of the corresponding multivariate normal variables. For a general discussion of variance function estimation see Davidian and Carroll (1987). The results in this section are from Heyde (1994b). For a discussion of consistency and asymptotic normality of REML estimators see Jiang (1996). 8.3 Estimating Parameters in Diffusion Type Processes The standard method of estimation for parameters in the drift coefficient of a diffusion process involves calculation of a likelihood ratio (Radon-Nikodym derivative) and thence the maximum likelihood estimator(s). This is less than straightforward for more complicated models and indeed it is often not available at all because of the nonexistence of the Radon-Nikodym derivative. The methods of quasi-likelihood , however, allow estimators to be obtained straightforwardly under very general conditions. They can deal, in particular, with the situation in which the Brownian motion in a diffusion is replaced by a general square-integrable martingale. The approach, which is based on selection of an optimal estimating function from within a specified class of such functions, involves assumptions on only the first two conditional moments of the underlying process. Nevertheless, the quasi-likelihood estimators will ordinarily be true maximum likelihood estimators in a context where the Radon-Nikodym derivative is available. Furthermore, they will generally be consistent, asymptotically normally distributed, and can be used to construct minimum size asymptotic confidence zones for the unknown parameters among estimators coming from the specified class. In this section we illustrate these ideas through a general discussion and application to the Cox-Ingersoll-Ross model for interest rates and to a modification of the Langevin model for dynamical systems. The results are from Heyde (1994a). The models which we shall consider in this section can all be written in the semimartingale form dXt = dAt() + dMt(), (8.6) 132 CHAPTER 8. BYPASSING THE LIKELIHOOD where the finite variation process {At} can be interpreted as the signal and the local martingale {Mt} can be interpreted as the noise, as has been dicussed in Section 2.6. Then, the quasi-score estimating function based on the family of local martingale estimating functions H = T 0 at() dMt(), {at} predictable is easily seen from Theorem 2.1 to be T 0 E d ˙Mt() Ft- (d M() t)dMt(), (8.7) where {Ft} is a filtration of past-history -fields, M() t is the quadratic characteristic and the - denotes the generalized inverse. Now (8.7) can be rewritten as T 0 E d ˙At() Ft- (d M() t)(dAt() - dXt) from which it is clear that the quasi-score estimating function is unaffected if the local martingale noise {Mt()} is replaced by another whose quadratic characteristic is the same. The precise distributional form of the noise does not need to be known. In the commonly met situation where Mt() = Wt with > 0 and Wt being standard Brownian motion, the results are robust to the extent that {Mt()} could be replaced by any local martingale {Zt()}, for example, one with independent increments, for which Z t = 2 t without changing the estimators. In the particular case of a diffusion process the components on the righthand side of the representation (8.6) can be written as dAt() = a(t, Xt, ) dt, dMt = b 1 2 (t, Xt) dWt, (8.8) where a and b are known vector and matrix functions, respectively, and {Wt} is standard Brownian motion. Then, an appropriate Radon-Nikodym derivative of the measure induced by the process {Xt, 0 t T} with parameter with respect to the corresponding measure for parameter 0 can be calculated and is given by exp T 0 C(t, Xt) dXt - T 0 D(t, Xt) dt , (8.9) where b(t, x) C(t, x) = a(t, x, ) - a(t, x, 0), D(t, x) = (a(t, x, 0)) C(t, x) + 1 2 (C(t, x)) b(t, x) C(t, x) (e.g., Basawa and Prakasa Rao (1980, p. 219)). 8.3. PARAMETERS IN DIFFUSION TYPE PROCESSES 133 From (8.9) it is easily checked that the likelihood score function (derivative of the logarithm of the Radon-Nikodym derivative with respect to ) is given by (8.7). This means that the quasi-likelihood estimator is the maximum likelihood estimator for the model (8.8). However, as indicated above, the quasi-likelihood estimator is available much more generally. The explanation for the quasi-score corresponding to the likelihood score for the model (8.8) is not hard to discern. A likelihood score is a martingale under modest regularity conditions and all square integrable martingales living on the same probability space as the Brownian motion in the noise term of the model (8.8) can be described as stochastic integrals with respect to the Brownian motion (see, e.g., Theorem 5.17 of Liptser and Shiryaev (1977)). The likelihood score will be one such martingale and will therefore be included in the relevant family H over which optimization takes place and it solves the optimization problem. As a concrete illustration of the methodology we shall discuss the stochastic differential equation dXt = ( - Xt) dt + Xt dWt, (8.10) where X0 > 0, > 0, > 0, > 0. This form was proposed by Cox, Ingersoll and Ross (1985) as a model for interest rates and it has been widely used in finance. In considering the model (8.10) we shall be concerned with the estimation of = (, ) . The parameter can be regarded as known whenever Brownian motion and continuous sampling are involved. Indeed, can be calculated with probability one on the basis of knowledge of a path of the process on any finite time interval. This follows from the definitions of the quadratic variation process and stochastic integrals with respect to Brownian motion (e.g., Rogers and Williams (1987, Chapter IV, Section 4)) from which one obtains that, writing t (n) i = min(T, 2-n i), lim n i=0 Xt (n) i+1 - Xt (n) i 2 = 2 T 0 Xt dt a.s. and lim n i=0 Xt (n) i t (n) i+1 - t (n) i = T 0 Xt dt a.s. For the model (8.10) the Radon-Nikodym derivative of the measure based on (, ) with respect to that based on (0, 0) is easily seen from (8.9) to be exp -2 T 0 X-1 t [( - Xt) - a0(0 - Xt)] dXt - 1 2 -2 T 0 X-1 t [2 ( - Xt)2 - a2 0(0 - Xt)2 ] dt 134 CHAPTER 8. BYPASSING THE LIKELIHOOD and differentiating the logarithm of this likelihood ratio with respect to = (, ) gives the likelihood score UT = -2 T 0 - Xt X - 1 2 t dWt, (8.11) which is also the quasi-score given by (8.7). If we modify the model to the form dXt = ( - Xt) dt + (, ) Xt dWt where (, ) reflects a possibly rate dependent noise, then the likelihood ratio does not exist in general. Indeed, when (1, ) = (2, ) the supports of the distributions of the two processes are disjoint. The quasi-likelihood methodology, however, is unaffected by this change. The quasi-score estimating function continues to be given by (8.11) and the asymptotic properties of the estimators are also unaffected. From (8.11) the maximum likelihood/quasi-likelihood estimators ^T , ^T are given by T 0 ^T - Xt X-1 t dXt - ^T ^T - Xt dt = 0, T 0 X-1 t dXt - ^T ^T - Xt dt = 0, and putting IT = T 0 X-1 t dXt, JT = T 0 X-1 t dt, KT = T 0 Xt dt, we find that ^T = (IT T - JT (XT - X0))/(JT KT - T2 ), ^T = (IT KT - T(XT - X0))/(IT T - JT (XT - X0)). For the model with 2 2 there is a strictly positive ergodic solution to (8.10) as T whose distribution has gamma density (2/2 , 2/2 ) (see, e.g., Kloeden and Platen (1992, p. 38)). Suppose X has this density; then EX = , EX-1 = 2/(2 - 2 ). (8.12) Using the ergodic theorem we obtain T-1 JT = T-1 T 0 X-1 t dt a.s. - EX-1 , (8.13) T-1 KT = T-1 T 0 Xt dt a.s. - EX, (8.14) 8.3. PARAMETERS IN DIFFUSION TYPE PROCESSES 135 and, since IT = T 0 X-1 t dXt = log X-1 0 XT + 1 2 2 JT using Ito's formula, T-1 IT a.s. - 1 2 2 EX-1 as T . These results readily give the strong consistency of the estimators ^T , ^T . Asymptotic normality of T 1 2 (^T - , ^T - ) can be obtained by applying Theorem 2.1, p. 405 of Basawa and Prakasa Rao (1980). All these results continue to hold under substantially weakened conditions on the noise component in the model. For example, they hold if {Wt} is replaced by a square integrable martingale with stationary independent increments {Zt} for which Z t t. The details involve straightforward applications of the martingale strong law and central limit theorem and are omitted. It should be remarked that there is a considerable recent literature concerning estimation of parameters in diffusion type models, partly motivated by burgeoning applications in mathematical finance. See, for example, Bibby and Srensen (1995), Pedersen (1995), Kessler and Srensen (1995), Kloeden, Platen, Schurz and Srensen (1996) and references therein. In most realistic situations the diffusion cannot be observed continuously, so discrete time approximations to stochastic integrals or a direct approach using the discrete time observations is required. The formulation via (8.6) and (8.7) has to be used with care for problems with multiple sources of variation. Suppose, for example, that the Langevin stochastic differential equation (e.g., Kloeden and Platen (1992, pp. 104­105)) is augmented with jumps coming from a Poisson process and becomes dXt = Xt- dt + dWt + dNt, (8.15) Nt being a Poisson process with intensity . Then, the process may be written in semimartingale form as Mt = Wt + Nt - t. (8.16) Using (8.15) and (8.16) in (8.7), the quasi-score estimating function based on noise {Mt} is T 0 (Xt- 1) dMt, leading to the estimating equations T 0 Xt- dXt = ^T T 0 X2 t dt + ^T T 0 Xt dt, XT = ^T T 0 X2 t dt + ^T T. 136 CHAPTER 8. BYPASSING THE LIKELIHOOD These, however, are the maximum likelihood estimating equations for the model dXt = (Xt + ) dt + dWt, i.e., a version of (8.15) in which Nt has been replaced by its compensator t. In this model the entire stochastic fluctuation is described by the Brownian process and this is only realistic if 1. This problem, first noted by Srensen (1990), can be circumvented and the true maximum likelihood estimators for , obtained if we treat the sources of variation separately. Details are provided in Chapter 2, Section 5. Models of the above kind are quite common and the general message is to identify relevant (local) martingales that focus on the individual sources of variation. It is then possible to obtain quasi-score estimating functions for each, and to combine them, provided the appropriate sample information is available. 8.4 Estimation in Hidden Markov Random Fields Considerable recent interest has been shown in hidden Markov models or more generally, partially observed stochastic dynamical system models. The area embodies a wide class of problems in which a random process (or field) is not observed directly but instead is observed subject to noise through a second process (or field). The techniques that have been developed for this class of problem typically involve full distributional assumptions. Examples are the use of the E-M algorithm (e.g., Qian and Titterington (1990)) or the use of a measure transformation which changes all the signal and noise variables into independent and identically distributed random variables. The monograph Elliott et al. (1994) is devoted to this latter topic and includes a wide range of references to the area. In this section we shall feature a quite different technique, namely, that of optimal combination of estimating functions. This can be used to provide quick derivations of optimal estimating equations for problems of this type. A representative example taken from Heyde (1996) is given to illustrate the approach in the case of hidden Markov models. It is chosen for its clarity rather than its generality. Suppose that we observe continuous variables {yi} on a lattice of sites indexed by i. The dimension of the lattice is d 1 and the observation at site i depends on the value of its neighbors. In order to specify the dependence on neighbors we introduce a matrix W = (wij) for which wii = 0, while wij = 1 if i and j are neighbors and is zero otherwise. The matrix W is assumed to be symmetric and wi denotes the ith column of W . We shall write [i] to denote the neighbors of i. 8.4. ESTIMATION IN HIDDEN MARKOV RANDOM FIELDS 137 We shall consider the field {yi} that satisfies the spatial autoregression specification yi = wiy + i, (8.17) where E i y[i] = 0, var yi y[i] = 2 and is scalar. These processes are also known as conditional autoregressions (CAR). See Ripley (1988, p. 11) for a discussion. Now we are concerned with the context in which {yi} is not observed directly but is hidden in a second field {zi} that is observed on the same lattice. We suppose that the field {zi} satisfies the dynamics zi = k{i[i]} bkyk + i, (8.18) where the {i} are uncorrelated with zero mean and finite variance 2 and are uncorrelated with the { i}, while the {bk} are known. We now seek to estimate {yi}, assuming known and subsequently , on the basis of the observations {zi}. Optimal estimation in general is achieved by identifying the sources of variation and optimally combining estimating functions from each of these. Here we have two sources of noise, namely { i} and {i} to work with. As a prelude to obtaining the quasi-score estimating functions corresponding to (8.17) and (8.18), we rewrite both in matrix form as y = W y + (8.19) and z = By + , (8.20) respectively, where y = (yi), = ( i), z = (zi) and B = (bij), where bij = bj if j {i [i]}, 0 otherwise. Now, we find quasi-score estimating functions from the families {G = A , nonrandom matrix A} and {H = C, nonrandom matrix C} of estimating functions. Suppose that G = A , H = C are the required forms. Applying Theorem 2.1, we find that EG G = AE A and E ˙G = AE y (I - W )y = A(I - W ) so that (E ) A = (I - W ), while EHH = CE C = 2 CC 138 CHAPTER 8. BYPASSING THE LIKELIHOOD and E ˙H = CE y (z - By) = -CB so that C = -B /2 . Since we have taken and as orthogonal we then have the combined estimating function from sources (8.17) and (8.18) as the sum A + C = (I - W ) (E )-1 (I - W ) y - B (z - By)/2 (8.21) (see Section 6.3). Now the specification (8.17) gives Ey y = 2 I + W Ey y , i.e., (I - W ) Ey y = 2 I and E = 2 (I - W ) so that, assuming that (I - W ) is positive definite, (I - W ) (E )-1 (I - W ) = (I - W )/2 and (8.21) becomes ((I - W ) y/2 ) - (B (z - By)/2 ). The quasi-score estimating equation for estimation of y is then (I - W ) y/2 = B (z - By)/2 , i.e., y = (I - W )/2 + B B /2 -1 B z/2 . (8.22) The estimating equation (8.22) is of the same as the one derived by Elliott, Aggoun and Moore (1994, Section 9.3) using assumptions of Gaussianity and Girsanov transformation type arguments to change measures and obtain conditional probability densities. The difference in formulation is just that they have begun from a model in Gibbs field form. The quasi-likelihood approach has enabled the specific distributional assumptions and change of measure arguments to be avoided. Now we suppose that, rather than being known, is to be estimated. Without specific distributional assumptions ordinary likelihood based methods are not available. However, we can find a quasi-score estimating function from the family {G(S) = S - E( S ), S symmetric matrix} to be G(S ) = y W y - tr W (I - W )-1 (8.23) 8.5. EXERCISE 139 following the argument of Section 8.2 and under the supposition that each yi has kurtosis 3. The estimating function (8.23) is, in fact, the score-function under the assumption that y is normally distributed (see, e.g., Lindsay (1988, p. 226)). Estimation of y and can then, in principle, be achieved by simultaneous numerical solution (8.22) and (8.23). Other simpler, albeit less efficient, procedures can be also be found. For example, it is possible to replace (8.23) by the estimating function y W y - y W 2 y; (8.24) as has been mentioned in Section 6.1. A similar approach to that used can be employed to deal with a wide variety of models, including those for which the different sources of variability are not uncorrelated. In each case the combined quasi-score estimating function can be found via Theorem 2.1. For an example of optimal combination of correlated estimating functions see Chapter 6, Section 3. It is also worth remarking that the hidden Markov model problem is structurally similar to problems in statistics involving measurement errors where surrogate predictors are used. For a discussion of the use of quasi-likelihood based methods in this context see Carroll and Stefanski (1990). 8.5 Exercise Suppose that the parameter = ( , ) , where is of interest and is a nuisance component. For estimating functions G() H we partition into G = G G H = H H . Consider the case where the likelihood score U = (U, U) is available and write I = E(U U) where , are or . Show that the likelihood score based estimating function U - II-1 U for is a quasi-score estimating function in H if it belongs to H. (Hint. Use the results of Section 7.3.) Chapter 9 Hypothesis Testing 9.1 Introduction In this chapter we discuss hypothesis testing based on the quasi-score estimating function. This is done by developing analogues of standard procedures of classical statistics. The classical statistical setting for hypothesis testing involves a sequence of T, say, independent random variables whose distribution depends on a pdimensional parameter = (1, 2, . . . , p) belonging to a sample space , an open subset of p-dimensional Euclidean space p . A null hypothesis H0 usually involves specification of the i, i = 1, 2, . . . , p to be functions gi(1, 2, . . . , k) of k , or specifying restrictions Rj(i) = 0 for j = 1, 2, . . . , r, (r + k = p). Tests of H0 against the full model have typically involved one of the three test statistics: (a) Likelihood ratio statistic (Neyman-Pearson) T = 2 LT (^T ) - LT (~T ) ; (b) Efficient score statistic (Rao) T = ST (~T )(IT (~T ))-1 ST (~T ); (c) Wald test statistic T = T R (^T ) W (^T )(IT (^T ))-1 W(^T ) -1 R(^T ), which are ordinarily asymptotically equivalent. Here LT is the log likelihood, R is the vector of restrictions that define H0, W () is the matrix with elements Ri/j, ST () is the likelihood score function LT ()/, IT () is the Fisher information matrix, ^T is the MLE for the full model and ~T is the MLE under the restrictions imposed by the hypothesis H0. Note that the efficient score statistic depends only on the MLE for the restricted class of parameters under H0, while Wald's statistic depends only on the MLE over the whole parameter space. For further details see Rao (1973, Section 6e). In the case of hypothesis testing for stochastic processes there is now a reasonably well developed large sample theory based on the use of the likelihood score. See for example Basawa and Prakasa Rao (1980), particularly Chapter 141 142 CHAPTER 9. HYPOTHESIS TESTING 7 and Basawa and Scott (1983, Chapter 3). The use of what are termed generalized M-estimators is discussed in Basawa (1985) in extension of the efficient score and Wald test statistics and these ideas are further developed in Basawa (1991). This path is also followed in the present chapter. However, much of the discussion in the existing literature has involved testing relative to sequences of alternative hypotheses. We shall here avoid this formulation to focus on the special features of the quasi-likelihood setting. The discussion here is based on Thavaneswaran (1991) together with various extensions. 9.2 The Details Neither the efficient score statistic nor the Wald statistic involve the existence of a scalar objective function (the likelihood) from which an optimal estimating function is derived by differentiation. Each can therefore be generalized straightforwardly to an estimating function setting and this we shall do below. We shall not consider restricted settings, such as that of the conservative quasiscore (Li and McCullagh (1994)) in which a likelihood ratio test generalization is available. The asymptotic properties of such a likelihood ratio test will be the same as those of the efficient score and Wald test generalizations (see the exercise of Section 9.3). Suppose that Q() is a quasi-score estimating function, for , in standardized form, chosen from an appropriate family H of estimating functions. Here and below we drop the suffix T for convenience. As usual we take V = V () = E(Q()) = E(Q() Q ()) for the information matrix. In the following denotes the unrestricted quasi-likelihood estimator whereas ~ is the quasi-likelihood estimator calculated under hypothesis H0. Most testing problems for stochastic models concern linear hypotheses on the unknown parameter. We shall follow the ideas of Section 7.2 on constrained parameter estimation, and consider testing H0 : F = d against H1 : F = d where F is a p × q matrix not depending on the data or with full rank q p. Note that this framework includes the problem of testing a subvector of length q ( p), say 2, of = (1, 2) with H0 : 2 = 02 to be tested against H1 : 2 = 02. Just take the block matrix form F = O(p-q)×(p-q) O(p-q),q Oq,(p-q) Iq , where Or,s is the r × s matrix all of whose elements are zero and, of course, Iq is the q × q identity matrix. For the H0 : F = d setting and for the ergodic case (see Section 4.3) we have that the efficient score statistic analogue is = Q (~)(E(Q(~)))Q(~) (9.1) 9.2. THE DETAILS 143 and the Wald test statistic analogue is = (F - d) (F (E(Q( )))F )(F - d). (9.2) These test statistics are, of course, the classical ones in the case where the quasi-score estimating function is the likelihood score. We would envisage using the test statistics and in circumstances under which V - 1 2 () Q() d MV N(0, Ip) (9.3) and, using Taylor series expansion, V 1 2 () ( - ) d MV N(0, Ip). (9.4) Then, both and are approximately distributed as 2 q under H0. From the proof of Theorem 7.2 we have that Q(~) d P Q(), where P = F (F V -1 F )F V -1 , so that using (9.3), Q(~) MV N(0, P V P ), while E(Q(~)) P V P so that d (P Q) (P V P )P Q = V - 1 2 Q V 1 2 P (P V P )P V 1 2 (V - 1 2 Q), which is approximately distributed as 2 q in view of (9.3) and since V 1 2 P (P V P )P V 1 2 is idempotent with rank q. On the other hand, from (7.2) we have that ~ = - V -1 F (F V -1 F )(F - d) (9.5) and, using (9.4) and H0, F - d d MV N(0, F V -1 F ). Then, since E(Q( )) E(Q()) = V we have that is also approximately distributed as 2 q. 144 CHAPTER 9. HYPOTHESIS TESTING Under H1 we generally have that (9.3) and (9.4) still hold and, for large T, upon expanding Q() about to the first order, and assuming that ˙Q approximates -V (as in the discussion of the proof of Theorem 7.2), Q(~) -V (~ - ) = F (F V -1 F )(F - d) from (9.5). Thus, from (9.4), F d MV N(F , F V -1 F ) (9.6) and then Q(~) d MV N(F (F V -1 F )(F - d), F (F V -1 F )F ). (9.7) We also have E(Q( )) E(Q()) = V (9.8) and E(Q(~)) F (F V -1 F )F . (9.9) Then, from (9.7) and (9.9) in the case of and (9.6) and (9.8) in the case of we find that both and are approximately distributed as noncentral 2 q, with noncentrality parameter C(V ) = (F - d) (F V -1 F )(F - d). (9.10) The test based on or using the quasi-score Q is optimal in the sense of providing a maximum noncentrality parameter under H1 for estimating functions within the family H. If another estimating function were chosen from H, then, under appropriate regularity conditions, we would obtain (9.10) with V replaced by V 1 say, with V 1 V in the Loewner ordering. It is easily seen that C(V 1) C(V ). This justifies the choice of the quasi-score as a basis for hypothesis testing. These results can readily be extended to the non-ergodic case when {QT } is a martingale. In this setting we replace (9.1) by = QT (^) Q(^) T QT (^), (9.11) where Q() T is the quadratic characteristic (see Section 4.3) and replace (9.2) by = F - d F Q( ) T F F - d . (9.12) The role of V in the above discussion is now taken by the (random) quantity Q . 