American Economic Association
The Colonial Origins of Comparative Development: An Empirical Investigation: Comment
Author(s): David Y. Albouy
Source: The American Economic Review, Vol. 102, No. 6 (OCTOBER 2012), pp. 3059-3076
Published by: American Economic Association
Stable URL: http://www.jstor.org/stable/41724681
Accessed: 22-11-2017 12:30 UTC
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide
range of content in a trusted digital archive. We use information technology and tools to increase productivity and
facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms
American Economic Association is collaborating with JSTOR to digitize, preserve and extend
access to The American Economic Review
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
American Economic Review 2012, 102(6): 3059-3076
http://dx. doi. org/1 0. 1 25 7 /aer. 102.6. 3059
The Colonial Origins of Comparative Development:
An Empirical Investigation: Commenť
By David Y. Albouy*
Acemoglu, Johnson, and Robinson (2001) (hereafter, AJR) is a seminal article that
has reinvigorated debate over the relationship between property rights and economic
growth. Following Knack and Keefer (1995), Mauro (1995), La Porta et al. (1998),
Hall and Jones (1999), Rodrik (1999), and others, AJR endeavors to determine the
causal effect of institutions that protect property rights, measured by risk of capital
expropriation, on economic performance. This endeavor is complicated by the fact
that the correlation between institutional and economic measures may reflect the
reverse influence of economic growth on institutions or the simultaneous influence
of omitted variables on both economic output and institutions. To circumvent these
problems, AJR uses an instrumental variable (IV) for expropriation risk in an equation
determining GDP per capita across previously colonized countries.
AJR argues that during the colonial era, Europeans were more likely to settle in
places where they had a lower risk of dying from disease. Colonies where Europeans
settled developed institutions that protect property better than colonies where
Europeans did not settle. The article argues that, in the long run, the direct effects
of mortality and European settlement on national income faded, while the indirect
effect through property-rights institutions persisted. This argument motivates the use
of potential European settler mortality rates as an instrument for the risk of capital
expropriation. The AJR IV estimates of the effect of expropriation risk on GDP per
capita are large, explaining much of the variation in income across countries.
The historical sources containing information on mortality rates during colonial
times are thin, which makes constructing a series of potential European settler mortality
rates challenging. AJR constructs this series by combining the mortality rates
of soldiers (Curtin 1989, 1998), laborers (Curtin et al. 1995), and bishops (Gutierrez
1986) from different time periods, mostly prior to the twentieth century. Researchers
have been eager to use this new series, particularly given its promise as an instrumental
variable for institutions. Currently, over 20 published articles, and many more working
papers, use the AJR settler mortality data in their econometric analyses.
* Albouy: University of Michigan, Department of Economics, Lorch Hall, 611 Tappan St., Ann Arbor, MI 48109-
1220, and NBER (e-mail: albouy@umich.edu). I thank Raj Arunachalam, Raphael Auer, Pranab Bardhan, Christina
Berkley, Chris Blattman, David Card, Brad DeLong, Gregory Clark, William Easterly, Rob Gillezeau, Tarek
Hassan, Jim Hines, Chang-Tai Hsieh, Michael Jansson, Chad Jones, Annalisa Leibold, Ian McLean, Ted Miguel,
Kris Mitchener, Marcelo Moreira, Maurice Obstfeld, Rohini Pande, Gerard Roland, Christina Romer, David Romer,
Emmanuel Saez, Andrei Shleifer, Francesco Trebbi, and the participants at the Berkeley Development Lunch and
the Economic History and Macroeconomics Seminars for their help, input, and advice. I am very grateful to Daron
Acemoglu, Simon H. Johnson, and James A. Robinson for providing me with data, and for sharing with me a preliminary
response and later formal responses to my work. Any mistakes are my own.
Ť To view additional materials, visit the article page at http://dx.d0i.0rg/l 0. 1 257/aer. 1 02.6.3059. To view the Appendix
or a rejoinder to the reply to this comment, please visit http://www.nber.org/papers/wl4130.
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
3060 THE AMERICAN ECONOMIC REVIEW OCTOBER 2012
This comment argues that there are several r
comparability of their European settler mort
depend on them. First, out of 64 countries in
mortality rates that originate from within th
tries in the sample are assigned rates based
to which countries have similar disease enviro
erally unfounded and potentially contradictor
incorrect interpretation of former colonial nam
are extrapolated from thin bishop mortality d
(1986), using a "benchmarking" procedure tha
rates, depending on how the data are benchma
mortality rates across countries requires that s
(Moulton 1990). This correction alone noticeab
results. If, in the hope of reducing measureme
ity rates are dropped from the sample, the po
with expropriation risk become substantially sm
covariates, which often gain significance.
Second, the mortality rates never come from
some settler rates are available in the authors' s
ily from European and American soldiers in th
tries, rates apply to soldiers at peace in barrac
soldiers on campaign. As is well known, soldier
mortality from disease. This causes problems
often in countries with greater expropriation r
ing the article's hypothesis. In a few countries
rates of African laborers, but these are not compa
rates. Controlling for the source of the morta
tionship between expropriation risk and morta
if these controls are added and the conjecture
virtually disappears, suggesting that it is largely
Additional data provided by Acemoglu, Johns
reply to an earlier version of this comment, do n
Without a robust relationship between exprop
AJR IV estimates of the effect of expropriatio
weak instrument problems: point estimates are
intervals are often infinite.
