American Economic Review 2012, 102(6): 3077–3110
http://dx.doi.org/10.1257/aer.102.6.3077
3077
The Colonial Origins of Comparative Development:
An Empirical Investigation: Reply†
By Daron Acemoglu, Simon Johnson, and James A. Robinson*
Military returns [reports of disease and death] serve to indicate to the
restless wanderers of our race the boundaries which neither the pursuit
of wealth nor the dreams of ambition should induce them to pass, and to
proclaim in forcible language that man, like the elements, is controlled
by a Power which hath said, “Hither thou shalt come, but no further.”
—— Tulloch, (1847, p. 259)
In Acemoglu, Johnson, and Robinson, henceforth AJR, (2001), we advanced the
hypothesis that the mortality rates faced by Europeans in different parts of the world
after 1500 affected their willingness to establish settlements and choice of colonization
strategy. Places that were relatively healthy (for Europeans) were—when they
fell under European control—more likely to receive better economic and political
institutions. In contrast, places where European settlers were less likely to go were
more likely to have “extractive” institutions imposed. We also posited that this early
pattern of institutions has persisted over time and influences the extent and nature
of institutions in the modern world. On this basis, we proposed using estimates of
potential European settler mortality as an instrument for institutional variation in
former European colonies today.
Data on settlers themselves are unfortunately patchy—particularly because not
many went to places they believed, with good reason, to be most unhealthy. We
therefore followed the lead of Curtin (1989 and 1998) who compiled data on the
death rates faced by European soldiers in various overseas postings.1
Curtin’s
data were based on pathbreaking data collection and statistical work initiated by
the British military in the mid-nineteenth century. These data became part of the
foundation of both contemporary thinking about public health (for soldiers and for
civilians) and the life insurance industry (as actuaries and executives considered the
1 
The data are also appealing because—at the same point in time—soldiers tend to live under fairly similar conditions
in different countries, i.e., in a military cantonment or camp of some kind. Also, while conditions changed as
medical knowledge advanced, Curtin and other sources provide a great deal of detail regarding what military doctors
knew, when they knew it, and when they were able to get commanding officers to implement health-improving
reforms. Curtin (1998) is particularly good on such details.
* Acemoglu: Department of Economics, Massachusetts Institute of Technology, 50 Memorial Drive, Cambridge,
MA 02142 (e-mail: daron@mit.edu); Johnson: Sloan School of Management, Massachusetts Institute ofTechnology,
50 Memorial Drive, Cambridge, MA 02142 (e-mail: sjohnson@mit.edu); Robinson: Harvard University, 1737
Cambridge St., Cambridge, MA 02138 (e-mail: jrobinson@gov.harvard.edu). We thank Joshua Angrist, Victor
Chernozhukov, and Guido Imbens for useful suggestions, and Melissa Dell and Karti Subramanian for their help
and suggestions. We are also grateful to the staff at the library at the Institute of Actuaries in London for help with
archival material.
†
To view additional materials, visit the article page at http://dx.doi.org/10.1257/aer.102.6.3077. This paper was
earlier circulated as “Hither Thou Shalt Come, But No Further: Reply to ‘The Colonial Origins of Comparative
Development: Comment’.” NBER Working Paper w16966, April 2011.
3078 THE AMERICAN ECONOMIC REVIEW october 2012
risks inherent in overseas travel) and shaped the perceptions of Europeans—including
potential settlers and their medical advisers.2
In his Comment on AJR (2001), Albouy (2012) focuses on one part of our argument
and claims two major problems: first, our Latin American data and some of
our African data are unreliable; and second, we inappropriately mix information
from peacetime and “campaign” episodes.3
To deal with the first issue, he discards
completely almost 60 percent of our sample—effectively arguing there is no reliable
information on the mortality of Europeans in those parts of the world during
the colonial period. To deal with the second issue, he codes his own “campaign”
dummy. His Comment argues that each of these strategies separately weakens
our results and together they undermine our first-stage results sufficiently that our
instrument (potential European settler mortality) becomes completely unhelpful for
determining whether institutions affect income today.
Neither of Albouy’s claims is compelling. First, there is no justification for discarding
most of our data. Ordinary Europeans, military establishments, the medical
profession, and the rapidly developing life insurance industry collected, published,
and discussed an abundance of relevant information on European mortality rates in
many different parts of the world during the nineteenth century.4
Our original coding
for 64 modern countries and the additional robustness checks reported in AJR
(2005) draw on this information. Simply throwing out data is certainly not a reasonable
approach to deal with this wealth of information. We repeat below some of
the extensive robustness checks that were originally reported in our working paper
version, AJR (2000), and also show that the main results in AJR (2001) are robust
to incorporating existing information on mortality rates in various reasonable ways
(a point also made in AJR 2005, 2006, and 2008).
Albouy needs to discard almost 60 percent of our original sample in order to
undermine our results. And even Albouy’s preferred regression results turn out to
be largely driven by one outlier, Gambia, which has very high mortality—combined
with a relatively favorable coding of its institutions that stands at odds with its recent
history. Limiting the effect of high mortality outliers—by capping mortality at 250
per 1,000 per annum or by excluding Gambia as an extreme outlier—makes our
results robust even in Albouy’s smallest sample (i.e., with just 28 and 27 observations,
respectively).
Second, Albouy argues that military “campaigns” pushed up mortality rates
above what they would be in peacetime. However, there was little difference in
2 
We augmented the data from Curtin with estimates of bishops’ mortality from Gutierrez (1986) benchmarked
to overlapping mortality rates from Curtin. Using these approaches, we were able to compute estimates of potential
settler mortality for 72 countries. The base sample in AJR (2001) comprised 64 of these modern countries, which
also had available GDP per capita and institutional quality measures.
3 
While his current Comment differs considerably from the 2006 version (which in turn was quite different from
both the 2004 vintages), the conclusions remain the same (Albouy 2004a, 2004b, 2006). In previous work, we
rebutted those of his claims that do not appear in his AER Comment (see AJR 2005, 2006, and 2008), so this Reply
addresses only those critiques that remain in his AER Comment.
4 
The information was available in medical and public health discussions (see AJR 2005, 2006, and 2008; Hirsch
1883–1886 summarizes the state of relevant medical knowledge at the end of the nineteenth century—see online
Appendix B; much of this knowledge was built by Tulloch and his colleagues, starting earlier in the century, Tulloch
1838a, 1838b, 1838c, 1840, 1841, 1847). It was also manifest in the life insurance literature (see Institute of
Actuaries 1851–1852, and Hunter 1907). This information, together with the discussion in Curtin (1964), makes it
clear that Albouy’s claim that there is no reliable knowledge about mortality for Europeans in most of their colonial
empires is simply untenable.
3079Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
practice between the activities that soldiers were engaged in during most colonial
“campaigns” for which we have data and at other times. They marched, lived at
close quarters, and were exposed to local disease vectors—particularly mosquitoes
and contaminated water. Most of these campaigns did not involve much actual
fighting; in some instances the soldiers traveled in the (then) luxury of steamers or
on mules—or someone else carried their equipment. As a result, there appear to be
no systematic differences in mortality rates between episodes that can reliably be
classified as “campaigns” and the rest. We show that even with Albouy’s own coding,
which is problematic as we explain below, his “campaign” dummy is typically
far from significant in first- or second-stage regressions. Moreover, it is difficult or
impossible—and makes little sense—to systematically distinguish campaigns and
noncampaigns on the basis of existing information.5
In addition and perhaps more important, Albouy’s classification of campaigns
is inconsistent and seriously at odds with the historical record in at least ten cases.
His results depend on this inconsistent and inaccurate coding. When we make even
minor adjustments, it becomes even more apparent that there is no basis for his
claims. Moreover, limiting the effect of very high mortality rates restores the robustness
of our results even without correcting the inconsistencies in his coding (except
in the very smallest of his samples, where the inconsistency and inaccuracy of his
coding turn out to be somewhat more consequential).
We agree that robustness checks are important, particularly for any historical exercise
of this nature. Specifically, it is reasonable to ask: do the outliers in our data—
i.e., very high mortality rates—reflect unusual, unrepresentative spikes in mortality
that were not European experience on average, or not what they expected if they
migrated to a particular place? We emphasized and attempted to deal with any such
potential concerns in our original working paper version, AJR (2000).6
In particular,
some of our highest mortality estimates may be very high because of epidemics,
unusual idiosyncratic conditions, or small sample variation and, thus, potentially
unrepresentative of mortality rates that would ordinarily have been expected by
­soldiers or settlers. This concern was our main rationale for using the logarithm of
mortality rates—to reduce the impact of outliers (see AJR 2000, 2001).7
In AJR
(2001), we argued that very high mortality rates could be viewed as a form of
5 
Curtin, for example, mentions when some episodes involved campaigning, but this is far from providing the
basis for a systematic coding. Specifically, he does not provide detail on what soldiers were doing when not engaged
in a specific named campaign—probably because there are not good records of exactly how they spent their time or
under what exact conditions (including how much marching they did and what kind of shelter they had). In addition,
one of Curtin’s (1998) main arguments is how militaries were able to bring down campaign mortality in the late
nineteenth century, below what could be achieved if the troops stayed put (e.g., chapter 3, “The March to Kumasi”
in Curtin 1998.) We have also reviewed the historical literature on military campaigns in the nineteenth century,
including but not limited to the multivolume History of the British Army by Fortescue (1929). Most of the military
activities that generate information on mortality that is picked up in our dataset were too low level and inconsequential
to be covered in the standard histories. It is not possible to say how much of the soldiers’ time was spent
traveling rather than remaining stationary, or when they fired their weapons. The distinction that Albouy pursues is
illusory in these historical episodes.
6 
AJR (2000) contained a long list of robustness checks motivated by this and related issues, including on how to
best benchmark Latin American data to Curtin’s data (see, in particular, Table 5 there). These were not ultimately
published in AJR (2001) due to space constraints. Albouy’s initial comment on our paper did not cite AJR (2000)
and the robustness checks therein (Albouy 2004a). Though he now cites AJR (2000), there is less than full acknowledgment
that our original robustness checks dealt with many of the issues he raises.
7 
Other strategies we employed to deal with this issue in AJR (2000) included constructing alternative African
series, using information from “long” data series from Curtin. See Table 5 in AJR (2000).
3080 THE AMERICAN ECONOMIC REVIEW october 2012
­measurement error, and, provided that it did not significantly deviate from classical
measurement error, this would not create an asymptotic bias for our IV procedure.
Partly prompted by an earlier version of Albouy’s Comment,8
in AJR (2005) we
proposed the alternative strategy of capping mortality estimates at 250 per 1,000.9
The 250 per 1,000 estimate was suggested by A.M. Tulloch, the leading authority
of the day, as the maximum mortality in the most unhealthy part of the world
for Europeans (see Curtin 1990; Tulloch 1840).10
This capping strategy has several
attractive features. First, provided that settler mortality is a valid instrument,
a capped version of it is also a valid instrument.11
Second, on a priori grounds one
might expect that mortality rates above a certain level should not have much effect
on settler behavior.12
Third, it is an effective strategy for reducing the impact of
various types of measurement errors, which are likely to be present in settler mortality
data.13
Fourth, we show below that the specific level of cap we use has little
effect on the results. This strategy not only further establishes the robustness of the
results in AJR (2001) to reducing the effects of outliers, but also shows that, when
the effect of outliers is so limited, even extreme versions ofAlbouy’s other modifications
leave these results largely robust.
Albouy proposes a number of other changes to our data. None of these are consequential
for our results, but two are worth discussing—not only to dispel some of
Albouy’s claims, but also because they shed light on his approach. Albouy recodes
our Mali data, arguing that it is too high due to an epidemic. But as mentioned
above, systematic capping along these lines—to reduce the influence of outliers and
potentially unrepresentative observations in a way that is consistent across countries—considerably
strengthens our results. Albouy also objects to how we use the
8 
As we discuss further in Section IIF, the current version of Albouy’s Comment also recommends capping the
mortality rate for Mali at a lower level because he thinks the reported mortality rate is too high. In particular, in his
text, Albouy says that Mali should be reduced from 2,940 per 1,000 per annum to 478.2 per 1,000; but in his dataset
(online, last checked September 19, 2011) he appears to cap it at 280 per 1,000. But he does not do this for other
similarly high mortality rates. For example, he does not apply the same cap for Gambia—even though the very high
mortality rate (1,470 per 1,000 p.a.) there was presumably due to an epidemic, as was the case in Mali. As discussed
above, Gambia is a significant outlier that matters a great deal for Albouy’s hypothesis.
9 
We follow Curtin and the nineteenth century literature by reporting mortality per 1,000 mean strength (also
referred to as “with replacement” ), meaning that the mortality rate refers to the number of soldiers who would have
died in a year if a force of 1,000 had been maintained in place for the entire year.
10 
Note that 250 per 1,000 is still a dauntingly high mortality rate. Potential settlers were definitely deterred
by the prospect that about a quarter of their number would die within the first year, even if there was not a major
epidemic—for example, of yellow fever. After early attempts ended in tragedy for would-be settlers, Europeans
viewed much of Africa as the “White Man’s Grave” and did not seriously attempt to build settlements there. Curtin
(1964, p. 86), for example, writes: “It was known in any case that West Africa was much more dangerous than the
West Indies. The best medical opinion was, indeed, opposed to the kind of establishments that already existed there.
James Lind [in Diseases in Hot Countries in 1768] argued that European garrisons for the West African posts should
be reduced to the smallest possible numbers and moved to hulks anchored off shore.”
11 
The AJR (2001) assumption is that potential settler mortality is orthogonal to the second-stage error term. If
so, any monotone transformation thereof would also be orthogonal to this error term and, thus, a valid instrument.
12 
For example, it is reasonable to presume that no Europeans, perhaps except the very small group of the most
reckless adventurers, would have considered settling in places where they faced probabilities of death above 25
percent in the first year.
13 
Trimming (dropping observations with extreme values) and winsorizing (capping extreme values) are common
strategies for dealing with outliers, potentially contaminated data, as well as classical and nonclassical measurement
error (see, e.g., Angrist and Krueger 1999).A large literature in statistics investigates alternative strategies
for reducing the influence of outliers, measurement error, and contaminated data. A particularly popular strategy is
to reduce (or bound) the influence of outlying and less reliable observations (e.g., Huber 1964; Rousseeuw 1984).
Capping—winsorizing—is one way of achieving this (e.g., Wilcox 2001; Andersen 2008). We discuss how trimming
affects our results in Section IIA.
3081Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
term “French Soudan” and claims that we have made a mistake in our coding. This
is not only inconsequential, but our use of this term was quite explicitly explained in
Appendix B of AJR (2000) and again in AJR (2005).14
Overall, Albouy’s Comment amounts to a series of objections to our approach.
All of these objections, upon closer inspection, are far from compelling, are often
unfounded, and prove minor and largely inconsequential for the robustness of our
results. The big picture from AJR (2001) remains intact and remarkably robust:
Europeans were more likely to move to places that were relatively healthy, and when
they moved in larger numbers, they imposed better institutions, which have tended
to persist from the colonial period to today.
The rest of the paper is organized as follows. Section I reviews our hypothesis
and data we used to test it; we also briefly present our original results across a range
of specifications. Section II considers Albouy’s concerns in detail. Section III concludes.
Online Appendices A, B, and C, which are available on our websites (linked
to the AER website), provide further details on econometrics, data for contested
observations, and historical background.