9.3. EXERCISE 145 9.3 Exercise Consider the setting of Section 9.2 in which H0 : F = d is to be tested against H1 : F = d. Suppose that Q() is a quasi-score estimating function chosen from some family H of estimating functions and that there is a scalar function q such that Q() = q/. The natural analogue of the likelihood ratio statistic in this setting is = 2(q( ) - q()). Using the first order approximations of this chapter show that d ( - ~) V ( - ~) and that this is approximately distributed as 2 q under H0 or as noncentral 2 q with noncentrality parameter (9.10) under H1. Chapter 10 Infinite Dimensional Problems 10.1 Introduction Many nonparametric estimation problems can be regarded as involving estimation of infinite dimensional parameters. This is the situation, for example, in the neuronal membrane potential model dV (t) = (- V (t) + (t)) dt + dM(t) discussed earlier in Section 2.6 with parameters held constant, if changes in (t) over the recording interval 0 t T are important and the function needs to be estimated. Both the situation where n replica trajectories of {V (t), 0 t T} are observed and n or one trajectory is observed and T are of interest but most of the established results deal with the former case. Various possible estimation procedures have been developed for use in such problems including kernel estimation and the method of sieves. Efforts have also been made to develop the quasi-likelihood methodology in a Banach space setting to deal directly with infinite dimensional problems (e.g., Thompson and Thavaneswaran (1990)) but the results are of limited scope and no asymptotic analysis has been provided. The topic is of sufficient intrinsic importance to be dignified with a chapter, albeit brief and more suggestive than definitive. Here the quasi-likelihood methodology described elsewhere in the book does not play a central role. We shall first sketch the approach via the method of sieves and then offer some heuristic discussion related to particular problems. 10.2 Sieves The method of sieves was first developed by Grenander (1981) as a natural extension of the finite dimensional theory. It is ordinarily applied in the context where there is replication of some basic process. We use as a sieve a suitably chosen complete orthonormal sequence {i(t)}, say, and approximate the unknown function (t) by (n) (t) = n k=1 i i(t), where n increases to infinity as the sample size increases. Then, the i are estimated, say by quasi-likelihood, to give an optimal estimate (n) (t) of (n) (t). The remaining task is to show that as n , (n) (t) is consistent for (t) and, hopefully, asymptotically normally distributed. 147 148 CHAPTER 10. INFINITE DIMENSIONAL PROBLEMS Considerable recent literature has been devoted to such studies. For example, Nguyen and Pham (1982) established consistency of the sieve estimator of the drift function in a linear diffusion model. McKeague (1986) dealt with a general semimartingale regression model and obtained a strong consistency result for certain sieve estimators. Sieve estimation problems for point processes, such as in the multiplicative intensity model, have been studied by Karr (1987), Leskow (1989) and Leskow and Rozanski (1989). Kallianpur and Selukar (1993) have provided a general discussion of sieve estimation where the focus is on maximum likelihood for the nested finite dimensional problems. Rate of convergence results for sieve estimators are discussed in Shen and Wong (1994). The method is a valuable one but it does suffer from the disadvantage of lack of exlicit expressions and the problem in practice of trade-offs between approximation error and estimation error. Much remains to be done and a general discussion, based on the use of quasi-likelihood for the nested finite dimensional problems, is awaited. 10.3 Semimartingale Models Here we shall consider semimartingale models of the form dXt = dAt + dMt, where {Xt} is the observation process, {Mt} is a square integrable martingale and the predictable bounded variation process {At} is of linear form dAt = t dBt in the unknown function t of interest, {Bt} being an observable process, possibly a covariate or a function of the observation process, such as dBt = Xt-dt. Observations on the process typically involve a history over an extended time interval 0 t T, say, or we have copies X1t, . . . , Xnt, say, of the process over a fixed time interval, which we can take as [0, 1]. In any case, we seek estimators that, in a suitable sense, are consistent as T or n . For models of the type considered here, the semimartingale representation immediately suggests a simple estimation procedure. If we have n histories, write Xt = (X1t, . . . , Xnt) , Mt = (M1t, . . . , Mnt) and Bt = (B1t, . . . , Bnt) . Then, dXt = t dBt + dMt, so that if 1 is the n × 1 vector, all of whose elements are unity, T 0 s ds = T 0 (1 dBs)-1 ds (1 dXs) - T 0 (1 dBs)-1 ds (1 dMs), and T = T 0 s ds (10.1) 10.3. SEMIMARTINGALE MODELS 149 is unbiasedly estimated by ^T = T 0 (1 dBs)-1 ds (1 dXs). (10.2) In a certain formal sense, t is estimated by (1 dBt)-1 (1 dXt), the derivative of a generally (pointwise) non-differentiable process. This approach is convenient but of course it does not readily lead to an estimator of t, but rather of its indefinite integral t. Since ^T - T = T 0 (1 dBs)-1 ds (1 dMs) (10.3) is a square integrable martingale, we obtain from the Burkholder, Davis, Gundy inequality (e.g., Theorem 4.2.1, p. 93 of Rogers and Williams (1987)) that for C a universal constant, E sup T [0,1] (^T - T )2 C E(^1 - 1)2 = C 0 (1 dBs)-1 ds (1 dMs) 1 = C 1 0 (1 dBs)-2 (ds)2 d 1 M s so we have consistency if 1 0 (1 dBs)-2 (ds)2 d 1 M s 0 as n . Various central limit results can also be formulated on the basis of (10.3). On the other hand, if we have a single history it is usually the case that T 0 (dBs)-1 ds dXs T 0 s ds a.s. - 1 (10.4) as T . Note that T 0 (dBs)-1 ds dXs T 0 s ds = 1 + T 0 (dBs)-1 ds dMs T 0 s ds , and for the martingale Nt = T 0 (dBs)-1 ds dMs, 150 CHAPTER 10. INFINITE DIMENSIONAL PROBLEMS we have d N t = (dBt)-2 (dt)2 d M t and, so long as N T a.s. as T , the martingale strong law (e.g., Theorem 12.5) ensures that NT / N T a.s. - 0, i.e. (10.4) holds if lim sup T T 0 (dBs)-2 (ds)2 d M s T 0 s ds < a.s. (10.5) This is easily checked in practice for simple models. For example, if dBs = ds and d M s = ds we require lim sup T T T 0 s ds < a.s. Sometimes minor adjustments are necessary, as for example in the case of the Aalen (1978) model for nonparametric estimation of the cumulative hazard function using censored lifetime data. In this setup Xt is a counting process representing the number of deaths up to time t, dBt = t Yt dt, where Yt is the number at risk at time t (possibly zero) and t is the hazard function of the lifetime distribution. Then, d M s = s Ys ds, and using I to denote the indicator function, T 0 I(Ys > 0) Y -1 s dXs T 0 I(Ys > 0) s ds = 1 + T 0 I(Ys > 0) Y -1 s dMs T 0 I(Ys > 0) s ds , and provided that its denominator goes a.s. to infinity. T 0 I(Ys > 0) Y -1 s dMs 0 I(Ys > 0) Y -1 s dMs T = T 0 I(Ys > 0) Y -1 s dMs T 0 I(Ys > 0) Y -1 s s ds a.s. - 0 using the above mentioned martingale strong law. The consistency result T 0 I(Ys > 0) Y -1 s dXs T 0 I(Ys > 0) s ds a.s. - 1 as T then follows since Ys 1 on I(Ys > 0). It is possible to use the quasi-likelihood theory of this book if, for example, t is assumed to be a step function with jumps at 0 = t0 < t1 < . . . < tk-1 < tk = T, namely, t = j, tj-1 t < tj, j = 1, 2, . . . , k. Then, postulating the model d ~Xt = k j=1 j I(tj-1 t < tj) d ~Bt + d ~Mt, say, the Hutton-Nelson quasi-score for estimation of = (1, . . . , k) is T 0 diag (I(t0 t < t1), . . . , I(tk-1 t < tk)) d ~Bt d ~M t d ~Mt 10.3. SEMIMARTINGALE MODELS 151 from which we see that ^j = tj tj-1 d ~Bt d ~M t d ~Xt tj tj-1 (d ~Bt)2 d ~M t , 1 j k, (10.6) are the corresponding quasi-likelihood estimators if ~M t does not involve the 's. If, on the other hand, d ~M t = t d ~Bt as for counting process models, then we find that ^j = ~Xtj - ~Xtj-1 ~Btj - ~Btj-1 , 1 j k, (10.7) for the quasi-likelihood estimators. One can envisage approximating the model dXt = t dBt + dMt, (10.8) observed over 0 t T, by a sequence of models such as d ~Xt = N j=1 j I j - 1 N t < j N d ~Bt + d ~Mt where N . This idea is formalized in the theory of the histogram sieve. An alternative approach would be to use a discrete model approximation to (10.8), such as provided by the Euler scheme, say ~Xj - ~X(j-1) = j ~Bj - ~B(j-1) + ~Mj - ~M(j-1) , (10.9) j = 1, 2, . . . , [T/], [x] denoting the integer part of x. Here the Hutton-Nelson quasi-score leads to the estimators ^j = ~Xj - ~X(j-1) ~Bj - ~B(j-1) , 1 j [T/], (10.10) regardless of the form of the quadratic characteristic of the martingale { ~Mt}. It should be noted that (10.7) and (10.10) accord with the interpretation mentioned above of the estimator of t as dXt/dBt. They are straightforward to deal with numerically. For asymptotic results we need, for example, replication as indicated above and it is again worth remarking on the model approximation error and the estimation error. It can ordinarily be expected that for Euler schemes the error in model approximation is O( 1 2 ) (see, e.g., Section 9.6 of Kloeden and Platen (1992) which readily extends beyond the diffusion context). On the other hand, if n replicates are being used, the estimation error is O(n 1 2 ) (central limit rate). Chapter 11 Miscellaneous Applications 11.1 Estimating the Mean of a Stationary Process A method of last resort, when more sophisticated methods are not tractable or not available, is the method of moments, which typically relies on an ergodic theorem and involves no significant structural assumptions. Nevertheless, in various circumstances, such as for estimating the mean of a stationary process, it turns out that the method of moments (which estimates EXi = by T-1 T i=1 Xi) produces an estimator that has the same asymptotic variance as the best linear unbiased estimator (BLUE) under broad conditions on the spectral density or covariance function. This is of considerable practical significance. The discussion in this section follows the papers Heyde (1988b), (1992b). Let {Xt, t = . . . , -1, 0, 1, . . .} be a stationary ergodic process with EXt = , which is to be estimated, and EX2 t < . Write (t - s) = cov (Xs - , Xt - ) = - ei(s-t) f() d for the covariance function, where f() = (2)-1 (0) + 2 j=1 (j) cos j , = 0, || , is the spectral density, the spectral function being assumed to be absolutely continuous. This is the case for all purely nondeterministic (i.e., contain no component which is exactly predictable) processes (e.g., Hannan (1970, Theorem 3 , p. 154)). Our concern here is to estimate on the basis of observations (X1, . . . , XT ) and we shall first restrict consideration to the class of estimating functions H1 = T i=1 ai,T (Xi - ), ai,T constants, T i=1 ai,T = 1 . Then, if GT = T i=1 ai,T (Xi - ), G T = T i=1 a i,T (Xi - ), we find that EGT = -1, EGT G T = T i=1 T j=1 a i,T aj,T (j - i) 153 154 CHAPTER 11. MISCELLANEOUS APPLICATIONS and, using Theorem 2.1, G T is a quasi-score estimating function within H, if T i=1 (j - i) a i,T = cT (constant), j = 1, 2, . . . , T. (11.1) In this case the quasi-score estimating function leads to a quasi-likelihood estimator that is just the best linear unbiased estimator (BLUE) MV possessing minimum asymptotic variance. To see this, note that EG2 T - EG2 T = E(GT - G T )2 + 2EG T (GT - G T ) = E(GT - G T )2 0. Of course the calculation of the BLUE involves a full knowledge of the covariance structure of the process so that it is ordinarily not feasible to use it in practice. To obtain an expression for the minimum variance cT , we write (T) for the T × T matrix ((i - j)) and let (T) and e(T) denote the T-vectors ( 1,T , . . . , T,T ) and (1, 1, . . . , 1) , respectively, the prime denoting transpose. Then, (11.1) gives (T) (T) = cT e(T), while (e(T)) (T) = 1, and hence cT = [(e(T)) ( (T))-1 e(T)]-1 . (11.2) The explicit expressions for (T) and cT are, however, of limited practical value as, even when the covariances (i) are known, they involve calculation of the inverse of the Toeplitz matrix (T), which will usually be of high order. On the other hand, the method of moments estimator M = T-1 T i=1 Xi uses equally weighted observations and requires no direct knowledge of the underlying distribution. We shall examine conditions under which var MV = [(e(T)) ( (T))-1 e(T)]-1 (11.3) and var M = T-2 (e(T)) (T) e(T) (11.4) have the same asymptotic behavior. That is, the estimator M is an asymptotic quasi-score estimating function. 11.1. ESTIMATING THE MEAN OF A STATIONARY PROCESS 155 Theorem 11.1 Suppose that the density f() of {Xt} is continuous and positive at = 0. Then T var M 2f(0) as T . (11.5) If, in addition, - (|p()|2 /f()) d < for some trigonometric polynomial p(), then T var MV 2f(0) as T , (11.6) and T 1 2 (MV - M ) p - 0 as T , (11.7) p denoting convergence in probability. Proof The result (11.5) is a well-known consequence of the representation var M = 1 T2 - sin(T/2) sin(/2) 2 f() d; see, for example, Ibragimov and Linnik (1971, p. 322). The result (11.6) follows from Theorem 3 of Adenstadt and Eisenberg (1974). To obtain (11.7) we note that T E(MV - M )2 = T E T i=1 ( i,T - T-1 ) Xi 2 = T T i=1 ( i,T - T-1 ) T j=1 ( j,T - T-1 ) (i - j) = T T i=1 ( i,T - T-1 ) cT - T-1 T j=1 (i - j) = - T i=1 ( i,T - T-1 ) T j=1 (i - j) = T(var M - var MV ) 0 as T using (11.5) and (11.6) and the result follows. This completes the proof. It should be noted from (11.7) that if a central limit result holds for T 1 2 (M -), then the corresponding one also holds for T 1 2 (MV - ). Various conditions of asymptotic independence lead to the result T 1 2 (M - ) = T- 1 2 T i=1 (Xi - ) d - N(0, 2 f(0)), 156 CHAPTER 11. MISCELLANEOUS APPLICATIONS namely, convergence in distribution to the normal law with zero mean and variance 2 f(0). For example, this holds if there is a -field M0 such that X0 is M0-measurable and k=0 E|E(Xk | M0) - | < (Hall and Heyde (1980, Theorem 5.4, p. 136)). This is a very general result from which many special cases, such as for mixing sequences, follow. For the results of Theorem 11.1 to be useful it is essential that f(0) > 0. An example where this may not be the case is furnished by the moving average model Xt = + t - r t-1, |r| 1, where the i are stationary and orthogonal with mean 0 and variance 2 . Here f() = 2 (1 + r2 - 2r cos )/2, so that f(0) > 0 requires r < 1. If r = 1 it turns out that (Grenander and Szeg¨o (1958, p. 211)), var M 22 T-2 , var MV 122 T-3 as T , so that the method of moments estimator is far from efficient in this case. In fact, if f(0) = 0, it holds rather generally that var MV = o(var M ) as T (Adenstadt (1974), Vitale (1973)). Now Theorem 11.1 can be used to show that MV and M have the same asymptotic variance if the spectral density f() is everywhere positive and continuous. Such information is ordinarily not directly available but a useful sufficient condition in terms of covariances is given in the following theorem. Theorem 11.2 If the covariance (T) decreases monotonically to zero as T , then {Xt} has a spectral function that is absolutely continuous. The spectral density f() is nonnegative and continuous, except possibly at zero, being continuous at zero when j=1 (j) < . If {(n)} is convex and (0) + (2) > 2(1), then the spectral density is strictly positive. Proof The first two parts of the theorem are given in Theorem 3 of Daley (1969). For the last part it should be noted that when {(n)} is convex, namely, 2 (n) = (n) + (n + 2) - 2(n + 1) 0, n 0, the spectral density f() can be expressed in the form f() = (2)-1 n=0 (n + 1)2 (n) Kn(), || , 11.1. ESTIMATING THE MEAN OF A STATIONARY PROCESS 157 where Kn() = 2 n + 1 sin((n + 1)/2) 2 sin(/2) 2 0 is the Fej´er kernel (Zygmund (1959, p. 183)). Then, since K0() = 1 2 and 2 (n) Kn() 0 for n 1, f() > 0 if 2 (0) > 0 as required. Corollary 11.1 If the covariance (T) decreases monotonically to zero as T , j=1 (j) < and {(n)} is convex with (0) + (2) > 2(1), then var MV var M T-1 (0) + 2 j=1 (j) as T . This result follows immediately from Theorem 11.1 and 11.2. The particular case of the result of Corollary 11.1 for which the covariance function (n) is of the form (n) = K i=1 ai e-ri|n| (11.8) where ai > 0, i = 1, 2, . . . , K, 0 < r1 < r2 < . . . , rK, has been obtained by Halfin (1982). Covariance functions of the type (11.8) are common for Markovian queueing models with limited queue space and various applications are given in Halfin's paper. Corollary 11.1, however, is very much more widely applicable and many examples can be noted from the review article of Reynolds (1975) on the covariance structure of queues and related processes. As one example, we mention the queue length process in an M/G/ queue (Reynolds (1975, p. 386)) where (n) = (n)/(0) = n G(x) dx with -1 = ES, the mean service time and G(x) = P(S x). Then, it is easily checked that n=0 (n) < is equivalent to ES2 < , while convexity of {(n)} is automatically satisfied and (0) + (2) > 2(1) requires 1 0 G(x) dx > 2 1 G(x) dx. Another example involves the waiting time {Wn} in an M/G/1 queue where (Reynolds (1975, p. 401)) (n) = 1 - n(1 - ) - n r=1 Cr 2 2 , n 1, 158 CHAPTER 11. MISCELLANEOUS APPLICATIONS for certain constants Cr , with = EWn, 2 = var Wn, the traffic intensity and the mean arrival rate. Here it is again easily checked that {(n)} is monotone and convex with (0) + (2) > 2(1) while n=0 (n) < if and only if ES4 < . Now if the spectral density has zeros, Theorem 11.1 allows for their removal by |p()|2 . This can be done, for example, for the queue length process in an M/D/ queue where the constant service time T = 2. For the M/D/ queue with constant service time S, (j) = (j)/(0) has the particularly simple form (j) = 1 - |j|/S, |j| S, 0, otherwise (Reynolds (1975, p. 387)) and (j) = T2 (j) where denotes the mean arrival rate. When S = 2, the spectral density is given by 2f() = 4(1 + cos ) and we find that T var MV 8, T var M 8 (11.9) as T despite the fact that f(-) = f() = 0. Now a crucial property that underlies the parity of (11.3) and (11.4) is that of short-range dependence. It is said that short-range or long-range dependence holds according as j=1 (j) converges or diverges or, essentially equivalently, the spectral density f() converges or diverges at = 0. For long-range dependence, T var XT ordinarily diverges as T in contrast to the short-range dependence case where convergence usually holds. Indeed, rather typical of the behavior in the long-range dependence case is the convergence in distribution of T+1/2 ( XT - ) to a proper limit law as T where -1 2 < < 0. The limit law is not Gaussian in general, save when the stationary process {Xt} is itself Gaussian. Little is known in general about the asymptotic behavior of MV . This should be contrasted with the case of short-range dependence where T1/2 ( XT - ) converges to the normal law N(0, 2 f(0)) under very broad conditions and then a similar result holds for T1/2 (MV - ) as readily follows from (11.7). These results are vitally important as they provide the basis for asymptotic confidence statements about . Efficiency comparisons of X and MV for the long-range dependence case appear in Samarov and Taqqu (1988), while Beran (1989) has provided confidence intervals (under Gaussian assumptions) that take the estimation of the parameter into account. The relevance of all this, however, rests critically on the assumption of Gaussianity because efficiency calculations based on variances may be of little use in assessing sizes of confidence zones save when 11.2. ESTIMATION FOR A HETEROSCEDASTIC REGRESSION 159 Gaussian limits obtain. Gaussian limits would not be typical for queueing models under long-range dependence. The parity of (11.5) and (11.6) for short-range dependence also breaks down in the case where the spectral density f() has a zero at = 0. Again efficiency comparisons can sometimes be made, for example in the moving-average model considered earlier, but the problem remains that asymptotic normality does not hold for the estimators in general. 11.2 Estimation for a Heteroscedastic Regression The model we consider in this section is the general linear regression model y = X + u, (11.10) where y is an n × 1 vector of observations, X is an n × k matrix of known constants of rank k < n, is a k × 1 vector of unknown parameters and u is an n×1 vector of independent residuals with zero mean and covariance matrix = diag (g1(), . . . , gn()), where gi() = 2 (Xi)2(1-) , Xt = (Xt1, . . . , Xtk), (t = 1, . . . , n), and X = (X1 . . . . . . Xn) . Variances proportional to a power of expectations have been widely observed, for example in cross-sectional data on microeconomic units such as firms or households. The object is the efficient estimation of the (k + 2) × 1 vector = ( 2 ) . The results here are from Heyde and Lin (1992). The obvious martingales to use for the model (11.10) are {Ut} and {Vt} given by Ut = t s=1 us, Vt = t s=1 (u2 s - gs()). Then, the quasi-score estimating function based on {Ut} is n t=1 (Xt 0 0) g-1 t () ut = u -1 X . . . 