Section I below discusses problems with the se
interest researchers using them, or any reade
Section n uses the same IV regression model us
and sensitivity of their hypothesis to problems in
1 Albouy (2008) also discusses AJR's inconsistent mortality rat
demonstrates how the empirical results are sensitive to these cho
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
VOL 102 NO. 6 ALBOUY: COLONIAL ORIGINS: COMMENT 3061
I. Problems with the Settler Mortality Data
The mortality rate data are constructed in four steps, a
Appendix. In their first step, the data include average m
Curtin (1989, p. 7-8) of European soldiers from disease
to mid-nineteenth century. In step two, the data also
selection of military campaigns in Curtin (1998), main
century. The Appendix states that when more than one
rate is chosen. Step three incorporates the peak mortali
who were moved to foreign disease environments in th
seen in Curtin et al. (1995). Also in step three, mortality
boring countries on the premise that they have similar d
in the fourth step, three mortality rates of Latin American
and eighteenth centuries from Gutierrez (1986) are mu
to benchmark them to a rate taken from a French cam
1863, and applied to 16 countries.
Mortality rates are expressed in the number of death
risk, and are cataloged in Table Al. In order to keep the
siderable detail is left to an Appendix on my website.
A. The Matching of Mortality Rates to Neighborin
Thirty-six countries out of 64 have mortality rates that o
own borders. The authors state in their data Appendix (p. 3
ity number to a country if it neighbors a country for whi
same disease environment." The text, however, does not
environments are determined. In regions such as sub-S
Asia, neighboring countries in the data have widely diff
is very sensitive to how neighboring countries are chose
The authors' data Appendix argues (p.l) that large diff
between neighboring countries "because there exists sub
environment, particularly for malaria, even in neighbor
in microclimates.2 Yet substantial variations in disease e
justification for assigning the same mortality rates to neigh
paucity of documentation presented, it is difficult to defen
ing very different rates to some neighboring countries, an
across others. If instead disease environments vary little ac
then much of the variation seen in the data is due to measu
tality rates are likely collinear with other variables suspect
GDP. Either horn of this dilemma poses serious problems
One set of mortality assignments, illustrated in Figu
ity rates that are all from French campaigns in wester
(1998). A close reading of the text reveals the geogr
2This passage arises when the authors assign a rate of 17.7 to Malaysia
fact, Curtin (1989, p. 17-18) does not ascribe this difference to microclimat
were at war in Indonesia.
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
THE AMERICAN ECONOMIC REVIEW OCTOBER 2012
All rates took
place in
campaigns here
Rate of 2,940 (Sep -Oct 1878)
Rate of 400 (1880-1883)
Rate of 280 (1883-1884)
Figure 1. Assignment of Mortality Rates from Mali
making the AJR assignments difficult to explain. They appear to originate from a
misguided interpretation of the changing geographic names for Mali, as explained
in my Appendix. Summarizing briefly,
• Mali is assigned a rate of 2,940 from an acute yellow fever epidemic that killed
49 percent of an expeditionary force from September to October 1878 (Curtin
1998). AJR annualizes the rate, multiplying it by 6.3
• Niger is assigned a rate of 400 from 1880 to 1883 (Curtin 1998, p. 85; this rate
is taken from a table labeled "Haut-Senegal-Niger," a territory that once held
Niger as well as Mali).
• Burkina Faso, Cameroon, Gabon, Angola, and Uganda are assigned a rate of
280 from 1883 to 1884 (Curtin 1998, p. 238; this rate is taken from an entry for
the "French Soudan," a territory that once held Burkina Faso as well as Mali).
There are two fundamental problems with these assignments. First, since all three
rates come from western Mali, there is no possible logical basis for assigning each
3 According to Curtin (1998, p. 81), the rate of 2,940 is an overestimate: because of acquired immunity, "the
annual rate and the rate of loss over two months [490] would have been about the same." Averaging the mortality
rates for Mali over time produces a rate of 478.2. As shown in the Appendix, replacing the rate of 2,940 with 478.2
lowers the significance of the results substantially.
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
VOL 102 NO. 6 ALBOUY: COLONIAL ORIGINS: COMMENT 3063
of these rates to different countries. Second, there is
assigning rates from Mali to countries as far away a
countries with rates taken from Mali have neighbors wit
78.2 in Algeria (which borders Niger) to 2,004 in N
and Cameroon). This large variation illustrates how as
neighboring countries may be very sensitive to choic
The procedure for assigning mortality rates to 16 La
bishop data in Gutierrez (1986) also raises question
mortality rates by country: rather, he categorizes c
medium-, and high-temperature regions, and assumes
peratures have similar disease environments, but nev
environments within these regions are similar.4
The bishop rates (Gutierrez 1986) are based on 4, 5,
populations of 24, 28.5, and 30.5 bishops in each regi
mortality rates of 16.7, 17.5, and 32.8. These rates are no
each other, or from mortality rates of similarly aged
of 18.32 (Sundbärg 1905), or from soldiers in barrack
(20.17) (Curtin 1989).5 In his abstract, Gutierrez (198
tancy at age 40 for bishops was 20.3 years in Latin A
France, implying that mortality was about 43 percent hig
difference accounted for only by deaths in the high-tem
America, bishops born in Europe died at rates slightly
New World. This evidence suggests that settler morta
was not much higher than in Europe.