I. Background
A. Theory and Data
The main focus of AJR (2001) was to estimate the causal effect of a broad cluster
of institutions on long-run development. We argued that there were various types
of colonization policies which created different sets of institutions. At one extreme,
European powers set up “extractive states,” which introduced neither any significant
protection for private property nor any checks against expropriation. In these cases,
the main purpose of the extractive state was to transfer resources of the colony
to the colonizer. At the other extreme, Europeans settled in a number of colonies
and settlers tried to replicate or extend European institutions, with great emphasis
on private property and checks against government and elite power. These colonial
institutions have tended to persist. This choice of colonization strategy was in turn
naturally influenced by the feasibility of settlements—in places where the mortality
rate from disease for Europeans was relatively high, the odds were against the
creation of settler colonies with better institutions, and the formation of an extractive
state was more likely. Based on this reasoning, we suggested that the potential mortality
rates expected by early European settlers in the colonies could be an instrument
for current institutions in these countries.
Of course, by its nature, potential settler mortality is often not observed.15
In
places where the potential settler mortality was high, large numbers of settlers did
not go, and it is difficult to obtain comparable measures of their mortality. Moreover,
14 
The Appendix B to AJR (2000) was not only explicit that the term “French Soudan” refers to a large area of
Western and Central Africa but also clearly named the countries this covered (see Section IIF for details). Albouy
prefers another definition of French Soudan, which is limited to a small part of West Africa. “Bilad al-Sudan,” from
which the term Sudan originates, means “the land of black people” in Arabic.
15 
Albouy still complains that data do not come from actual settlers. But AJR (2000, 2001) were very clear that
these were potential settler mortality rates, and of course, Europeans did not settle and should not have settled in
places where the annual mortality rates run in the range of 20 percent or higher—and in places where they perceived
mortality to be in this range. The available evidence not only enables us to have a fairly good idea about which parts
3082 THE AMERICAN ECONOMIC REVIEW october 2012
in the critical early periods for settlements and institutional development, data on
mortality rates of European settlers are sometimes hard to find—and we should
worry about whether these groups were demographically similar (e.g., in terms of
age structure or social background). Our strategy was therefore to use a homogeneous
group of Europeans in these colonies to form an estimate of settler mortality
rates. This strategy was made possible by the fact that Philip Curtin, in a series of
works (Curtin 1989 and 1998, but also Curtin 1961, 1964), reported comparable
data on the disease mortality rates of European soldiers stationed in various colonies.
Curtin also took a view on how Europeans perceived mortality in various parts
of the world and discussed how this view was shaped by the available data over time.
As a practical matter our approach was straightforward. We began with Table 1.1
of Curtin (1989), which is entitled, “Mortality of European Troops Overseas,
1817–38.” This is a summary of Curtin’s base data from around the world. Curtin’s
book is focused on the relocation costs for Europeans, i.e., exactly the issue we are
interested in, and he is very careful with data, so it made sense to take these estimates
without any editing or selectivity. Note that while these data are for soldiers,
for whom there is always likely to be some military activity (marching, engaging in
exercises, traveling on ships, etc.), they are definitely not from major wars involving
mass armies and large-scale casualties. Curtin (1989, 1998) emphasized that mortality
rates declined through the nineteenth century as European militaries became better
at managing health issues. In particular, after 1850 there were dramatic declines
in military mortality from disease in the tropics (see, e.g., the contrast between
Tables 1.1 and 1.2 in Curtin 1989). Curtin’s work therefore focused our attention on
taking the earliest possible data from periods without major wars (preferably before
1850), and we tried to stick to this throughout.16
While Curtin’s Table 1.1 spanned most of the world, it did not report specific estimates
for all countries. We therefore adopted the following coding rule. In each case
we took the estimate from Table 1.1 if available. We then took the earliest number
from Curtin when such data were available—excluding episodes with significant
wars.17
The mortality estimates came from Curtin (1989) or, if nothing relevant was in
that source, from Curtin (1998). In addition, if it was likely on the basis of other information
that Europeans faced similar mortality rates in two countries but only one of
them had an estimate, we assigned the mortality rate from one country to the other.18
of the globe had such very high mortality rates, but also establishes that Europeans were at the time well aware of
these mortality rates.
16 
From the perspective of our theoretical framework, we really needed potential settler mortality before 1800—
during the formative period of colonization for most of these places. But such data are not generally available, and
in his estimates before 1850 Curtin offered data from before the improvement in European public health management
(both in general and for the tropics in particular).
17 
Some of these data are from expeditions—but these were a common occurrence in the colonial era. An expedition
is a group of men, often soldiers, traveling together for a particular purpose. This could be exploration,
to open trade routes, to demonstrate force against a local ruler, or some combination of these activities. Curtin
(1998) reviews data from a number of these experiences in Africa, including against the Ashanti and in Ethiopia.
These were all small expeditions and none of them get more than a passing mention in the multivolume History
of the British Army by Fortescue (1929); they were not seen as significant wars by the British military or its historians.
Examples of significant conflicts from which we did not use data include the siege of Cartegena in 1742
(Curtin 1989) and the invasion of Egypt in 1882 (Curtin 1998).
18 
In constructing our dataset we preferred simplicity and transparency. Albouy contends that we do not have
any information about countries to which mortality is “assigned” from neighbors. This is incorrect, as we showed
in our earlier replies (AJR 2005, 2006, 2008). We summarize this additional information in Section II. Online
Appendix B discusses the various sources available to evaluate whether the disease ecologies are sufficiently similar
3083Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
InAJR (2000), we provided a detailed analysis of an alternative series without this
type of assignment and also some other robustness checks (see, e.g., Table 5, columns
1–4). Since we followed this coding rule rather than make arbitrary judgment
calls, some of the mortality rates in West Africa were extremely high, especially
when the soldiers encountered a yellow fever epidemic (though other mortality estimates,
such as for Ethiopia, were very low). Our use of logarithm of mortality rates
was in part motivated by these very high mortality rates.
The most important gap in Curtin’s data is for Latin America. Curtin reported
estimates for the Caribbean, but in Central and South America, his work contained
estimates only for Mexico.19
To supplement the numbers from Curtin, we used an
article by Gutierrez (1986) on the mortality rates of bishops in Latin America (i.e.,
Central and South America, including some data on the Caribbean).20
Naturally, the
mortality rates of bishops and soldiers were unlikely to be the same: bishops presumably
resided in more comfortable and sanitary conditions than soldiers in barracks;
they could escape epidemics more easily; and overall they must have had a much
higher standard of living. When the series overlap, the Gutierrez mortality estimates
are lower than the Curtin estimates. To create a comparable series, we therefore
benchmarked the mortality rates of bishops to those of soldiers.21
Gutierrez provides
an estimate for Mexico (for which we had a Curtin estimate) and also for the
Dominican Republic, which we assumed had a similar mortality rates to Jamaica
(again, for which there is a Curtin estimate). Since we had two points of overlap,
we could benchmark using either number, or some combination of the numbers.
We decided to use the Mexican number, which was lower and therefore reduced the
mortality rates in Latin America—making for mortality estimates that were more
plausible, given the available qualitative evidence.22
In AJR (2000), these issues
were extensively discussed, and we reported that our results were robust using either
type of benchmarking (see again below).
B. Baseline Results
The first-stage relationship in AJR (2001) is the link between settler mortality,
in logs, and a measure of institutions. Here we focus on our main measure of
institutions, which is protection against the risk of expropriation. This is an OLS
regression, with one observation per country.
to ­reasonably assign a mortality rate from one country to its modern neighbor. We use the historical and contemporary
literature on historical geography, in the form of both text and maps. We also include the relevant medical
literature—because this speaks to the issue of conditions under which some diseases, such as malaria, become
prevalent. The nineteenth century literature on life insurance is also helpful on some key points.
19 
There was a reference in Curtin (1989) to an English attack on Cartegena in 1742. But the Gutierrez data for
Colombia are for Bogota, and there is good reason to think this was not as unhealthy for Europeans as the Caribbean
coast, so Curtin’s information on Cartegena did not help us merge the Gutierrez and Curtin series.
20 
Specifically, we used data on bishops aged 40–49. Many of these bishops were born in Europe, so they would
not have an acquired or inherited immunity to local diseases.
21 
Namely, we combined the two series by using Gutierrez’s relative mortality rates for bishops to impute mortality
levels that are consistent with Curtin’s data. This lets us calculate levels for Latin America.
22 
This choice seemed less favorable to our hypothesis and thus also preferable on these grounds. Our checks
using the Dominican Republic/Jamaica number indicated slightly stronger results for us. Also using rates from
Mexico in benchmarking the Gutierrez/Curtin series does not involve any assignment of mortality to neighbors.
See Section II below for the alternative results, using the Dominican Republic/Jamaica for the benchmarking, in
Tables 1A and 1B (columns 3 and 4).
3084 THE AMERICAN ECONOMIC REVIEW october 2012
For the sake of brevity, Table 1A is structured to show results only for the log
mortality variable. Each set of rows shows a different specification, with covariates
and alternative samples that were presented in AJR (2001). The first set of rows has
no additional covariates in the regression, the second set of rows includes latitude,
the third set drops the neo-Europes (the USA, Canada,Australia, and New Zealand),
the fourth set drops all of Africa, the fifth set includes continent dummies, the sixth
set includes continent dummies and latitude, the seventh set includes the percent
of the population in 1975 that was of European descent, and the eighth set of rows
includes malaria. These are the specifications which Albouy also discusses—our
rows match the columns in his Tables 2 and 3, with the exception that we also report
results without any African data.23
We should note that as discussed in AJR (2000, 2001), the specification in the last
row that includes current prevalence of malaria is highly problematic—and likely to
bias results against finding both a significant first-stage and second-stage relationship—because
the current prevalence of malaria is endogenous, generally driven
by institutional and income per capita differences. We included this specification in
AJR (2000, 2001) for completeness but emphasized the potential bias that it would
create against our hypothesis was a serious concern.24
As a matter of fact, this was
the least robust specification in AJR (2001) as the results in Table 1A and Table
1B here also show. In what follows, unsurprisingly, this will be the specification in
which Albouy’s strategies sometimes lead to less robust results.
For each set of rows we show five numbers: the coefficient on log settler mortality,
the homoskedastic standard error, the clustered standard error, the number of clusters,
and the number of observations. The number of clusters is less than the number
of observations because about half the potential settler mortality estimates in the
AJR (2001) sample are inferred from mortality rates in neighboring countries.25
Table 1A begins with first-stage results using the original AJR data (column 1),
corresponding to column 1 of Table 4 of AJR (2001).26
The coefficient is −0.61 and
the standard error is 0.13; when we cluster the standard error, it rises to 0.17, and the
coefficient remains highly significant.27
Across the broad range of other specifications in Table 1A our first-stage results
are similar. The parameter point estimate does not move much across rows. When
we drop the neo-Europes, the estimated coefficient is smaller, but the standard
23 
We drop the African data because in an earlier comment Albouy did the same.
24 
In particular, we wrote: “Since malaria was one of the main causes of settler mortality, our estimate may be
capturing the direct effect of malaria on economic performance. We are skeptical of this argument since malaria
prevalence is highly endogenous; it is the poorer countries with worse institutions that have been unable to eradicate
malaria” (p. 1,391). We also provided examples of richer countries with better institutions successfully eradicating
malaria, including the United States eliminating it from the Panama Canal zone and Australians from Queensland.
Acemoglu and Johnson (2007) provide additional evidence that differences in malaria prevalence today are unlikely
to account for significant differences in income per capita across countries. In addition, Albouy uses a malaria variable
which is different from the one in AJR (2000, 2001) and the provenance of which is unclear. In what follows,
we consistently use the original data from AJR (2000, 2001).
25 
Such clustering may be viewed as somewhat conservative since we have quantitative and qualitative corroborating
evidence from other sources on mortality rates on all the countries in our sample (for example, from the
literature on life insurance, part of which was discussed in AJR 2005).
26 
This matches column 9 in Table 3 of AJR (2001).
27 
In the original AJR series, we used the relative rates of 1, 1.1, and 2.3 between the Gutierrez regions. This was
based on an approximate formula that converted Gutierrez’s mortality rates into mortality rates “with replacement”
comparable with the base data from Curtin. In Appendix 2 of AJR (2005) we showed that the exact ratios should
be 1, 1.1, and 2.2. This does not make any difference, within two significant figures, to our results in column 1.
3085Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
Table 1A—First-Stage Regressions, Alternative Mortality Series
Alternative series for settler mortality
Original
AJR
series
Original
AJR series,
capped
at 250
Benchmarking
to
Caribbean
Benchmarking
to Caribbean,
capped
at 250
Using
naval
stations,
method 1
Using naval
stations,
method 1,
capped at 250
Using
naval
stations,
method 2
Using naval
stations,
method 2,
capped at 250
Dependent variable is protection against risk of expropriation
(1) (2) (3) (4) (5) (6) (7) (8)
No covariates −0.61 −0.94 −0.59 −0.91 −0.54 −0.77 −0.58 −0.88
(standard error) (0.15) (0.16) (0.15) (0.17) (0.14) (0.16) (0.15) (0.16)
(clustered standard error) (0.17) (0.18) (0.17) (0.18) (0.16) (0.18) (0.17) (0.17)
Number of clusters 36 36 36 36 35 35 36 36
Observations 64 64 64 64 53 53 64 64
With latitude −0.52 −0.86 −0.5 −0.83 −0.43 −0.66 −0.49 −0.79
(standard error) (0.17) (0.18) (0.16) (0.19) (0.15) (0.19) (0.16) (0.18)
(clustered standard error) (0.19) (0.20) (0.19) (0.20) (0.17) (0.20) (0.18) (0.20)
Number of clusters 36 36 36 36 35 35 36 36
Observations 64 64 64 64 53 53 64 64
Without neo-Europes −0.4 −0.66 −0.38 −0.64 −0.35 −0.52 −0.38 −0.61
(standard error) (0.15) (0.18) (0.15) (0.18) (0.13) (0.16) (0.14) (0.18)
(clustered standard error) (0.17) (0.20) (0.16) (0.19) (0.14) (0.16) (0.16) (0.18)
Number of clusters 33 33 33 33 32 32 33 33
Observations 60 60 60 60 49 49 60 60
Without Africa −1.21 −1.21 −1.23 −1.23 −0.82 −0.82 −1.11 −1.11
(standard error) (0.20) (0.20) (0.23) (0.23) (0.28) (0.28) (0.22) (0.22)
(clustered standard error) (0.18) (0.18) (0.21) (0.21) (0.29) (0.29) (0.20) (0.20)
Number of clusters 19 19 19 19 18 18 19 19
Observations 37 37 37 37 26 26 37 37
With continent dummies −0.44 −0.81 −0.42 −0.78 −0.32 −0.56 −0.41 −0.73
(standard error) (0.19) (0.25) (0.19) (0.25) (0.17) (0.21) (0.18) (0.23)
(clustered standard error) (0.20) (0.25) (0.20) (0.25) (0.18) (0.21) (0.19) (0.22)
Number of clusters 36 36 36 36 35 35 36 36
Observations 64 64 64 64 53 53 64 64
With continent dummies
  and latitude
−0.35 −0.72 −0.33 −0.68 −0.25 −0.46 −0.33 −0.63
(standard error) (0.20) (0.26) (0.20) (0.25) (0.18) (0.23) (0.19) (0.24)
(clustered standard error) (0.21) (0.26) (0.21) (0.25) (0.20) (0.22) (0.20) (0.23)
Number of clusters 36 36 36 36 35 35 36 36
Observations 64 64 64 64 53 53 64 64
With percent of European
  descent in 1975
−0.42 −0.73 −0.39 −0.7 −0.31 −0.5 −0.39 −0.67
(standard error) (0.17) (0.19) (0.17) (0.20) (0.17) (0.21) (0.16) (0.19)
(clustered standard error) (0.19) (0.20) (0.19) (0.21) (0.19) (0.24) (0.19) (0.21)
Number of clusters 36 36 36 36 35 35 36 36
Observations 64 64 64 64 53 53 64 64
With malaria −0.43 −0.81 −0.4 −0.8 −0.13 −0.39 −0.39 −0.74
(standard error) (0.21) (0.22) (0.22) (0.24) (0.20) (0.22) (0.21) (0.22)
(clustered standard error) (0.24) (0.24) (0.25) (0.26) (0.21) (0.24) (0.23) (0.23)
Number of clusters 35 35 35 35 34 34 35 35
Observations 62 62 62 62 51 51 62 62
Notes: OLS regressions, with one observation per country. Coefficients and standard errors for covariates, where
included, are not reported to save space. All variables are from AJR (2001). Dependent variable is protection
against risk of expropriation; independent variable is log settler mortality. Column 1 uses original settler mortality
series from AJR (2001) as independent variable. Column 2 uses original settler mortality series, capped at 250
per 1,000 per annum. Column 3 uses alternative settler mortality series, benchmarking Latin American data to
Jamaica/Dominican Republic. Column 4 uses same series as column 3, but capped at 250 per 1,000. Column 5
uses mortality data directly from naval stations, without benchmarking. Column 6 uses same series as column 5, but
capped at 250 per 1,000. Column 7 uses mortality data from naval stations, with life insurance data, without benchmarking.