0 0 , (11.11) which focuses on , provides no information on 2 and involves as a nuisance parameter. On the other hand, the quasi-score estimating function based on {Vt} is n t=1 (˙gt()) (var u2 t )-1 (u2 t - gt()) = vec vec F -1 ( - u u ), (11.12) 160 CHAPTER 11. MISCELLANEOUS APPLICATIONS where F = diag (var u2 1, . . . , var u2 n), which allows estimation of all components of provided F is a known function of . However, efficiency in the estimation is enhanced if the martingales {Ut} and {Vt} are used in combination and, using results of Chapter 6 (e.g., (6.5)), the combined quasi-score estimating function is n t=1 (Xt 0 0) [ut + Rt (u2 t - gt())] gt()(1 - RtSt) + n t=1 (˙gt()) [St ut + (u2 t - gt())] (var u2 t )(1 - RtSt) , (11.13) where Rt = Eu3 t / var u2 t , St = Eu3 t /gt(). This simplifies considerably when Eu3 t = 0 for each t, for then the martingales {Ut} and {Vt} are orthogonal and (11.13) is a sum of the quasi-score estimating functions (11.11) and (11.12). In this case we may separate the corresponding estimating equation into the two components X u + vec vec F -1 ( - u u ) = 0, (11.14) vec vec F -1 ( - u u ) = 0, where = (, 2 ) . If the ut are normally distributed, F = 2 and the quasi-likelihood estimator for coincides with the maximum likelihood estimator. The equation (11.14) then corresponds to the true score equations (4.2), (4.3) of Anh (1988). The martingale information (see Section 6.3) contained in (11.11) is IQS(U) = X -1 X Ok×2 O2×k O2×2 , which clearly contributes only to the estimation of while that in (11.12) is IQS(V ) = n t=1 (var u2 t )-1 (˙gt()) (˙gt()) = A F -1 A, where A = (gi()/j), and the combined information in (11.13) is IQS(U,V ) = X -1 (I - R S)-1 X . . . Ok×2 O2×k . . . O2×2 + A F -1 (I - R S)-1 A + X -1 (I - R S)-1 RA O2×(k+2) (11.15) 11.2. ESTIMATION FOR A HETEROSCEDASTIC REGRESSION 161 + A R(I - RS)-1 -1 X . . . O(k+2)×2 , where R = diag (R1, . . . , Rn), S = diag (S1, . . . , Sn), which of course is IQS(U)+ IQS(V ) when Eu3 t = 0 for each t. In contrast to the above estimating functions, Anh (1988) used nonlinear least squares based on minimization of n t=1 (e2 t - Ee2 t )2 , (11.16) where the et are the (observable) elements of the vector e = (I - X(X X)-1 X ) u = P u = P y, where P = I - X(X X)-1 X . Note that (11.16) amounts to using the estimating function n t=1 ˙ft()(e2 t - ft()), where ft() = Ee2 t . Anh showed that this leads to a strongly consistent and asymptotically normal estimator ^A, say, satisfying ^A - d MV N(0, (A A)-1 A F A(A A)-1 ) for large n under his Conditions 1­3. However, under the same conditions, minor modifications of the same method of proof can be used to establish that ^QS(V ) - d MV N(0, (A F -1 A)-1 ) and ^QS(U,V ) - d MV N(0, I-1 QS(U,V )), where ^QS(V ) and ^QS(U,V ) are, respectively, the quasi-likelihood estimators based on (11.12) and (11.13). By construction we have that I-1 QS(V ) - I-1 QS(U,V ) in nonnegative definite, while (A A)-1 A F A(A A)-1 - (A F -1 A)-1 is also nonnegative definite. This last result holds because, for a covariance matrix F and n×p matrix A, the Gauss-Markov Theorem gives nonnegativity of BF B - (A F -1 A)-1 for every p × n matrix B satisfying BA = Ip (e.g., Heyde (1989)), and the result holds in particular for B = (A A)-1 A . Thus, from the point of view of asymptotic variance, the order of preference is ^QS(U,V ), ^QS(V ), ^A. 162 CHAPTER 11. MISCELLANEOUS APPLICATIONS 11.3 Estimating the Infection Rate in an Epidemic The discussion in this section is based on the results of Watson and Yip (1992). We shall begin by considering the simple stochastic epidemic model for a closed population of size N and where the infection rate is . Let S(t) denote the number of susceptibles and I(t) the number of infectives as time t, so that S(t) + I(t) = N, and suppose that I(0) = a. The simple epidemic process (e.g., Bailey (1975, p. 39)) is a Markov process for which, when I(t) infectives are present at time t, the chance of an infective contact with a specified individual in the time interval (t, t + t) is I(t) t + o(t), where is the infection rate, which is taken as constant. Then, since there are N - N(t) susceptibles present at time t, the probability of an infection in time (t, t + t) is I(t)(N - I(t)) t + o(t). It is assumed that the probability of more than one infection in (t, t + t) is in o(t). Then, if Ft denotes the -field of history up to time t, we have in an obvious notation, E d I(t) Ft- = I(t)(N - I(t)) dt and the basic zero mean martingale defined on the infectives process {I(t)} is M(t) = I(t) - I(0) - t 0 E d I(s) Fs= I(t) - I(0) - t 0 I(s)(N - I(s)) ds. (11.17) Then, the Hutton-Nelson family of estimating based on the martingale (11.17) is M = t 0 (s) dI(s) - t 0 (s) I(s)(N - I(s)) ds, (s) predictable from which the quasi-score estimating function for estimation of is obtained by choosing (s) 1 for the predictable weight function. This follows (e.g., using results of Section 2.5) since it is easily seen that E d ˙M(t) Ft- = -I(t)(N - I(t)) dt, d M t = E (dM(t))2 Ft- = E dM(t) Ft= I(t) (N - I(t)) dt. The quasi-likelihood estimator of based on complete observation of the process {I(t)} over [0, T] is then ^T = (I(T) - a) T 0 I(s)(N - I(s)) ds (11.18) 11.3. ESTIMATING THE INFECTION RATE IN AN EPIDEMIC 163 and it should be noted that this coincides with the maximum likelihood estimator. It is not difficult to check that ^T is a strongly consistent estimator of and (I(t) - a)- 1 2 M(T) d - N(0, 1) as T so that (I(t) - a) 1 2 ^-1 T (^T - ) d - N(0, 1) from which confidence intervals for could be constructed. In practice, one would not expect complete observation of the process. The available data might, for example, be of the form {(tk, ik), k = 0, 1, . . . , m}, where m 1, 0 t0 < t1 < . . . < Tm T are fixed time points and ik is the observed number of infectives at time tk. Then, a reasonable approximation to the estimator (11.18) is provided by (I(tm)-a) 1 2 (I(t0)(N - I(t0)) + I(tm)(N - I(tm))) + m-1 r=1 I(tr)(N - I(tr)) which is obtained by using the simple trapezoidal rule to approximate the stochastic integral. This estimator is due to Choi and Severo (1988). A rather more realistic model is provided by the general stochastic epidemic model that allows for removals from the population (e.g., Bailey (1975, pp. 88- 89)). Here the probability of one infection in the time interval (t, t + t) is I(t) S(t) t + o(t) and of one removal is, say, I(t) (t) + o(t), while the probability of more than one change is o(t). Suppose S(0) = n and I(0) = a. Since the removals come only from the infectives, we have that E dS(t) Ft- = - I(t) S(t) dt and the basic zero-mean martingale defined on the process {S(t)} is N(t) = n - S(t) + t 0 E dS(s) Fs= n - S(t) - t 0 I(s) S(s) ds. Estimation of can then be achieved along similar lines to those described above for the case of the simple epidemic. The quasi-score estimating function for the Hutton-Nelson family based on the martingale {N(t)} is n - S(T) - T 0 I(s) S(s) ds, so that the quasi-likelihood estimator is ^T = (n - S(T)) T 0 I(s) S(s) ds. (11.19) 164 CHAPTER 11. MISCELLANEOUS APPLICATIONS Again ^T is strongly consistent for and asymptotically normally distributed, this time with (n - S(T)) 1 2 ^-1 T (^T - ) d - N(0, 1) as T . Similar ideas can be applied for the estimation of the removal rate . If R(t) denotes the number of removals up to time t, we have that E dR(t) Ft- = I(t) dt and the basic zero-mean martingale defined on the process {R(t)} is K(t) = R(t) - t 0 E dR(s) Fs= R(t) - s 0 I(s) ds. We have E d ˙K(t) Ft- = I(t) dt, the dot now referring to differentiation with respect to , and d K t = E (dR(t))2 Ft- = I(t) dt, so that, as in the previous cases, the basic zero-mean martingale is itself the quasi-score for the Hutton-Nelson family. The quasi-likelihood estimator of is then ^T = R(T) T 0 I(s) ds. (11.20) Again we would not expect complete observation of the epidemic process. The available information might be of the form {(tk, sk, ik), k = 0, 1, . . . , m}, which provides progressive information on both the numbers of susceptibles and infectives. This would allow for approximation to the quasi-likelihood estimator ^T given in (11.19), but the removal rate would not be able to be estimated via (11.20) without information on the number of removals. 11.4 Estimating Population Size from Multi- ple Recapture Experiments Consider a population in which there are animals, where is unknown. A multiple recapture experiment is conducted in which animals are marked at the time of their first capture. At each subsequent capture the mark is correctly 11.4. ESTIMATING POPULATION SIZE 165 recorded. Apart from time allowed for marking or noting of marks, each captured animal is released immediately. The aim is to make inferences about from the set of observed captures of marked and unmarked animals. The results in this section follow Becker and Heyde (1990) and are derived using a continuous-time formulation. Analogous results also hold for multiple recapture experiments in discrete time. We label the animals by 1, 2, . . . , and let Ni(t) denote the number of times animal i has been caught in [0, t]. Write Nt for (N1(t), . . . , N(t)) . Each {Ni(t); t 0} is a continuous-time counting process with right-continuous sample paths and common intensity function defined by t = lim h0 1 h Pr Ni(t + h) = Ni(t) + 1 Ft , where Ft is the -field generated by {Nx : 0 x t}. The dependence of on t can reflect the dependence of the animal's behavior on time as well as variations in the intensity of trapping. Consider observations from a multiple capture experiment over [0, ]. The total number of captures by time t is given by Nt = i=1 Ni(t). Write Mt for the number of animals that are marked by time t. Then Mt is the number of captures of unmarked animals, while Kt = Nt - Mt gives the number of captures of marked animals by time t. Maximum likelihood estimation of based on observation of {Mt, Nt, 0 t } is complicated by the fact that the intensity function x is generally unknown. We find that the log-likelihood function is L = 0 log(x) dNx - 0 x dx + 0 log( - Mx-) dMx = 0 log( x) dNx - 0 x dx - N log + 0 log( - Mx-) dMx, where x = x is an arbitrary positive value if x is arbitrary positive. The likelihood function is maximized with respect to by the solution of - L = -1 N - 0 dMx - Mx= 0, (11.21) irrespective of the form of , which provides the maximum likelihood estimator for . See Samuel (1969), and references quoted by her, for more detailed consideration of related maximum likelihood estimation. Estimation of via equation (11.21) requires the use of numerical computation because of the term - Mx- in the denominator. However, there are alternative estimators for that have explicit expressions, are easy to compute, and have high asymptotic efficiency relative to the benchmark provided by (11.21). To facilitate avoidance of the computationally troublesome (-Mx-) in the denominator of (11.21) it is sensible to work with the martingale Ht defined by dHt = Mt- dMt - ( - Mt-) dKt, (11.22) 166 CHAPTER 11. MISCELLANEOUS APPLICATIONS which is suggested by the right-hand alternative representation for the martin- gale N - 0 dMx - Mx= K - 0 Mx- dMx - Mx, (11.23) and to seek effective estimating functions from within the class of martingale estimating functions 0 fx dHx, fx predictable . (11.24) The quasi-score estimating function from this class is 0 (d Hx)(d H x)-1 dHx, (11.25) where d Hx = E d(dHx/d) Fx- = -E dKx Fx- = -Mx- dx and H t is the quadratic characteristic of the martingale Ht, so that d H x = E (dHx)2 Fx- = E M2 x- dMx + ( - Mx-)2 dKx Fx= Mx-( - Mx-) dx, and (11.25) corresponds to (11.23). Thus the quasi-likelihood estimator is the maximum likelihood estimator (MLE), ^ say. Now the martingale information in (11.23) (see Chapter 6), whose reciprocal is essentially an asymptotic variance, is IML = -1 0 Mx dx - Mx , (11.26) and the central limit result Theorem 12.6 or, for example, those of Aalen (1977) and Rebolledo (1980) can be used to show that I 1/2 ML(^ - ) tends in distribution to the standard normal law as . Maximum likelihood provides a benchmark but other estimating functions from the family (11.24) of the type 0 gx dHx, (11.27) where gx is of the form g(Nx-), produce simple and easily computable estima- tors ^g = 0 gx Mx- dNx 0 gx dKx, (11.28) 11.4. ESTIMATING POPULATION SIZE 167 whose performance relative to the benchmark is well worth considering. We shall examine a number of choices for gx and, in particular, determine whether gx 1 is a good choice. The martingale information in the estimating function (11.27) is Ig = 0 gx Mx dx 2 0 g2 x Mx( - Mx) dx (11.29) and, subject to suitable restrictions on g, the above cited central limit results can be used to show that I1/2 g (^g - ) tends in distribution to a standard normal law as . To investigate the relative behavior of the random quantities (11.26) and (11.29) as increases, information on the distributions of Mx and Nx, x > 0, is needed and one can show that the probability generating function of (Mx, Nx) is E(yMx zNx ) = e-x (1 - y + y exx ) . Thus, Mx is distributed as Binomial (, 1 - e-x ) and Nx as Poisson ( x). In particular, -1 Mx a.s. - 1 - e-x and -1 Nx a.s. - x as . Then, if the function g is well behaved, and, in particular, there exists a sequence of constants {c} such that c gx p - gd(x), x [0, ], (11.30) as , where gd(x) is a deterministic function, then it follows that IML a.s. - 0 (ex - 1) dx = e - 1 - and Ig p - 0 gd(x)(1 - e-x ) dx 2 0 g2 d(x) e-x (1 - e-x ) dx. The asymptotic relative efficiency of ^g to the benchmark of the MLE is the ratio ARE(^g) = 0 gd(x)(1 - e-x ) dx 2 (e- - 1 - ) 0 g2 d(x) e-x (1 - e-x ) dx . (11.31) That ARE(^g) is less than or equal to unity follows from the Cauchy-Schwarz inequality and equality holds iff gd(x) = exp x. There is no obvious choice for g which satisfies the requirement (11.30) for gd(x) = exp x and does not involve . Using (11.28), we consider the convenient estimators ^1 = 0 Mx- dNx 0 dKx, ^2 = 0 M2 x- dNx 0 Mx- dKx, 168 CHAPTER 11. MISCELLANEOUS APPLICATIONS and ^3 = 0 Mx- Nx- dNx 0 Nx- dKx for which the form gd(x) in (11.30) is 1, 1 - e-x and x, respectively. By direct integration, ARE(^1) = 2( + e- 1)2 (e - 1 - )(1 - e- )2 , ARE(^2) = (2 + 1 - (2 - e- )2 )2 (e - 1 - )(1 - e- )4 , and ARE(^3) = (2 + 2 e- 2 + 2e- )2 (e -1- ){3 + 2e- ( +1)( e- -2 -2)+ (2-e- )2} . It is convenient to compare these efficiencies against the scale 1 - e- which represents the expected fraction of animals marked by time and therefore has direct experimental relevance. Graphs are given in Becker and Heyde (1990). All three estimators are very efficient over the range 0 < 1 - e< 0.5, which should contain the range of most practical importance. The efficiency of ^1 is highest over this range, which suggests that it is better to have a function g in (11.28) that is not zero initially. The efficiency of ^1 drops significantly once the expected fraction of marked animals exceeds 0.5, as might be expected, since gx = 1 is not an increasing function. These observations encourage us to consider the estimator ^4 = 0 exp Nx- Mx- 1{Mx->0} Mx- dNx 0 exp Nx- Mx- 1{Mx->0} dKx. Corresponding to this estimator we have gd(x) = exp x 1 - e-x , which is an increasing function starting at the value e and is close to the optimal ex for large x. The efficiency of this estimator, ARE(^4) is 0.995 over the range 0 < 1 - e< 0.5 and never falls below 0.9774. It should finally be remarked that the argument leading to estimators given by (11.28) remains valid even when is random. For example, random fluctuations in environmental factors might affect the trapping rate. If a single trap is used that captures just one animal at a time, then the estimators given by (11.28) remain valid even when immediate release of the animals is not possible. Occupancy of the trap simply has the effect of suspending the entire capture process temporarily. 11.5. ROBUST ESTIMATION 169 11.5 Robust Estimation The results in this section are from Kulkarni and Heyde (1987). Let Y1, . . . , Yn be a sample of n observations from a discrete time stochastic process whose distribution depends on a real valued parameter belonging to an open interval of the real line. It is known that when there are outliers in the observations, the standard methods of estimation (maximum likelihood, method of moments etc.) may be seriously affected with adverse results. Robustified versions of some of these procedures are discussed in quite a number of sources, for example, Huber (1981), Gastwirth and Rubin (1975), Denby and Martin (1979), Martin (1980, 1982), Martin and Yohai (1985), K¨unsch (1984), Basawa, Huggins and Staudte (1985), Bustos (1982) and references therein. These authors, however, concentrate on particular problems and we shall here discuss a general procedure of broad applicability based on robustifying the quasi-likelihood framework. As usual this produces an estimating function, which has certain optimality properties, within a specified class of estimating functions. First note, that it is generally possible to specify a set of functions hi = hi(Y1, . . . , Yi, ), i = 1, 2, . . . , n, that are martingale differences, namely, E hi Fi-1 a.s., 1 i n, (11.32) where Fk denotes the past history -field generated by Yj, 1 j k, k 1, and F0 is the trivial -field. A natural choice for hi in the case of integrable Yi is just Yi - E(Yi | Fi-1). Now in this book we have given special consideration to the class M of square integrable martingale estimating functions Gn = n i=1 ai-1 hi, (11.33) where the hi are specified and the coefficients ai-1 are functions of Yi, . . . , Yi-1 and , i.e., are Fi-1-measurable. We have shown in Chapter 2 how to choose from M a quasi-score estimating function G n = n i=1 a i-1 hi with certain optimality properties. However, although the hi can be chosen to be robust functions of the observations, outliers can enter the estimating function through the weights ai-1 and hence adversely affect the properties of the estimates. Thus the class M of estimating functions is not resistant to outliers. To overcome this difficulty we constrain the ai-1's in (11.33) to be of robust form by considering a subset S of M whose elements are of the form Sn = n i=1 hi i j=1 bj,i ki-j, (11.34) where the bj,i, 1 j i, are constants and the ki-j, 1 j i, i 1, are specified functions of Y1, . . . , Yi-j and such that k0 = 1 and E(ki-j Fi-j-1) = 0, i > j. (11.35) 170 CHAPTER 11. MISCELLANEOUS APPLICATIONS That is, the ki-j, i > j, are martingale differences. Note that, unlike the ai-1's in (11.33), the bj,i's in (11.34) are constants and free of the observations. Also, the ki-j's in (11.34) being specified functions, can be chosen to be robust, just as with the hi's. We may take, for example ki-j = hi-j. Thus, the estimating functions in S can be chosen to be robust and we shall show how to select the optimum S within S. In these introductory remarks we have described the case of a scalar parameter for clarity. The results given below, however, deal with a vector parameter and vector valued estimating functions. In Section 11.5.1 we shall develop a general theory of robust quasi-score estimating functions and discuss the loss of efficiency due to robustification. The analysis is carried out formally without specific attention to whether the resultant quasi-score estimating functions are strictly allowable in the sense of involving only the data and the parameter to be estimated. Indeed, they often involve unobservable quantities but are nevertheless amenable to iterative solution. In Section 11.5.2 we shall illustrate by considering the example of a regression model with autoregressive errors. 11.5.1 Optimal Robust Estimating Functions Suppose that Y 1, Y 2, . . . , Y n is a vector process and that = (1, . . . , p) is a parameter taking values in an open subset of p dimensional Euclidean space. As usual we consider the class G of zero mean, square integrable p dimensional vector estimating functions Gn = Gn(Y 1, . . . , Y n, ) that are a.s. differentiable with respect to the components of and such that E ˙Gn = (E Gn,i/j) and EGn Gn are nonsingular. Now let hi and ki-j be specified vectors of dimension q satisfying vector forms of (11.32) and (11.35) for 1 j i - 1, i 1 and k0 = (1, 1, . . . , 1) . Usually q p. We consider the subset S of G having elements Sn = n i=1 i j=1 bj,i ki-j hi (11.36) where the bj,i are constant vectors of dimension p. Then, with the aid of Theorem 2.1 we shall obtain the following theorem. Theorem 11.3 The quasi-score estimating function S n within the class S defined by (11.36) is given by S n = n i=1 i j=1 b j,i ki-j hi, 11.5. ROBUST ESTIMATION 171 where the b j,i, 1 i n, satisfy E(ki-1 hi hi ki-1) E(ki-1 hi hi k0) E(ki-2 hi hi ki-1) E(ki-2 hi hi k0) ... ... E(k0 hi hi ki-1) E(k0 hi hi k0) b 1,i b 2,i ... b i,i = E(ki-1 ˙hi) E(ki-2 ˙hi) ... E(k0 ˙hi) . (11.37) For the special case when E(ki hm hm kj) = 0, i = j and all m (which holds, for example, if E(hm hm | Fm-1) = cm, a constant, for each m), b j,i = (E(ki-j hi hi ki-j))-1 E(ki-j ˙hi), 1 j i, 1 i n. (11.38) Proof We have, using the fact that the hi's are martingale differences, E ˙Sn = E n i=1 i j=1 bj,i ki-j ˙hi, while ESn S n = E n i=1 i j=1 bj,i ki-j hi hi i l=1 ki-l b l,i . Then, the result (11.37) follows from Theorem 2.1 since E ˙Sn = ESn S n for all Sn S when E ki-j hi hi i l=1 ki-l b l,i = E(ki-j ˙hi), 1 j i, and this gives (11.37). The important special case (11.38) follows immediately from (11.37). Now it is important to be able to assess the effect on efficiency of choosing a quasi-score estimating function from S rather than the broader class M of martingales of the form Gn = n i=1 ai-1 hi with the ai-1 being matrices of dimension p × q depending on Y 1, . . . , Y i-1 and , 1 i n. Efficiency may conveniently be assessed by comparing the martingale information in the quasi-score estimating functions within G and S. In particular, it determines the size of the confidence zone in asymptotic confidence statements about the unknown parameter . 