Yet in the data the authors multiply the bishop mor
mark them to a mortality rate in Mexico, from
from 1862 to 1863. Here, 71 out of a thousand died f
low-temperature bishop rate of 16.7. With the many
there are many other alternatives to benchmark the
that "alternative methods produce remarkably simil
my Appendix, alternative methods in fact produce r
most of them lower. Across areas, the ratio of actual sold
varies from 0.98 to 10.80, rather than staying constan
ing system adopted for Latin America implies that Ch
4 times deadlier than the United States, as the latter
soldiers in the North from 1829 to 1838, a period of
Mexican rate of 7 1 AJR uses is not annualized; based
in Reynaud (1898), I annualize the rate to 61. This ra
4 A map showing the AJR assignments is given in my Appendix Figu
translation) "we cannot study in a profound way the influence of climate
ops in the seventeenth and eighteenth centuries, given the small number
mental situations of which we do not know well the characteristics, and fi
which could affect adults having survived the perils of diseases in infancy
5 An F-test that all three regions have the same mortality rate is not rej
6 In Acemoglu, Johnson, and Robinson (2005, p. 35), the authors propos
a mortality rate for low-temperature regions of 15.4, close to the original
7 In particular, see Appendix Table A2. The Mexican extrapolation itself
much time in Veracruz, a high-temperature area (Reynaud 1898, p. 102-2
the high-temperature area lowers the benchmarking factor from 4.25 to 1
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
3064 THE AMERICAN ECONOMIC REVIEW OCTOBER 2012
mortality rate of white Union soldiers of 53.4 d
(1861-1865), reported in Adams (1952).
The countries with mortality rates inferred from
the 36 countries with conjectured rates. There are
For example, the rate of 14.9 for Hong Kong a
1,200 miles away in Northern China, close to Beiji
annualized, the rate is 50.6 (Army Medical Depa
in Acemoglu, Johnson, and Robinson (2005), Br
during peacetime died at a rate of 285 from 1
the original AJR rate - justifying one characteriz
pestilential, unprofitable, and barren rock" (Ca
method of assigning rates to neighboring coun
deeply flawed, generating rates that may be far t
B. Campaigning Soldiers and African
The cited works by Curtin are concerned prim
of soldiers during the European conquests of th
he took as given the current circumstances an
when comparing their mortality rates. These r
proxy for potential European settler mortality,
with similar living conditions, subject to the c
ments. Living conditions have a large effect on
(1989) discusses how clean water and adequ
lower mortality rates from waterborne disease
intestinal infections. Adequate shelter, nutritio
long known to protect against malaria, also low
Variation in disease due to living conditions s
One reason for this is that AJR combines the m
soldiers in barracks with rates from soldiers o
Curtin emphasizes differences between what
paign rates" (this exact terminology is used re
that "one of the fundamental facts of military
barracks are much healthier than troops on cam
combat," (Curtin 1989, p. 4). Soldiers on camp
disease and were less likely to have safe water,
disposal. Consequently, "[t]he disease toll for s
higher than it was in peacetime" (Curtin 1998,
Curtin (1998) documents how during campaign
increases by more than 100 percent, from gastroi
percent, and from typhoid by more than 600 perc
8 Many valuable sources are cited in Acemoglu, Johnson, a
Cantlie (1974), Balfour (1845), and others mentioned in the Ap
9 This is evident in Curtin (1989, p. xiii): "This book is a qu
European soldiers in the tropics between about 1815 and 1914,"
The Health of European Troops in the Conquest of Africa.
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
VOL. 102 NO. 6 ALBOUY: COLONIAL ORIGINS: COMMENT 3065
2,000 percent higher than barracks rates.10 While AJR emp
diseases such as malaria and fever, much of the variation in
digestive diseases that can occur outside the tropics when
Europe, where barracks rates are usually below 25 (Curti
as high as 332, seen by the British in the Netherlands in 180
Curtin often discusses whether a mortality rate is from a
it possible to code a variable indicating which rates are
not discussed by Curtin, I code a rate as from a campaig
time was spent campaigning. Details of my coding are g
Except in highly unusual circumstances (e.g., at M
rates tend to be higher than barracks rates in a given c
stable relationship between the two. The distinction bet
rates affects the analysis as AJR uses campaign rates mo
high risk of capital expropriation and low GDP per capit
States and Canada are given barracks rates of 15 and 16
than campaign mortality rates during the Civil and Rev
lower than the initial mortality rates of actual European
century. Latin American countries are given campaign r
Mexico, making them appear comparatively much d
suggests. Thus, measured mortality rates are endogenou
security of property rights and lower output per capita
tive measurement error in their mortality rates. This creat
hypothesis that mortality is negatively correlated with
per i - capita.12 - ri - - ri - - r
The effects of c
ing to Curtin (
tions. This is se
below Curtin's r
paigns are abou
and Morocco. M
tive diseases, w
10 Curtin's distincti
with the terms "cam
distinction between
rates are comparable
and seems contrary
and wartime are prim
Indonesia, Mexico, an
authors' claims that no wartime rates are used in the data.
11 This source is cited in Acemoglu, Johnson, and Robinson (2005), although it does not mention these rates.
12 AJR (footnote 17) admits that the data contain measurement error, but that it "does not lead to inconsistent
estimates of the effect of institutions on performance." This is true only if measurement error is uncorrected with
the error term in the equation determining log GDP per capita, which does not appear to be the case.