Column 8 uses same series as column 7, but capped at 250 per 1,000.
3086 THE AMERICAN ECONOMIC REVIEW october 2012
Table 1B—Second-Stage Regressions, Alternative Mortality Series
Alternative series for settler mortality
Original
AJR
series
Original
AJR series,
capped
at 250
Benchmarking
to
Caribbean
Benchmarking
to Caribbean,
capped
at 250
Using
naval
stations,
method 1
Using naval stations,
method 1,
capped at 250
Using
naval
stations,
method 2
Using naval
stations,
method 2,
capped at 250
Dependent variable is log GDP per capita in 1995
(1) (2) (3) (4) (5) (6) (7) (8)
No covariates 0.93 0.82 0.96 0.85 1.01 0.94 0.98 0.86
AR confidence
 interval
[0.68,  1.40] [0.62,  1.14] [0.71,  1.47] [0.64, 1.20] [0.74,  1.63] [0.70, 1.40] [0.72, 1.50] [0.65, 1.23]
AR confidence
  interval, clustered
[0.66, 1.72] [0.60, 1.19] [0.69, 1.85] [0.63, 1.29] [0.72, 1.90] [0.69, 1.50] [0.70, 1.85] [0.64, 1.31]
F-stat, first stage 23.34 35.55 22.06 33.53 18.26 24.52 21.95 32.36
F-stat, first stage,
 clustered
12.45 28.09 11.72 25.31 11.96 19.00 12.05 25.11
With latitude 0.96 0.79 1.01 0.85 1.07 0.96 1.03 0.87
AR confidence
 interval
[0.65, 1.78] [0.55, 1.24] [0.68, 1.94] [0.59, 1.36] [0.70, 2.44] [0.65, 1.79] [0.70, 1.99] [0.61, 1.41]
AR confidence
  interval, clustered
[0.64, 2.49] [0.55, 1.20] [0.68, 2.86] [0.59, 1.35] [0.71, 3.44] [0.67, 1.82] [0.69, 2.90] [0.62, 1.40]
F-stat, first stage 13.48 21.82 12.67 20.37 9.66 13.44 12.52 19.46
F-stat, first stage,
 clustered
7.30 19.26 6.89 17.14 6.10 10.37 6.93 16.32
Without neo-Europes 1.24 1.04 1.30 1.11 1.32 1.20 1.31 1.13
AR confidence
 interval
[0.78, 3.09] [0.67, 1.99] [0.82, 3.35] [0.73, 2.18] [0.82, 3.81] [0.77, 2.74] [0.83, 3.37] [0.74 2.25]
AR confidence
  interval, clustered
[0.76, 5.43] [0.65, 2.10] [0.78, 5.97] [0.70, 2.35] [0.83, 4.72] [0.78, 2.61] [0.80, 5.60] [0.72, 2.36]
F-stat, first stage 8.89 13.22 8.61 12.74 7.77 10.16 8.70 12.46
F-stat, first stage,
 clustered
5.54 11.27 5.43 10.77 6.19 10.38 5.64 11.09
Without Africa 0.61 0.61 0.65 0.64 0.93 0.93 0.68 0.67
AR confidence
 interval
[0.41, 0.87] [0.41, 0.87] [0.45, 0.94] [0.44, 0.94] [0.59, 2.26] [0.59, 2.26] [0.47, 1.01] [0.47, 1.01]
AR confidence
  interval, clustered
[0.44, 0.85] [0.45, 0.85] [0.46, 0.94] [0.47, 0.94] [0.57, 2.32] [0.57, 2.32] [0.48, 0.99] [0.48, 0.99]
F-stat, first stage 30.62 30.62 27.62 27.62 8.64 8.64 24.26 24.26
F-stat, first stage,
 clustered
45.98 45.98 36.16 36.16 8.16 8.16 32.41 32.41
With continent
 dummies
0.97 0.78 1.00 0.81 1.21 0.96 1.04 0.84
AR confidence
 interval
[0.59, 3.20] [0.52, 1.42] [0.60, 3.95] [0.52, 1.53] [−∞, −9.76]
⋃ [0.64, ∞]
[0.56, 3.51] [0.63, 4.02] [0.55, 1.64]
AR confidence
  interval, clustered
[0.52, 4.87] [0.45, 1.43] [0.55, 6.14] [0.46, 1.51] [−∞, −20.92]
⋃ [0.63, ∞]
[0.53, 2.14] [0.58, 4.97] [0.49, 1.52]
F-stat, first stage 6.49 13.32 5.89 12.10 3.34 6.22 5.96 11.59
F-stat, first stage,
 clustered
4.68 10.61 4.42 10.03 3.20 7.35 4.79 10.90
With continent
  dummies and
 latitude
1.07 0.80 1.12 0.84 1.39 1.04 1.17 0.88
AR confidence
 interval
[−∞, −27.23]
⋃ [0.57, ∞]
[0.48, 1.93][−∞, −9.26]
⋃ [0.59, ∞]
[0.49, 2.22] [−∞, −1.86]
⋃ [0.63, ∞]
[−∞, −44.57]
⋃ [0.53, ∞]
[−∞, −10.11]
⋃ [0.61, ∞]
[0.52, 2.54]
AR confidence
  interval, clustered
[−∞, −4.72]
⋃ [0.44, ∞]
[0.30, 1.53][−∞, −3.62]
⋃ [0.47, ∞]
[0.32, 1.64] [−∞, −1.57]
⋃ [0.58, ∞]
[0.47, 4.79] [−∞, −4.94]
⋃ [0.49, ∞]
[0.39, 1.72]
F-stat, first stage 3.71 8.52 3.36 7.67 1.87 3.80 3.37 7.25
F-stat, first stage,
 clustered
2.72 7.74 2.52 7.38 1.57 4.25 2.66 7.83
(Continued)
3087Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
error is also reduced. Without Africa, the results become significantly stronger.
Table 1B shows the equivalent second-stage results, in which we regress log GDP
per capita in 1995 on institutions, with log settler mortality as the instrument.
In AJR (2001), we followed standard practice at the time and reported standard
errors. Here we instead report the Anderson-Rubin (AR) 95 percent confidence
set (allowing for nonspherical error structure due to clustering and heteroskedasticity),
which is consistent when the first stage may be weak (see, for example,
Chernozhukov and Hansen 2005).28
For the baseline estimate with clustering, this
confidence set has a lower bound of 0.66 and an upper bound of 1.72, around a
point estimate of 0.93.
The two exceptions are the specification with continent dummies and latitude and
the one with malaria. In the former case, with clustered standard errors, the coefficient
28 
We do this mostly because Albouy has emphasized the importance of using Anderson and Rubin (1949) confidence
sets and reports only these in his comment. In fact, since there is only one endogenous regressor and one instrument,
these make little difference relative to the more standard Wald confidence intervals that also allow nonspherical
errors.
Our AR confidence intervals do not always match those reported by Albouy. This seems to be a consequence of
his use of an insufficiently fine grid. Our procedure is described in online Appendix A.
Table 1B—Second-Stage Regressions, Alternative Mortality Series (Continued)
Alternative series for settler mortality
Original
AJR
series
Original
AJR series,
capped
at 250
Benchmarking
to
Caribbean
Benchmarking
to Caribbean,
capped
at 250
Using
naval
stations,
method 1
Using naval
stations,
method 1,
capped at 250
Using
naval
stations,
method 2
Using naval
stations,
method 2,
capped at 250
Dependent variable is log GDP per capita in 1995
(1) (2) (3) (4) (5) (6) (7) (8)
With percent of
 European
  descent in 1975
0.92 0.71 0.99 0.77 1.23 1.03 1.02 0.79
AR confidence
 interval
[0.55, 2.31] [0.44, 1.27] [0.59, 2.92] [0.48, 1.47] [0.66, 30.44] [0.58, 4.05] [0.61, 3.13] [0.49, 1.56]
AR confidence
 interval,
 clustered
[0.54, 4.32] [0.37, 1.21] [0.57, 9.08] [0.42, 1.42] [−∞, −4.29]
⋃ [0.66, ∞]
[0.56, 6.22] [0.58, 9.67] [0.44, 1.45]
F-stat, first stage 8.67 15.32 7.45 13.27 4.17 6.12 7.19 12.38
F-stat, first stage,
 clustered
4.92 12.92 4.20 10.60 2.61 4.44 4.17 10.30
With malaria 0.67 0.52 0.74 0.56 2.03 1.08 0.79 0.61
AR confidence
 interval
[0.29, 2.93] [0.27, 0.95] [0.32, 10.24] [0.29, 1.09] [−∞, −0.36]
⋃ [0.54, ∞]
[−∞, −2.46]
⋃ [0.45, ∞]
[0.37, 8.59] [0.33, 1.21]
AR confidence
  interval, clustered
[−∞, −3.75]
⋃ [0.25, ∞]
[0.23, 0.89][−∞, −1.46]
⋃ [0.28, ∞]
[0.25, 1.06] [−∞, −0.37]
⋃ [0.62, ∞]
[−∞, −3.58]
⋃ [0.48, ∞]
[−∞, −2.71]
⋃ [0.34, ∞]
[0.30, 1.12]
F-stat, first stage 5.38 13.95 4.27 11.90 0.46 2.45 4.41 11.43
F-stat, first stage,
 clustered
3.11 11.45 2.50 9.18 0.41 2.68 2.77 10.00
Notes: 2SLS regressions, with one observation per country, corresponding to first-stage regressions in Table 1A.
Coefficients and standard errors for covariates, where included, are not reported to save space.All variables are from
AJR (2001). Dependent variable is log GDP per capita in 1995. Right-hand side variable is protection against expropriation,
instrumented by log settler mortality. Column 1 uses original settler mortality series from AJR (2001) as
the instrument. Column 2 uses original settler mortality series, capped at 250 per 1,000 per annum. Column 3 uses
alternative settler mortality series, benchmarking Latin American data to Jamaica/Dominican Republic. Column 4
uses same series as column 3, but capped at 250 per 1,000. Column 5 uses mortality data directly from naval stations,
without benchmarking. Column 6 uses same series as column 5, but capped at 250 per 1,000. Column 7 uses
mortality data from naval stations, with life insurance data, without benchmarking. Column 8 uses same series as
column 7, but capped at 250 per 1,000.
3088 THE AMERICAN ECONOMIC REVIEW october 2012
on settler mortality is −0.35 and the standard error is 0.19 in the first stage. In the
second stage, the AR confidence interval is the union of two disjoint and unbounded
intervals: [−∞, −4.72] and [0.44+, ∞] (or [−∞, −27.23] and [0.57, +∞] without
clustering).As also suggested by Chernozhukov and Hansen (2005), the lower interval
is irrelevant: not only does it not even include the point estimate, 1.07, but such
large negative estimates make neither economic nor econometric sense. Therefore,
we interpret this as evidence that the 95 percent confidence set excludes zero and
reasonable negative estimates, allowing us to statistically reject the hypothesis that
institutions have no effect on GDP per capita. To be sure, such a confidence interval
is still a sign of imprecise estimates and cause for concern, since it is much wider
than the confidence sets in our other specifications—though it still enables us to
reject the hypothesis that the second-stage coefficient is zero. The pattern is similar
with malaria—the confidence set consists of two disjoint intervals but still rejects a
zero coefficient.
In summary, the different specifications in column 1 of Table 1B confirm the
results in AJR (2001) that institutions have a significant positive effect on income
per capita, though in specifications that include continent dummies and latitude
together or malaria, the confidence sets are quite wide.
II.  Are the AJR (2001) Results Fragile?
A. Concerns about Very High Mortality Rates
As we noted in AJR (2000), some of the data, particularly from Africa, may have
had excessively high mortality rates. In the Appendix to AJR (2000), we discussed
the source of these data and flagged clearly when they were due to epidemics. If
epidemics occurred with some regularity—or if they were rare and yet still affected
European perceptions of mortality for settlers—such mortality data are highly relevant.
But if these epidemics were one-off or seen as rare, then they introduce additional,
perhaps significant measurement error. We first investigate whether extreme
mortality rates and outliers drive the results inAJR (2001). Our strategy, as explained
in the introduction, is to cap the settler mortality rates so as to reduce the influence
of extreme observations. This strategy not only helps us limit the influence of very
high mortality rates (cf. footnote 13), but is also reasonable on a priori grounds,
since we may expect that European settlement behavior was not very sensitive to
variations in mortality rates above a certain threshold.
In column 2 of Table 1A we show the effects of capping mortality at 250 per
1,000 per annum. This is the rate that Tulloch, the pioneer in this area, estimated
to be average European soldier mortality rate “for West Africa in general” from
1792 through 1840 (Curtin 1990, p. 67; see Tulloch 1840, p. 7).29
Tulloch and his
colleagues also regarded that region as the most unhealthy part of the world for
Europeans in the early nineteenth century. And of course, 250 per 1,000 per annum
is still a very high mortality rate, sufficient to discourage anybody but the most
29 
This rate of 250 per 1,000 is also close to the rate of 209 per 1,000 per annum for officers stationed in Sierra
Leone and Cape Coast Command, 1819–1836, in Balfour (1849); ordinary soldiers had a higher death rate.
3089Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
­reckless from ­permanent settlement (see footnote 10). In our data, there are 13 countries
with mortality rates above 250 per 1,000 per annum.
The use of this cap is also consistent with the assessment of Curtin, the leading
historian of comparative colonial mortality rates. At the start of Curtin’s 1998 book,
Disease and Empire: The Health of European Troops in the Conquest of Africa, he
says of West Africa: “At that time [before the 1860s], the disease mortality of newly
arrived Europeans was higher than it was anywhere else in the world. It was so high,
in fact, that one historical problem is to discover why people were willing to go to
a place where the probability of death was about 50 percent in the first year and 25
percent a year thereafter” (p. 1).