172 CHAPTER 11. MISCELLANEOUS APPLICATIONS Let G n be a quasi-score estimating function within G. Then, G n = n i=1 a i-1 hi (11.39) with a i-1 = E ˙hi Fi-1 E hi hi Fi-1 -1 , the inverse being assumed to exist a.s. and the martingale information in G n is IG n = n i=1 E ˙hi Fi-1 E hi hi Fi-1 -1 E ˙hi Fi-1 (11.40) (see Section 6.3). On the other hand, the corresponding result for the quasiscore estimating function from within S is IS n = n i=1 BiE ˙hi Fi-1 n i=1 BiE hi hi Fi-1 Bi -1 n i=1 BiE ˙hi Fi-1 , (11.41) where Bi = i j=1 b j,i ki-j and the b j,i are given by (11.37). It should be noted that if ^G and ^S are the estimators obtained from the estimating equations G n = 0 and S n = 0, respectively, then, under regularity conditions that are usually satisfied in cases which are of practical relevance, (^G - ) IG (^G - ) d - 2 p, (11.42) (^S - ) IS (^S - ) d - 2 p, (11.43) as n (Godambe and Heyde (1987, Section 4)). Here 2 p is the chi-squared distribution with p degrees of freedom. Of course the matrices IG and IS are of dimension p × p and their comparison is not straightforward in general. However, comparison can often be confined to scalar criteria; see Section 2.3. Furthermore, in special cases, such as that treated in the next section, useful results can be obtained from examination of the diagonal elements of interest. It should be remarked that statistics based on optimal robust estimating functions usually involve quantities which cannot be calculated explicitely in terms of the data provided. However, in specific cases iterative computations can be carried out beginning from preliminary estimates of the unknown parameters. For information on similar procedures see Martin and Yohai (1985). 11.5. ROBUST ESTIMATION 173 11.5.2 Example A regression model with autoregressive errors. Suppose that (Y1, . . . , Yn) comes from a model of the form Yi = Ci + Xi, Xi = Xi-1 + i, i = 1, 2, . . . , n, where = (1, . . . , r), Ci = (C1i, . . . , Cri), || < 1, X0 = 0 and the i are independent and identically distributed random variables (i.i.d. r.v.'s) having zero mean and unit variance. Here the Xi are not directly observed, the Ci are fixed regressors and = (, ) are unknown parameters. Let hi = (Xi - Xi) = ( i) and ki-j = hi-j where () is a bounded function chosen so that E( i) = 0. If the i have a symmetric distribution, then a convenient choice for is Huber's function (u) = u if |u| < m, m sgn u if |u| m, m being any specified constant. Here E(h2 i Fi-1) = E2 ( ) is constant and from Theorem 11.3 we readily find (see (11.38)) that b j,i = K1(j-1 , 0 ) , j < i, K2(0, Ci - Ci-1) , j = i, where 0 = (0, . . . , 0) (of dimension r) while K1 and K2 are (scalar) constants and the quasi-score robust estimating function for within S based on {hi} is given by S n = (S n1, S n2) with S n1 = K1 N i=1 ( i) i-1 j=1 j-1 ( i-j) = K1 n i=1 ( i) ~Xi-1, S n2 = K2 n i=1 ( i)(Ci - Ci-1), where ~Xt is given by ~Xt = ~Xt-1 + ( t) = t j=1 j-1 ( t-j+1). This shows that the estimating function approach discussed by Basawa et al. (1985, Example 1), under the assumption of normally distributed i, gives a quasi-score estimating function within S whatever the distribution of the i. Now using the same set of hi, the estimating function which is optimal within G is, from (11.39), G n = (G n1, G n2) 174 CHAPTER 11. MISCELLANEOUS APPLICATIONS with G n1 = K3 n i=1 ( i) Xi-1, G n2 = K3 n i=1 ( i) (Ci - Ci-1), K3 being a scalar constant. Clearly the ( i) will not be directly observable in practice and iterative calculations will be necessary to deal with all these estimating functions. For example, the basic strategy for is as follows. Start from ~X0(^0, 0) where (^0, ^0) are robust preliminary estimates. Then calculate ~0,i = (Yi - ^0 Ci - ^0Yi-1 + ^0Ci-1), i = 1, 2, . . . and obtain { ~X0,i, i = 1, 2, . . .} recursively from ~X0,i = ^0 ~X0,i-1 + ~0,i taking ~X0,0 = 0. Next compute ^1 = n i=1 ~X0,i ~X0,i-1 n i=1 ~X2 0,i-1 and repeat the procedure. Continue the iterations until a stable value for ^ is obtained. We note that S n2 and G n2 are the same, except for the constant multiplier, both being robust, and it is of special interest to compare the performance of the robust estimating function S n1 with that of the nonrobust G n1. For the purpose of this comparison the regression part of the model is irrelevant. Consequently, we now focus attention on the autoregression and suppose that the data is (X1, . . . , Xn). From (11.40) and (11.41) we obtain IG 1 = d n i=1 X2 i-1, (11.44) IS 1 = d n i=1 ~Xi-1 Xi-1 2 n i=1 ~X2 i-1 -1 , (11.45) where d = (E ˙( ))2 /E2 ( ). Now suppose that || < 1, i.e., the autoregression is (asymptotically) stationary. Then, from the martingale strong law of large numbers applied in turn to the martingales n i=1 X2 i - E X2 i Fi-1 , 11.5. ROBUST ESTIMATION 175 n i=1 Xi ~Xi - E Xi ~Xi Fi-1 and n i=1 ~X2 i - E ~X2 i Fi-1 , we obtain after some straightforward analysis that n-1 n i=1 X2 i-1 a.s. - (1 - 2 )-1 , (11.46) n-1 n i=1 ~Xi-1Xi-1 a.s. - E( ( ))(1 - 2 )-1 , (11.47) n-1 n i=1 ~X2 i-1 a.s. - E2 ( )(1 - 2 )-1 (11.48) as n . Consequently, IG 1 nd (1 - 2 )-1 a.s., IS 1 nd (E( ( )))2 (E2 ( ))-1 (1 - 2 )-1 , which gives IS 2 I-1 G 1 a.s. - corr2 ( , ( )) 1 where corr denotes correlation. The quantity corr2 ( , ( )) is the asymptotic relative efficiency of the estimator ^S obtained from the estimating equation S n1 = 0 compared with the estimator ^G obtained from the estimating equation G n1 = 0. This follows along the lines of (11.42), (11.43), which are readily formalized in the present context. We also note that the standard (nonrobust) estimator for is ^QL = n i=1 XiXi-1 n i=1 X2 i-1, which is a quasi-score estimating function within the class n i=1 ai-1(Xi - Xi-1), ai-1 being Fi-1-measurable. It is easily shown that n1/2 (^QL - ) d - N(0, 1 - 2 ), so that the asymptotic efficiency of ^S relative to ^QL is d corr2 ( , ( )). 176 CHAPTER 11. MISCELLANEOUS APPLICATIONS 11.6 Recursive Estimation There are important applications in which data arrive sequentially in time and parameter estimates need to be updated expeditiously, for example, for on-line signal processing. We shall illustrate the general principles in this section. To focus discussion we shall, following Thavaneswaran and Abraham (1988), consider a time series model of the form Xt = ft-1 + t, (11.49) where is to be estimated, ft-1 is measurable w.r.t. Ft-1, the -field generated by Xt-1, . . . , X1 and the { t, Ft} are martingale differences. This framework covers many non-linear time series models such as random coefficient autoregressive and threshold autoregressive models. However, we shall just treat the scalar case here for simplicity. For the family of estimating functions HT = T t=1 at(Xt - ft-1), at Ft-1 measurable we have that the quasi-score estimating function is G T = T t=1 a t (Xt - ft-1), where a t = E ˙ht Ft-1 E h2 t Ft-1 = ft-1 E 2 t Ft-1 . The optimal estimator for based on HT and the first T observations is then given by ^T = T t=1 a t Xt T t=1 a t ft-1. (11.50) When the (T +1)st observation becomes available, we now carry out estimation within HT +1. The estimator for based on the first (T +1) observations, ^T +1, is given by (11.50) with T replaced by (T + 1) and ^T +1 - ^T = KT +1 T +1 t=1 a t Xt - ^T K-1 T +1 , where K-1 T +1 = T +1 t=1 a t ft-1. After a little more algebra we find that KT +1 = KT 1 + fT a T +1 KT (11.51) 11.6. RECURSIVE ESTIMATION 177 and ^T +1 = ^T + KT a T +1 1 + fT a T +1 KT XT +1 - ^T fT . (11.52) The algorithm given by (11.52) provides the new estimate at time (T + 1) in terms of the old estimate at T plus an adjustment. This is based on the prediction error since the forecast of XT +1 at time T is ^T fT . From starting values 0 and K0 the estimator can be calculated recursively from (11.51) and (11.52). In practice 0 and K0 are often chosen on the basis of an additional data run. Various extensions are possible. A similar recursive procedure may be developed for a rather more complicated model that (11.49) proposed by Aase (1983). For details see Thavaneswaran and Abraham (1988). Chapter 12 Consistency and Asymptotic Normality for Estimating Functions 12.1 Introduction Throughout this book it has at least been implicit that consistency and asymptotic normality will ordinarily hold, under appropriate regularity conditions, just as for the case of ordinary likelihood. Sometimes we have been quite explicit about this expectation, such as with the meta theorem enunciated in Chapter 4. We have chosen not to attempt a substantiation of these principles because of the elusiveness of a satisfying general statement and the delicacy of the results as exercises in mathematics as distinct from the reality of practically relevant examples. Also, we have made it clear already that it is generally preferable to check directly, in any particular example, for consistency and asymptotic normality, rather than trying to check the conditions of a special purpose theorem. When the quasi-score (or more generally, the estimating function) is a martingale, consistency and asymptotic nomality are usually straightforward to check directly using the strong law of large numbers (SLLN) and central limit theorem (CLT), respectively, for martingales. In this chapter we shall give some general consistency results and versions of the SLLN and CLT for martingales which enable most typical examples to be treated. There are, of course, applications for which the martingale paradigm is not a satisfactory general tool. The most notable context in which this is the case is that of random fields. Sometimes causal representations of fields are possible for which martingale asymptotics works well but other limit theorems are necessary in general. Consistency results are usually straightforward to obtain, especially in the context of stationary fields where ergodic theorems can be used, but asymptotic normality poses more problems. A standard approach to the central limit theorem here has been to develop results based on the use of mixing conditions to provide the necessary asymptotic independence (e.g., Rosenblatt (1985), Doukhan (1994), Guyon (1995, Chapter 3)). However, mixing conditions are difficult, if not impossible, to check, so this approach is not entirely satisfactory. Fortunately, recent central limit results for random fields based on conditional centering (e.g., Comets and Janˇzura (1996)) seem very promising. Non-random field examples of estimating functions that are not martingales can often be replaced, for asymptotic purposes, by martingales. Thus, for 179 180 CONSISTENCY AND ASYMPTOTIC NORMALITY example, if GT () = T t=1 (Xt - ) where Xt is the moving average Xt = + t - r t-1, |r| < 1, the 's being i.i.d., then GT () = (1 - r) T t=1 t + r( T - 0), which is well approximated by the martingale {(1-r) T t=1 t} for the purposes of studying the asymptotics of ^T = T-1 T t=1 Xt. 12.2 Consistency A range of consistency results for quasi-likelihood estimators, and more generally for the roots of estimating functions, can be developed to parallel corresponding ones for the maximum likelihood estimator (MLE). We therefore begin with a brief discussion of methods that have been established for this context. There are two classical approaches for the MLE. These are the one due to Wald (1949), which yields consistency of global maximizers of the likelihood, and the one due to Cram´er (1946), which yields a consistent sequence of local maximizers. The Wald approach is not available in general in a quasi-likelihood setting. Certainly its adoption requires the quasi-score to be the derivative with respect to the parameter of a scalar objective function. However, there are certainly cases in which the principles can be used. For a discussion of the approach see, for example, Cox and Hinkley (1974, pp. 288­289). Similar comments on applicability also apply to some recent necessary and sufficient results of Vajda (1995). There is, however, a newly developed minimax approach to consistency of quasi-likelihood estimators (Li (1996a)) that has some of the advantages of the Wald approach. The Cram´er approach is based on local Taylor series expansion of the likelihood function and an analogous route can easily be followed for estimating functions. It may be remarked that the law of large numbers plays a central role in both approaches. We shall use a stochastic process setting to describe the ideas behind the Cram´er approach. Let LT ({Xt}; ) denote the likelihood that we suppose is differentiable with respect to , here assumed to be scalar for clarity, and for which the derivative of the log likelihood is square integrable. Then, under modest regularity conditions which involve differentiability under an integral sign, {d log LT ({Xt}; )/d, FT , T 1} is a martingale, Ft denoting the field generated by X1, . . . , Xt, t 1, e.g., Hall and Heyde (1980, Chapter 6, p. 157). 12.2. CONSISTENCY 181 We write d d log LT ({Xt}; ) = T t=1 ut(), where the ut() are martingale differences, and take the local Taylor series expansion for . This gives d d log LT ({Xt}; ) = T t=1 ut( ) = T t=1 ut() - ( - ) IT () +( - )(JT ( T ) - IT ()), (12.1) where IT () = T t=1 E u2 t () Ft-1 , JT () = T t=1 d d ui(), and T = + ( - ) with = ({Xt}; ) satisfying || < 1. Now the martingale strong law of large numbers (e.g., using Theorem 2.18 of Hall and Heyde (1980); see pp. 158, 159, or Theorem 12.4 herein) gives (IT ())-1 T i=1 ui() a.s. - 0 (12.2) provided that IT () a.s. - as T . Thus, using (12.1) and (12.2), the score function has a root ^T , which is strongly consistent for the true parameter if IT () a.s. - and lim sup T |IT () + JT ( T ) | IT () < 1 a.s. (12.3) Various sufficient conditions can be imposed to ensure that (12.3) holds but these are not illuminating. The approach discussed above can be paralleled in many quasi-likelihood contexts. If convenient families of martingale estimating functions are available for a particular context, they would ordinarily be used. Then the quasi-score estimating function will be a martingale, local Taylor series expansion can be used just as above, and the martingale strong law of large numbers retains its role. The case of vector valued can be dealt with in the same way. The different approaches to establishing consistency results depend fundamentally on whether the estimating function is the derivative with respect to of a scalar objective function. If such a representation is possible, as it is in the case of the score function (which is the derivative of the log likelihood), then the consistency question can be translated into the problem of checking 182 CONSISTENCY AND ASYMPTOTIC NORMALITY for local maxima. However, the representation is not possible in general for quasi-score estimating functions; a scalar objective function (quasi-likelihood) may not exist. If there is a scalar objective function and the parameter space is a compact subset of p-dimensional Euclidean space, then an estimator obtained by maximization always exists. The situation, however, is more complicated if a scalar objective function does not exist or is not specified. Example (Chen (1991)). Consider the quadratic estimating function g = T i=1 a(Xi - ) + b (Xi - )2 - 2 , where the Xi's are i.i.d. with EXi = , var Xi = 2 , a, b and 2 being known and is the parameter to be estimated. This is a quasi-score estimating function if 3 a + (4 - 4 ) b = 0, where k = E(X1 - )k . It is easily seen that there are two roots for g() = 0, namely, ^1,T = a 2b + 1 T T i=1 Xi - 1/2 T , ^2,T = a 2b + 1 T T i=1 Xi + 1/2 T , where T = a 2b + 1 T T i=1 Xi 2 - a Tb T i=1 Xi + 1 T T i=1 X2 i - 2 . The roots only exist on the sequence of events {T 0}, but it follows from the strong law of large numbers that T a.s. - (a/2b)2 as T , so that they exist almost surely. Note that ^1,T a.s. - , but ^2,T a.s. - + (a/b). The point here is that both the existence and the consistency of estimators can at most be expected in the almost sure sense. For various problems we have to deal with situations in which the asymptotic results on which we rely hold only on a set E, say, with P(E) < 1. This is the case, for example, for estimating the mean of the offspring distribution for a supercritical branching process; the asymptotic results hold on the nonextinction set whose probability is less than one in general. Thus, in view of the above considerations, and given a sequence of estimating functions {GT ()}, we shall say that existence and consistency of the estimator ^ that solves GT () = 0 holds if: for all sufficiently small > 0, a.s. on E, and for T sufficiently large, GT () = 0 has a root in the sphere { : - 0 }, 0 denoting the "true value". We shall now give a consistency criterion that does not require the estimating function to be the derivative with respect to of a scalar function. This 12.2. CONSISTENCY 183 is an adaption of a result originally due to Aitchison and Silvey (1958), which uses the Brouwer fixed-point theorem. Theorem 12.1 (Consistency Criterion) Let {GT ()} be a sequence of estimating functions that are continuous in a.e. on E for T 1, and for all small > 0 a.e. on E, there an > 0 so that lim sup T sup -0 = ( - 0) GT () < - , (12.4) then there exists a sequence of estimators ^T such that for any E, ^T () 0 and GT (^T ()) = 0 when T > T, a constant depending on . On the other hand, if instead we have an in probability version of (12.4), namely, lim T P sup -0 = ( - 0) GT () < - E = 1 (12.5) for any > 0, then there is a sequence of estimators ^T such that ^T converges to 0 in probability on E and P GT (^T ) = 0 E - 1 as T . Proof For any suitable small positive and let ET = : sup -0 = ( - 0) GT () - . For ET , define ^() to be a root of GT () = 0 with ^ - 0 . The existence of such a root is ensured by Brouwer's fixed point theorem. Next, define ^T () for Ec T , the complementary event, to be an arbitrary point in the parameter space. Then, from the definition of ^T , we have { ^T - 0 < } ET so that { ^T - 0 } Ec T and hence P( ^T - 0 i.o. | E) P(Ec T i.o. | E) = 0, i.o. denoting infinitely often and the right-hand equality following from condition (12.4). This gives a.s. convergence of ^ to 0 on E. Furthermore, since P(Ec T i.o. | E) = 0, we have GT (^T ()) = 0 on E when T > T. 184 CONSISTENCY AND ASYMPTOTIC NORMALITY In the case where (12.5) holds, the results follow from P( ^T - 0 < | E) P(ET | E) 1 as T . Theorem 12.1 is not usually easy to use in practice. The principal difficulty is in showing that (12.4) holds uniformly in . Some sufficient conditions for ensuring applicability are given in Theorem 12.1 of Hutton, Ogunyemi and Nelson (1991). However, direct checking is usually preferable, albeit tedious. See Exercises 12.1, 12.2 for example. Next we shall consider the case where GT () is the gradient of a scalar potential function q(), GT () = q()/. For this to be the case the matrix GT ()/ must be symmetrical (e.g., McCullagh and Nelder (1989)). The results that can then be obtained parallel these which have been developed for maximum likelihood estimators. The scalar potential function takes the place of the likelihood. In the following we will use x for vector x to denote the Euclidean norm (x x)1/2 and A for matrix A to denote the corresponding matrix norm (maxA A)1/2 , max(resp. min) being the maximum (resp. minimum) eigen- value. For a sequence {AT } of p × p matrices, possibly random or depending on the true parameter value 0, we shall define the sets MT () = p : AT ( - 0) (AT )GT (0) , T = 1, 2, . . ., 0 < < , the minus denoting generalized inverse. Then, we shall make use of the following consistency conditions, W (for weak) and S (for strong). These conditions refer to the true probability measure and hence they must usually be checked for all possible 0, with c perhaps depending on 0. Condition W (i) For all 0 < < , sup MT () - 0 p - 0. (ii) For some random variable c, 0 < c < , and = 2/c, P(- ˙GT () - cAT AT 0 for all MT () ) 1, being the parameter space, an open subset of p . Condition S (i) For all 0 < < , sup MT () - 0 a.s. - 0. (ii) For some random variables T1, c, 0 < c < , and = 2/c, - ˙GT () cAT AT 0 for all MT () and all T T0, AT is nonsingular, T T0, for some (random) T0 a.s., and (AT )GT (0) 2 = o (min(AT AT )) a.s. 12.2. CONSISTENCY 185 Some explanations are in order. The Condition W(i) can be thought of as MT () shrinking to 0 in probability and it is equivalent to P(AT nonsingular) 1, (AT )GT (0) 2 = op (min(AT AT )), op meaning smaller order in probability, from which the parallel with Condition S is evident. Theorem 12.2 Under Condition W, there exists a weakly consistent sequence {^T } of estimators. Actually, with from W(ii), we have P(^T MT ()) 1 and ^T p - 0. To facilitate comparison with published results, assume that (AT )GT () = Op(1), order one in probability having the obvious meaning. Then, it is natural to consider the simpler sets NT () = { p : AT ( - 0) }, T = 1, 2, . . . 0 < < and it is not difficult to see that the Condition W(i) is implied by the analogous assumption with NT () replacing MT (). However, no such simple substitution is possible for W(ii). The usual Cram´er type results on weak consistency in the literature (e.g., Basawa and Scott (1983)) are essentially covered by this formulation. A plausible choice for the norming matrix AT is usually (-EG ˙GT ())1/2 . For the case of strong consistency we have the following result. Theorem 12.3 Under Condition S, there exists a strongly consistent sequence of estimators {T }. We have ^T MT () a.s. for all T T2 with some random T2 and ^T a.s. - 0. Proof of Theorem 12.2 If Condition W(i) holds for all constants , 0 < < , it also holds if is a random variable with 0 < < . Let c, = 2/c be the random variables of W(ii) and denote by ET the combined event: AT is nonsingular, MT () , - ˙GT () - c AT AT 0 for all MT () . 