13 "Climatically the south shore of the Mediterranean was much like the north shore in Italy or southern
France... The high Algerian figure [78.2] in the 1830s was certainly the result of campaigning in the conquest
period. Within a decade or so, the Algerian death rate was close to the rates of the Mediterranean islands" (Curtin
1989, p. 17).
14 Deaths from digestive diseases also play a large role in the rates for Mexico, India, and Vietnam. This may
have more to do with preexisting poverty than with climate: Curtin (1998, p. 113) writes "Typhoid had become
a 'tropical disease' - because the tropical world is poor, not because of climate." Earle (1979) estimates that in
Virginia from 1618 to 1624, British settlers suffered a mortality rate of 283, primarily from dysentery and typhoid.
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
3066 THE AMERICAN ECONOMIC REVIEW OCTOBER 2012
Table 1 - Relationship of Main Variables to Campaign and Laborer Indicators
Expropriation risk
(2)Dependent variable
Original sample (64 countries)
Campaign indicator 1.51
(0.30)
1.68
(0.27)
0.30
Laborer indicator
Correlation with log mortality
Full
Partial, controlling for indicators
Notes: Expropriation risk is "Average protection against expropriation risk 1985-1995"
as measured on a scale from 0 to 10, where a higher score represents greater protection by
Political Risk Services. The original log mortality is the logarithm of European settler mortality
rates from AJR. Heteroskedasticity-robust standard errors in parentheses. Partial correlations
control for campaign and laborer indicators.
Another source of incomparability comes from the use of mortality rates from
African laborers, coerced to move to foreign environments under harsh conditions
(Curtin et al. 1995). Comparing rates from Africa in Curtin (1968), the AJR Data
Appendix argues that the laborer rates provide a lower bound for soldier rates, as
black soldiers had lower average mortality rates than white soldiers. Yet the rates
used are from harshly treated black laborers, not soldiers. Second, all of the rates
taken from Curtin et al. (1995) are maximum rates, and not average rates, as in
Curtin (1968): in the Congo, the maximum rate was 240, while the average rate listed
in the same paragraph - was 100; in Kenya the maximum rate was 145, and
no
Th
ar
m
wo
sig
tro
cia
Fig
in
co
it
sol
15 P
16
Tab
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
VOL 102 NO. 6 ALBOUY: COLONIAL ORIGINS: COMMENT 3067
II. Sensitivity of the Empirical Results
The above discussion raises questions about any empir
constructed potential settler mortality data. For the sake of
the original article, AJR (2001), are examined here.17
The econometric model can be written as the combinat
second-stage equations
(1) r¡ = ßttii +
(2) y i = ami + e¡,
where i indexes colonial countries, y¡ is log G
m, m, is is log log potential potential settler settler mortality, mortality, and and v¡ v¡ and and e, e, are are error error terms, terms, with with E[m,v¡' E[m,v¡' = = 0 0m, m, is is log log potential potential settler settler mortality, mortality, and and v¡ v¡ and and e, e, are are error error terms, terms, with with E[m,v¡' E[m,v¡' = = 0 0m, m, is is log log potential potential settler settler mortality, mortality, and and v¡ v¡ and and e, e, are are error error terms, terms, with with E[m,v¡' E[m,v¡' = = 0 0m, m, is is log log potential potential settler settler mortality, mortality, and and v¡ v¡ and and e, e, are are error error terms, terms, with with E[m,v¡' E[m,v¡' = = 0 0
by construction.18 IV estimates require an instrument that is relevant ( ß ^ 0) and
excludable (£'[m,£i] = 0). Letting n = aß and £, = av¡ + e¡, the reduced form of
the second-stage equation is given by y¡ = irmi + £,. By the principle of indirect
least squares, the IV estimator of a is the ratio of the ordinary least squares (OLS)
estimates of ir and /3; i.e., aIV = í r0Ls/ßoLS- The analysis here first considers the
Because mortality rates are shared by countries, the residuals are correlated because
of clustering effects (see Moulton 1990). This invalidates the conventional standard
errors and test statistics in the original paper, which assumes independent, homoscedastic
errors. The standard procedure used to correct for these clustering effects, as
well as heteroscedasticity (Froot 1989, Wooldridge 2002), is applied below.
More fundamentally, it is worthwhile to examine how sensitive the results are
to robustness checks that account for the weaknesses in the data documented
above. One check is to drop the 36 countries with conjectured mortality rates
that originate from outside their own borders - including the benchmarked Latin
American data - leaving a sample of 28 countries.
This check is similar, but not identical, to the check reported in columns 3 and 4
of the AJR Appendix Table A5 labeled "Earliest Available Data," with 30 countries
(3 1 in AJR 2000), which is supposed to retain rates derived from their first two dataconstruction
steps. Yet the AJR check retains Niger, Burkina Faso, Gabon, Guyana,
and Singapore even though those rates are conjectured from elsewhere: the source of
Niger, Burkina Faso, and Gabon's rate from Mali is already explained in Figure 1;
Guyana's rate is extrapolated from French Guiana (Curtin 1989); Singapore's rate
is extrapolated Í from the city J of Penang, C' Malaysia, J 7 well to the north (Curtin ' 1989). / 19Í J C' J 7 ' /Í J C' J 7 ' /
The A
from
borde
l7The E
Levine
18 Cont
the orig
19 Gab
one observation.