In column 2, for the base specification in the first set of rows, the coefficient on
log settler mortality in the first stage increases in absolute value to −0.94 (compared
with −0.61 in column 1), while the clustered standard error increases from 0.17 (in
column 1) to 0.18. There is a similar pattern in all other rows, except the row without
Africa (as the capping only affects African rates).30
Now in all cases, the AR
confidence sets for the second stage are much more precisely estimated and never
extend to infinity and always exclude zero. These results are not specific to capping
the potential settler mortality rate at 250—they apply with a range of reasonable
caps as we show next.
Online Appendix Table 1 reports results using alternative caps. Columns 1 and
2 repeat the results with the original data from AJR (2001) and with the 250 cap.
Columns 3–5 show that the results are very similar if we instead cap settler mortality
at 150, 280, or 350 (see also column 6 of Table 1B in AJR 2005). The rate of 150 per
1,000 is suggested by our reading of the contemporary life insurance literature.31
A
cap of 350 reflects the fact that, as the Curtin quote above illustrates, mortality for
newly arrived Europeans could be higher than for people who survived the first year.
The rate of 280 is another obvious alternative, since 5 of the 13 countries above 250
in our data are at 280. The results in each case are very similar to those in column 2
and generally considerably more stable than those in column 1, where outliers play
a more consequential role.
As discussed in footnote 13, one might also attempt to deal with concerns related
to unrepresentative data and measurement error by trimming the outliers rather than
capping them. In this context, capping is a more attractive strategy both because the
sample is already small, and because, as we discuss below, a wealth of other information
suggests that places with very high mortality rates in our sample are indeed those
with the highest mortality rates for Europeans—even though, most probably, these
rates did not exceed 20 percent or 25 percent mortality by much. Nevertheless, for
completeness, the remaining columns of online Appendix Table 1 show the implications
of trimming the very high mortality rates. In particular, column 6 drops observations
with mortality above 250, and columns 7 and 8 do the same for 150 and 350,
30 
Note that a few of the highest mortality rates in AJR (2001) were used in the raw form reported in Curtin and
are not “with replacement” rates. Capping mortality rates means that this definitely does not matter—with or without
replacement, these rates would be above the level of the cap.
31 
The Institute of Actuaries (1851–1852) provides comparative data that are broadly consistent with our original
estimates (see online Appendix B). But our reading of this source suggests that the highest expected mortality rates
for insured civilians may have been around 150 per 1,000 per annum; see, for example, Sprague (1895) and Hunter
(1907).
3090 THE AMERICAN ECONOMIC REVIEW october 2012
respectively.32
Notably, the results are very similar to the other columns of the table,
and the confidence intervals are more precisely estimated despite the smaller numbers
of observations. This bolsters the case that reducing the influence of the very high
mortality rates is a useful strategy for dealing with potential measurement error.
In summary, some of the mortality estimates from Curtin are very high, partly
driven by unusual conditions, the impact of epidemics, or small samples. In AJR
(2000, 2001), we discussed this issue at length and used logarithms to reduce the
impact of these very high mortality rates. In AJR (2005), we went one step further,
and, following the information in Curtin’s original sources (in particular Tulloch’s
mid-nineteenth century research), we capped mortality rates at 250 (per 1,000 per
annum). In the analysis below, for all relevant specifications we also show results
including the mortality cap at 250. This mortality cap, or other reasonable caps,
reduce the effect of extreme settler mortality observations. We will show below that
capping mortality also demonstrates that, rather than the AJR (2001) results being
fragile, it is Albouy’s results that are far from robust.
B. Does Discarding Data Make Sense? Latin America
Albouy claims that we lack any reliable data for 36 countries in our base AJR
sample. He drops those countries completely in panels B and D in his Table 2 (first
stage) and Table 3 (second stage), running regressions with just 28 countries. Of the
36 countries which Albouy drops, 16 are in Central and South America. These were
coded using the Gutierrez procedure discussed above. In this section, we discuss
these 16 countries, returning to the remaining 20 countries in the next section.
Albouy is concerned that our LatinAmerican data are not reliable because he does
not like the particular way we benchmark Gutierrez data with Curtin data.33
We
agree that results using this procedure should be subject to robustness checks. This
was the approach in AJR (2000) and in all our subsequent work.
Column 3 of Table 1A reports results using an alternative series. This was discussed
but not explicitly shown in AJR (2001). It was later shown in detail in
AJR (2005). In this series, we offer an alternative way of linking the Curtin and
Gutierrez datasets. Specifically, instead of benchmarking using Mexico, we use
Jamaica/Dominican Republic.34
We continue to assign countries to mortality
32 
Trimming observations above 280 is the same as trimming those above 350, since there are no observations
with settler mortality rates between 280 and 350. Naturally, capping at 280 versus 380 gives different results as
columns 4 and 5 show.
33 
He also argues, for example, around Appendix Table A2, that we simply have no idea about relative mortality
in South and Central America. But as we have discussed, in addition to the evidence from Gutierrez, there is quantitative
evidence on relative mortality in South and Central America from British South American naval stations and
from life insurance rates for sailors in South America from Institute of Actuaries (1851–1852). In addition, Institute
of Actuaries (1851–1852) and Hunter (1907) indicate that the life insurance industry took a clear view on mortality
in this region relative to other regions and also on how mortality varied between countries. This view is entirely
consistent with our benchmarked data.
34 
In the original AJR (2001) series, we assumed that the mortality rate in the Dominican Republic was the same
as in Jamaica, and Albouy does not take issue with this point. Using the Mexico estimates as the benchmark implies a
mortality rate (per 1,000 per annum) of 71 (low), 78.1 (medium), and 163.3 (high) in Gutierrez’s three LatinAmerican
mortality regions (these numbers are used in the data series of column 1). If we use the Jamaica/Dominican Republic
estimates, this gives rates for the three regions of 56.5 (low), 62.2 (medium), and 130 (high); these numbers are used
in the data series of columns 3 and 4. As in AJR (2001), we use the relative mortality ratios of 1, 1.1, and 2.3 between
Gutierrez’s three regions (see footnote 27).
3091Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
regions as in AJR (2001). With this alternative benchmarking, the results are almost
identical in all specifications to those in column 1.35
The second-stage results in
column 3 of Table 1B are also very similar to those in column 1. The AR clustered
confidence sets in the specifications that control for continent dummies and latitude
and for malaria are again fairly wide (consisting of two disjoint intervals) but
exclude zero.
Column 4 of Table 1A shows first stage results with the same measure of mortality
(as in column 3) but now capped at 250 per 1,000. The results are now stronger,
more precisely estimated, and more robust. The AR confidence sets in all cases
comfortably exclude zero (and never extend to infinity).
As an alternative to using the Gutierrez data, we can also use information on mortality
directly from British “South American” naval stations in modern Argentina, Brazil,
Chile, Peru, and Panama; Bryson (1847) gives this as 7.7 per 1,000.36
These data can
be used without any benchmarking to Gutierrez’s data, though naturally they do need to
be converted into what they imply for soldier mortality—as the death rate for soldiers
was typically higher than for sailors when the two types of forces were stationed in the
same area. From Tulloch (1841), we know the mortality of the British naval force (in
the Mediterranean) from disease was 9.2 per 1,000 and the mortality of the military
force (on the ground in that region) from disease was 18.We use this ratio (1.96) for the
SouthAmerican station to convert naval mortality of 7.7 per 1,000 to military mortality
of 15.07.37
This is a conservative—i.e., low—mortality coding for Latin America.38
In
addition, again erring on the conservative side, we assume in this approach that settler
mortality is missing for the remaining countries of South America and for those parts
of Central America for which we previously used the Gutierrez data. This drops 11
countries from our sample, leaving us with only 53 observations.39
Column 5 in Table 1A reports results using this series, which is labeled as “Naval
stations, method 1.” Compared with our baseline results in column 1, the coefficient
is now smaller in absolute value (−0.54). Settler mortality is robustly significant in
the basic regressions (the first four sets of rows), but in the last four specifications
confidence sets become wider (e.g., when we introduce continent dummies) and
this is reflected in Table 1B where the clustered AR confidence sets are quite wide
now and consist of two disjoint intervals extending to infinity. Nevertheless, in all of
these cases these confidence sets exclude zero.
35 
The number of clusters falls by 2. In the original AJR series, Argentina and Chile’s estimates were based on
naval stations. In the revised series they are derived just from bishops’ mortality zones (both are in the low zone;
see previous footnote).
36 
These naval stations were in Rio de Janeiro, Buenos Aires, Bahia, Pernambuco, Para, Valparaiso, Callao,
Coquimbo, and San Blas (Statistical Reports on the Health of the Navy 1841). There is also a San Blas in Mexico,
but our assessment is that the station was in San Blas, Panama. Curtin (1964) cites Bryson (footnote 16); we have
also checked Bryson (1847) directly.
37 
This would put the low end of Latin American mortality almost exactly at the same level as for the United
States, which is 15 per 1,000 per annum (directly from Curtin 1989, Table 1.1, p. 7, for “Northern United States” ).
Note that the ratio of military to civilian mortality may have changed in the second half of the nineteenth century;
this point is examined further in Acemoglu, Johnson, and Subramanian (2012).
38 
Institute of Actuaries (1851–1852, pp.169–170) suggests that mortality rates for civilians within 15 degrees of
the equator in South America were close to those of military personnel in “East Indies and China.” In the healthier
Southern Cone, mortality rates were deemed close to those of Mauritius (which is 30.5 in Curtin 1989, Table 1.1).
39 
To be clear, we are not using any information from Gutierrez in this series. In our baseline series, we use
Gutierrez for 16 countries. We are dropping these 11 countries in this case not because we believe that the data for
them are not reliable (as we have explained this is definitely not the case). Instead, we are doing this as a highly
conservative robustness check.
3092 THE AMERICAN ECONOMIC REVIEW october 2012
Column 6 in Table 1A and Table 1B shows parallel results using the same series
as in column 5, but now with mortality capped at 250 per 1,000. This mortality cap
again strengthens our results, and now confidence sets consist of two disjoint intervals
only in the specification with malaria but continue to comfortably exclude a
zero effect in the second stage.
In column 7 we use data from naval stations in a different way. We compare
life insurance rates for sailors on the “South American Station” from Institute of
Actuaries (1851–1852, p.170), with the rates for places that are also covered by
the earliest Curtin mortality estimates (1989, Table 1.1). According to the same
life insurance source, the healthiest parts of Latin America were determined to
have the same mortality rates as Mauritius while the least healthy parts were
slightly below the West Indies.40
In Curtin’s data (1989, Table 1.1), Mauritius has
a mortality rate of 30.5 per 1,000, while the West Indies average is 93.25.41
We use
these rates for Latin America.
With this alternative series, our first-stage results are robust and very similar to
what we find with the original AJR data. Table 1B shows that in the specifications
with continent dummies and latitude and with malaria (but not in the other specifications),
the AR clustered confidence sets are again wide and consist of two disjoint
intervals. Nevertheless, as is the case in all of these specifications, they do comfortably
exclude zero.
Column 8 reports results for the same series if we cap maximum mortality at 250
per 1,000. Now the results are again more precise, and all AR confidence intervals
are more tightly estimated and never consist of two disjoint intervals extending to
infinity. In fact, the results are very consistent with and confirm those in AJR (2001),
as a comparison with columns 1 and 2 shows.
In summary, there is no basis whatsoever to discard all of our Latin American data
as Albouy does. In contrast to Albouy’s claims, there are several alternative sources
of information on mortality in Latin America. Using our original source, Gutierrez
(1986), with different benchmarking procedures or these alternative data sources,
produces similar mortality rates, which are also consistent with available qualitative
evidence. Different sources of data for Latin America and different benchmarking
procedures lead to very similar and robust results.
C. Does Discarding Data Make Sense? Remaining Countries
The previous subsection discussed Latin American data. Here we discuss the
remaining 20 countries that Albouy drops, which include 12 in Africa, four in the
40 
Specifically, in the language of life insurance, the “extra premium” for mortality above the British death
rate recommended for the South American Station was 40 shillings (so we apply this to Argentina, Brazil, Chile,
Peru, and Panama as these were part of the Station). For the rest of Latin America we use the extra premium
for the North American and West Indian Station, which was 80 shillings. In the same data, the extra premium
for Mauritius was 40 shillings. This approach gives a plausible estimate for parts of the continent closer to the
West Indies, but it is probably on the higher side for Uruguay. See Institute of Actuaries (1851–1852), which
provides the earliest comprehensive assessment of comparative mortality rates. The life insurance literature
from this period developed rapidly, and views were revised and refined subsequently; this is discussed further in
Acemoglu, Johnson, and Subramanian (2012).
41 
These West Indies data points (mortality rates per 1,000 per annum) are: 130 (Jamaica), 85 (Windwards and
Leewards), 106.87 (Guadeloupe), 112.18 (Martinique) and 32.18 (French Guiana).
3093Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
Caribbean, three in Asia, and Australia.42
Albouy’s proposition is that there is no
reliable knowledge about nineteenth century mortality for Europeans in those places.
Our contention is that for each of the countries under discussion, both Europeans at
the time and we presently have information on potential settler mortality—although
undoubtedly there is measurement error in both.
In this section, we summarize the state of knowledge about disease and mortality
and briefly document that for each observation Albouy wishes to drop, there is
considerable evidence supporting the mortality estimates used in AJR (2001). More
details for each of these observations are provided in Appendix B. The discussion
here is short both because of space constraints and because, as the next section
shows, even dropping so many observations has little effect on the robustness of the
results in AJR (2001).
Our main procedure was to assign mortality rates from one country to its neighbors,
based on our reading of the relevant disease ecologies, i.e., taking a position
that the climatic and other environmental conditions for disease were similar in the
country for which we had direct data and the country to which we were making the
assignment. Curtin (1964, 1989, 1998) shows that differential rates of mortality for
Europeans in the early nineteenth century were due primarily to local conditions
for malaria, yellow fever, typhoid, dysentery, cholera, and other so-called “tropical”
diseases—though there was also a great deal of variation even within the tropics. In
assessing disease environments and the knowledge about disease in the nineteenth
century, we use the definitive work of that time by Hirsch (1883–1886) and its
modern-day equivalent, Kiple (1993).43
A brief summary of countries dropped by Albouy is as follows. On Australia, to
which the New Zealand rate was assigned in AJR (2001), Tulloch (1847) provides
an almost identical and independent mortality estimate (see also AJR 2005). On
Singapore, AJR (2001) used the Straits Settlements information. Our numbers are
confirmed by Statistical Society of London (1841), as reported in AJR (2005),
and by Kiat (1978). On Guyana, AJR (2001) used the mortality rate from French
Guiana. This is consistent with the public health literature (Roberts 1948, Mandle
1970), and, in addition, there is independent information from Tulloch (1838a)
and the life insurance literature (Meikle 1876; Hunter 1907). On the Dominican
Republic and Haiti, AJR (2001) used mortality information from Jamaica, and
there is independent confirmation from Institute of Actuaries (1851–52) and
Hunter (1907). On the Bahamas, AJR (2001) used information from the Windward
and Leeward Command; there is independent confirmation from Tulloch (1838b)
and Meikle (1876). On Hong Kong, AJR (2001) used the China Field Force rate
from the British Army; this is backed up by Army Medical Department (1862)
42 
In sub-Saharan Africa, Albouy drops 11 countries: Angola, Burkina Faso, Cote d’Ivoire, Cameroon, Gabon,
Guinea, Niger, Tanzania, Togo, Uganda, Zaire; and in North Africa he drops Morocco. In the Caribbean, Albouy
drops the Bahamas, Dominican Republic, Guyana, and Haiti. InAsia he drops Hong Kong, Singapore, and Pakistan.