186 CONSISTENCY AND ASYMPTOTIC NORMALITY Conditions W(i),(ii) and open ensure P(ET ) 1. On the subset of ET where GT (0) = 0, MT () is degenerate to {0}. Taking ^T = 0, we have GT (^T ) = 0, - ˙GT (^T ) > 0 and ^T is a (locally unique) maximizer of q with ^T MT (). On the subset of ET where GT (0) = 0, MT () is a compact ellipsoidal neighborhood of 0. For on the boundary of MT (), set v = AT ( - 0). Then, taking the Taylor expansion qT () - qT (0) = v (AT )-1 GT () - 1 2 v (AT )-1 ˙GT (~) A-1 T v with ~ MT (), we obtain qT () - qT (0) v (AT )-1 GT () - 1 2 v 2 minA-1 T ˙GT (~)(AT )-1 v (AT )-1 GT () - 1 2 (AT )-1 GT () c = 0. Let M denote the subset of MT () where supMT () T () is attained. Since MT () is compact, M is nonempty. Also, since qT () qT (0) for all on the boundary of MT (), if M were to contain the whole line segment between and 0, it would also contain the whole line segment between and 0. But - ˙GT (~) > 0, ~ MT () makes this impossible and we conclude that M consists of a unique point in the interior of MT (). This point defines a (measurable) local maximizer of q(), on the set ET {GT (0) = 0}. Proof of Theorem 12.3 Let {ET } be defined as in the proof of Theorem 12.2. Using the fact that lim infT ET requires ET to occur for all T T2, say, it is seen that Condition S implies lim infT ET = 1 with probability one. Arguing as above, ^T MT (), T T2, and ^T a.s. - 0 follow. The general results given above obscure the simplicity of many applications. Very often the estimating functions {GT ()} of interest form a square integrable martingale. Then, direct application of an appropriate strong law of large numbers (SLLN) for martingales will usually give strong consistency of the corresponding estimator without requiring the checking of uniwieldy conditions. Useful general purpose SLLN's for martingales are given in Section 12.3 and a corresponding central limit result is given in Section 12.4. 12.3 The SLLN for Martingales The first result given here is a multivariate version of the standard SLLN for martingales. It is from Lin (1994b, Theorems 2 and 3) and it extends work of Kaufmann (1987) to allow for random norming. The martingales here will be assumed to be vectors of dimension p 1. As usual we shall use the norm x = (x x)1/2 for vector x and, for matrix A we 12.3. THE SLLN FOR MARTINGALES 187 shall write max(A) and min(A) for the maximum and minimum eigenvalues, tr A for the trace and A = (max(A A))1/2 for the norm. A sequence of matrices {An} will be called monotonically increasing if An+1 An+1 - An An is nonnegative definite for all n. Theorem 12.4 Let {Sn = n i=1 Xi, Fn, n 1} be a p-dimensional square integrable martingale and {An} a monotone increasing sequence of nonsingular symmetric matrices such that Ai is Fi-1-measurable for each i, and min(An) a.s. as n . If n=1 E A-1 n Xn 2 < (12.6) or n=1 E A-1 n Xn 2 Fn-1 < a.s., (12.7) then A-1 n Sn a.s. - 0 as n . Remarks (1) Since A-1 n Xn 2 = tr A-1 n Xn Xn A-1 n , we have E A-1 n Xn 2 = tr A-1 n E(Xn Xn) A-1 n , E A-1 n Xn 2 Fn-1 = tr A-1 n E Xn Xn Fn-1 A-1 n . (2) Theorem 12.4 is for the case of discrete time but corresponding results are available for continuous time with the conditions (12.6) and (12.7) replaced by tr 0 A-1 t d (E S t) A-1 t < (12.8) and tr 0 A-1 t d S t A-1 t < a.s., (12.9) respectively, { S t} being the quadratic characteristic process for {St}. In some applications it may be convenient to use one dimensional martingale results on each component in the vector rather than the multivariate results quoted above. A comprehensive collection of one dimensional stong law results for martingales in discrete time are given in Chapter 2 of Hall and Heyde (1980). Again these results have natural analogues in continuous time. However, an additional one dimensional result for continuous time which is particularly useful is given in the following theorem due to L´epingle (1977). 188 CONSISTENCY AND ASYMPTOTIC NORMALITY Theorem 12.5 Let {St, Ft, t 0} be a locally square integrable martingale that is right continuous with limits from the left (cadlag) and h(x) be an increasing nonnegative function such that 0 (h(x))-2 dx < . Then, St/h( S t) 0 a.s. on the set { S t }. Proof of Theorem 12.4 Suppose supn E( n k=1 A-1 k Xk 2 ) < . Then, since x + y 2 - y 2 = x 2 + 2x y, we have, upon setting x = A-1 k Xk and y = A-1 k Sk-1 and using the martingale property, E A-1 k Sk 2 - E A-1 k Sk-1 2 = E A-1 k Xk 2 . Write zn = n k=1 A-1 k Sk 2 - A-1 k Sk-1 2 . Now supn Ezn < and since {zn, Fn, n 1} is a nonnegative submartingale we must have the a.s. convergence of zn = A-1 n Sn 2 + n-1 k=1 A-1 k Sk 2 - A-1 k+1 Sk 2 . But, the monotonicity condition for the A's ensures that n-1 k=1 A-1 k Sk 2 - A-1 k+1 Sk 2 = zn - A-1 n Sn 2 is nondecreasing and it is a.s. bounded by supn zn, so that it must converge a.s. This forces A-1 n Sn 2 to converge a.s. Hence, it suffices to check that A-1 n Sn 2 p - 0 to ensure that A-1 n Sn 2 a.s. - 0 (and, consequently, A-1 n Sn a.s. - 0). Now, for any > 0 and N < n, P A-1 n Sn 2 > P A-1 n SN 2 > 1 2 + P An (Sn - SN ) 2 > 1 2 = I1 + I2, (12.10) 12.3. THE SLLN FOR MARTINGALES 189 say. But, using Chebyshev's inequality and the monotonicity of the A's 1 2 I2 E A-1 n n k=N+1 Xk 2 E n k=N+1 Xk A-2 k Xk + n k=N+2 Xk A-2 k k-1 j=N+1 Xj + n k=N+2 Xk A-2 k n j=k+1 Xj = E n k=N+1 Xk A-2 k Xk = E n k=N+1 A-1 k Xk 2 , since Ak is Fk-1-measurable. Also, since E k=1 A-1 k Xk 2 < , for given > 0 there is an N0 > 0 such that I2 E n k=N0+1 A-1 k Xk 2 ( /2) < ( /2) (12.11) for all n > N0. Now fix N0. In view of the condition that min(An) a.s. as n there is an N > N0 such that if n > N, I1 P A-1 n 2 SN0 2 > /2 < /2 (12.12) and hence, from (12.10), (12.11) and (12.12), P A-1 n Sn 2 > < for n > N. This gives A-1 n Sn 2 p - 0 and hence A-1 n Sn a.s. - 0 and completes the proof of the first part of the theorem. Now suppose that (12.7) holds. Given a real C > 0, let TC = max n : n k=1 E A-1 k Xk 2 Fk-1 C . Note that TC is a stopping time since, for any n0, {TC < n0} = n0 k=1 E A-1 k Xk 2 Fk-1 > C Fn0-1. Therefore, for each C we can define a new random norming, such as I(n TC) A-1 n , I denoting the indicator function, which is still Fn-1-measurable. Clearly n k=1 E I(k TC) A-1 k Xk 2 Fk-1 C 190 CONSISTENCY AND ASYMPTOTIC NORMALITY for all n and hence n k=1 E I(k TC) A-1 k Xk 2 C. Also, {I(n TC) A-1 n } is monotonically decreasing, so applying the result of the first part of the theorem we obtain I(n TC) A-1 n Sn a.s. - 0 and hence A-1 n Sn a.s. - 0 and {TC = }. The required result follows since C can be any positive number. Proof of Theorem 12.5 The result basically follows from the martingale convergence theorem and a Kronecker lemma argument. Put At = S t, cu = inf{t : At > u} and define the martingale Zt = t 0 (h(Au))-1 dSu. As Acu u for all u < A, Z = 0 dAu (h(Au))2 = A 0 du (h(Acu ))2 0 dt (h(t))2 < . It follows that Z is square integrable and converges a.s. Now, using integration by parts, St = t 0 h(Au) dZu = Zt h(At) - t 0 Zu- dh(Au) = t 0 (Zt - Zu-) dh(Au). Furthermore, if 0 < v < t < , St h(At) 1 h(At) v 0 (Zt - Zu-) dh(Au) + 1 h(At) (h(At) - h(Av)) sup v 0, then H-1 T (0)(- ˙GT (0))(^T - 0) d - Z. This last result is the basis for confidence zones for 0 as discussed in Chapter 4 and, as indicated in Section 4.2, - ˙GT (0) may often be replaced asymptotically by equivalent normalizers [G(0)]T or G(0) T . The usual practice is for a martingale CLT to be used to provide (12.13), while (12.14) is obtained from ad hoc use of law of large number results and inequalities. The multivariate central limit result given here is an adaption of a result of Hutton and Nelson (1984) to deal with more general norming sequences. A similar, but slightly weaker, result appears in Srensen (1991). Our proof follows that of Hutton and Nelson (1984). The result is phrased to deal with the continuous time case but it also covers the discrete time one. This is dealt with by replacing a discrete time process {Xn, n 0} by a continuous version {Xc t , t 0} defined through Xc t = Xn, n t n + 1. Here all the processes that are considered are assumed to be right continuous with limits from the left (cadlag) and defined on a complete filtered probability space (, F, {Ft}, P) satisfying the standard conditions. As usual we denote by Ip the p × p identity matrix. We shall make use of the concepts of stable and mixing convergence in distribution and these are defined below. Let Y n = (Y1n, . . . , Ypn) d - Y = (Y1, . . . , Yp) . Suppose that y = (y1, . . . , yp) is a continuity point for the distribution function of Y and E F. Then, we say that the convergence holds stably if lim n P(Y1n y1, . . . , Ypn yp, E) = Qy(E) 192 CONSISTENCY AND ASYMPTOTIC NORMALITY exists and Qy(E) P(E) as yi , i = 1, 2, . . . , p. On the other hand, we say that the convergence holds in the mixing sense if lim n P(Y1n y1, . . . , Ypn yp, E) = P(Y1 y1, . . . , Yp yp) P(E). For details see Hall and Heyde (1980, pp. 56­57) and references therein. Theorem 12.6 Let {St = (S1t, . . . , Spt) , Ft, t 0} be a p-dimensional martingale with quadratic variation matrix [S]t. Suppose that there exists a non-random vector function kt = (k1t, . . . , kpt) with kit > 0 increasing to infinity as t for i = 1, 2, . . . , p such that as t : (i) k-1 it supst | Sis| p - 0, i = 1, 2, . . . , p, where Sis = Sis - Sis-; (ii) K-1 t [S]t K-1 t p - 2 where Kt = diag (k1t, . . . , kpt) and 2 is a random nonnegative definite matrix; (iii) K-1 t E(St St) K-1 t where is a positive definite matrix. Then, K-1 t St d - Z (stably) (12.15) where the distribution of Z is the normal variance mixture with characteristic function E(exp(-1 2 u 2 u)), u = (u1, . . . , up) , Kt [S]-1 t Kt 1/2 K-1 t (St det2 > 0) d - N(0, Ip) (mixing) (12.16) det denoting determinant, and St [S]-1 t St det2 > 0 d - 2 p (mixing) (12.17) as t . Remark It should be noted that {[S]t - S t, Ft, t 0} is a martingale, S t being the quadratic characteristic matrix. Each of the quantities [S]t, S t goes a.s. to infinity as t and ordinarily [S]t, S t are asymptotically equivalent. Proof of Theorem 12.6 First we deal with the case p = 1 and for this purpose we shall use an adaption of Theorem 3.2 of Hall and Heyde (1980). As it stands, this theorem deals with finite martingale arrays (Sni, Fni, 1 i kn, n 1) with differences Xni and it gives sufficient conditions under which Yn kn i=1 Xni d - Z (stably) 12.4. THE CLT FOR MARTINGALES 193 where Z is as defined in (12.15). But, if {kn} increases monotonically to infinity in such a way that E i>kn X2 ni - 0 as n , the theorem extends to infinite sums. That i=1 Xni = kn i=1 Xni + i>kn Xni d - Z (stably) under the condition i>kn Xni L1 - 0 follows from the definition of stable convergence. To prove (12.15) in the p = 1 case it suffixes to show that for any sequence {Tn} diverging monotonically to infinity sufficiently fast that KTn+1 /KTn 4, we have K-1 Tn STn d - Z (stably). (12.18) Indeed, if K-1 t St does not converge stably we can choose a subsequence along which the convergence is not stable and then a sufficiently sparse subsequence {Tn} for which KTn+1 /KTn 4 and K-1 Tn STn does not converge stably thus giving a contradiction. To show (12.18), define for each n the stopping times: T0 n = 0, Tk+1 n = inf t > Tk n : K-1 Tn |St - ST k n | 2-n = Tn if no such t Tn exists. We shall consider the martingale differences Xik = K-1 Tn ST k n - ST k-1 n , Fnk = FT k n . Note that Tk n = Tn ultimately in k for n = 1, 2, . . . so that k Xnk = K-1 Tn STn . Now, to establish (12.15) we need to check (a) supn E ( k Xnk) 2 < , (b) supk |Xnk| p - 0, (c) k X2 nk p - 2 , (d) E supk X2 nk is bounded in n, (e) for all n and k, Fnk = (Xn1, . . . , Xnk) Fn+1,k = (Xn1, . . . , X(n+1)k). 194 CONSISTENCY AND ASYMPTOTIC NORMALITY Now (a) is satisfied because K-2 Tn E S2 Tn is bounded, (b) is satisfied since ST k n - ST k-1 n 2-n + sup tTn St, (d) is satisfied since K-2 Tn E sup k ST k n - ST k-1 n 2 K-2 Tn E sup k 2S2 T k n + 2S2 T k-1 n 4K-2 Tn E S2 Tn and (c) holds, while (e) is satisfied since by induction on k, we can show that Tk n Tk n+1 a.s. To check (c), it suffices to show that k X2 nk - K-2 Tn [S]Tn p - 0. Now, by Ito's formula, [S]Tn = S2 Tn - S2 0 - 2 Tn 0 Su- dSu, and similarly, k X2 nk = K-2 Tn S2 Tn - S2 0 - 2 k=0 ST k n (ST k+1 n - ST k n ) . Then, setting n(u) = k=0 ST k n I(Tk n < u Tk+1 n ), we have k X2 nk - K-2 Tn S2 Tn = 2 K-2 Tn Tn 0 (Su- - n(u)) dSu = In, say. Now, from the definitions of Tk n and n, we have K-1 Tn (Su - u(u)) 2-n a.s., so that for > 0, P(|In| > ) -2 EI2 n = 4 -2 K-4 Tn E Tn 0 (Su- - n(u))2 d[S]u 4 -2 K-2 Tn 2-2n E[S]Tn - 0 12.5. EXERCISES 195 as n . The completes the proof in the case p = 1. To deal with the general case we use the so-called Cram´er-Wold device. Let c = (c1, . . . , cp) be a nonzero vector. By applying the theorem to the martingale {Rt = c St}, we have (c E(St St) c)-1/2 (c St) d - Zc (stably) where E(exp(i u Zc)) = E exp - 1 2 u2 2 c with c given by (c E(St St) c)-1 [c S]t p - 2 c . But, using Conditions (ii) and (iii) of the theorem, 2 c = (c 2 c)/(c c). It then follows that c K-1 t St d - (c c)1/2 Zc (stably) and since the characteristic function of (c c)1/2 Zc is E[exp(-1 2 c 2 c u2 )], the result (12.15) follows. The remaining results (12.16) and (12.17) are easy consequences of the definition of stable convergence together with (12.15). This completes the proof. A broad range of one-dimensional CLT results may be found in Chapter 3 of Hall and Heyde (1980). These may be turned into multivariate results using the Cram´er-Wold device as in the proof of Theorem 12.6 above. 12.5 Exercises 1. Suppose that {Zi} is a supercritical, finite variance Galton-Watson branching process with = E(Z1 | Z0 = 1), the mean of the offspring distribution to be estimated from a sample {Z1, . . . , ZT }. Show that QT () = T i=1 (Zi - Zi-1) may be obtained as a quasi-score estimating function for from an appropriate family of estimating functions. Establish the strong consistency of the corresponding quasi-likelihood estimator on the non-extinction set (i) using Theorem 12.1, and (ii) using the martingale SLLN, Theorem 12.4. (Hint. Note that -n Zn a.s. - W as n where W > 0 a.s. on the nonextinction set.) 196 CONSISTENCY AND ASYMPTOTIC NORMALITY 2. Let {Mt, Ft} be a square integrable martingale with quadratic characteristic M T = T a.s. Thus, {Mt} could be a process with stationary independent increments and need not be Brownian motion. Consider the model dXt = 1 2 + 1(c tc-1 )1/2 d + dMt, 1 > 0, 2 > 0, 1 2 < c < 1, and obtain the Hutton-Nelson solution (Section 2.3) for the quasi-score estimating function for = (1, 2) . Show that the quasi-likelihood estimator is strongly consistent for using Theorem 12.1. (Hint. Use Taylor expansion and L´epingle's version of the SLLN for martingales (Theorem 12.5).) (Hutton and Nelson (1986).) 3. Suppose that {Xk} is a process and {Fk} an increasing sequence of -fields such that E(Xk | Fk-1) = , E((Xk - )2 | Fk-1) = 1 + uk-1, k 1, with {uk} a family of non-negative random variables adapted to {Fk}. For a sample {X1, . . . , XT }, obtain the estimating function QT () = T k=1 Xk - 1 + uk-1 (12.19) as a quasi-score. Noting the difficulty of working with this quasi-score, it is interesting to make a comparison with the simple, non-optimal, estimating function HT () = T k=1 (Xk - ) u-1 k-1 I (uk-1 > 0). (12.20) In the particular case where (u2 k, Fk) is a Poisson process with parameter 1 sampled at integer times, and Xk = + wk (1 + uk-1)1/2 , k 1, the {wk} being i.i.d. with finite fourth moment and E(wk | Fk-1) = 0, E(w2 k | Fk-1) = 1, k 1, investigate the consistency and asymptotic normality of the estimators obtained from (12.19) and (12.20) (Hutton, Ogunyemi and Nelson (1991)). 4. Consider the stochastic differential equation dXt = (1 + 2 Xt) dt + dWt + dZt where Wt denotes standard Brownian motion and Zt is a compound Poisson process Zt = Nt i=1 i, 12.5. EXERCISES 197 Nt being a Poisson process with parameter 3 and the i's being i.i.d random variables with mean 4, which are independent of Nt. This s.d.e has been used to model the dynamics of soil moisture; see Mtundu and Koch (1987). For a sample {Xt, 0 t T} investigate the consistency and asymptotic normality, of quasi-likelihood estimators for = (1, 2, 3, 4). Take as known (Srensen (1991)). Chapter 13 Complements and Strategies for Application 13.1 Some Useful Families of Estimating Functions 13.1.1 Introduction A broad array of families of estimating functions has been presented in this monograph but many others are available and may be of particular use in special circumstances. The focus of the theory has usually been on polynomial functions of the data and these will not always provide ideal families of estimating functions. It must be remembered that the general theory requires estimating functions which are at least square integrable and if the underlying variables of interest do not have finite second moments then transformations are required. A number of useful approaches are indicated below. 13.1.2 Transform Martingale Families Suppose we have real valued observations X1, . . . , Xn whose distribution involves a parameter . Let Fj be the -field generated by X1, . . . , Xj, j 1. Then, following Merkouris (1992), write Fj(x | Fj-1) = P(Xj x | Fj-1), ^Fj(x) = I(Xj x), I being the indicator function. Now introduce a suitably chosen index set of {gt(x), t T }, which is real or complex valued and such that the integral transform j(t) = E(gt(Xj) | Fj-1) = gt(x) dFj(x | Fj-1) exists and is finite for all and all t T. Transforms such as the characteristic function and moment generating function, for which gt(x) is eitx and etx , respectively, are clearly included. Now let ^j(t) = gt(Xj) = gt(x) d ^Fj(x) and write hj(t) = ^j(t) - j(t) = gt(Xj) - E(gt(Xj) | Fj-1). 199 200 COMPLEMENTS AND STRATEGIES For fixed t T the {hj(t), Fj} are martingale differences and the family of martingale estimating functions H = n j=1 aj hj(t), aj's Fj-1-measurable will be useful in various settings. For an application to estimation of the scale parameter in a Cauchy distribution using gt(x) = cos tx, see Section 2.8, Exercise 5. Estimation of parameters of progression time distributions in multi-stage models is treated by this methodology, modulo minor adaption to the context, in Schuh and Tweedie (1970) and Feigin, Tweedie and Belyea (1983). It is shown that Laplace transform based estimators may be very efficient when maximum likelihood estimators are available. 13.1.3 Use of the Infinitesimal Generator of a Markov Process A number of useful approaches to estimation are based on the use of the infinitesimal generator of a continuous time Markov process {Xt, t > 0} with states in a sample space , say a locally compact metric space. Suppose that X has distribution P, . Let L = L be the infinitesimal generator of the Markov process X. That is, for suitable functions f, Lf is given by lim 0 -1 [E(f(X) | X0 = x) - f(x)] provided that the limit exists (see, e.g., Karlin and Taylor (1981, p. 294)). Often a bounded limit is prescribed. Now let D be the domain of L. Then, for an eigenfunction D with eigenvalue , L = and it is not difficult to prove, using the Markov property, that L E((X) | X0 = x) = - E((X) | X0 = x) and hence that E((X) | X0 = x) = e- (x). This last result is a basis for showing that {Yt = e-t (Xt), t > 0} is a martingale and hence for the defining of families of estimating functions. Kessler and Srensen (1995) have used these ideas for the estimation of parameters in a discretely observed diffusion process with observations (Xt0 , . . . , Xtn ) and ti - ti-1 = for each i. In this context the family of martingale estimating functions n i=1 (Xti-1 ; ) (Xti ; ) - e-() (Xti-1 ; ) 13.2. SOLUTION OF ESTIMATING EQUATIONS 201 with 's arbitrary functions, is considered. Also, more generally, when k eigenfunctions 1(x; ), . . . , k(x; ) with distinct eigenvalues 1(), . . . , k() are available, the family of estimating functions n i=1 k j=1 j(Xti-1 ; ) j(Xti ; ) - e-j () j(Xti-1 ; ) , with j's arbitrary functions, can be used. Kessler and Srensen (1995) have discussed quasi-likelihood estimators from these families and their consistency and asymptotic normality. Another approach to estimation which makes use of the infinitesimal generator has been suggested by Baddeley (1995). Here a parametric model given by a family of distributions {P, } is assumed and the aim is to estimate on the basis of x drawn from P. The idea is to find a continuous time Markov process {Xt, t > 0} for which P is an equilibrium distribution for each . Now let L be the infinitesimal generator of {Yt} and choose a family of statistics {S(x)} belonging to the domain D of L. Each of these produces an estimating function (LS)(x) and it is easily checked that these have zero mean. Baddeley calls the corresponding estimating equations time invariant. As an example we consider the discrete random fields X = (Xi, i G) in which the set of "sites" G is an arbitrary finite set and the site "labels" Xi take values in an arbitrary finite set. Let P(X = x) = (x), x and assume that (x) > 0 for all x , . If {Yt, t > 0} is the Gibbs sampler for (e.g., Guyon (1995, p. 211)), then the infinitesimal generator of {Yt} operating on the statistic S turns out to be (L S)(x) = iG {E [S(X) | XG\i = xG\i] - S(x)}, where XB = (Xi, i B) denotes the restriction of X to B G. The corresponding time invariant estimator is thus the solution of 1 |G| iG E [S(X) | XG\i = xG\i] = S(x) where |G| is the number of "sites" in G. For more details see Baddeley (1995). Questions of choice of family and of optimality are largely open. 13.2 Solution of Estimating Equations Iterative numerical methods are very often necessary for obtaining estimators from estimating equations. Detailed discussion of computational procedures is outside the scope of this monograph. However, there are various points which can usefully be noted. 