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
THE AMERICAN ECONOMIC REVIEW OCTOBER 2012
i i
4 6
Logarithm
From countr
Conjectured: O Barracks
Figure 2A. Expropriation Risk and Settler Mortality According to
Mortality Rate Characteristics
o «0
•d d
2 ad D g
I I
4 6
Logarithm
From countr■ Campaign A Laborer
□ Campaign A LaborerConjectured: O Barracks
Figure 2B. Income per Capita and Settler Mortality According to
Mortality Rate Characteristics
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
VOL 102 NO. 6 ALBOUY: COLONIAL ORIGINS: COMMENT 3069
The AJR check also retains Congo but not Kenya, alth
since they originate from AJR's third data step using lab
century. My check retains Congo and Kenya, since th
these countries and are controlled for in the next check.
The second robustness check adds two control variables that indicate when a mortality
rate is taken from campaigning soldiers or from imported African laborers to
deal with the comparability issues of different data sources. As is standard, these
variables are included in both the first- and second-stage regressions.
The third check adds new data introduced by Acemoglu, Johnson, and Robinson
(2005) . These data provide native rates for Australia, Bahamas, Guyana, Hong Kong,
Honduras, and Singapore - expanding the sample size of nonextrapolated rates to
34 - as well as unique rates for Sierra Leone and Trinidad and Tobago. These rates,
as well as indicators for soldier campaign rates, laborer rates, and nonextrapolated
rates are reported in Appendix Table Al .
A. First-Stage Estimates
Table 2 presents the first-stage estimates of ß obtained when one applies the
two checks described above, using controls in the original paper. The point estimates
and standard errors for the control and indicator variables are reported in
Table A4. Columns 2 through 5 use geographic controls: latitude (measured in absolute
degrees); omitting "Neo-Europes" (Australia, Canada, New Zealand, and the
United States) ; continent indicators (Asia, Africa, and "Other," with the Americas as
the reference); and combining latitude and continent indicators. These correspond to
columns 2, 3, 7, and 8 in Table 4 of AJR. Column 6 controls for the percentage of the
population of European descent in 1975, like Table 6 of AJR, column 3. Column 7
controls for the percentage of the population living where falciporum malaria is
endemic in 1994, as in Table 7 of AJR, column 1.
The first-stage results with the original data in panel A report that log mortality
is usually a significant predictor of expropriation risk. The clustering adjustment
does increase the size of the standard errors, making ß insignificant at the
10 percent level in column 5.
Panel B applies the first robustness check, dropping conjectured rates, which
causes the standard errors to widen and the point estimates of ß to fall, albeit only
noticeably with controls.20 Interestingly, the controls generally become more significant,
despite the smaller sample size. With the original sample, most control
variables are not significant and lower estimates of ß by less than half. Accordingly,
the authors only consistently use latitude as a control variable. Yet in the smaller,
more reliable subsample, all of the control variables grow appreciably in significance,
while the point estimates of ß fall considerably more. Thus, the conjectured
mortality rates appear to mask the collinearity between the the controls and the more
20 Although the smaller sample does reduce the power of statistical tests, its greater accuracy should raise the
expected value of the point estimate ß by reducing attenuation, at least in the case of classical measurement error
and controls uncorrected with mortality.
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
THE AMERICAN ECONOMIC REVIEW OCTOBER 2012
Table 2 - First-Stage Estimates
(Dependent variable: expropriation risk)
Continent
indicators
and
latitude
(5)Control variables
Panel A. Original data (64 countries, 36 mortality rates)
Log mortality (ß) -0.61 -0.52
{homoscedastic standard error} {0.13} {0.14}
(heteroscedastic-clustered SE) (0.17) (0. 1 9)
p- value of log mortality
p- value of controls
0.001 0.01
- 0.17
Panel B. Removing conjectured mortality rates (28 co
Log mortality (ß) -0.59 -0.42 -0.32 -0.31 -0.22 -0.29
(heteroscedastic standard error) (0.19) (0.22) (0.19) (0.20) (0.23) (0.21)
p- value of log mortality
p- value of controls
0.10 0.13
- 0.01
Panel C. Original data, adding campaign and labor
Log mortality (ß) -0.45 -0.39 -0.31 -0.37 -0.30 -0.27
(heteroscedastic-clustered SE) (0.18) (0.20) (0.17) (0.22) (0.23) (0.19)
p- value of log mortality 0.020 0.06 0.09 0.09 0.20 0.17
p- value of indicators 0.16 0.22 0.31 0.26 0.35 0.19
p- value of controls - 0.27 - 0.75 0.66 0.02
Panel D. Removing conjectured mortality and adding campaign and labo
(28 countries and mortality rates)
Log mortality (ß) -0.35 -0.21 -0.18 -0.25 -0.14 -0.20
(heteroscedastic standard error) (0.22) (0.25) (0.22) (0.23) (0.26) (0.23)
p- value of log mortality 0.12 0.42 0.42 0.28 0.60 0.39
p- value of indicators 0.03 0.06 0.08 0.34 0.44 0.14
p- value of controls - 0.07 - 0.03 0.01 0.05
Panel E. Removing conjectured rates, adding campaign and laborer indic
(34 countries and rates)
Log mortality (ß) -0.41 -0.30 -0.19 -0.31 -0.19 -0.24 -0.30
(heteroscedastic standard error) (0.20) (0.21) (0.21) (0.21) (0.22)
p- value of log mortality 0.05 0.17 0.36 0.16 0.39 0.28 0.20
p- value of indicators 0.02 0.04 0.04 0.30 0.41 0.07 0.06
p- value of controls - 0.13 - 0.20 0.22 0.05 0.33
Notes : Standard errors, assuming uncorrected homoscedastic errors, are sho
standard errors and tests adjust for heteroscedasticity and clustering effects
tries sharing the same mortality rate; p-value of controls are probability va
the controls are significant in the regression; p- value of indicators refers to an
campaign and laborer indicators. See Appendix Table Al for indicators of wh
or is a rate from campaigning soldiers or laborers. "Neo-Europes" consist of
the United States, and are based off of three mortality rates. The three contin
and Other, taken from AJR, consists of Australia, Malta, and New Zealand.