43 
Kiple’s team has the benefit of hindsight and contains today’s leading medical historians but might be considered
somewhat distant from events and perceptions of the nineteenth century. Between Hirsch (1883–1886)
and Kiple (1993), there is Clemow (1903), whose volume benefits from the medical advances at the end of the
nineteenth century but who is still close to the major mortality events of that century. There is no indication in
Clemow (1903) that our assessments based on Kiple (1993) are off the mark on anything that matters for our
analysis. We also checked the assessments in Kiple (2003) against Kuczynski (1948), Lancaster (1990), and, for
malaria, Giles and Warrell (1993).
3094 THE AMERICAN ECONOMIC REVIEW october 2012
and Tulloch (1847). On Pakistan, AJR (2001) used the information from Bombay;
there is independent confirmation from Institute of Actuaries (1851–1852) and
Hunter (1907). On Morocco, AJR (2001) used the mortality rate from Algeria.
The mortality rates from Tunisia and Egypt were also similar, and the Institute of
Actuaries (1851–1852) confirms that these countries had roughly the same mortality
level.44
In summary, for all of the non-African observations dropped by Albouy, there is
independent information supporting the rates used in AJR (2001). We documented
this in a detailed manner in AJR (2005). There is no reasonable argument for dropping
these data.
On West Africa, AJR (2001) used data primarily from Curtin (1989), specifically,
mortality rates of soldiers from Sierra Leone, Senegal, Gambia, Gold Coast, Mali,
and Nigeria. The general approach is supported qualitatively and quantitatively by
Curtin (1964), and Bruce-Chwatt and Bruce-Chwatt (1977). There is additional
confirmation from Institute of Actuaries (1851–1852) and Kuczynski (1948).
For Central Africa, we used data from Curtin (1998) and Curtin et al. (1995). Our
use of these data is consistent with evidence in Kiple (1984), and our estimates are
supported by assessments in the life insurance literature—e.g., Institute of Actuaries
(1851–1852), Sprague (1895), and Hunter (1907). On the basic similarity of disease
ecology between West and Central Africa there is general support in Patterson
(1993) and infectious disease-by-disease confirmation in Hirsch (1883–1886) and
American Geographical Society (1951a, b, c, d, e).
While the overall patterns of European mortality in West and Central Africa
reflected in AJR (2000, 2001) are well supported by contemporary and modern
sources, some of this validation is more qualitative than for our non-African observations.
Below we report regressions that drop West and Central African data as an
additional robustness check.
D. Albouy’s Preferred Sample
Albouy proposes to use a sample of just 28 countries (panel B in his Tables 2
and 3). First-stage results with this sample are shown in column 1 of Table 2A,
which has the same set of rows as Table 1B. Second-stage results are shown in
Table 2B.
In Albouy’s preferred sample, our first stage is weakened as soon as covariates are
added. In particular, the confidence sets in specifications without neo-Europes, with
continent dummies, with continent dummies and latitude, with percent of European
descent in 1975, and with malaria are very wide and consist of two disjoint intervals,
extending to infinity on either side. Though in most cases a zero coefficient of
institutions in the second-stage regression can be rejected, such wide (and disjoint)
confidence sets are a clear cause for concern.
44 
Morocco has an extra premium for mortality over British levels of 40 shillings. The category “Mediterranean,
Barbary and Tripoli” has an extra charge of 20 shillings over British levels; this implies a mortality rate within 5–10
per 1,000 of the Morocco level.
3095Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
Figures 1A and 1B, however, show that the first stage in Albouy’s preferred sample
is at least partly weakened by a single significant outlier, Gambia.45
Gambia has
a very high institutions score (8.77, compared to an average of 6.51 in the whole
sample and 5.88 in Africa) and was always an outlier in terms of mortality (1,470
45 
Figure 1A is for the first-stage specification without covariates, and Figure 1B is for the specification with
continent dummies and latitude. Gambia is similarly an outlier in the other specifications.
Table 2A—First Stage Regressions, Using Albouy Preferred Sample
Alternative samples for settler mortality, using original AJR mortality series
Albouy
sample
of 28
Albouy
sample
of 28,
without
Gambia
Albouy
sample
of 28,
mortality
capped
at 250
Albouy
sample
of 28,
without
Gambia,
mortality
capped at 250
Original AJR
series,
without
contested
observations
in West and
Central Africa
Original AJR
series, without
contested
observations in
West and Central
Africa, mortality
capped at 250
Dependent variable is protection against risk of expropriation
(1) (2) (3) (4) (5) (6)
No covariates −0.59 −0.74 −0.95 −1.06 −0.66 −1.02
(standard error) (0.19) (0.15) (0.24) (0.22) (0.18) (0.18)
(clustered standard error) (0.19) (0.15) (0.24) (0.22) (0.19) (0.19)
Number of clusters 28 27 28 27 34 34
Observations 28 27 28 27 53 53
With latitude −0.42 −0.59 −0.74 −0.88 −0.57 −0.94
(standard error) (0.22) (0.15) (0.29) (0.26) (0.19) (0.21)
(clustered standard error) (0.22) (0.15) (0.29) (0.26) (0.20) (0.21)
Number of clusters 28 27 28 27 34 34
Observations 28 27 28 27 53 53
Without neo-Europes −0.32 −0.48 −0.52 −0.66 −0.43 −0.7
(standard error) (0.19) (0.12) (0.23) (0.20) (0.18) (0.21)
(clustered standard error) (0.19) (0.12) (0.23) (0.20) (0.18) (0.22)
Number of clusters 25 24 25 24 31 31
Observations 25 24 25 24 49 49
Without Africa −1.00 −1.00 −1.00 −1.00 −1.21 −1.21
(standard error) (0.28) (0.28) (0.28) (0.28) (0.20) (0.20)
(clustered standard error) (0.28) (0.28) (0.28) (0.28) (0.18) (0.18)
Number of clusters 13 13 13 13 19 19
Observations 13 13 13 13 37 37
With continent dummies −0.31 −0.48 −0.63 −0.75 −0.5 −0.89
(standard error) (0.20) (0.13) (0.22) (0.18) (0.21) (0.26)
(clustered standard error) (0.20) (0.13) (0.22) (0.18) (0.21) (0.26)
Number of clusters 28 27 28 27 34 34
Observations 28 27 28 27 53 53
With continent dummies
  and latitude
−0.22 −0.4 −0.52 −0.66 −0.41 −0.78
(standard error) (0.23) (0.15) (0.27) (0.22) (0.21) (0.27)
(clustered standard error) (0.23) (0.15) (0.27) (0.22) (0.22) (0.27)
Number of clusters 28 27 28 27 34 34
Observations 28 27 28 27 53 53
(Continued)
3096 THE AMERICAN ECONOMIC REVIEW october 2012
per 1,000). But in the 64-country sample of AJR (2001), it did not have a consequential
impact on results. It becomes much more of an outlier when Albouy drops
36 other observations. In addition, there is strong reason to suspect that this institutions
score is not a true reflection of institutional quality in Gambia. For example,
there were military coups in Gambia in 1981 and 1994, and other political turmoil
was manifest the late 1980s. There has not been a return to free and fair elections
since 1994. Moreover, between 1981 and 1989, Gambia and Senegal were united
as part of the Senegambia Confederation, but throughout, Gambia receives a much
higher institutions score than Senegal.46
Column 2 shows that dropping Gambia also from the sample (thus reducing it
to 27 countries) restores the results back to a pattern very similar to those in AJR
(2001). The impact of institutions in the second stage is estimated more precisely,
and none of the clustered AR confidence sets now consists of two disjoint intervals
extending to infinity. Confidence intervals in all specifications except the one with
malaria comfortably exclude a zero effect in the second stage.
46 
The military leader of the 1994 coup,Yahya Jammeh, has reinvented himself as a civilian president but remains
in power through elections that are judged as corrupt. Even before 1994, Gambia had serious political problems.
In 1981, there was a military coup against the independence leader Sir Dawda Jawara, who asked for help from the
Senegalese, and the next year they formed the Senegambia Confederation between the two countries, which lasted
until 1989 (see, e.g., Hughes and Perfect 2008). Throughout this period Senegal has a low institutions score, so
Gambia’s high score is truly puzzling.
Table 2A—First-Stage Regressions, Using Albouy Preferred Sample (Continued)
Alternative samples for settler mortality, using original AJR mortality series
Albouy
sample
of 28
Albouy
sample
of 28,
without
Gambia
Albouy
sample
of 28,
mortality
capped
at 250
Albouy
sample
of 28,
without
Gambia,
mortality
capped at 250
Original AJR
series,
without
contested
observations
in West and
Central Africa
Original AJR
series, without
contested
observations in
West and Central
Africa, mortality
capped at 250
Dependent variable is protection against risk of expropriation
(1) (2) (3) (4) (5) (6)
With percent of European
  descent in 1975
−0.29 −0.46 −0.49 −0.64 −0.48 −0.81
(standard error) (0.21) (0.15) (0.29) (0.25) (0.19) (0.20)
(clustered standard error) (0.21) (0.15) (0.29) (0.25) (0.20) (0.21)
Number of clusters 28 27 28 27 34 34
Observations 28 27 28 27 53 53
With malaria −0.28 −0.44 −0.65 −0.7 −0.49 −0.87
(standard error) (0.26) (0.22) (0.32) (0.30) (0.22) (0.21)
(clustered standard error) (0.26) (0.22) (0.32) (0.30) (0.24) (0.23)
Number of clusters 27 26 27 26 33 33
Observations 27 26 27 26 51 51
Notes: OLS regressions, with one observation per country. Coefficients and standard errors for covariates, where
included, are not reported to save space. Variables are from AJR (2001). Dependent variable is protection against risk
of expropriation; independent variable is log settler mortality. Column 1 uses original settler mortality series fromAJR
(2001) as independent variable but Albouy’s preferred sample of 28 countries. Column 2 is the same as column 1 but
drops Gambia. Column 3 is the same as column 1 but caps mortality at 250 per 1,000 per annum. Column 4 is the same
as column 1 but drops Gambia and caps mortality at 250. Column 5 is the same as column 1 but drops contested observations
for West and Central Africa. Column 6 is the same as column 5 but caps mortality at 250.
3097Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
Column 3 shows that capping mortality rates at 250 also has a major impact on
Albouy’s results. Column 4 shows the results without Gambia and with the 250
mortality cap, which are again very similar and confirm the robustness of the AJR
(2001) estimates.
Columns 5 and 6 follow up on the discussion in the previous two subsections and
add back the Latin American, Caribbean, Asian, and Australian data that Albouy
dropped—thus excluding only the West and Central African data that Albouy would
like to drop. This gives us a sample of 51. Column 5 reports results without capping
and column 6 with the 250 cap. In both columns, the results are very similar to those
in AJR (2000, 2001), and in all cases the second-stage estimates are fairly precise,
the clustered AR confidence intervals never consist of two disjoint intervals, and a
zero effect can be rejected at 5 percent.
Table 2B—Second-Stage Regressions, Using Albouy Preferred Sample
Alternative samples for settler mortality, using original AJR mortality series
Albouy sample
of 28
Albouy sample
of 28, without
Gambia
Albouy sample
of 28, mortality
capped at 250
Albouy sample
of 28, without
Gambia,
mortality
capped at 250
Original AJR
series, without
contested
observations
in West and
Central Africa
Original AJR
series, without
contested
observations
in West and
Central Africa,
mortality
capped at 250
Dependent variable is protection against risk of expropriation
(1) (2) (3) (4) (5) (6)
No covariates 0.87 0.74 0.83 0.75 0.87 0.77
AR confidence set [0.57, 1.64] [0.50, 1.12] [0.56, 1.40] [0.52, 1.13] [0.63, 1.32] [0.56, 1.08]
AR confidence set, clustered [0.60, 1.82] [0.55, 1.02] [0.59, 1.35] [0.56, 1.07] [0.62, 1.62] [0.57, 1.12]
F-stat, first stage 12.47 22.31 17.13 24.38 22.55 33.34
F-stat, first stage, clustered 9.24 24.16 15.63 24.28 12.30 29.22
With latitude 0.82 0.63 0.73 0.62 0.89 0.75
AR confidence set [0.40, 5.79] [0.32, 1.31] [0.35, 2.54] [0.30, 1.31] [0.59, 1.56] [0.51, 1.17]
AR confidence set, clustered [0.42, 19.00] [0.35, 0.97] [0.30, 1.63] [0.24, 0.98] [0.60, 2.09] [0.51, 1.14]
F-stat, first stage 4.93 10.85 6.44 10.26 14.96 22.56
F-stat, first stage, clustered 3.62 14.42 6.50 11.84 8.02 19.49
Without neo-Europes 1.15 0.84 1.13 0.91 1.15 1.00
AR confidence set [−∞, −33.55]
⋃ [0.52, ∞]
[0.41, 1.98] [−∞, −205.15]
⋃ [ 0.51, ∞]
[0.43, 2.83] [0.71, 2.87] [0.61, 2.03]
AR confidence set, clustered [−∞, −10.47]
⋃ [ 0.52, ∞]
[0.44, 1.51] [0.51, 5.61] [0.42, 1.90] [0.68, 5.50] [0.60, 2.14]
F-stat, first stage 3.82 9.90 3.98 7.29 8.78 11.50
F-stat, first stage, clustered 3.00 16.18 4.99 10.76 5.29 9.90
Without Africa 0.90 0.90 0.90 0.90 0.61 0.61
AR confidence set [0.52, 2.09] [0.52, 2.09] [0.52, 2.09] [0.52, 2.09] [0.41, 0.87] [0.41, 0.87]
AR confidence set, clustered [0.63, 1.61] [0.63, 1.61] [0.63, 1.61] [0.63, 1.61] [0.45, 0.85] [0.45, 0.85]
F-stat, first stage 9.87 9.87 9.87 9.87 30.62 30.62
F-stat, first stage, clustered 12.30 12.30 12.30 12.30 45.98 45.98
With continent dummies 1.12 0.81 0.90 0.77 0.93 0.81
AR confidence set [−∞, −4.30]
⋃ [ 0.50, ∞]
[0.41, 2.15] [0.44, 4.74] [0.40, 1.75] [0.59, 2.26] [0.55, 1.40]
AR confidence set, clustered [−∞, −5.70]
⋃ [ 0.47, ∞]
[0.37, 1.38] [0.39, 1.93] [0.34, 1.27] [0.54, 3.04] [0.51, 1.46]
F-stat, first stage 2.85 8.57 5.65 10.89 8.30 15.02
F-stat, first stage, clustered 2.48 12.83 8.06 16.83 5.61 11.77
(Continued)
3098 THE AMERICAN ECONOMIC REVIEW october 2012
E. Albouy’s “Campaign” Dummy
Albouy’s second major concern is that some of our data are taken from military
campaigns, while others are not. To deal with this, he proposes to introduce a coding
for whether or not our data are drawn from a “campaign” and to include that
dummy in the first-stage regression. Despite Albouy’s claims, except during times
of major wars (which are excluded from the data), there is little difference in practice
between what soldiers were engaged in during “campaigns” and at other times.
As a result, it does not in general make sense, and in fact it is not possible, to systematically
distinguish campaigns and noncampaigns, and Curtin does not do so.47
47 
Curtin does mention some campaigns as part of his historical discussion, but he does not offer a systematic
noncampaigns versus campaign distinction. Albouy quotes very selectively from Curtin (1989).