202 COMPLEMENTS AND STRATEGIES Preliminary transformation of the parameter so that the covariance matrix is approximately diagonal can be very helpful. Also, if the dimension of the problem can be reduced via expressing some of the components explicitly in terms of others, then this is usually advantageous. For details on computational methods the reader is referred to Thisted (1988). In particular, Chapter 3, Section 10 of this reference deals with GLIM and generalized least squares methods and Chapter 4 with the Newton-Raphson method and Fisher's method of scoring. These are perhaps the most widely used methods. For going beyond the primary tool of GLIM for possibly nonlinear models see Gay and Welsh (1988) and Nelder and Pregibon (1987). Let G(^) = 0 be an estimating equation. Then, the Newton-Raphson and scoring iterative schemes can be defined by ^ (r+1) = ^ (r) - ( ˙G(^ (r) ))-1 G(^ (r) ) and ^ (r+1) = ^ (r) - (E ˙G(^ (r) ))-1 G(^ (r) ), respectively. The former uses the observed - ˙G() and the latter the expected -E ˙G() which, when G() is a standardized estimating function, equals the covariance matrix EG() G () These schemes often work well, with rapid convergence, especially if good starting values are chosen. A simple method of estimation, such as the method of moments, if available, can provide suitable starting values. An alternative approach which has the attraction of not involving matrices of partial derivatives of G has been suggested by Mak (1993). The idea here is as follows. For any , let H(, ) = E G(). Then, if ^ (0) is a given starting value, the sequence {^ (r) , r = 0, 1, . . . , } is defined via G(^ (r) ) = H(^ (r+1) , ^ (r) ). The properties of this scheme, which is equivalent to the method of scoring when H is linear in , have been investigated in Mak (1993). In particular, it is shown, under the usual kinds of regularity conditions, that if ^ is the required estimator, then ^ (r) - ^ = Op(n-r/2 ), n being the sample size and Op denoting order in probability. 13.3 Multiple Roots 13.3.1 Introduction For maximum likelihood estimation the accepted practice is to use the likelihood itself to discriminate between multiple roots. However, in cases where 13.3. MULTIPLE ROOTS 203 the estimating equation G() = 0, say, has been derived from other principles, there has been uncertainty about how to choose the appropriate root (e.g., Stefanski and Carroll (1987, Section 2.3), McCullagh (1991)). This has provided motivation for the search for scalar potential functions from which the estimating equation can be derived and, in particular, the development of a conservative quasi-score (Li and McCullagh (1994)), which, albeit with some loss in efficiency in general, is always the derivative of a scalar function. Recently Li (1993) and Hanfelt and Liang (1995) have developed approximate likelihood ratio methods for quasi-likelihood settings. In each of these papers it is shown that, under various regularity conditions the approximate likelihood ratio can be used to discriminate asymptotically, with probability one, to allow the choice of the correct root. To use these methods it is necessary that the estimating functions, G(s) (), say, are standardized to have the likelihood score property E - ˙G (s) () = E G(s) () G(s) () . However, unless has small dimension, the process of standardization may be tedious if explicit analytical expressions are desired. For example, the method of moments, or the eliminiation of incidential parameters, typically lead to equations which are not standardized. Furthermore, a wide range of estimating functions are used in practice, and many of these have no physically convincing interpretation. For example, if is 1-dimensional and G has an even number of roots, then its integral must increase steadily at one or other end of the range of . In many examples the integral is unbounded above, implying that the appropriate root is at a local rather than a global maximum. For these reasons we propose three simple direct methods which do not involve the construction of a potential function. This is clearly advantageous, for example, when (i) the estimating function is not optimal (for example being derived by the method of moments), or (ii) many incidental parameters are to be eliminated (e.g. for functional relationships). Also, conceptually unsatisfying issues such as the arbitrariness of paths for integrals defining approximate likelihood ratios in non-conservative cases are then avoided. The results of this section are from Heyde and Morton (1997). The methods which we advocate for choosing the correct root of G() = 0 are: (1) examining the asymptotics to see which root provides a consistent result; (2) picking the root for which ˙G() behaves asymptotically as its expected value E{ ˙G()}; and (3) using a least squares or goodness of fit criterion to select the best root. 204 COMPLEMENTS AND STRATEGIES The motivation for Method 2 comes from generalization of the usual technique for identifying the maximum of a quasi-likelihood where - ˙G() and -E ˙G() have interpretations as empirical and expected information respectively. The principles which underly the advocated methods are simple and transparent and we give three examples to illustrate their use. This is followed by a discussion of the underlying theoretical considerations. 13.3.2 Examples We begin with some simple examples, the first of which concerns the estimation of the mean of a non-normal distribution using the first two moments. Let X1, . . . , Xn be i.i.d. with mean (- < < ) and known variance 2 and consider the estimating function G() = n-1 n i=1 {a(Xi - ) + (Xi - )2 - 2 }, where a is a constant. Certainly we can calculate, say, () = 0 G(u) du, but to regard this cubic in as an appropriate likelihood surrogate requires a leap of faith. Nevertheless, the multiple root issue is simply resolved by each of the approaches mentioned above. Let = 1 4 a2 + X2 - n-1 n i=1 X2 i + 2 , X being the sample mean. Then, on the set { > 0} the roots of G are ^1 = 1 2 a + X - 1/2 , ^2 = 1 2 a + X + 1/2 , and the strong law of large numbers gives ^1 a.s. - 0, ^2 a.s. - 0 + a, 0 being the true value. Thus, the root ^1 is the correct one. This is Method 1. Using Method 2, we see that - ˙G() = a + 2( X - ), and E ˙G() = -a so that the ratio ˙G()/E ˙G() tends to 1 when = ^1 and to -1 when = ^2, identifying ^1 as the correct root. For Method 3 we use the criterion S() = n-1 n i=1 (Xi - )2 , and see that S(^2) - S(^1) = 2a 1 2 a2 n 13.3. MULTIPLE ROOTS 205 as n , from which we again prefer ^1 to ^2. The next example concerns estimation of the angle in circular data. This problem, which has been studied by McCullagh (1991), concerns the model in which (X1i, X2i), i = 1, . . . , n are i.i.d. multivariate normal, with mean vector (r cos , r sin ) and covariance matrix I. First we consider r = 1 and use the quasi-score (also the score) function Q() = -( X1 - cos ) sin + ( X2 - sin ) cos = - X1 sin + X2 cos . Then Q() = 0 possesses two solutions ^1 and ^2 with ^2 = ^1 + , ^1 = tan-1 ( X2/ X1) being the root in (-/2, /2). Assume that -/2 < 0 < /2 is the true value. Then the law of large numbers shows that ^1 is the consistent root. This is the approach of Method 1. For Method 2, we see that ˙Q() = - X2 sin - X1 cos , while E ˙Q() = -1 and ˙Q(^1)/E^1 ˙Q(^1) 1, ˙Q(^2)/E^2 ˙Q(^2) -1, so we choose ^1. Finally, for Method 3 we can take the sum of squares criterion S() = n i=1 (X1i - cos )2 + (X2i - sin )2 , and S(^2) - S(^1) = 4n( X1 cos ^1 + X2 sin ^2) 4n if ^1 is in the same half circle as 0. There is also an interesting sequel to this example. If r is also to be estimated, then the constraint r > 0 removes the ambiguity in and there is only one solution, which is ^1. In this case, two parameters are simpler than one! These have been toy examples and we now move to something more substantial. Here we follow up on the example on a functional normal regression model treated in Section 2.3 of Stefanski and Carroll (1987). The setting is one in which Y has a normal distribution with mean + u and variance 2 . It is supposed that u cannot be observed but that independent measurements X on it are available and var (X | u) = 2 where is known. For this setting, Stefanski and Carroll derive the estimating function (cf. their equation (2.15)) G() = - SY X + (SY Y - SXX) + SY X 206 COMPLEMENTS AND STRATEGIES for where SY Y = n-1 n i=1 (Yi - Y )2 , SY X = (SY X1 , . . . , SY Xp ) , SXX = (SXiYj ), with SY Xi = n-1 n k=1 (Yk - Y )(Xik - Xi), SXiXj = n-1 n k=1 (Xik - Xi)(Xjk - Xj). In contrast to Stefanski and Carroll we shall not specialize to the case where is scalar (p = 1). For Method 1 we need to examine the asymptotics. We suppose that SU,U = n-1 n i=1 (ui - u)(ui - u) var u, say, as n and then, using the law of large numbers in an obvious notation which covers the structural model case as well as the functional one, SY X a.s. - cov (X, Y ), SXX a.s. - var X, SY Y a.s. - var Y. Also var Y = E(var (Y | u)) + var (E(Y | u)) = 2 + (var u) , var X = E(var (X | u)) + var (E(X | u)) = 2 + var u, cov (X, Y ) = E(cov (X, Y ) | u)) + cov (E(X | u), E(Y | u)) = cov (u, + u) = (var u) . Thus, asymptotically, G(^) -^ (var u)^ + ( (var u) - var u)^ + (var u) = (-^ - I)(var u)(^ - ). (13.1) One root of (13.1) is ^ = and any other requires that (I +^ ) be singular and hence that ^ = -1 since, otherwise, (I + ^ )-1 = I - ^ 1 + ^ . 13.3. MULTIPLE ROOTS 207 Any such other root satisfies ^ - = (var u)-1 [(I + ^ )(I + ^ ) - I]Z, where Z is an arbitrary vector and, noting that we can take (I + ^ )= I + ^ , ^ - = (var u)-1 ^ Z. (13.2) Next, using ^ = -1 in (13.2), we have (var u)(^ - ) = - Z, (13.3) and, setting Z = s, (13.2) can be rewritten as ^ = [I - (var u)-1 s]-1 , (13.4) provided s det{(var u)-1 } = 1 so that I - (var u)-1 s is nonsingular. Note that, using (13.4) in (13.3), (I - (var u)-1 s)-1 = -1, (13.5) and s is obtained from this equation. The equations (13.4), (13.5) then specify the second root. When solving the estimating equation G() = 0, it is clear that the correct root is close to (SXX - ^2 )-1 SY X which consistently estimates when ^2 = SY Y - SY X(SXX - ^2 )-1 SY X. If we have an incorrect root ^, then SY X(SXX - ^2 )-1 ^ -1. Thus, Method 1 should pick the correct root given a suitably large sample. To use Method 2 in this context we find ˙G() = -SY X - SY X + (SY Y - SXX) -0(var u) - (var u)0 +(0(var u)0 - var u) almost surely, 0 denoting the true value, and E ˙G() -( + I)var u almost surely, so that if ^0 is a root which is consistent for 0, ˙G(^0)(E^0 ˙G(^0))-1 p - I, 208 COMPLEMENTS AND STRATEGIES while if ^1 is a root which is consistent for 1 = 0, ˙G(^1)(E^1 ˙G( ^1))-1 p - (10 + I)(11 + I) + (1 - 0) (var u) 0 (var u)-1 (11 + I)-1 = I + 10 + (1 - 0) (var u) 0 (var u)-1 (11 + I)-1 = I - s(11 + I) (var u)(11 + I)-1 = I since s = 0. Here we have used, in the algebra, the intermediate results (1 - 0) (var u) 0 = -s and 10(11 + I) var u = -s11. The ratio thus detects the correct root. 13.3.3 Theory Informal general explanations of the properties observed in the examples treated in Section 13.3.2 are straightforward to provide. The asymptotics of an estimating function G may usually be examined directly. Suppose, for example, that the data are {Xt, t T}, T being some index set, and that G is of the form GT = G(AT ({Xt, t T}); ) with AT ({Xt, t T}) a.s. - A(0) in the limit, 0 denoting the true value of . Then, the equation G(A(0); ) = 0 gives a clear indication of the large sample behavior of the roots of G() = 0. Results of the kind AT ({Xt, t T}) a.s. - A(0) are typically established using law of large number or ergodic theorem arguments. This is the basis of Method 1. To use Method 2, we assume that the matrix derivative ˙G() H(0, ) in probability, where H(0, ) = E0 ( ˙G()). Both quantities should be unbounded as sampling continues and their difference has zero expectation. Then, if ^0 is a consistent estimator of 0, (H(^0, ^0))-1 ˙G(^0) I in probability. But, if ^1 is a consistent estimator of 1 = 0, then (H(^1, ^1))-1 ˙G(^1) (H(1, 1))-1 H(0, 1), 13.3. MULTIPLE ROOTS 209 which in most problems is different from Is the main exception being where H(0, 1) is constant in 0. Apart from any such exception, (H(^, ^))-1 ˙G(^) can be used to identify the correct root. The matrix may not need to be examined in detail in order to discard incorrect roots. We could, for example, inspect the signs of the diagonal elements of ˙G(^) and H(^, ^). If we have a quasiscore, Q(), the asymptotic positive definiteness of - ˙Q(^) for ^ a consistent estimator of 0 may be a useful diagnostic. We note that for likelihood scores, - ˙Q(^) > 0 identifies a local maximum of the likelihood. Thus, Method 2 could be regarded as a generalization of this idea to non-standardized and nonoptimal estimating functions. The positive definiteness of - ˙Q is particularly useful when ˙Q depends on further variance parameters which are estimated through its magnitude. Finally, for the goodness of fit approach of Method 3 we assume that we have some acceptable criterion for the goodness of fit of the data. A modest requirement might be that the expectation of the criterion is a minimum when = 0. While we do not use the minimization of the criterion for estimating 0, it can be used to determine preference between the roots of G. Least squares, or some generalization of it, seems appropriate for this purpose. An informal justification can easily be given. We shall consider the situation in which there are consistent roots and possibly others. Let 0 denote the true value, for which there is a consistent estimator ^0 and suppose that ^1 p - 1 = 0. Then, for a least squares type setting with minimization criterion S() = n i=1 (Xi - f()) (Xi - f()), say, where X1, . . . , Xn is the data and EXi = f(), we have n-1 [S(^1) - S(^0)] = n-1 (f(^0) - f(^1)) n i=1 (2Xi - f(^0) - f(^1)) p - (f(0) - f(1)) (f(0) - f(1)), using law of large number considerations. If, on the other hand, ^2 is not a consistent root, then n-1 [S(^2) - S(^0)] = n-1 (f(^0) - f(^2)) n i=1 (2Xi - f(^0) - f(^2)) (f(0) - f( ^2)) (f(0) - f( ^2)) in probability, which is Op(1) and the discrimination between roots should still be clear. Note also that n-1 S(^0) will typically converge in probability to a constant. 210 COMPLEMENTS AND STRATEGIES 13.4 Resampling Methods Resampling methods, especially the bootstrap and jackknife, have proved to be highly useful in practice for obtaining confidence intervals for unknown , particularly when the sample size is limited. The methods were originally developed for i.i.d. random variables but considerable effort has been expended recently in providing adaptions to various dependent variable contexts. Nevertheless, the ideas behind resampling always require a focus on variables within the model of interest which do not depart drastically from stationarity. A detailed discussion on this subject is outside the scope of this monograph, in part because only a small proportion of the current literature involves estimating functions, but the prospects for future work are clear. Both jackknife and bootstrap methods of providing confidence intervals for estimating functions which include martingales are discussed by Lele (1991b). The jackknife results come from Lele (1991a) and the bootstrap results, which are rather less complete, adapt work of Wu (1986) and Liu (1988). One may expect that the technology associated with the moving block bootstrap to be able to deal with estimating functions based on stationary variables (e.g. K¨unsch (1989) for weakly dependent variables; Lahiri (1993), (1995) for strongly dependent ones). The idea behind the moving block bootstrap is to resample blocks of observations, rather than single observations as with the ordinary bootstrap, thereby adequately preserving the dependence of the data within blocks. References Aalen, O. O. (1977). Weak convergence of stochastic integrals related to counting processes. Z. Wahrsch. Verw. Geb. 38, 261­277. Aalen, O. O. (1978). Nonparametric inference for a family of counting processes. Ann. Statist. 6, 701­726. Aase, K. K. (1983). Recursive estimation in nonlinear time series of autoregressive type. J. Roy. Statist. Soc. Ser. B 45, 228­237. Adenstadt, R. K. (1974). On large-sample estimation for the mean of a stationary random sequence. Ann. Statist. 2, 1095­1107. Adenstadt, R. K., and Eisenberg, B. (1974). Linear estimation of regression coefficients. Quart. Appl. Math. 32, 317­327. Aitchison, J., and Silvey, S. D. (1958). Maximum likelihood estimation of parameters subject to restraint. Ann. Math. Statist. 29, 813­828. Andersen, P. K., Borgan, O., Gill, R. D., and Keiding, N. (1982). Linear nonparametric tests for comparison of counting processes, with applications to censored survival data. Internat. Statist. Rev. 50, 219­258. Anh, V. V. (1988). Nonlinear least squares and maximum likelihood estimation for a heteroscedastic regression model. Stochastic Process. Appl. 29, 317­333. Baddeley, A. J. (1995). Time-invariance estimating equations. Research Report. Dept. Mathematics, Univ. Western Australia. Bailey, N. T. J. (1975). The Mathematical Theory of Infectious Diseases. Griffin, London. Barndorff-Nielsen, O. E., and Cox, D. R. (1994). Inference and Asymptotics. Chapman and Hall, London. Barndorff-Nielsen, O. E., and Srensen, M. (1994). A review of some aspects of asymptotic likelihood theory for stochastic processes. Internat. Statist. Rev. 62, 133­165. 212 REFERENCES Basawa, I. V. (1985). Neyman-Le Cam tests based on estimating functions. In L. Le Cam and R. A. Olshen Eds., Proceedings of the Berkeley Conference in Honor of Jerzy Neyman and Jack Kiefer 2, Wadsworth, Belmont, CA, 811­825. Basawa, I. V. (1991). Generalized score tests for composite hypotheses. In V. P. Godambe, Ed. Estimating Functions, Oxford Science Publications, Oxford, 121­131. Basawa, I. V., and Brockwell, P. J. (1984). Asymptotic conditional inference for regular non-ergodic models with an application to autoregressive processes. Ann. Statist. 12, 161­171. Basawa, I. V., Huggins, R. M., and Staudte, R. G. (1985). Robust tests for time series with an application to first-order autoregressive processes. Biometrika 72, 559­571. Basawa, I. V., and Koul, H. L. (1988). Large-sample statistics based on quadratic dispersion. Internat. Statist. Rev. 56, 199­219. Basawa, I. V., and Prakasa Rao, B. L. S. (1980). Statistical Inference for Stochastic Processes. Academic Press, London. Basawa, I. V., and Scott, D. J. (1983). Asymptotic Optimal Inference for NonErgodic Models. Lecture Notes in Statistics 17, Springer, New York. Becker, N. G., and Heyde, C. C. (1990). Estimating population size from multiple recapture-experiments. Stochastic Process. Appl. 36, 77­83. Beran, J. (1989). A test of location for data with slowly decaying serial correlations. Biometrika 76, 261­269. Berliner, L. M. (1991). Likelihood and Bayesian prediction for chaotic systems. J. Amer. Statist. Assoc. 86, 938­952. Besag, J. E. (1975). Statistical analysis of non-lattice data. The Statistician 24, 179­195. Bhapkar, V. P. (1972). On a measure of efficiency in an estimating equation. Sankhya Ser. A 34, 467­472. Bhapkar, V. P. (1989). On optimality of marginal estimating equations. Technical Report No. 274, Dept. Statistics, Univ. Kentucky. Bibby, B. M., and Srensen, M. (1995). Martingale estimation functions for discretely observed diffusion processes. Bernoulli 1, 17­39. Billingsley, P. (1968). Convergence of Probability Measures. Wiley, New York. Bradley, E. L. (1973). The equivalence of maximum likelihood and weighted least squares estimates in the exponential family. J. Amer. Statist. Assoc. 68, 199­200. REFERENCES 213 Bustos, O. H. (1982). General M-estimates for contaminated p-th order autoregressive processes: consistency and asymptotic normality. Robustness in autoregressive processes. Z. Wahrsch. Verw. Geb. 59, 491­504. Carroll, R. J., and Stefanski, L. A. (1990). Approximate quasi-likelihood estimation in models with surrogate predictors. J. Amer. Statist. Assoc. 85, 652­663. Chan, N. H., and Wei, C. Z. (1988). Limiting distributions of least squares estimates of unstable autoregressive processes. Ann. Statist. 16, 367­401. Chandrasekar, B., and Kale, B. K. (1984). Unbiased statistical estimation for parameters in presence of nuisance parameters. J. Statistical Planning Inf. 9, 45­54. Chen, K., and Heyde, C. C. (1995). On asymptotic optimality of estimating functions. J. Statistical Planning Inf. 48, 102­112. Chen, Y. (1991). On Quasi-likelihood Estimations. PhD thesis, University of Wisconsin-Madison. Cheng, R. C. H., and Traylor, L. (1995). Non-regular maximum likelihood problems (with discussion). J. Roy. Statist. Soc. Ser. B 57, 3­44. Choi, Y. J., and Severo, N. C. (1988). An approximation for the maximum likelihood estimator for the infection rate in a simple stochastic epidemic. Biometrika 75, 392­394. Chung, K. L., and Williams, R. J. (1983). Introduction to Stochastic Integration. Birkh¨auser, Boston. Comets, F., and Janˇzura, M. (1996). A central limit theorem for conditionally centered random fields with an application to Markov fields (Preprint). Cox, D. R., and Hinkley, D. V. (1974). Theoretical Statistics. Chapman and Hall, London. Cox, J. S., Ingersoll, J. E., and Ross, S. A. (1985). A theory of term structure of interest rates. Econometrica 53, 363­384. Cram´er, H. (1946). Methods of Mathematical Statistics. Princeton University Press, Princeton. Crowder, M. (1987). On linear and quadratic estimating functions. Biometrika 74, 591­597. Cutland, N. J., Kopp, P. E., and Willinger, W. (1995). Stock price returns and the Joseph effect: a fractional version of the Black-Scholes model. Seminar on Stochastic Analysis, Random Fields and Applications (Ascona 1993), Progr. Probab. 36, Birkh¨auser, Basel, 327­351. 214 REFERENCES Dahlhaus, R. (1988). Efficient parameter estimation for self-similar processes. Ann. Statist. 17, 1749­1766. Daley, D. J. (1969). Integral representations of transition probabilities and serial covariances of certain Markov chains. J. Appl. Prob. 6, 648­659. Davidian, M., and Carroll, R. J. (1987). Variance function estimation. J. Amer. Statist. Assoc. 82, 1079­1091. Davis, R. A., and McCormick, W. P. (1989). Estimation for first order autoregressive processes with positive or bounded innovations. Stochastic Process. Appl. 31, 237­250. Dempster, A. P., Laird, N. M., and Rubin, D. B. (1977). Maximum likelihood from incomplete data via the E-M algorithm. J. Roy. Statist. Soc. Ser. B 39, 1­38. Denby, L., and Martin, R. L. (1979). Robust estimation of the first order autoregressive parameter. J. Amer. Statist. Assoc. 74, 140­146. Desmond, A. F. (1991). Quasi-likelihood, stochastic processes and optimal estimating functions. In V. P. Godambe, Ed., Estimating Functions, Oxford Science Publications, Oxford, 133­146. Desmond, A. F. (1996). Optimal estimating functions quasi-likelihood and statistical modelling. J. Statistical Planning Inf., in press. Dion, J.