the percent of the population of European descent in 1975 from AJR. Malari
ulation with endemic malaria in 1994 in Sachs and Gallup (2001), which do
Bahamas. Revisions with new data from Acemoglu, Johnson, and Robinson
and given in Table AI. See Table 1 and the text for more detail.
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
VOL 102 NO. 6 ALBOUY: COLONIAL ORIGINS: COMMENT 3071
accurately measured rates.21 Altogether, ß is only significa
the specification without controls.22
Note that using the AJR "Earliest Available Data" sam
A5, causes the point estimates of ß to be almost the sa
ple. The standard errors do rise, however, lowering th
percent in specifications 3 through 7, raising some of
lems discussed below.
Using the original sample again, panel C demonstrates that controlling for whether
a mortality rate comes from soldiers on campaign or from African laborers makes
log mortality insignificant at the 5 percent level in all specifications with controls.
This reduction in significance is the result of lower point estimates for ß, as well as
larger standard errors, and thus does not just come from the indicators using additional
degrees of freedom. The campaign and laborer indicators, whose coefficients
are reported in Table A4 of the Appendix, have negative signs, but tend to have limited
statistical significance W in the original U sample.23 1W U 1W U 1
Panel
signif
indica
the sp
cant a
insign
Panel
check
somew
also m
21 Appe
Albouy
first-sta
22 Acem
sample.
Also, th
Mediter
using th
estimate
IV appro
based on
model as
an assum
Neo-Eur
Fails and
mortalit
of the v
appear s
once thi
23 The r
data fro
produce
24 To en
Append
Order T
25 In my
be weake
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
3072 THE AMERICAN ECONOMIC REVIEW OCTOBER 2012
Instrumental Variable Estimates
When the first-stage estimate of ß is not significantly different from zero - a common
occurrence in the results seen so far - the relevance assumption needed for IV estimates
(ß 0) is not guaranteed, causing a weak instrument problem. This introduces a
number of statistical pathologies to the IV estimates. Most importantly, inference based
on the IV estimate using conventional asymptotic confidence regions (point estimate ±
t X standard error), based on the Wald statistic, can be grossly incorrect (Dufour 1997).
Confidence regions for a of the correct size can be built by inverting the AR statistic
proposed by Anderson and Rubin (1949). While using the AR statistic seems unorthodox
- producing asymmetric, and sometimes disjointed and unbounded, confidence
regions - in the presence of a single weak instrument, it provides correct inference and
an exact test as appropriate as a f-statistic in OLS. When the instrument is strong, AR
and Wald confidence regions are similar, as the latter is not grossly incorrect.26
Table 3 presents the IV estimates and confidence regions corresponding to the
first-stage results in Table 2. In panel A with the original data, weak instrument
problems appear despite the stability of the point estimates. In columns 1 and 2,
where the first stage is fairly strong, the AR and Wald 95 percent confidence regions
are fairly similar. As the instrument weakens in columns 3 and 4, however, the AR
confidence regions widen, until in column 5 it becomes unbounded: as the indirect
least squares formula a = ir/ß implies, once zero cannot be rejected for ß, infinity
cannot be rejected for a.
As the robustness checks are applied in panels B through D, these weak instrument
problems are aggravated: point estimates become unstable and the confidence
regions expand until most of the regions in panels D and E equal the entire real line.
Nevertheless, the point estimates of a get larger, which can also be understood through
the formula a = ir/ß, as the checks have a greater effect in reducing ß than 7r.27 The
estimates of a are sometimes implausibly large, often approaching 2 in panel E: this
would imply some incredible conclusions: e.g., if Mexico and the United States had
the same property rights (a 2.5 point difference) then the GDP per capita ratios of the
two countries would go from less than one-third to over 40 in Mexico's favor.28
The volatile estimates and unbounded confidence regions for a reveal how
instrumental variable inference is frustrated when the first-stage estimate of ß
is not highly significant, which becomes quite an issue when problems with the
mortality data are accounted for.
26 Moreira (2009) proves that, in the exactly identified case, AR tests are the uniformly most powerful among
unbiased tests. The AR confidence regions are said to have "95 percent confidence" because they have 5 percent
size. It does not mean that the true a is within this region 95 percent of the time, but that the AR statistic computed
is within the first 95 percent of the cumulative distribution of the statistic under the null hypothesis. With a weak
instrument, Staiger and Stock (1997) show that conventional F-tests of significance for exogenous variables and
over-identification tests (e.g., Sargan 1958) for the second stage are invalid. Correctly specified tests depend on
parameters that cannot be estimated. Since mortality is a weak instrument in most cases, these test statistics are not
reported ivriivu to w save space.ivriivu wivriivu w
27 This is seen for the second check in Table 1, as the partial corr
by controlling for the data indicators than that for log GDP per cap
28 As shown in Albouy (2008), when the Mali rate is also lowered
a sometimes becomes large and negative, as the estimate of ß beco
estimate of 7 r remains negative. Results are also sensitive to incon
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
VOL. 102 NO. 6 ALBOUY: COLONIAL ORIGINS: COMMENT 3073
Table 3 - Instrumental Variable Estimates and Confidence Regions
First-stage dependent variable: expropriation risk; second-stage dependent variable ,
log GDP per capita, 1995, PPP basis )
Panel A. Original mortality (64 countries, 36 mortality rates)
Expropriation risk (a) 0.93 0.96 1.