Table 2B—Second-Stage Regressions, Using Albouy Preferred Sample (Continued)
Alternative samples for settler mortality, using original AJR mortality series
Albouy sample
of 28
Albouy sample
of 28, without
Gambia
Albouy sample
of 28, mortality
capped at 250
Albouy sample
of 28, without
Gambia,
mortality
capped at 250
Original AJR
series, without
contested
observations
in West and
Central Africa
Original AJR
series, without
contested
observations
in West and
Central Africa,
mortality
capped at 250
Dependent variable is protection against risk of expropriation
(1) (2) (3) (4) (5) (6)
With continent dummies and
 latitude
1.25 0.77 0.87 0.70 0.99 0.83
AR confidence set [−∞, −0.86]
⋃ [0.37, ∞]
[0.27, 5.26] [−∞, −3.11]
⋃ [ 0.28, ∞]
[0.24, 2.96] [0.56, 5.99] [0.52, 1.80]
AR confidence set, clustered [−∞, −0.39]
⋃ [0.12, ∞]
[0.09, 1.71] [−0.35, 4.35] [−0.13, 1.35] [0.45, 20.44] [0.42, 1.56]
F-stat, first stage 1.24 5.12 2.82 6.17 5.02 9.89
F-stat, first stage, clustered 0.91 7.06 3.61 9.05 3.57 8.61
With percent of European
  descent in 1975
0.94 0.65 0.82 0.65 0.84 0.67
AR confidence set [−∞, −2.35]
⋃ [ 0.33, ∞]
[0.24, 2.07] [−∞, −2.45]
⋃ [ 0.22, ∞]
[0.16, 3.38] [0.50, 1.84] [0.40, 1.17]
AR confidence set, clustered [−∞, −1.46]
⋃ [ 0.33, ∞]
[0.25, 1.36] [−∞, −5.04]
⋃ [ −0.02, ∞]
[−0.03, 1.71] [0.50, 2.72] [0.33, 1.12]
F-stat, first stage 2.53 7.20 2.80 5.32 10.28 16.42
F-stat, first stage, clustered 1.83 9.78 2.99 6.27 5.85 14.80
With malaria 0.90 0.65 0.72 0.67 0.64 0.56
AR confidence set [−∞, ∞] [−∞, −21.20]
⋃ [−0.09, ∞]
[-∞, −61.31]
⋃ [0.21, ∞]
[0.19, 2.41] [0.30, 1.80] [0.32, 0.97]
AR confidence set, clustered [−∞, ∞] [−0.17, 3.80] [0.04, 3.56] [0.07, 1.79] [0.28, 4.59] [0.32, 0.91]
F-stat, first stage 1.24 3.92 3.95 6.12 7.13 16.72
F-stat, first stage, clustered 1.09 3.93 4.07 5.39 4.17 14.62
Notes: 2SLS regressions, with one observation per country, corresponding to first-stage regressions in Table 2A.
Coefficients and standard errors for covariates, where included, are not reported to save space. Variables are from
AJR (2001). Dependent variable is log GDP per capita in 1995. Right-hand side variable is protection against expropriation,
instrumented by log settler mortality. Column 1 uses original settler mortality series from AJR (2001) as
the instrument but Albouy’s preferred sample of 28 countries. Column 2 is the same as column 1 but drops Gambia.
Column 3 is the same as column 1 but caps mortality at 250 per 1,000 per annum. Column 4 is the same as ­column 1
but drops Gambia and caps mortality at 250. Column 5 is the same as column 1 but drops contested observations for
West and Central Africa. Column 6 is the same as column 5 but caps mortality at 250.
3099Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
BGD
CAN
COG
DZA
EGY
ETH
GHA
GMB
IDNIND
JAM
KEN
LKA
MDG
MEX
MLI
MLT
MYS
NGA
NZL
SDN
SEN SLE
TTO
TUN
USA
VNM
ZAF
−2
0
2
Residualexpropriationrisk
−2 −1 0 1 2 3
Residual log settler mortality
Figure 1B. Controlling for Latitude and Continent Dummies
BGD
CAN
COG
DZA
EGY
ETH
GHA
GMB
IDN
IND
JAM
KEN
LKA
MDG
MEX
MLI
MLT
MYS
NGA
NZL
SDN
SEN SLE
TTO
TUN
USA
VNM
ZAF
4
6
8
10
Expropriationrisk
2 4 6 8
Log settler mortality
Figure 1A. Without Controls
3100 THE AMERICAN ECONOMIC REVIEW october 2012
In fact, a major emphasis of Curtin (1998) is that some nineteenth century military
expeditions could have low mortality for some of the reasons we discuss here.
For example, explaining the low mortality for British soldiers on the Magdala campaign
(chapter 2 in his book) and the Asante campaign (chapter 3) in the 1870s,
Curtin (1998, p. 30) writes: “In fact, the Magdala campaign was the engineer’s war.
It was commanded by an engineering officer and hailed by observers as a triumph of
logistical planning. The Asante campaign was the doctors’ war, perceived as the first
evidence that modern medicine made it possible for European troops to act safely in
the tropical world” (italics in original).
The historical record does not indicate systematic differences in mortality
between what can be classified as “campaign” and the rest. Sometimes periods of
fighting brought even less sanitary conditions and higher mortality rates. But at
other times, the organizers of campaigns were able to hold down mortality below
what would have been experienced by civilian settlers—for example, when the
troops were in country for just a short while during a relatively healthy time of
year, or when the doctors in charge happened to be of the opinion that large doses
of quinine must be taken (the general medical consensus, unfortunately, took a long
time to converge on this point).
At other times, campaigning proved a good way to reduce military mortality
compared to sitting in barracks. The mosquito vector for yellow fever has a very
short range—less than a hundred yards (Curtin 1998). In many instances, physically
changing location—for campaign purposes or in a deliberate evacuation—
was enough to reduce the impact of an epidemic. The third major cause of death
was disease due to contaminated water, particularly typhoid, dysentery, and related
diseases. Most of the contamination was caused by poor sanitation, in the sense that
human waste was allowed to find its way into drinking water. Generally speaking,
this contamination was a slow process due to poor design of waste removal and
storage (Curtin 1998, chapter 5)—again, this was more likely to be a problem when
people were in a fixed location for a prolonged period (i.e., not campaigning). If an
expedition camps in one place and moves on the next day, fecal contamination of
the water supply is less likely.
Even more important Albouy’s procedure for coding this dummy is inconsistent
and extremely selective. For example, Albouy decides, very consequentially for his
results, that New Zealand is a noncampaign rate. But British troops were campaigning
in New Zealand against Maori tribes consistently from the mid-1840s—with an intensity
of confrontation that varied over time. There was definitely campaigning during
the time period covered by our data.48
As another example of inconsistency, consider Hong Kong (data from the China
Field Force). As the name suggests, the China Field Force was a field force engaged
in fighting (and Curtin says so explicitly—see Table A8.2 in 1998). But Albouy
chooses to code this as a “noncampaign” rate.
These and other inconsistencies in Albouy’s coding (and the general point that
such a distinction has little meaning) are further discussed in greater detail in
online Appendix C. The rest of this subsection reviews Albouy’s results with his
48
See Belich (1986); Sinclair (1957); and Cowan (1922).
3101Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
“campaign” dummy and how they change significantly once either the impact of
high mortality outliers is limited by capping mortality estimates at 250 per 1,000 or
when minimal corrections for inconsistency are made to his coding.
The first columns of Tables 3A and 3B show Albouy’s first-stage and secondstage
results for the full sample and include his campaign dummies—as well as a
dummy for “slave labor” that he introduces. Although the inclusion of the campaign
and slave labor dummies leads to wide confidence intervals in several specifications
(in particular, in rows 3, 5, 6, 7, and 8), it is notable that these dummies themselves
are not significant individually or jointly in either the first- or the second-stage estimation
in any of these specifications.49
This is, of course, unsurprising given the historical
record discussed above indicating that there were no systematic differences
between “campaigns” and noncampaigns.
For example, in column 1 of Table 3A, the estimates (standard errors in parentheses)
of the campaign dummy in rows 1–8, respectively, are −0.72 (0.46), −0.71
(0.46), −0.43 (0.42), −0.31 (0.38), −0.61 (0.49), −0.59 (0.49), −0.72 (0.43),
and −0.88 (0.53). The corresponding numbers for the slave dummy in rows 1–3
and 5–8 (since the slave dummy is dropped for the regression without Africa) are
−1.61 (0.89), −1.39 (0.91), −1.34 (0.89), −1.45 (0.93), −1.20 (0.94), and −1.40
(0.98). In none of these cases are the two dummies jointly statistically significant at
15 percent or less. They are also very far from statistical significance in the secondstage
regressions reported in Table 3B. Therefore, there does not appear to be any
support from the data that these dummies, even in their miscoded form, belong in
either the first- or the second-stage regressions. To the extent that they are affecting
the AJR (2001) results, this is likely to be due to their spurious presence.
Column 2 shows that simply capping potential European settler mortality at 250
again restores the results essentially back to those obtained in AJR (2001). Once
again, the second stage is estimated more precisely and the clustered AR confidence
sets consist of a single interval excluding a zero effect of institutions, except in the
specification with malaria.50
Column 3 implements the minimal corrections to Albouy’s “campaign” dummy—
just for Hong Kong and New Zealand, which are clearly inconsistently coded as noted
above. This minor correction also leads to more precisely estimated second-stage
results. Column 4 shows that if, in addition, we introduce the mortality capping at 250
per 1,000 per annum, the results are fairly precisely estimated and very similar to those
inAJR (2000, 2001), as can be seen by comparing the estimates and the standard errors
to those in column 1 of Table 1B.51
In both situations, clustered AR confidence sets
49 
Albouy reports the estimates and the (lack of) statistical significance of these dummies in his Appendix Table
A4 but not in his published text.
50 
The campaign and slave dummies are even further from statistical significance in this case, suggesting that they
may have been affecting the results in conjunction with the outliers. In particular, the coefficients and standard errors
for the campaign dummy in rows 1-8 in this case are: −0.46 (0.40), −0.47 (0.41), −0.31 (0.38), −0.31 (0.38), −0.42
(0.42), −0.43 (0.43), −0.52 (0.39), and −0.54 (0.45). For the slave dummy the numbers in rows 1–3 and 5–8 are: −1.07
(0.82), −0.99 (0.84), −1.03 (0.85), −1.01 (0.86), −0.91 (0.89), −1.07 (0.86), and −0.97 (0.89). In none of these cases
are the two dummies jointly statistically significant at 36 percent or less. The pattern is similar in other columns except
that these dummies do become significant in some specifications in Albouy’s preferred sample of 28, but they again
become insignificant in our column 8 when the campaign dummy is corrected and settler mortality is capped at 250.
51 
When the campaign dummy is corrected, it becomes even less significant both in the first and the second
stages, making it even clearer that there is no support from the data for this dummy (or for the slave dummy) to be
included in the regressions.
3102 THE AMERICAN ECONOMIC REVIEW october 2012
again consist of a single interval, excluding zero effects in the second stage (except in
the specification with malaria, where they exclude zero marginally without clustering
and do not exclude it with clustering). The campaign and slave dummies continue to
be insignificant in both the first and second stages in these regressions.
Column 5 considers the more extensively corrected campaign dummy (see online
Appendix C for details). Column 6 reports results from this extensively corrected campaign
dummy together with the 250 per 1,000 mortality cap. The results are once again
very much consistent with those inAJR (2000, 2001); the clusteredAR confidence sets
are now more tightly estimated intervals, always comfortably excluding zero effects.
Finally, column 7 presents Albouy’s results when all his strategies are combined
(only 28 observations and his coding of the campaign dummy). These results, of
course, are highly imprecise with very wide confidence sets, often not excluding
zero. Column 8, however, shows that dropping Gambia, correcting the inconsistencies
in Albouy’s campaign dummy, and capping mortality at 250 leads to results
Table 3A— First-Stage Regressions, with Corrections to Albouy’s “Campaign Dummy”
Alternative codings for campaign dummy, using original AJR mortality series
AJR mortality
series, Albouy
campaign
dummy
AJR mortality
series, capped
at 250; Albouy
campaign
dummy
AJR mortality
series, minimal
correction
to Albouy
campaign
dummy
AJR mortality
series,capped
at 250;
minimal
correction
to Albouy
campaign
dummy
AJR mortality
series,
extended
correction
to Albouy
campaign
dummy
AJR mortality
series, capped
at 250;
extended
correction
to Albouy
campaign
dummy
AJR mortality
series, Albouy
sample of 28;
campaign
dummy
AJR mortality
series, capped
at 250; Albouy
sample of 28;
extended
correction
to Albouy
campaign
dummy;
dropping
Gambia
Dependent variable is protection against expropriation
(1) (2) (3) (4) (5) (6) (7) (8)
No covariates −0.45 −0.77 −0.52 −0.84 −0.6 −0.91 −0.35 −0.96
(standard error) (0.17) (0.21) (0.17) (0.19) (0.15) (0.17) (0.22) (0.26)
(clustered standard error) (0.18) (0.21) (0.18) (0.19) (0.18) (0.18) (0.22) (0.26)
Number of clusters 36 36 36 36 36 36 28 27
Observations 64 64 64 64 64 64 28 27
With latitude −0.39 −0.72 −0.45 −0.79 −0.53 −0.86 −0.21 −0.86
(standard error) (0.19) (0.23) (0.18) (0.21) (0.17) (0.19) (0.25) (0.27)
(clustered standard error) (0.20) (0.23) (0.20) (0.21) (0.19) (0.19) (0.25) (0.27)
Number of clusters 36 36 36 36 36 36 28 27
Observations 64 64 64 64 64 64 28 27
Without neo-Europes −0.31 −0.54 −0.33 −0.57 −0.39 −0.63 −0.18 −0.5
(standard error) (0.16) (0.22) (0.16) (0.20) (0.16) (0.20) (0.22) (0.25)
(clustered standard error) (0.17) (0.20) (0.17) (0.19) (0.17) (0.20) (0.22) (0.25)
Number of clusters 33 33 33 33 33 33 25 24
Observations 60 60 60 60 60 60 25 24
Without Africa −1.11 −1.11 −1.16 −1.16 −1.22 −1.22 −0.88 −0.98
(standard error) (0.28) (0.28) (0.22) (0.22) (0.20) (0.20) (0.32) (0.29)
(clustered standard error) (0.23) (0.23) (0.19) (0.19) (0.17) (0.17) (0.32) (0.29)
Number of clusters 19 19 19 19 19 19 13 13
Observations 37 37 37 37 37 37 13 13
With continent dummies −0.37 −0.7 −0.41 −0.74 −0.46 −0.8 −0.25 −0.73
(standard error) (0.20) (0.27) (0.20) (0.26) (0.20) (0.26) (0.23) (0.20)
(clustered standard error) (0.22) (0.27) (0.21) (0.25) (0.21) (0.25) (0.23) (0.20)
Number of clusters 36 36 36 36 36 36 28 27
Observations 64 64 64 64 64 64 28 27
(Continued)
3103Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
broadly similar to those in the AJR (2001) baseline—even with almost 60 percent of
the sample discarded and the campaign and slave labor dummies included.
We therefore conclude that none of Albouy’s strategies has a major impact on
the results in AJR (2001) once one limits the impact of very high, outlier mortality
rates. Most of the results are remarkably robust. The only specification in which the
second-stage estimates are sometimes insignificant is the one that includes current
prevalence of malaria, which is a specification that biases results against finding
­significant effects as pointed out in AJR (2001) and discussed above. Moreover,
even modest corrections to Albouy’s strategies also lead to similar results.