-P., and Ferland, R. (1995). Absolute continuity, singular measures and asymptotics for estimators. J. Statistical Planning Inf. 43, 235­242. Doukhan, P. (1994). Mixing. Lecture Notes in Statistics 85, Springer, New York. Duffie, D. (1992). Dynamic Asset Pricing Theory. Princeton Univ. Press, Princeton. Durbin, J. (1960). Estimation of parameters in time series regression models. J. Roy. Statist. Soc. Ser. B 22, 139­153. Efron, B., and Hinkley, D. V. (1978). Assessing the accuracy of the maximum likelihood estimator: observed versus Fisher information. Biometrika 65, 457­487. Elliott, R. J. (1982). Stochastic Calculus and Applications. Springer, New York. Elliott, R. J., Aggoun, L., and Moore, J. B. (1994). Hidden Markov Model Estimation and Control. Springer, New York. Feigin, P. D. (1977). A note on maximum likelihood estimation for simple branching processes. Austral. J. Statist. 19, 152­154. REFERENCES 215 Feigin, P. D. (1985). Stable convergence of semimartingales. Stochastic Process. Appl. 19, 125­134. Feigin, P. D., Tweedie, R. L., and Belyea, C. (1983). Weighted area techniques for explicit parameter estimation in multi-stage models. Austral. J. Statist. 25, 1­16. Firth, D. (1987). On efficiency of quasi-likelihood estimation. Biometrika 74, 233­246. Firth, D. (1993). Bias reduction for maximum likelihood estimates. Biometrika 80, 27­38. Firth, D., and Harris, I. R. (1991). Quasi-likelihood for multiplicative random effects. Biometrika 78, 545­555. Fitzmaurice, G. M., Laird, N. M., and Rotnitzky, A. G. (1993). Regression models for discrete longitudinal responses. Statistical Science 8, 284­309. Fox, R., and Taqqu, M. S. (1986). Large sample properties of parameter estimates for strongly dependent stationary Gaussian time series. Ann. Statist. 14, 517­532. Fox, R., and Taqqu, M. S. (1987). Central limit theorems for quadratic forms in random variables having long-range dependence. Probab. Th. Rel. Fields 74, 213­240. F¨urth, R. (1918). Statistik und Wahrscheinlichkeitsnachwirkung. Physik Z. 19, 421­426. Gastwirth, J. I., and Rubin, H. (1975). The behavior of robust estimators on dependent data. Ann. Statist. 3, 1070­1100. Gauss, C. F. (1880). Theoria Combationis Observatorium Erroribus Minimus Obnaie Part 1, 1821, Part 2, 1823; Suppl. 1826. In Werke 4, G¨ottingen, 1­108. Gay, D. M., and Welsh, R. E. (1988). Maximum-likelihood and quasi-likelihood for nonlinear exponential family regression models. J. Amer. Statist. Assoc. 83, 990­998. Glynn, P., and Iglehart, D. L. (1990). Simultaneous output analysis using standardized time series. Math. Oper. Res. 15, 1­16. Godambe, V. P. (1960). An optimum property of regular maximum-likelihood estimation. Ann. Math. Statist. 31, 1208­1211. Godambe, V. P. (1976). Conditional likelihood and unconditional optimum estimating equations. Biometrika 63, 277­284. Godambe, V. P. (1985). The foundations of finite sample estimation in stochastic processes. Biometrika 72, 419­428. 216 REFERENCES Godambe, V. P. (Ed.) (1991) Estimating Functions. Oxford Science Publications, Oxford. Godambe, V. P. (1994). Linear Bayes and optimal estimation. Technical Report STAT-94-11, Dept. Statistical & Actuarial Sciences, University of Waterloo. Godambe, V. P., and Heyde C. C. (1987). Quasi-likelihood and optimal estimation. Internat. Statist. Rev. 55, 231­244. Godambe, V. P., and Kale, B. K. (1991). Estimating functions: an overview. In V. P. Godambe, Ed., Estimating Functions, Oxford Science Publications, Oxford, 3­20. Godambe, V. P., and Thompson, M. E. (1974). Estimating equations in the presence of a nuisance parameter. Ann. Statist. 2, 568­571. Godambe, V. P., and Thompson, M. E. (1986). Parameters of superpopulation and survey population: their relationship and estimation. Internat. Statist. Rev. 54, 127­138. Godambe, V. P., and Thompson, M. E. (1989). An extension of quasilikelihood estimation. J. Statistical Planning Inf. 22, 137­152. Greenwood, P. E., and Wefelmeyer, W. (1991). On optimal estimating functions for partially specified counting process models. In V. P. Godambe, Ed., Estimating Functions, Oxford Science Publications, Oxford, 147­ 168. Grenander, U. (1981). Abstract Inference. Wiley, New York. Grenander, U., and Szeg¨o, G. (1958). Toeplitz Forms and their Applications. Univ. Calif. Press, Berkeley and Los Angeles. Guyon, X. (1982). Parameter estimation for a stationary process on a ddimensional lattice. Biometrika 69, 95­105. Guyon, X. (1995). Random Fields on a Network, Modeling, Statistics and Applications. Springer, New York. Halfin, S. (1982). Linear estimators for a class of stationary queueing processes. Oper. Res. 30, 515­529. Hall, P., and Heyde, C. C. (1980). Martingale Limit Theory and its Application. Academic Press, New York. Hanfelt, J. J., and Liang, K.-Y. (1995). Approximate likelihood ratios for general estimating functions. Biometrika 82, 461­477. Hannan, E. J. (1970). Multiple Time Series. Wiley, New York. REFERENCES 217 Hannan, E. J. (1973). The asymptotic theory of linear time series models. J. Appl. Prob. 10, 130­145. Hannan, E. J. (1974). Correction to Hannan (1973). J. Appl. Prob. 11, 913. Hannan, E. J. (1976). The asymptotic distribution of serial covariances. Ann. Statist. 4, 396­399. Harville, D. A. (1977). Maximum likelihood approaches to variance components estimation and related problems. J. Amer. Statist. Assoc. 72, 320­338. Heyde, C. C. (1986). Optimality in estimation for stochastic processes under both fixed and large sample conditions. In Yu. V. Prohorov, V. A. Statulevicius, V. V. Sazonov and B. Grigelionis, Eds., Probability Theory and Mathematical Statistics. Proceedings of the Fourth Vilnius Conference, Vol. 1, VNU Sciences Press, Utrecht, 535­541. Heyde, C. C. (1987). On combining quasi-likelihood estimating functions. Stochastic Process. Appl. 25, 281­287. Heyde, C. C. (1988a). Fixed sample and asymptotic optimality for classes of estimating functions. Contemp. Math. 80, 241­247. Heyde, C. C. (1988b). Asymptotic efficiency results for the method for moments with application to estimation for queueing processes. In O. J. Boxma and R. Syski, Eds., Queueing Theory and its Application. Liber Amicorum for J. W. Cohen. CWI Monograph No. 7, North-Holland, Amsterdam, 405­412. Heyde, C. C. (1989a). On efficiency for quasi-likelihood and composite quasilikelihood methods. In Y. Dodge, Ed., Statistical Data Analysis and Inference, Elsevier, Amsterdam, 209­213. Heyde, C. C. (1989b). Quasi-likelihood and optimality for estimating functions: some current unifying themes. Bull. Internat. Statist. Inst. 53, Book 1, 19­29. Heyde, C. C. (1992a). On best asymptotic confidence intervals for parameters of stochastic processes. Ann. Statist. 20, 603­607. Heyde, C. C. (1992b). Some results on inference for stationary processes and queueing systems. In U. N. Bhat and I. V. Basawa, Eds., Queueing and Related Models. Oxford Univ. Press, Oxford, 337­345. Heyde, C. C. (1993). Quasi-likelihood and general theory of inference for stochastic processes. In A. Obretenov and V. T. Stefanov, Eds., 7th International Summer School on Probability Theory and Mathematical Statistics, Lecture Notes, Science Culture Technology Publishing, Singapore, 122­152. 218 REFERENCES Heyde, C. C. (1994a). A quasi-likelihood approach to estimating parameters in diffusion type processes. In J. Galambos and J. Gani, Eds., Studies in Applied Probability, J. Applied Prob. 31A, 283­290. Heyde, C. C. (1994b). A quasi-likelihood approach to the REML estimating equations. Statistics & Probability Letters 21, 381­384. Heyde, C. C. (1996). On the use of quasi-likelihood for estimation in hiddenMarkov random fields. J. Statistical Planning Inf. 50, 373­378. Heyde, C. C., and Gay, R. (1989). On asymptotic quasi-likelihood. Stochastic Process. Appl. 31, 223­236. Heyde, C. C., and Gay, R. (1992). Thoughts on modelling and identification of random processes and fields subject to possible long-range dependence. In L. H. Y. Chen, K. P. Choi, K. Hu and J.-H. Lou, Eds., Probability Theory, de Gruyter, Berlin, 75­81. Heyde, C. C., and Gay, R. (1993). Smoothed periodogram asymptotics and estimation for processes and fields with possible long-range dependence. Stochastic Process. Appl. 45, 169­182. Heyde, C. C., and Lin, Y.-X. (1991). Approximate confidence zones in an estimating function context. In V. P. Godambe, Ed., Estimating Functions, Oxford Science Publications, Oxford, 161­168. Heyde, C. C., and Lin, Y.-X. (1992). On quasi-likelihood methods and estimation for branching processes and heteroscedastic regression models. Austral. J. Statist. 34, 199­206. Heyde, C. C., and Morton, R. (1993). On constrained quasi-likelihood estimation. Biometrika 80, 755­761. Heyde, C. C., and Morton, R. (1995). Quasi-likelihood and generalizing the E-M algorithm. J. Roy. Statist. Soc. Ser. B 58, 317-327. Heyde, C. C., and Morton, R. (1997). Multiple roots and dimension reduction issues for general estimating equations (Preprint). Heyde, C. C., and Seneta, E. (1972). Estimation theory for growth and immigration rates in a multiplicative process. J. Appl. Prob. 9, 235­256. Heyde, C. C., and Seneta, E. (1977). I. J. Bienaym´e: Statistical Theory Anticipated. Springer, New York. Hoffmann-Jrgensen, J. (1994). Probability With a View Towards Statistics. Vol. II. Chapman and Hall, New York. Hotelling, H. (1936). Relations between two sets of variables. Biometrika 28, 321­377. REFERENCES 219 Huber, P. J. (1981). Robust Statistics. Wiley, New York. Hutton, J. E., and Nelson, P. I. (1984). A mixing and stable central limit theorem for continuous time martingales. Technical Report No. 42, Kansas State University. Hutton, J. E., and Nelson, P. I. (1986). Quasi-likelihood estimation for semimartingales. Stochastic Process. Appl. 22, 245­257. Hutton, J. E., Ogunyemi, O. T., and Nelson, P. I. (1991). Simplified and twostage-quasi-likelihood estimators. In V. P. Godambe, Ed., Estimating Functions, Oxford Science Publications, Oxford, 169­187. Ibragimov, I. A., and Linnik, Yu. V. (1971). Independent and Stationary Sequences of Random Variables. Wolters-Noordhoff, Groningen. Jiang, J. (1996). REML estimation: asymptotic behaviour and related topics. Ann. Statist. 24, 255­286. Judge, G. G., and Takayama, T. (1966). Inequality restrictions in regression analysis. J. Amer. Statist. Assoc. 61, 166­181. Kabaila, P. V. (1980). An optimal property of the least squares estimate of the parameter of the spectrum of a purely non-deterministic time series. Ann. Statist. 8, 1082­1092. Kabaila, P. V. (1983). On estimating time series parameters using sample autocorrelations. J. Roy. Statist. Soc. Ser. B 45, 107­119. Kallianpur, G. (1983). On the diffusion approximation to a discountinuous model for a single neuron. In P. K. Sen, Ed., Contributions to Statistics: Essays in Honor of Norman L. Johnson. North-Holland, Amsterdam, 247­258. Kallianpur, G., and Selukar, R. S. (1993). Estimation of Hilbert space valued random variables by the method of sieves. In J. K. Ghost et al., Eds., Statistics and Probability, A Raghu Rag Bahadur Festschrift, Wiley Eastern, New Delhi, 325­347. Karlin, S., and Taylor, H. M. (1981). A Second Course in Stochastic Processes. Academic Press, New York. Karr, A. F. (1987). Maximum likelihood estimation in the multiplicative intensity model via sieves. Ann. Statist. 15, 473­490. Karson, M. J. (1982). Multivariate Statistical Methods. Iowa State Univ. Press, Ames, IA. Kaufmann, H. (1987). On the strong law of large numbers for multivariate martingales. Stochastic Process. Appl. 26, 73­85. 220 REFERENCES Kaufmann, H. (1989). On weak and strong consistency of the maximum likelihood estimator in stochastic processes (Preprint). Kessler, M., and Srensen, M. (1995). Estimating equations based on eigenfunctions for a discretely observed diffusion process. Research Report No. 332, Dept. Theoretical Statistics, Univ. Aarhus. Kimball, B. F. (1946). Sufficient statistical estimation functions for the parameters of the distribution of maximum values. Ann. Math. Statist. 17, 299­309. Kloeden, P. E., and Platen, E. (1992). Numerical Solution of Stochastic Differential Equations. Springer, Berlin. Kloeden, P. E., Platen, E., Schurz, H., and Srensen, M. (1996). On effects of discretization on estimators of drift parameters for diffusion processes. J. Appl. Prob. 33, 1061-1076. K¨unsch, H. (1984). Infinitesimal robustness for autoregressive processes. Ann. Statist. 12, 843­863. K¨unsch, H. R. (1989). The jackknife and bootstrap for general stationary observations. Ann. Statist. 17, 1217­1241. Kulkarni, P. M., and Heyde, C. C. (1987). Optimal robust estimation for discrete time stochastic processes. Stochastic Process. Appl. 26, 267­ 276. Kulperger, R. (1985). On an optimal property of Whittle's Gaussian estimate of the parameter of the spectrum of a time series. J. Time Ser. Anal. 6, 253­259. Kutoyants, Yu., and Vostrikova, L. (1995). On non-consistency of estimators. Stochastics and Stochastic Reports 53, 53­80. Lahiri, S. N. (1993). On the moving block bootstrap under long range dependence. Statistics & Probability Letters 18, 405­413. Lahiri, S. N. (1995). On the asymptotic behaviour of the moving block bootstrap for normalized sums of heavy-tail random variables. Ann. Statist. 23, 1331­1349. Lai, T. L., and Wei, C. Z. (1982). Least squares estimates in stochastic regression models with applications to identification of dynamic systems. Ann. Statist. 10, 154­166. Laird, N. (1985). Missing information principle. In N. L. Johnson and S. Kotz, Eds., Encyclopedia of Statistical Sciences, Wiley, New York, 5, 548­552. Le Cam, L. (1990a). On the standard asymptotic confidence ellipsoids of Wald. Internat. Statist. Rev. 58, 129­152. REFERENCES 221 Le Cam, L. (1990b). Maximum likelihood: an introduction. Internat. Statist. Rev. 58, 153­171. Lele, S. (1991a). Jackknifing linear estimating equations: asymptotic theory and applications in stochastic processes. J. Roy. Statist. Soc. Ser. B 53, 253­267. Lele, S. (1991b). Resampling using estimating equations. In V. P. Godambe, Ed., Estimating Functions, Oxford Science Publications, Oxford, 295­ 304. Lele, S. (1994). Estimating functions in chaotic systems. J. Amer. Statist. Assoc. 89, 512­516. L´epingle, D. (1977). Sur le comportement asymptotique des martingales locales. Springer Lecture Notes in Mathematics 649, 148­161. Leskow, J. (1989). A note on kernel smoothing of an estimator of a periodic function in the multiplicative model. Statistics & Probability Letters 7, 395­400. Leskow, J., and Rozanski, R. (1989). Histogram maximum likelihood estimation in the multiplicative intensity model. Stochastic Process. Appl. 31, 151­159. Li, B. (1993). A deviance function for the quasi-likelihood method. Biometrika 80, 741­753. Li, B. (1996a). A minimax approach to the consistency and efficiency for estimating equations. Ann. Statist. 24, 1283­1297. Li, B. (1996b). An optimal estimating equation based on the first three cumulants (Preprint). Li, B., and McCullagh, P. (1994). Potential functions and conservative estimating equations. Ann. Statist. 22, 340­356. Liang, K. Y., and Zeger, S. L. (1986). Longitudinal data analysis using generalized linear models. Biometrika 73, 13­22. Lin, Y.-X. (1992). The quasi-likelihood method. PhD thesis, Australian National University. Lin, Y.-X. (1994a). The relationship between quasi-score estimating functions and E-sufficient estimating functions. Austral. J. Statist. 36, 303­311. Lin, Y.-X. (1994b). On the strong law of large numbers of multivariate martingales with random norming. Stochastic Process. Appl. 54, 355­360. Lin, Y.-X. (1996). Quasi-likelihood estimation of variance components of heteroscedastic random ANOVA model (Preprint). 222 REFERENCES Lin, Y.-X., and Heyde, C. C. (1993). Optimal estimating functions and Wedderburn's quasi-likelihood. Comm. Statist. Theory Meth. 22, 2341­2350. Lindsay, B. (1982). Conditional score functions: some optimality results. Biometrika 69, 503­512. Lindsay, B. G. (1988). Composite likelihood methods. Contemp. Math. 80, 221­239. Liptser, R. S., and Shiryaev, A. N. (1977). Statistics of Random Processes I. General Theory. Springer, New York. Little, R. J. A., and Rubin, D. B. (1987). Statistical Analysis with Missing Data. Wiley, New York. Liu, R. Y. (1988). Bootstrap procedures under some non-i.i.d. models. Ann. Statist. 16, 1696­1708. Mak, T. K. (1993). Solving non-linear estimating equations. J. Roy. Statist. Soc. Ser. B 55, 945­955. Martin, R. D. (1980). Robust estimation of autoregressive models (with discussion). In D. R. Brillinger and G. C. Tiao, Eds, Directions of Time Series, Inst. Math. Statist., Hayward, CA, 228­262. Martin, R. D. (1982). The Cram´er-Rao bound and robust M-estimates for autoregressions. Biometrika 69, 437­442. Martin, R. D., and Yohai, V. J. (1985). Robustness in time series and estimating ARMA models. In E. J. Hannan, P. R. Krishnaiah and M. M. Rao, Eds., Handbook of Statistics 5, Elsevier Science Publishers, New York, 119­155. McCullagh, P. (1983). Quasi-likelihood functions. Ann. Statist. 11, 59­67. McCullagh, P. (1991). Quasi-likelihood and estimating functions. In D. V. Hinkley, N. Reid and E. J. Snell, Eds., Statistical Theory and Modelling. In Honour of Sir David Cox, FRS. Chapman and Hall, London, 265­286. McCullagh, P., and Nelder, J. A. (1989). Generalized Linear Models, 2nd Ed., Chapman and Hall, New York. McKeague, I. W. (1986). Estimation for a semimartingale model using the method of sieves. Ann. Statist. 14, 579­589. McLeish, D. L., and Small, C. G. (1988). The Theory and Applications of Statistical Inference Functions. Lecture Notes in Statistics 44, Springer, New York. REFERENCES 223 Merkouris, T. (1992). A transform method for optimal estimation in stochastic processes: basic aspects. In J. Chen, Ed. Proceedings of a Symposium in Honour of Professor V. P. Godambe, University of Waterloo, Waterloo, Canada, 42 pp. Morton, R. (1981a). Efficiency of estimating equations and the use of pivots. Biometrika 68, 227­233. Morton, R. (1981b). Estimating equations for an ultrastructural relationship. Biometrika 68, 735­737. Morton, R. (1987). A generalized linear model with nested strata of extraPoisson variation. Biometrika 74, 247­257. Morton, R. (1988). Analysis of generalized linear models with nested strata of variation. Austral. J. Statist. 30A, 215­224. Morton, R. (1989). On the efficiency of the quasi-likelihood estimators for exponential families with extra variation. Austral. J. Statist. 31, 194­ 199. Mtundu, N. D., and Koch, R. W. (1987). A stochastic differential equation approach to soil moisture. Stochastic Hydrol. Hydraul. 1, 101­116. Mykland, P. A. (1995). Dual likelihood. Ann. Statist. 23, 386­421. Naik-Nimbalkar, U. V., and Rajarshi, M. B. (1995). Filtering and smoothing via estimating functions. J. Amer. Statist. Assoc. 90, 301­306. Nelder, J. A., and Lee, Y. (1992). Likelihood, quasi-likelihood and pseudolikelihood: some comparisons, J. Roy. Statist. Soc. Ser. B 54, 273­284. Nelder, J. A., and Pregibon, D. (1987). An extended quasi-likelihood function. Biometrika 74, 221­232. Nguyen, H. T. and Pham, D. P. (1982). Identification of the nonstationary diffusion model by the method of sieves. SIAM J. Optim. Control 20, 603­611. Osborne, M. R. (1992). Fisher's method of scoring, Internat. Statist. Rev. 60, 99­117. Parzen, E. (1957). On consistent estimates of the spectrum of a stationary time series. Ann. Math. Statist. 28, 329­348. Pedersen, A. R. (1995). Consistency and asymptotic normality of an approximate maximum likelihood estimator for discretely observed diffusion processes. Bernoulli 1, 257­279. Pollard, D. (1984). Convergence of Stochastic Processes. Springer, New York. 224 REFERENCES Prentice, R. L. (1988). Correlated binary regression with covariates specific to each binary observation. Biometrics 44, 1033­1048. Priestley, M. B. (1981). Spectral Analysis and Time Series. Academic Press, London. Pukelsheim, F. (1993). Optimal Design of Experiments. Wiley, New York. Qian, W., and Titterington, D. M. (1990). Parameter estimation for hidden Markov chains. Statistics & Probability Letters 10, 49­58. Rao, C. R. (1973). Linear Statistical Inference and its Applications, 2nd Ed., Wiley, New York. Rao, C. R., and Mitra, S. K. (1971). Generalized Inverse of Matrices and its Applications, Wiley, New York. Rebolledo, R. (1980). Central limit theorems for local martingales. Z. Wahrsch. Verw. Geb. 51, 269­286. Reynolds, J. F. (1975). The covariance structure of queues and related processes - a survey of recent work. Adv. Appl. Prob. 7, 383­415. Ripley, B. D. (1988). Statistical Inference for Spatial Processes. Cambridge Univ. Press, Cambridge. Rogers, L. C. G., and Williams, D. (1987). Diffusions, Markov Processes and Martingales, Vol. 2, Ito Calculus. Wiley, Chichester. Rosenblatt, M. (1985). Stationary Sequences and Random Fields. Birkh¨auser, Boston. Samarov, A., and Taqqu, M. S. (1988). On the efficiency of the sample mean in long-memory noise. J. Time Series Anal. 9, 191­200. Samuel, E. (1969). Comparison of sequential rules for estimation of the size of a population. Biometrics 25, 517­527. Schuh, H.-J., and Tweedie, R. L. (1979). Parameter estimation using transform estimation in time-evolving models. Math. Biosciences 45, 37­67. Shen, X., and Wong, W. H. (1994). Convergence rate of sieve estimates. Ann. Statist. 22, 580­615. Shiryaev, A. N. (1981). Martingales, recent developments, results and applications. Internat. Statist. Rev. 49, 199­233. Shumway, R. H., and Stoffer, D. S. (1982). An approach to time series smoothing and forecasting using the E-M algorithm. J. Time Series Anal. 3, 253­264. REFERENCES 225 Small, C. G., and McLeish, D. L. (1989). Projection as a method for increasing sensitivity and eliminating nuisance parameters. Biometrika 76, 693­703. Small, C. G., and McLeish, D. L. (1991). Geometrical aspects of efficiency criteria for spaces of estimating functions. In V. P. Godambe, Ed., Estimating Functions, Oxford Science Publications, Oxford, 267­276. Small, C. G., and McLeish, D. L. (1994). Hilbert Space Methods in Probability and Statistical Inference, Wiley, New York. Smith, R. L. (1985). Maximum likelihood estimation in a class of non-regular cases. Biometrika 72, 67­92. Smith, R. L. (1989). A survey of nonregular problems. Bull. Internat. Statist. Inst. 53, Book 3, 353­372. Srensen, M. (1990). On quasi-likelihood for semimartingales. Stochastic Process. Appl. 35, 331­346. Srensen, M. (1991). Likelihood methods for diffusions with jumps. In N. U. Prabhu and I. V. Basawa, Eds., Statistical Inference in Stochastic Processes, Marcel Dekker, New York, 67­105. Stefanski, L. A., and Carroll, R. J. (1987). Conditional score and optimal scores for generalized linear measurement-error models. Biometrika 74, 703­716. Sweeting, T. J. (1986). Asymptotic conditional inference for the offspring mean of a supercritical Galton-Watson process. Ann. Statist. 