Wald 95% conf. [0.52,1.34] [0.42,1.50] [0.35
region
AR "95%" conf. [0.66,1.83] [0.64,2.39] [0.73
region
Panel i urici- u. B. i'GUiwuig Removing conjectured itiut mortality i muy rutcj rates (28 countries and u/»u mortality mist litiny rate ti urici- u. Removing i'GUiwuig conjectured mortality itiut i muy rutcj u/»u mist litiny t y
Expropriation Expropriation risk risk (a) (a) 0.87 0.87 0.82 0.82 1.15 1.15 1.12 1.25 0.94 0.71Expropriation Expropriation risk risk (a) (a) 0.87 0.87 0.82 0.82 1.15 1.15
Wald Wald 95% 95% conf. conf. [0.43,1.31] [0.43,1.31] [0.13,1.51] [0.13,1.51] [-0.10,2.40] [-0.10,2.40] [- [-0.17,2.42] [-1.18,3.67] [-0.33,2.21] [-0.53,1.96]Wald Wald 95% 95% conf. conf. [0.43,1.31] [0.43,1.31] [0.13,1.51] [0.13,1.51] [-0.10,2.40] [-0.10,2.40] [i
urici- u. Removing i'GUiwuig conjectured mortality itiut i muy rutcj u/»u mist litiny t y
Expropriation Expropriation risk risk (a) (a) 0.87 0.87 0.82 0.82 1.15 1.15 1.12 1.25 0.94 0.71Expropriation Expropriation risk risk (a) (a) 0.87 0.87 0.82 0.82 1.15 1.15
Wald Wald 95% 95% conf. conf. [0.43,1.31] [0.43,1.31] [0.13,1.51] [0.13,1.51] [-0.10,2.40] [-0.10,2.40] [- [-0.17,2.42] [-1.18,3.67] [-0.33,2.21] [-0.53,1.96]Wald Wald 95% 95% conf. conf. [0.43,1.31] [0.43,1.31] [0.13,1.51] [0.13,1.51] [-0.10,2.40] [-0.10,2.40] [-
region
AR "95%" conf. [0.58,2.01] (-oo,-7.92]U (-oo,-5.14]U (-oo,-2.25]U (-oo,+oo) (-oo,-0.94]U (-oo,+oo)
region [0.38, +oo) [0.49, +oo) [0.37, +oo) [0.27, +oo)
Panel C. Original data, adding campaign and laborer indicators (64 countries, 36 mortalit
Expropriation risk (a) 1.09 1.15 1.45 1.06 1.19 1.18 0.66
Wald 95% conf. [0.32,1.87] [0.12,2.18] [-0.01,2.91] [0.07,2.05] [-0.30,2.67] [-0.29,2.66] [-0.50,1.81]
region
AR "95%" conf. [0.62,5.07] (-00,- 17.59] U (-00, -8.05] U (-00, -3.28] U (-00, -0.67] U (-00,- 1.67] U (-00, +00)
region U [0.60, +00) [0.69, +00) [0.45, +00) [0.29, +00) [0.44, +00) (-00, +00)
Panel D. Removing conjectured mortality and adding campaign and laborer indicators (28 countries and mortality
Expropriation risk (a) 1.02 0.90 1.51 1.23 1.44 1.13 0.64
Wald 95% conf. [-0.04,2.08] [-1.01,2.81] [-1.89,4.91] [-0.83,3.29] [-3.93,6.80] [-1.22,3.49] [-2.60,3.87]
region
AR "95%" conf. (-00, - 1.82] U (-00, +00) (-00, +00) (- 00, +00) (-00, +00) (-00, +00) (-co, + 00)
region [0.36, +00)
Panel E. Removing conjectured rates, adding campaign and laborer indicators, and revising with new data (34 c
Expropriation risk (a) 1.31 1.11 1.91 1.66 1.36 1.72 1.52
Wald 95% conf. [-0.19,2.80] [-1.35,3.57] [-2.62,6.45] [-1.23,4.55] [-3.61,6.32] [-2.07,5.51] [-2.22,5.25]
region
AR "95%" conf. (-00, -2.86] U (-00, +00) (-00, -0.29] U (-00, -0.24] U (-00, +00) (-00, +00) (-00, +00)
region [0.41, +00) (- 00, +00) [0.17, +00) [-0.14, +00) (- 00, +00) (-00, +00) (-00, +00)
Notes: Panels present the instrumental variable estimates of Expropriation Risk on Log GDP per Capita,
basis, using Log Mortality as an instrument, and the control variables and sample selection described in
Wald 95% Conf. Region are the standard (erroneous) IV confidence regions based on the Wald statistic. A
Conf. Region are the confidence regions calculated from the Anderson-Rubin (1949) statistic as descr
text. Heteroscedasticity and clustering effects are corrected for all confidence regions. See text and Tab
for more details.