F. Minor Points
Albouy also presents results using a small modification of the series from AJR
(2001), partly based on AJR (2005). Use of this slightly modified series makes
Table 3A— First-Stage Regressions, with Corrections to Albouy’s “Campaign Dummy” (Continued)
Alternative codings for campaign dummy, using original AJR mortality series
AJR mortality
series, Albouy
campaign
dummy
AJR mortality
series, capped
at 250; Albouy
campaign
dummy
AJR mortality
series, minimal
correction
to Albouy
campaign
dummy
AJR mortality
series,
capped at
250; minimal
correction
to Albouy
campaign
dummy
AJR mortality
series,
extended
correction
to Albouy
campaign
dummy
AJR mortality
series, capped
at 250;
extended
correction
to Albouy
campaign
dummy
AJR mortality
series, Albouy
sample of
28; campaign
dummy
AJR mortality
series, capped
at 250; Albouy
sample of 28;
extended
correction
to Albouy
campaign
dummy;
dropping
Gambia
Dependent variable is protection against expropriation
(1) (2) (3) (4) (5) (6) (7) (8)
With continent dummies
  and latitude
−0.3 −0.63 −0.34 −0.67 −0.38 −0.72 −0.14 −0.66
(standard error) (0.22) (0.29) (0.21) (0.28) (0.20) (0.26) (0.26) (0.23)
(clustered standard error) (0.23) (0.29) (0.22) (0.27) (0.21) (0.26) (0.26) (0.23)
Number of clusters 36 36 36 36 36 36 28 27
Observations 64 64 64 64 64 64 28 27
With percent of European
descent in 1975
−0.27 −0.55 −0.34 −0.63 −0.42 −0.71 −0.2 −0.61
(standard error) (0.17) (0.21) (0.17) (0.20) (0.17) (0.19) (0.23) (0.27)
(clustered standard error) (0.19) (0.22) (0.19) (0.20) (0.19) (0.20) (0.23) (0.27)
Number of clusters 36 36 36 36 36 36 28 27
Observations 64 64 64 64 64 64 28 27
With malaria −0.23 −0.62 −0.35 −0.73 −0.47 −0.83 −0.1 −0.6
(standard error) (0.22) (0.24) (0.23) (0.23) (0.22) (0.22) (0.28) (0.33)
(clustered standard error) (0.24) (0.26) (0.25) (0.25) (0.24) (0.24) (0.28) (0.33)
Number of clusters 35 35 35 35 35 35 27 26
Observations 62 62 62 62 62 62 27 26
Notes: OLS regressions, with one observation per country. Coefficients and standard errors for covariates, where
included, are not reported to save space. Variables are from AJR (2001). Dependent variable is protection against
risk of expropriation; independent variable is log settler mortality; all regressions include Albouy’s “slave labor”
dummy. Column 1 uses original settler mortality series from AJR (2001) as independent variable but includes
Albouy’s campaign dummy. Column 2 is the same as column 1 but caps mortality at 250 per 1,000 per annum.
Column 3 is the same as column 1 but uses our minimal correction of the campaign dummy. Column 4 is the same
as column 3 but caps mortality at 250. Column 5 is the same as column 1 but uses our extended correction of the
campaign dummy. Column 6 is the same as column 5 but caps mortality at 250. Column 7 is the same as column 1
but uses Albouy’s preferred sample of 28. Column 8 is the same as column 7 but uses our extended correction of the
campaign dummy, drops Gambia, and caps mortality at 250.
3104 THE AMERICAN ECONOMIC REVIEW october 2012
little difference as documented in AJR (2005)—unless, of course, the sample is
reduced to 28 observations and the miscoded campaign dummy is included (panel E
of Albouy’s Table 3), but the real issue is appropriate sample size and the miscoded
dummy. We therefore do not provide the details and refer interested readers
to AJR (2005) on this issue.
Table 3B—First-Stage Regressions, with Corrections to Albouy’s “Campaign Dummy”
Alternative codings for campaign dummy, using original AJR mortality series
AJR mortality
series, Albouy
campaign
dummy
AJR mortality
series, capped
at 250; Albouy
campaign
dummy
AJR mortality
series, minimal
correction
to Albouy
campaign
dummy
AJR mortality
series,
capped at
250; minimal
correction
to Albouy
campaign
dummy
AJR mortality
series, extended
correction
to Albouy
campaign
dummy
AJR mortality
series, capped
at 250;
extended
correction
to Albouy
campaign
dummy
AJR mortality
series, Albouy
preferred
sample;
campaign
dummy
AJR mortality
series, capped
at 250; Albouy
preferred
sample;
extended
correction
to Albouy
campaign
dummy;
dropping
Gambia
Dependent variable is log GDP per capita in 1995
(1) (2) (3) (4) (5) (6) (7) (8)
No covariates 1.09 0.86 1.01 0.84 0.93 0.80 1.02 0.83
AR confidence set [0.69, 2.61] [0.56, 1.54] [0.69, 1.85] [0.58, 1.31] [0.67, 1.45] [0.59, 1.16] [−∞, −4.58]
⋃ [ 0.42, ∞]
[0.53, 1.44]
AR confidence set,
 clustered
[0.65, 3.96] [0.51, 1.54] [0.65, 2.52]⋃ [0.56, 1.35] [0.65, 1.78] [0.59, 1.18] [−∞, −4.21]
⋃ [ 0.44, ∞]
[0.58, 1.34]
F-stat, first stage 9.21 15.17 13.89 21.65 20.91 30.53 3.11 15.17
F-stat, first stage,
 clustered
5.9 13.78 8.02 19.2 11.49 26.27 2.57 13.17
With latitude 1.15 0.85 1.06 0.82 0.96 0.79 0.9 0.66
AR confidence set [0.66, 4.87] [0.50, 1.81] [0.66, 2.64] [0.52, 1.48] [0.64, 1.76] [0.54, 1.25] [−∞, ∞] [0.33, 1.49]
AR confidence set,
 clustered
[0.60, 34.78] [0.41, 1.69] [0.62, 4.35] [0.48, 1.41] [0.64, 2.30] [0.55, 1.18] [−∞, ∞] [0.28, 1.08]
F-stat, first stage 5.91 10.52 8.69 14.61 13.81 21.09 0.89 8.94
F-stat, first stage,
 clustered
3.67 9.43 5.31 13.71 7.82 19.7 0.67 9.83
Without neo-Europes 1.45 1.13 1.4 1.12 1.24 1.03 1.51 1.2
AR confidence set [0.78, 22.39] [0.62, 4.15] [0.79, 7.44] [0.65, 3.15] [0.76, 3.42] [0.64, 2.20] [−∞, −0.56]
⋃ [ 0.27, ∞]
[−∞, −6.96]
⋃ [0.46, ∞]
AR confidence set,
  clustered
[−∞, −19.88]
⋃ [ 0.75, ∞]
[0.56, 3.07] [0.77, 20.94] [0.62, 2.50] [0.74, 5.99] [0.64, 2.10] [−∞, −0.38]
⋃ [ 0.26 ∞]
[0.48, 6.67]
F-stat, first stage 4.34 6.39 5.44 7.83 8.09 11 0.93 3.28
F-stat, first stage,
 clustered
3.17 6.89 3.94 8.57 5.11 10.28 0.67 4.11
Without Africa 0.66 0.66 0.64 0.64 0.61 0.61 0.92 0.96
AR confidence set [0.41, 1.08] [0.41, 1.08] [0.42, 0.96] [0.42, 0.96] [0.41, 0.88] [0.41, 0.88] [0.39, 14.18] [0.53, 2.93]
AR confidence set,
 clustered
[0.45, 1.02] [0.45, 1.02] [0.45, 0.93] [0.45, 0.93] [0.44, 0.84] [0.44, 0.84] [0.54, 2.03] [0.61, 1.78]
F-stat, first stage 17.88 17.88 24.5 24.5 30.42 30.42 4.65 7.55
F-stat, first stage,
 clustered
23.03 23.03 37.47 37.47 51.1 51.1 7.52 11.86
With continent
 dummies
1.06 0.81 1.03 0.82 0.96 0.79 1.23 0.81
AR confidence set [0.58, 20.72] [0.48, 2.00] [0.60, 5.34] [0.51, 1.74] [0.59, 2.86] [0.51, 1.47] [−∞, −1.09]
⋃ [0.40, ∞]
[0.41, 2.11]
AR confidence set,
 clustered
[−∞, −9.54]
⋃ [0.51, ∞]
[0.37, 1.77] [0.51, 8.77] [0.41, 1.61] [0.54, 3.68] [0.45, 1.46] [−∞, −0.93]
⋃ [0.29, ∞]
[0.37, 1.38]
F-stat, first stage 4.2 8.35 5.29 10.09 7.08 12.67 1.48 9.06
F-stat, first stage,
 clustered
2.98 6.82 3.96 8.63 5.04 9.88 1.21 12.96
(Continued)
3105Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
Albouy feels that the Mali observation is too high, and he recodes this to 280 per
1,000 in his alternative data series.52
The Mali datapoint we used, 2,940 per 1,000
per annum, reflected an epidemic; we were transparent about this in AJR (2000)
52 
As noted in footnote 8, we find 280 per 1,000 in his online dataset (last checked September 19, 2011). In his
text, he says this value should be 478.2 per 1,000 per annum.
Table 3B—First-Stage Regressions, with Corrections to Albouy’s “Campaign Dummy (Continued)
Alternative codings for campaign dummy, using original AJR mortality series
AJR mortality
series, Albouy
campaign
dummy
AJR mortality
series, capped
at 250; Albouy
campaign
dummy
AJR mortality
series, minimal
correction
to Albouy
campaign
dummy
AJR mortality
series,
capped at
250; minimal
correction
to Albouy
campaign
dummy
AJR mortality
series, extended
correction
to Albouy
campaign
dummy
AJR mortality
series, capped
at 250;
extended
correction
to Albouy
campaign
dummy
AJR mortality
series, Albouy
preferred
sample;
campaign
dummy
AJR mortality
series, capped
at 250; Albouy
preferred
sample;
extended
correction
to Albouy
campaign
dummy;
dropping
Gambia
Dependent variable is log GDP per capita in 1995
(1) (2) (3) (4) (5) (6) (7) (8)
With continent
  dummies and
 latitude
1.19 0.83 1.15 0.84 1.05 0.8 1.44 0.68
AR confidence set [−∞, −2.71]
⋃ [0.56, ∞]
[0.44, 3.52] [−∞, −5.72]
⋃ [0.58, ∞]
[0.47, 2.60] [0.57, 32.57] [0.47, 1.96] [−∞, ∞] [0.22, 3.51]
AR confidence set,
  clustered
[−∞, −1.16]
⋃ [0.37, ∞]
[0.14, 2.20] [−∞, −2.80]
⋃ [0.42, ∞]
[0.24, 1.81] [−∞, −27.45]
⋃ [0.48, ∞]
[0.31, 1.48] [−∞, ∞] [−0.10, 1.28]
F-stat, first stage 2.45 5.64 3.06 6.75 4.09 8.33 0.37 5.62
F-stat, first stage,
 clustered
1.73 4.81 2.31 6.13 3.19 7.64 0.29 8.3
With percent of
  European descent
  in 1975
1.18 0.73 1.03 0.72 0.91 0.7 1.13 0.69
AR confidence set [−∞, −5.13]
⋃ [ 0.54, ∞]
[0.31, 2.88] [0.55, 6.82] [0.38, 1.69] [0.54, 2.37] [0.41, 1.30] [−∞, −0.33]
⋃ [ 0.13, ∞]
[0.17, 17.78]
AR confidence set,
  clustered
[−∞, −2.35]
⋃ [ 0.54, ∞]
[0.13, 1.95] [−∞, −37.42]
⋃ [0.50, ∞]
[0.21, 1.39] [0.52, 3.78] [0.32, 1.18] [−∞, ∞] [−0.01, 1.90]
F-stat, first stage 2.88 6.01 5.02 9.35 8.49 13.94 1.01 4.22
F-stat, first stage,
 clustered
2.01 6.19 3.38 9.85 5.02 12.76 0.78 5.05
With malaria 0.84 0.48 0.75 0.52 0.68 0.53 1.17 0.8
AR confidence set [−∞, ∞] [0.00, 1.91] [−∞, −2.42]
⋃ [0.21, ∞]
[0.20, 1.19] [0.31, 2.44] [0.27, 0.97] [−∞, ∞] [0.21, 9.14]
AR confidence set,
 clustered
[−∞, ∞] [−0.41, 1.08] [−∞, ∞] [0.07, 0.98] [0.27, 11.30] [0.23, 0.85] [−∞, ∞] [0.09, 12.90]
F-stat, first stage 1.15 5.48 2.98 9.09 5.97 13.67 0.14 4.46
F-stat, first stage,
 clustered
0.95 5.75 2.00 8.54 3.69 12.01 0.13 3.28
Notes: 2SLS regressions, with one observation per country, corresponding to first-stage regressions in Table
3A. Coefficients and standard errors for covariates, where included, are not reported to save space. Variables
are from AJR (2001). Dependent variable is log GDP per capita in 1995. Right-hand side variable is protection
against expropriation, instrumented by log settler mortality. Column 1 uses original settler mortality series
from AJR (2001) as the instrument but includes Albouy’s campaign dummy. Column 2 is the same as column
1 but caps mortality at 250 per 1,000 per annum. Column 3 is the same as column 1 but uses our minimal correction
of the campaign dummy. Column 4 is the same as column 3 but caps mortality at 250. Column 5 is the
same as column 1 but uses our extended correction of the campaign dummy. Column 6 is the same as column 5
but caps mortality at 250. Column 7 is the same as column 1 but uses Albouy’s preferred sample of 28. Column
8 is the same as column 7 but uses our extended correction of the campaign dummy, drops Gambia, and caps
mortality at 250.
3106 THE AMERICAN ECONOMIC REVIEW october 2012
and at all subsequent points. As explained above, capping the mortality data series
to reduce the effect of very high rates is a useful exercise—as long as this is done in
a systematic and evenhanded manner. Just capping Mali, but not others, is ad hoc,
selective, and unsatisfactory.53
In addition, Albouy claims that we mistakenly thought countries such as Gabon
and Cameroon border Mali. But we never made any such statement. In Appendix
B of AJR (2000), we wrote “Angola, Cameroon, Rwanda, and Uganda receive the
[mortality] estimate from French Soudan...” The French Soudan was, during the
colonial period, a large area that included Mali but also other parts of West Africa,
much of central Africa, and some of eastern Africa well down into what is now
the Congo.54
We were quite explicit on this usage of the term, stating on page 33:
“For example, Burkina Faso, Central African Federation, Chad, French Congo, and
Mauritania were part of French Soudan.” We used this large area—also known as
French West Africa and French Equatorial Africa—to infer estimates for countries
that were in or neighboring this region. This was clearly explained again in AJR
(2005), where we also provided additional evidence supporting the mortality rates
we originally used in this area, and online Appendix B to this paper recaps some of
the details. Unfortunately, Albouy has chosen to ignore all of this.
In any case, all of this discussion is completely inconsequential. Once again as
shown in detail in AJR (2005)—and as ignored by Albouy in his Comment—the
modification to these rates and changing the rate for Mali in general makes no difference
to the results, and we refer to reader to AJR (2005) for details.