14, 925­ 933. Thavaneswaran, A. (1991). Tests based on an optimal estimate. In V. P. Godambe, Ed., Estimating Functions, Oxford Science Publications, Oxford, 189­197. Thavaneswaran, A., and Abraham, B. (1988). Estimation for non-linear time series using estimating equations, J. Time Ser. Anal. 9, 99­108. Thavaneswaran, A., and Thompson, M. E. (1986). Optimal estimation for semimartingales. J. Appl. Prob. 23, 409­417. Thisted, R. A. (1988). Elements of Statistical Computing. Chapman and Hall, New York. Thompson, M. E., and Thavaneswaran, A. (1990). Optimal nonparametric estimation for some semimartingale stochastic differential equations. Appl. Math. Computation 37, 169­183. Tjstheim, D. (1978). Statistical spatial series modelling. Adv. Appl. Prob. 10, 130­154. 226 REFERENCES Tjstheim, D. (1983). Statistical spatial series modelling II: some further results on unilateral lattice processes. Adv. Appl. Prob. 15, 562­584. Vajda, I. (1995). Conditions equivalent to consistency of approximate MLE's for stochastic processes. Stochastic Process. Appl. 56, 35­56. Verbyla, A. P. (1990). A conditional derivation of residual maximum likelihood. Austral. J. Statist. 32, 227­230. Vitale, R. A. (1973). An asymptotically efficient estimate in time series analysis. Quart. Appl. Math. 30, 421­440. Wald, A. (1949). Note on the consistency of the maximum likelihood estimate. Ann. Math. Statist. 20, 595­601. Watson, R. K., and Yip, P. (1992). A note on estimation of the infection rate. Stochastic Process. Appl. 41, 257­260. Wedderburn, R. W. M. (1974). Quasi-likelihood functions, generalized linear models, and the Gauss-Newton method. Biometrika 61, 439­447. Wei, C. Z., and Winnicki, J. (1989). Some asymptotic results for branching processes with immigration. Stochastic Process. Appl. 31, 261­282. Whittle, P. (1951). Hypothesis Testing in Time Series Analysis. Almqvist and Wicksell, Uppsala. Whittle, P. (1952). Tests of fit in time series. Biometrika 39, 309­318. Whittle, P. (1953). The analysis of multiple time series. J. Roy. Statist. Soc. Ser. B 15, 125­139. Whittle, P. (1954). On stationary processes in the plane. Biometrika 41, 434­449. Winnicki, J. (1988). Estimation theory for the branching process with immigration. Contemp. Math. 80, 301-322. Wu, C. F. J. (1986). Jackknife, bootstrap and other resampling methods in regression analysis. Ann. Statist. 14, 1261­1295. Yanev, N. M., and Tchoukova-Dantcheva, S. (1980). On the statistics of branching processes with immigration. C. R. Acad. Sci. Bulg. 33, 463­ 471. Zeger, S. L., and Liang, K.-Y. (1986). Longitudial data analysis for discrete and continuous outcomes. Biometrics 44, 1033­1048. Zehnwirth, B. (1988). A generalization of the Kalman filter for models with state-dependent observation variance. J. Amer. Statist. Assoc. 83, 164­ 167. Zygmund, A. (1959). Trigonometric Series, Vol. 1, 2nd Ed., Cambridge Univ. Press, Cambridge. Index Aalen, O.O., 18, 19, 150, 166, 211 Aase, K.K., 177, 211 Abraham, B., 176, 177, 225 acceptable estimating function, 74 Adenstadt, R.K., 155, 156, 211 Aggoun, L., 136, 138, 214 Aitchison, J., 183, 211 algorithm, 127, 177 E-M, 116­119, 123, 127 P-S, 116­119, 123, 127 ancillary, 43, 60, 107 (see also E- ancillary) Anderson, P.K., 18, 211 Anh, V.V., 160, 161, 211 ARMA model, 23 asymptotically non-negative definite, 71, 72 asymptotic mixed normality, 27, 60, 62­64 asymptotic normality, 26, 27, 40, 41, 54­56, 60, 62­64, 110, 136, 147, 159, 161, 163, 164, 179, 191, 196, 197, 201 asymptotic first order efficiency, 72 asymptotic quasi-likelihood (see quasilikelihood, asymptotic) asymptotic relative efficiency, 74, 125, 126, 167, 168, 175 autoregressive process, 19, 23, 31, 56, 79, 98, 170, 173, 174, 176 spatial, 137 conditional, 137 random coefficient, 176 threshold, 176 Baddeley, A.J., 201, 211 Bailey, N.T.J., 162, 163, 211 Barndorff-Nielsen, O.E., 35, 56, 59, 62, 86, 211 Banach space, 147 Basawa, I.V., 54, 60, 63, 132, 135, 141, 142, 169, 173, 185, 212 Becker, N.G., 165, 168, 212 Belyea, C., 200, 215 Beran, J., 158, 212 Berliner, L.M., 37, 212 Bernoulli distribution, 87 Besag, J.E., 91, 212 best linear unbiased estimator (BLUE), 153, 154 beta distribution, 124 Bhapkhar, V.P., 115, 212 bias correction, 61, 67 Bibby, B.M., 135, 212 binary response, 25 Billingsley, P., 65, 212 binomial distribution, 122, 167 birth and death process, 18 birth process, 56 Black-Scholes model, 31 bootstrap, 9, 210 Borgan, O., 18, 211 Bradley, E.L., 7, 212 branching process, 182 Galton-Watson, 2, 15, 27, 31, 35, 36, 69, 87, 195 Brockwell, P.J., 60, 212 Brouwer fixed point theorem, 183 Brownian motion, 17, 31, 33, 34, 131­133, 136, 196 Burkholder, Davis, Gundy inequality, 149 Bustos, O.H., 169, 213 227 228 INDEX cadlag, 191 Carroll, R.J., 131, 139, 203, 205, 213, 214, 225 Cauchy distribution, 40, 200 Cauchy-Schwarz inequality, 5, 24, 167 censored data, 150 central limit results, 4, 9, 26, 27, 64, 149, 155, 166, 167, 179, 190­ 195 Chan, N.H., 56, 213 Chandrasekar, B., 19, 213 characteristic function, 199 Chebyshev inequality, 189 Chen, K., 88, 213 Chen, Y., 182, 213 Cheng, R.C.H., 54, 213 chi-squared distribution, 58, 110, 143­ 145, 172 Choi, Y.J., 163, 213 Cholesky square root matrix, 71, 72 Chung, K.L., 97, 213 coefficient of variation, 121, 124 colloidal solution, 69 Comets, F., 179, 213 complete probability space, 92 conditional centering, 179 conditional inference, 107 confidence intervals (or zones), 1, 4, 8, 9, 24, 27, 53­67, 69, 71, 88, 110, 131, 158, 163, 171, 191, 210 average, 56, 66 simulteneous, 58 consistency, 6, 8, 10, 26, 38, 40, 54, 55, 63, 64, 70, 131, 135, 147­150, 161, 163, 179­186, 190, 196, 197, 201, 203 constrained parameter estimation, 8, 107­112, 142 control, 41 convergence stable, 191, 193, 195 mixing, 57, 191, 192 convex, 14, 46, 156­158 correlation, 6, 13, 25 counting process, 18, 112, 150, 151, 165 covariate, 25, 148 Cox, D.R., 35, 41, 54, 58, 61, 62, 107, 180, 211, 213 Cox, J.S., 131, 133, 213 Cox-Ingersoll-Ross model, 131, 133 Cram´er, H., 2, 7, 180, 185, 195, 213 Cram´er-Rao inequality, 2, 7 Cram´er-Wold device, 195 cross-sectional data, 159 Crowder, M., 95, 105, 213 cumulant spectal density, 159 cumulative hazard function, 9, 150 curvature, 62 Cutland, N.J., 31, 213 Dahlhaus, R., 86, 214 Daley, D.J., 156, 214 Davidian, M., 131, 214 Davis, R.A., 19, 214 demographic stochasticity, 36 Dempster, A.P., 116, 214 Denby, L., 169, 214 dererminant criterion for optimality, 19 Desmond, A.F., 2, 25, 214 differentiability (stochastic), 56 differential geometry, 62 diffusion, 8, 17, 129, 131, 132, 135, 148, 151, 200 Dion, J.-P., 10, 214 discrete exponential family, 2, 16 dispersion, 12, 22, 119, 124, 129 distance, 7, 12 Doob-Meyer decomposition, 27 drift coefficient, 131, 148 Doukhan, P., 179, 214 Duffie, D., 31, 214 Durbin, J., 1, 214 dynamical system, 37, 38, 40, 131, 136 E-ancillary, 8, 43­51, 113, 114 Edgeworth expansion, 62 efficient score statistic, 9, 141, 142 INDEX 229 Efron, B., 59, 214 eigenfunction, 200, 201 Eisenberg, B., 155, 211 Elliott, R.J., 50, 136, 138, 214 ellipsoids (of Wald), 58 E-M algorithm, 8, 103, 107, 116, 117, 119­123, 127 epidemic, 9, 162, 163 equicontinuity (stochastic), 56 ergodic, 134, 142, 153 ergodic theorem, 134, 153, 208 error approximation, 148, 151 contrasts, 139 estimation, 148, 151 estimating functions, 1, 2 combination, 2, 8, 35, 103, 137, 138 optimal, 4­6 robust, 9, 169­175 standardized, 3­5, 118, 138 estimating function space, 3­5, 8, 11, 13, 15­19, 22­25, 28, 32, 36­40, 43­51, 71, 75, 88, 89, 92, 94­96, 99, 100, 104, 105, 111, 130, 132, 137, 138, 153, 162­164, 166, 169­ 171, 176, 195, 199­201 Hutton-Nelson, 32, 33, 162­164 convex, 14, 46­48 estimation constrained parameter, 8, 107­ 112, 142 function, 147­151 recursive, 176­177 robust, 169­175 E-sufficient, 8, 43­51, 113, 114 Euclidean space, 11, 92, 94, 141, 182 Euler scheme, 151 exponential distribution, 41, 121, 125 exponential family, 2, 7, 16, 24, 38, 53, 79, 125 failure, 38 F-distribution, 110 Feigin, P.D., 17, 200, 214, 215 Fejer kernel, 157 Ferland, R., 10, 214 field (see random field) filtering, 8, 103 filtration, 27, 30, 54, 94, 132 finance, 31, 135 finite variation process, 31, 148 first order efficient, 72, 73 Firth, D., 61, 67, 95, 102, 105, 215 Fisher, R.A., 1, 2, 12, 40, 59, 72, 97, 107, 141, 202 Fisher information, 2, 12, 40, 59, 72, 141 conditional, 97 Fisher method of scoring, 120, 202 Fitzmaurice, G.M., 25, 26, 215 Fourier methods, 85 Fox, R., 86, 215 fractional Brownian motion, 31, 86 function estimation, 9, 147­151 functional relationships, 112, 205 F¨urth, R., 87, 215 Galton-Watson process, 2, 15, 27, 31, 36, 69, 195 gamma distribution, 56, 134 Gastwirth, J.L., 169, 215 Gauss, C.F. i, 1, 3, 25, 83, 84, 86, 127, 138, 158, 159, 161, 215 Gaussian distribution, 83, 86, 127, 138, 158, 159 (see also normal distribution) Gauss-Markov theorem, 3, 4, 161 Gay, D.M., 202, 215 Gay, R., 74, 82, 88, 218 generalized estimating equation (GEE), 8, 25, 26, 89 generalized inverse, 30, 108, 109, 130, 132, 184 generalized linear model (GLIM), 21, 22, 79, 100, 104, 202 geometric distribution, 38 Gibbs field, 138 Gibbs sampler, 201 Gill, R.D., 18, 211 230 INDEX Girsanov transformation, 138 Glynn, P., 64, 215 Godambe, V.P., 1, 2, 21, 38, 48, 97, 105, 107, 116, 172, 215, 216 Gram-Schmidt orthogonalization, 93, 96 Greenwood, P.E., 56, 216 Grenander, U., 147, 156, 216 Guyon, X., 179, 201, 216 Hajek convolution theorem, 63 Halfin, S., 157, 216 Hall, P., 54, 57, 63, 87, 156, 180, 181, 187, 192, 195, 216 Hanfelt, J.J., 203, 216 Hannan, E.J., 83, 85, 86, 153, 216, 217 Harris, I.R., 102, 215 Harville, D., 130, 217 hazard function, 150 Hermite-Chebyshev polynomials, 97 heteroscedastic autoregression, 79 regression, 9, 159­161 Heyde, C.C., 1, 2, 13, 38, 48, 54, 57, 62, 63, 69, 72, 74, 82, 87, 88, 92, 94, 97, 107, 116, 126, 131, 136, 153, 156, 159, 161, 165, 168, 169, 172, 180, 181, 187, 192, 195, 203, 212, 213, 216­218, 220, 222 hidden Markov models, 9, 93, 136, 139 Hinkley, D.V., 35, 41, 54, 58, 59, 61, 107, 180, 213, 214 Hilbert Space, 13, 44 Hoffmann-Jrgensen, J., 56, 218 Holder inequality, 86 Hotelling, H., 13, 218 Huber, P.J., 169, 173, 219 Huber function, 173 Huggins, R.M., 169, 173, 212 Hutton, J.E., 32­36, 56, 61, 97, 150, 151, 162­164, 184, 191, 196, 219 Hutton-Nelson estimating function, 32­36, 150, 151, 162­164, 196 hypothesis testing, 9, 141­145 Ibragimov, I.A., 155, 219 idempotent matrix, 99, 100, 143 Iglehart, D.L., 64, 215 immigration distribution, 69, 70, 87 Ingersoll, J.E., 131, 133, 213 infection rate, 9, 162 infectives, 162, 164 infinitesimal generator, 9, 200, 201 information, 2, 7, 8, 12, 40, 41, 55, 72, 92, 108, 113, 114, 118, 126, 142, 159 empirical, 59, 204 expected, 59, 204 Fisher, 2, 12, 40, 59, 72, 97, 141 martingale, 28, 96­98, 160, 166, 167, 172 observed, 59 integration by parts, 190 intensity, 18, 34, 56, 135, 148, 165 interest rate, 133 invariance, 2, 58 Ito formula, 32, 97, 135, 194 Janˇzura, M., 179, 213 jackknife, 9, 210 Jensen inequality, 66 Jiang, J., 131, 219 Judge, G.G., 111, 219 Kabaila, P.V., 85, 87, 219 Kale, B.K., 2, 19, 211, 216 Kallianpur, G.K., 33, 148, 219 Kalman filter, 8, 103 Karlin, S., 200, 219 Karr, A., 148, 219 Karson, M.J., 130, 219 Kaufmann, H., 186, 219, 220 Keiding, N., 18, 211 kernel estimation, 147 Kessler, M., 135, 200, 201, 220 INDEX 231 Kimball, B.F., 1, 220 Kloeden, P.E., 134, 135, 151, 220 Koch, R.W., 197, 223 Kopp, P.E., 31, 213 Kronecker lemma, 190 Kronecker product, 80 K¨unsch, H., 169, 210, 220 Kulkarni, P.M., 169, 220 Kulperger, R., 85, 220 Kunita-Watanabe inequality, 50 Kutoyants, Yu., 62, 220 kurtosis, 62, 98, 104, 130, 139 Lagrange multiplier, 108, 111­113 Lahiri, S.N., 208, 220 Laird, N.M., 25, 26, 116, 122­124, 212, 215, 220 Langevin model, 131, 135 Laplace, P.S. de, 1 Laplace transform, 200 lattice, 81, 136 law of large numbers, 120, 204, 205, 207, 208 (see also martingale strong law) least squares, 1, 2, 3, 5, 7, 10, 21, 87, 161, 202­205, 209 Le Cam, L., 53, 58, 220, 221 Lee, Y., 23, 223 Legendre, A., 1 Lele, S., 37, 38, 210, 221 L´epingle, D., 187, 196, 221 Leskow, J., 148, 221 lexicographic order, 101 Li, B., 62, 142, 180, 203, 221 Liang, K.Y., 21, 25, 221, 226 lifetime distribution, 150 likelihood, 6, 8, 10, 16, 25, 35, 40, 41, 53, 54, 58, 72, 83, 91, 111, 117, 118, 122, 123, 127, 129, 133, 138, 141, 142, 165, 179, 180, 184, 202, 203 conditional, 107 constrained non-regular cases, 40, 41 partial, 107 likelihood ratio, 58, 63, 131, 134, 141, 142, 145, 203 Lin, Y.-X., 43, 88, 105, 159, 186, 218, 221 Lindeberg-Feller theorem, 4 Lindsay, B., 91, 107, 139, 222 Linnik, Yu.V., 155, 219 Liptser, R.S., 133, 222 Little, R.J.A., 127, 222 Liu, R.Y., 208, 222 Loewner optimality, 12, 20 Loewner ordering, 12, 55, 118, 144 logistic map, 37, 40 logit link function, 25 longitudinal data, 8, 25 long-range dependence, 82, 86, 158, 159 Mak, T.K., 202, 222 Markov, A.A. i, 3, 9, 93, 139, 157, 161, 162, 200, 201 Markov process, 9, 157, 162, 200, 201 Martin, R.D., 169, 172, 214, 222 martingale, 15, 17, 18, 26­28, 35, 48, 49, 51, 59, 61, 62, 69, 70, 93, 94­98, 131­133, 135, 136, 148­150, 159, 160, 162­ 167, 169­171, 176, 180, 181, 186­196, 200, 210 central limit theorem, 55, 179, 186, 190­195 continuous part, 34, 54 information (see information, mar- tingale) strong law, 55, 150, 174, 179, 181, 182, 186­190, 195, 196 maximum likelihood, 1, 2, 16­19, 21, 34, 35, 38, 39, 41, 53, 54, 57, 58, 61, 70, 92, 98, 116, 124, 129, 131, 134, 136, 141, 148, 160, 163, 165, 166, 169, 180, 200, 202 non-parametric, 17 regularity conditions, 40, 41, 54, 62 232 INDEX restricted, 141 McCormick, W.P., 19, 214 McCullagh, P., 21, 22, 125, 142, 184, 203, 205, 221, 222 McKeague, I.W., 148, 222 McLeish, D.L., 1, 3, 8, 13, 43­45, 49, 127, 128, 222, 225 measurement errors, 139 membrane potential, 33, 147 Merkouris, T., 13, 199, 223 M-estimation, 1, 142 method of moments, 1, 153, 154, 156, 169, 202 metric space, 200 minimum chi-squared, 1 minimal sufficient, 1 missing data, 8, 107, 116, 117, 127 Mitra, S.K., 131, 224 mixed normality, 27, 62, 63, 192 mixing conditions, 84, 156, 179 mixture densities, 127 models branching (see branching pro- cess) epidemic, 162­164 hidden Markov, 93 interest rate, 133 logistic, 37, 38, 40 membrane potential, 33, 147 multi-stage, 200 nearest neighbour, 91 nested, 99­102 particles in a fluid, 87 physician services, 127 population, 35­37 queueing, 157, 158 recapture, 164­168 risky asset, 31 soil moisture, 196­197 moment generating function, 199 Moore, J.B., 136, 138, 214 Moore-Penrose inverse, 30 Morton, R., 21, 23, 100­102, 107, 112, 116, 126, 203, 218, 223 Mtundu, N.D., 197, 223 multiple roots, 202­209 multiplicative errors, 100, 102 mutual quadratic characteristic, 26 Mykland, P.A., 62, 223 Naik-Nimbalkar, U.V., 103, 104, 222 Nelder, J.A., 21, 22, 125, 184, 202, 222, 223 Nelson, P.I., 32­36, 56, 61, 97, 150, 151, 162­164, 184, 191, 196, 219 nested strata, 8, 99 neurophsiological model, 32, 147 Newton-Raphson method, 202 Nguyen, H.T., 148, 223 noise, 17, 30, 31, 33, 127, 132, 133, 135, 136 additive, 36, 37 multiplicative, 36 non-ergodic models, 59, 60, 63, 144 nonparametric estimation, 147, 150 norm (Euclidean), 55, 71, 184, 186 normal distribution, 4, 27, 28, 41, 55­57, 62­66, 88, 91, 105, 111, 116, 121, 129­131, 139, 156, 161, 163, 164, 166, 173 nuisance parameter, 8, 57, 59, 60, 70, 71, 88, 107, 113­115, 139, 159 offspring distribution, 16, 27, 69, 70, 87, 182, 195 Ogunyemi, O.T., 56, 184, 196, 219 on-line signal processing, 9, 176 optimal asymptotic (OA), 1, 2, 7, 8, 11, 12, 29, 30 experimental design, 8, 12, 20 fixed sample (OF ), 1, 5, 7, 11­ 21, 28, 30, 55 orthogonal, 13, 71, 93, 100, 102­ 104, 138, 156, 160 orthonormal sequence, 147 Osborne, M.R., 120, 223 outliers, 169 parameters INDEX 233 constrained, 107­112, 142 incidental, 112 nuisance (see nuisance param- eters) scale, 200 parametric model, 1, 11 parameter space, 41, 43, 50, 53 partial likelihood, 107 Parzen, E., 84, 223 Pearson, K., 1 Pedersen, A.R., 135, 223 periodogram, 82 Pham, D.P., 148, 223 pivot, 66 Platen, E., 134, 135, 151, 220 point process, 148 (see also counting process, Poisson pro- cess) Poisson distribution, 33, 67, 70, 87, 100, 121, 167 Poisson process, 19, 31, 34, 35, 111, 112, 135, 196, 197 compound, 35, 196 Pollard, D., 56, 223 population process, 35 (see also branching process) power series family, 2, 16 Prakasa-Rao, B.L.S., 54, 132, 135, 141, 212 predictable process, 18, 31, 51, 94, 97, 162 prediction variance, 82 Pregibon, D., 202, 223 Prentice, R.L., 25, 224 Priestley, M.B., 84, 224 probability generating function, 167 projected estimating function, 108­ 110, 143 projected estimator, 13, 108, 143, 144 projection, 13, 47, 107­109, 114, 125, 129 P-S algorithm, 107, 116 Pukelsheim, F., 12, 20, 38, 224 purely nondeterministic, 153 Qian, W., 136, 224 quadratic characteristic, 27, 29, 59, 94, 132, 144, 151, 166, 187, 192, 196 quadratic form, 130 quadratic variation process, 33, 54, 133 quasi-likelihood, 1, 7­9, 12, 16, 18, 19, 21, 22, 36, 38, 40, 53, 61, 69, 70, 73, 87, 107, 116, 121, 129, 130, 131, 134, 138, 139, 142, 147, 148, 150, 151, 160­165, 169, 180­182, 195­ 197 asymptotic, 8 composite, 91, 92 quasi-score, 7­9, 12, 13, 22, 23, 26, 35, 39, 43, 46, 48, 50, 51, 55, 58, 62, 67, 69­72, 80, 82, 83, 85, 87, 92­96, 98­ 100, 102, 104, 105, 108, 109­ 112, 116­118, 122, 125, 127, 128, 130, 132, 134, 135, 137­ 139, 142, 144, 145, 154, 159, 162, 163, 166, 170­173, 175, 176, 179, 182, 195, 196 asymptotic, 26, 35, 37, 58, 69, 72--76, 78­80, 88, 154 combined, 98, 99, 103, 160 conservative, 142, 203 existence of, 15, 21, 36­38 Hutton-Nelson, 33­36, 61, 150, 151, 162­164 robust, 170­174 sub-, 48, 49 queueing models, 157, 158 Radon-Nikodym derivitive, 129, 131­ 133 Rajarshi, M.B., 103, 104, 223 random coefficient autoregression, 176 random effects model, 105 random environment, 35, 36 random field, 8, 82, 93, 136, 137, 179, 201 random norming, 8, 56, 63, 189 234 INDEX Rao, C.R., 2, 7, 20, 28, 29, 72, 73, 95, 108, 128, 129, 131, 141, 224 Rao-Blackwell theorem, 128 Rebolledo, R., 166, 224 recapture experiment, 9, 164 recursive estimation, 9, 176, 177 regression, 9, 21, 23, 25, 60, 89, 111, 113, 148, 159, 170, 173, 174, 205 REML estimation, 8, 129­131 removal rate, 164 resampling, 9, 210 residuals, 99 Reynolds, J.F., 157, 158, 224 Riemann zeta function, 126 Ripley, B.D., 137, 224 robust methods, 9, 169 Rogers, L.C.G., 26, 31, 54, 133, 149, 224 Rosenblatt, M., 82, 84, 85, 179, 224 Rotnitzky, A.G., 25, 26, 215 Rozanski, R., 148, 221 Rubin, D.B., 116, 127, 214, 222 Rubin, H., 169, 215 Samarov, A., 158, 224 sample space, 2, 43 Samuel, E., 165, 224 Schuh, H.-J., 200, 224 score function, 2, 4, 6­8, 17, 24, 41, 62, 67, 91, 93, 94, 107, 111, 113, 115­118, 122, 127, 134, 139, 141, 160, 181, 205 Scott, D.J., 63, 142, 185, 212 screening tests, 122 Selukar, R.S., 148, 219 semimartingale, 8, 9., 30­36, 54, 69, 93, 97, 131, 135, 148 semiparametric model, 1, 9 Seneta, E., 69, 87, 218 service time, 157 Severo, N., 163, 213 Shen, X., 148, 224 Shiryaev, A.N., 26, 133, 222, 224 short-range dependence, 82, 158, 159 Shumway, R.H., 127, 224 sieve, 9, 147, 148, 151 signal, 30­32, 132, 136 Silvey, S.D., 183, 211 simulation, 64 simultaneous reduction lemma, 20, 21 skewness, 62, 98, 104 Slutsky theorem, 65 Small, C.G., 1, 3, 8, 13, 43­45, 49, 127, 128, 222, 225 Smith, R.L., 54, 225 smoothed periodogram, 82 smoothing, 8, 103, 104 smoothing function, 82, 84 soil moisture, 197 Srensen, M., 34, 35, 56, 59, 135, 136, 191, 197, 200, 201, 212, 220, 225 sources of variation, 33­35, 101, 136, 137, 139 spatial process, 93 spectral density, 82, 86, 153, 155, 156, 158, 159 square root matrix, 71, 72 standardization, 3­5, 11, 71 state space models, 103 stationary process, 1, 153, 156, 158, 210 Staudte, R.G., 169, 173, 212 Stefansky, L.A., 139, 203, 205, 206, 213, 225 stochastic differential equation, 32, 49, 133, 135, 196, 197 stochastic disturbance (see noise) stochastic integral, 133, 163 Stoffer, D.S., 127, 224 stopping time, 189, 193 strata, 99, 100, 102 strong consistency, 56, 70, 135, 148, 161, 163, 181, 182, 185, 186, 190, 195 structural relationships, 112, 206 sufficiency, 2, 3, 43 (see also E-sufficient) surrogate predictors, 139 Sweeting, T.J., 60, 225 INDEX 235 Szeg¨o, G., 156, 216 Takayama, T., 111, 219 Taqqu, M.S., 86, 158, 215, 224 Taylor, H.M., 200, 219 Tchoukova-Dantcheva, S., 70, 226 Thavaneswaran, A., 18, 142, 147, 176, 177, 225 Thisted, R.A., 202, 225 Thompson, M.E., 18, 105, 116, 147, 216, 225 time series, 8, 103, 176 moving average, 156, 159, 180 (see also autoregressive pro- cess) standardized, 64 Titterington, D.M., 136, 224 Toeplitz matrix, 154 trace criterion (for optimality), 19 traffic intensity, 158 transform martingale families, 199 trapping rate, 168 Traylor, L., 54, 213 treatment, 4 t-statistic, 64, 66, 168 Tweedie, R.L., 200, 215, 22 unacceptable estimating function, 74, 75 uniform convergence, 57 uniform distribution, 41 Vajda, I., 180, 226 Verbyla, A.P., 130, 226 Vitale, R.A., 156, 226 Vostrikova, L., 62, 220 Wald, A., 9, 58, 141­143, 180, 226 Wald test, 9, 141­143 Watson, R.K., 162, 226 Wedderburn, R.W.M. i, 7, 21, 23­ 25, 101, 226 Wedderburn estimating function, 24 Wefelmeyer, W., 56, 216 Wei, C.Z., 56, 70, 87, 211, 226 Welsh, R.E., 202, 215 Whittle, P. ii, 82, 83, 86, 226 Whittle estimator, 83, 86 Williams, D., 26, 31, 54, 133, 149, 224 Williams, R.J., 97, 213 Willinger, W., 31, 213 Winnicki, J., 69, 70, 87, 226 Wong, W.H., 148, 224 Wu, C.F.J., 208, 226 Yanev, N.M., 70, 226 Yip, P., 162, 226 Yohai, V.J., 169, 172, 222 Yule-Walker equations, 23 Zeger, S.L., 21, 25, 200, 226 Zehnwirth, B., 103, 226 Zygmund, A., 157, 226