III.  Concluding Comments
The AJR (2001) results on the relationship between potential settler mortality and
institutions are highly robust to a range of checks and variations. First, limiting the
effect of high mortality outliers has no impact on the main results in AJR (2001).
Capping mortality rates at 250 (or 150 or 280 or 350) per 1,000 per annum not only
leaves our results unchanged, but, by reducing the effect of outliers, it increases their
robustness. In fact, this procedure to limit the effect of outliers is by itself sufficient
to make the results in AJR (2001), to a very large extent, immune to Albouy’s two
main critiques—that all the data from Latin America and much of the data from
Africa, making up almost 60 percent of our sample, should be discarded; and that a
“campaign” dummy coded by himself should be included in the first stage.
Second, Albouy’s critiques are simply unfounded. His arguments that there is
no reliable information on settler mortality for much of the world are at odds with
the historical record. We have summarized here—and shown at greater length in
53 
Albouy’s recoding of Mali, by itself, hardly makes any difference in the full sample or even in his preferred
sample of 28. For example, with the full sample and no covariates, the coefficient on settler mortality is −0.59 and
the clustered standard error is 0.24. This is the same point estimate and just a slightly higher standard error compared
with using our original Mali estimate.
54 
Much of this territory was contested with other colonial powers, including the British. We have reviewed colonial-era
maps, and the precise markings and terminology vary for this region. This is why we were explicit about
which countries were in the French Soudan. When writing the first version of his Comment, Albouy (2004a) apparently
understood French Soudan to be a much smaller area, confined to West Africa. We clarified our usage of this
term to him in a private letter in 2004 and have followed up in our public responses (AJR 2005, 2006, 2008).Albouy
still insists on another definition of French Soudan, which is limited to part of WestAfrica. But this is just semantics.
3107Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
AJR (2000, 2001, 2005, 2006, 2008)—that there is a great deal of well-documented
comparable information on the mortality of Europeans in those places during the
relevant period. This information is consistent with the mortality rate estimates used
in AJR (2001). There is no basis for discarding most of our data. Using a variety of
sources, most of which were referenced in our earlier work, one can obtain several
alternative series, which all give results very similar results to those in AJR (2001).
Moreover, even if one were to follow Albouy and discard 60 percent of the AJR
(2001) sample, his results turn out to be driven in large part by an outlier in his sample
of 28: Gambia. There are good reasons to suspect that the very high institutions
score that Gambia receives is not accurate. Also dropping this observation (thus
reducing the sample to 27) leads to results quite similar to those in AJR (2001).
Albouy’s second critique is similarly unfounded. His arguments and coding of the
“campaign” dummy are at odds with the historical record, and his coding procedure
is implemented inconsistently. Even modest corrections to these inconsistencies
overturn his results and show that the main findings in AJR (2001) are robust.
Albouy’s other concerns about Mali are minor, are based on a misreading of our
work—as clearly explained in AJR (2005)—and in any case have no meaningful
effect on our results. Similarly, his slight modification of the data in AJR (2001)
based on AJR (2005) is also not consequential.
References
Acemoglu, Daron, and Simon Johnson. 2007. “Disease and Development: The Effect of Life Expectancy
on Economic Growth.” Journal of Political Economy 115 (6): 925–85.
Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2000. “The Colonial Origins of Comparative
Development: An Empirical Investigation.” National Bureau of Economic Research Working
Paper 7771.
Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2001. “The Colonial Origins of Comparative
Development: An Empirical Investigation.” American Economic Review 91 (5): 1369–1401.
Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2005. “A Response to Albouy’s ‘A Reexamination
Based on Improved Settler Mortality Data.’” Unpublished.
Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2006. “Reply to the Revised (May 2006)
Version of David Albouy’s ‘The Colonial Origins of Comparative Development: An Investigation of
the Settler Mortality Data.’” Unpublished.
Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2008. “Reply to the Revised (2008) Version
of David Albouy’s ‘The Colonial Origins of Comparative Development: An Investigation of the
Settler Mortality Data.’” Unpublished.
Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2011. “Hither Thou Shalt Come, But
No Further: Reply to ‘The Colonial Origins of Comparative Development: Comment’.” National
Bureau of Economic Research Working Paper w16966.
Acemoglu, Daron, Simon Johnson, and James A. Robinson. 2012. “The Colonial Origins of Comparative
Development: An Empirical Investigation: Reply: Dataset.” American Economic Review. http://
dx.doi.org/10.1257/aer.102.6.3077.
Acemoglu, Daron, Simon Johnson, and Karti Subramanian. 2012. “Death and Migration: The View
from Life Insurance Data.” Unpublished.
Albouy, David. 2004a. “The Colonial Origins of Comparative Development: A Reinvestigation of the
Data.” Unpublished.
Albouy, David. 2004b. “The Colonial Origins of Comparative Development: A Reexamination Based
on Improved Settler Mortality Data.” Unpublished.
Albouy, David. 2006. “The Colonial Origins of Comparative Development:An Investigation of the Settler
Mortality Data.” Unpublished.
Albouy, David. 2008. “The Colonial Origins of Comparative Development:An Investigation of the Settler
Mortality Data.” Unpublished.
3108 THE AMERICAN ECONOMIC REVIEW october 2012
Albouy, David. 2012. “The Colonial Origins of Comparative Development: An Empirical Investigation:
Comment.” American Economic Review 102 (6): 3059–76.
American Geographical Society. 1951a. World Distribution of Malaria Vectors. New York: American
Geographical Society.
American Geographical Society. 1951b. Distribution of Cholera, 1816–1950. New York: American
Geographical Society.
American Geographical Society. 1951c. World Distribution of Plague. NewYork:American Geographical
Society.
American Geographical Society. 1951d. World Distribution of Rickettial Diseases: Louse-Borne and
Flea-Borne Typhus. New York: American Geographical Society.
American Geographical Society. 1951e. Distribution of Dengue and Yellow Fever. New York: American
Geographical Society.
Andersen, Robert. 2008. Modern Methods for Robust Regression. Los Angeles: Sage Publications.
Anderson, T. W., and Herman Rubin. 1949. “Estimation of the Parameters of a Single Equation in a
Complete System of Stochastic Equations.” Annals of Mathematical Statistics 20 (1): 46–63.
Angrist, Joshua D., and Alan B. Krueger. 1999. “Empirical Strategies in Labor Economics.” In Handbook
of Labor Economics, Volume 3A, edited by Orley Ashenfelter and David Card, 1277–1366.
Handbooks in Economics, Vol. 5. New York: Elsevier Science.
Army Medical Department. 1838. Statistical Report on the Sickness, Mortality, and Invaliding among
the Troops in the West Indies. London: Army Medical Department.
Army Medical Department. 1839. Statistical Report on the Sickness, Mortality, and Invaliding among
the Troops in the United Kingdom, the Mediterranean, and British America. London: Army Medical
Department.
Army Medical Department. 1841. Statistical Report on the Sickness, Mortality, and Invaliding among
Her Majesty’s Troops Serving in Ceylon, the Tenasserim Provinces, and the Burmese Empire. London:
Army Medical Department.
Army Medical Department. 1853. Statistical Report on the Sickness, Mortality and Invaliding among
the Troops in the United Kingdom, the Mediterranean, and British America. London: Army Medical
Department.
Army Medical Department. 1862. Statistical Report on the Health of the Army in 1860. London: Army
Medical Department.
Balfour, Edward. 1849. “Additional Observations on the Means of Maintaining Troops in Health.”
Journal of the Statistical Society of London 12 (1): 33–42.
Belich, James. 1986. The New Zealand Wars and the Victorian Interpretation of Racial Conflict. Auckland:
Penguin Books.
Bruce-Chwatt, L. J., and Joan Bruce-Chwatt. 1977. “Malaria andYellow Fever.” In Health in Tropical
Africa during the Colonial Period, edited by E. E. Sabben-Clare, D. J. Bradley, and K. Kirkwood.
43–62. Oxford: Clarendon Press.
Bryson, A. 1847. Report on the Climate and Principal Diseases of the African Station. London: W.
Clowes and Sons.
Bulletin of the World Health Organization. 1954. “Malaria Control: Lutte Contre Le Paludisme.” Bulletin
of the World Health Organization 11 (4–5): entire issue.
Chernozhukov, Victor, and Christian Hansen. 2005. “A Simple and Practical Approach to Heteroskedasticity
and Autocorrelation Consistent Inference with Weak Instruments.” Unpublished.
Clemow, Frank G. 1903. The Geography of Disease. Cambridge: Cambridge University Press.
Cowan, James. 1922. The New Zealand Wars, A History of the Maori Campaigns and the Pioneering
Period, Vol. 1 and 2. Wellington: Government Printer.
Curtin, Philip D. 1961. “‘The White Man’s Grave’: Image and Reality, 1780–1850.” Journal of British
Studies 1 (1): 94–110.
Curtin, Philip D. 1964. The Image of Africa: British Ideas and Action, 1780–1850. Madison, WI: University
of Wisconsin Press.
Curtin, Philip D. 1989. Death by Migration: Europe’s Encounter with the Tropical World in the Nineteenth
Century. New York: Cambridge University Press.
Curtin, Philip D. 1990. “The End of the ‘White Man’s Grave?’ Nineteenth-Century Mortality in West
Africa.” Journal of Interdisciplinary History 21 (1): 63–88.
Curtin, Philip D. 1998. Disease and Empire: The Health of European Troops in the Conquest of Africa.
Cambridge: Cambridge University Press.
Curtin, Philip D., Steven Feierman, Leonard Thompson, and Jan Vansina. 1995. African History:
From Earliest Times to Independence. Second Edition. New York: Longman.
Expert Committee on Malaria. 1947. “Extract from the Report on the First Session.” Bulletin World
Health Organization 1 (1): 1947–48.
3109Acemoglu et al.: colonial origins: replyVOL. 102 NO. 6
Feinberg, H. M. 1974. “New Data on European Mortality in WestAfrica: The Dutch on the Gold Coast,
1719–1760.” Journal of African History 15 (3): 357–71.
Fortescue, John W. 1929. History of the British Army. 13 Vols. London: Naval & Military Press, 2004.
Giles, H. M., and D. A. Warrell. 1993. Bruce-Chwatt’s Essential Malariology. Third Edition. London:
Arnold.
Gutierrez, Hector. 1986. “La mortalite des eveques Latino-Americains aux XVIIe et XVIII siecles.”
Annales de Demographie Historique 53 (2): 29–39.
Hirsch, August. 1883–1886. Handbook of Geographical and Historical Pathology. London: New
Sydenham Society.
Huber, Peter J. 1964. “Robust Estimation of a Location Parameter.” Annals of Mathematical Statistics
35 (1): 73–101.
Hughes, Arnold, and David Perfect. 2008. A Political History of the Gambia, 1816–1994. Rochester:
University of Rochester Press.
Hunter,Arthur. 1907. “Mortality in Semi-Tropical and Tropical Countries: Rates of Premium Charged,
and Valuation Basis for these Countries.” Transactions of the American Society of Actuaries.
Institute of Actuaries. 1851–52. “On the Extra Premium Charged for Assurance of the Lives of Persons
Going Abroad.” Assurance Magazine 2 (2): 166–70.
Kiat, Lee Yong. 1978. Medical History of Early Singapore. Tokyo: Southeast Asian Medical Center.
Kiple, Kenneth F. 1984. The Caribbean Slave: A Biological History. New York: Cambridge University
Press.
Kiple, Kenneth F., ed. 1993. The Cambridge World History of Human Disease. Cambridge: Cambridge
University Press.
Kuczynski, R. R. 1948. Demographic Survey of the British Colonial Empire. 3 Vols. London: Oxford
University Press.
Lancaster, H. O. 1990. Expectations of Life: A Study in the Demography, Statistics, and History of
World Mortality. New York: Springer.
Lind, James. 1768. An Essay on Diseases Incidental to Europeans in Hot Climates with the Method of
Preventing their Fatal Consequences. London: Becket and Hondt.
Mandle, Jay R. 1970. “The Decline in Mortality in British Guiana, 1911–1960.” Demography 7 (3):
301–15.
Medical Department of the Navy. 1840. Statistical Reports on the Health of the Navy [Part I]. Parliamentary
paper 159. London: Medical Department of the Navy.
Medical Department of the Navy. 1841. Statistical Reports on the Health of the Navy for the Years
1830, 1831, 1832, 1833, 1834, 1835, and 1836. London: William Clowes and Sons.
Meikle, James. 1876. “On the Additional Premium Required for Residence in Foreign Climates.” Journal
of the Institute of Actuaries and Assurance Magazine 19 (4): 268–94.
Patterson, K. David. 1993. “Disease Ecologies of Sub-Saharan Africa.” In The Cambridge World History
of Human Disease, edited by Kenneth F. Kiple, 447–52. Cambridge: Cambridge University
Press.
Roberts, G. W. 1948. “Some Observations on the Population of British Guiana.” Population Studies 2
(2): 185–218.
Rousseeuw, Peter J. 1984. “Least Median of Squares Regression.” Journal of the American Statistical
Association 79 (388): 871–80.
Select Committee. 1866. Report from the Select Committee on Mortality of Troops (China), 24 July,
House of Commons, Parliament, Great Britain.
Showers, Victor. 1979. World Facts and Figures. New York: John Wiley and Sons.
Sinclair, Keith. 1957. The Origins of the Maori Wars. Auckland: Auckland University Press.
Sprague, A. E. 1895. “Distribution of Diseases in Africa.” Journal of the Institute of Actuaries 32 (1):
65–70.
Statistical Reports on the Health of the Navy for the Years 1830–1836. London: William Clowes and
Sons.
Statistical Society of London. 1840. “Report of a Committee of the Statistical Society of London,
Appointed to Collect and Enquire into Vital Statistics, upon the Sickness and Mortality among the
European and Native Troops Serving in the Madras Presidency, from the Year 1793 to 1838.” Journal
of the Statistical Society of London 3 (2): 113–43.
Statistical Society of London. 1841. “Second Report of a Committee of the Statistical Society of London,
Appointed to Collect and Enquire into Vital Statistics, upon the Sickness and Mortality among
the European and Native Troops Serving in the Madras Presidency.” Journal of the Statistical Society
of London 4 (2): 137–55.
Tulloch, A. M. 1838a. “On the Sickness and Mortality among the Troops in the West Indies.” Journal
of the Statistical Society of London 1 (3): 129–42.
3110 THE AMERICAN ECONOMIC REVIEW october 2012
Tulloch, A. M. 1838b. “On the Sickness and Mortality among the Troops in the West Indies.” Journal
of the Statistical Society of London 1 (4): 216–30.
Tulloch, A. M. 1838c. “On the Sickness and Mortality among the Troops in the West Indies.” Journal
of the Statistical Society of London 1 (7): 428–44.
Tulloch, A. M. 1840. Statistical Reports on the Sickness, Mortality, and Invaliding among the Troops
in Western Africa, St. Helena, The Cape of Good Hope, and Mauritius. London: Army Medical
Department.
Tulloch, A. M. 1841. “Comparison of the Sickness, Mortality, and Prevailing Disease among Seamen
and Soldiers, as Shewn by the Naval and Military Statistical Reports.” Journal of the Statistical
Society of London 4 (1): 1–16.
Tulloch, A. M. 1847. “On the Mortality among Her Majesty’s Troops Serving in the Colonies during
the Years 1844 and 1845.” Journal of the Statistical Society of London 10 (3): 252–59.
Wilcox, Rand R. 2001. Fundamentals of Modern Statistical Methods: Substantially Improving Power
and Accuracy. New York: Springer.