III. Conclusion
Given the paucity of instrumental variables in the cross-country growth lit
ture, it is regrettable that the AJR mortality series suffers from severe measureme
problems. While broad regions like West Africa and the Caribbean were cle
unhealthy for Europeans, the mortality differences in the series between neigh
ing countries are largely unreliable. Much of the variation in the mortality data is du
to questionable mortality assignments, which often reflect transitory fluctuatio
living conditions, rather than actual permanent differences among these countr
Given the limited data sources currently available, it seems unlikely that a c
vincing set of settler mortality rates can be constructed. As such, cross-count
growth regressions cannot disentangle the effect of settler mortality from tha
other variables that may explain institutions and growth, such as geography,
mate, culture, and preexisting development, leaving the AJR theoretical hypot
without a strong empirical foundation. Moreover, any researchers who have u
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
THE AMERICAN ECONOMIC REVIEW OCTOBER 2012
Table Al - Mortality Rates And Data Indicators
Original Soldier
mortality campaign rate
Rate from "Benchmarked" Latin- Revised
within country American ratea mortality1
Angola
Argentina
Australia
Burkina Faso
Bangladesh
Bahamas
Bolivia
Brazil
Canada
Chile
Côte d'Ivoire
Cameroon
Congo
Colombia
Costa Rica
Dominican Republic
Algeria
Ecuador
Egypt
Ethiopia
Gabon
Ghana
Guinea
Gambia
Guatemala
Guyana
Hong Kong
Honduras
Haiti
Indonesia
India
Jamaica
Kenya
Sri Lanka
Morocco
Madagascar
Mexico
Mali
Malta
Malaysia
Niger
Nigeria
Nicaragua
New Zealand
Pakistan
Panama
Peru
Paraguay
Sudan
Senegal
Singapore
Sierra Leone
El Salvador
Togo
Trinidad and Tobago
Hinisia
Tanzania
Uganda
Uruguay
US
Venezuela
Vietnam
South Africa
Zaire
280
68.9
8.55
280
71.41
85
71
71
16.1
68.9
668
280
240
71
78.1
130
78.2
71
67.8
26
280
668
483
1,470
71
32.18
14.9
78.1
130
170
48.63
130
145
69.8
78.2
536.04
71
2,940
16.3
17.7
400
2,004
163.3
8.55
36.99
163.3
71
78.1
88.2
164.66
17.7
483
78.1
668
85
63
145
280
71
15
78.1
140
15.5
240
Notes : The sample includes the 28 countries indicated by "Rate from within Country," plus 6 countries that have
revised, but not original, mortality rates from within: Australia, Bahamas, Guyana, Hong Kong, Honduras, and
Singapore. Revised rates for Sierra Leone and Trinidad and Tobago are from more geographically disaggregated
data, as they were previously based on data shared with Gambia and the Lesser Antilles. Honduras is recoded as a
non campaign rate when the data are revised. See the text and Appendix for further details.
a Column indicates the 28 countries included in the sample for panels B, D, and E in Tables 2 and 3.
b Results in panel E of Tables 2 and 3 use the eight "Revised Mortality" rates in place of the "Original Mortality" rates.
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
VOL. 102 NO. 6 ALBOUY: COLONIAL ORIGINS: COMMENT 3075
the AJR mortality series in their analyses may need t
in light of the data issues raised here.
REFERENCES
Acemoglu, Daron, and Simon H. Johnson. 2005. "Unbundling Institutions." Journal of Political
omy 113 (5): 949-95.
Acemoglu, Daron, Simon H. Johnson, and James A. Robinson. 2000. "The Colonial Origins o
parative Development: An Empirical Investigation." National Bureau of Economic Research
ing Paper 7771.
Acemoglu, Daron, Simon H. Johnson, and James A. Robinson. 2001. "The Colonial Origins o
parative Development: An Empirical Investigation." American Economic Review 91 (5): 136
Acemoglu, Daron, Simon H. Johnson, and James A. Robinson. 2005. "A Response to Alb
Reexamination Based on Improved Settler Mortality Data.'" Unpublished.
Acemoglu, Daron, Simon H. Johnson, and James A. Robinson. 2006. "Reply to the Revi
2006) version of David Albouy's 'The Colonial Origins of Comparative Development: An
gation of the Settler Mortality Data.'" Unpublished.
Acemoglu, Daron, Simon H. Johnson, and James A. Robinson. 2009. "Report on 'The Colon
gins of Comparative Development: An Investigation of the Settler Mortality Data,' by David
(2008)." Unpublished.
Adams, George W. 1952. Doctors in Blue: The Medical History of the Union Army in the C
New York: Henry Schuman.
Albouy, David Y. 2008. "The Colonial Origins of Comparative Development: An Investigatio
Settler Mortality Data." National Bureau of Economic Research Working Paper 14130.
Anderson, T. W., and Herman Rubin. 1949. "Estimation of the Parameters of a Single Equa
Complete System of Stochastic Equations." Annals of Mathematical Statistics 20 (1): 46-6
Army Medical Department. 1862. Statistical, Sanitary and Medical Reports of the Army M
Department for the Year 1860. London: Harrison and Sons.
Balfour, Edward. 1845. "Statistical Data for Forming Troops and Maintaining them in Healt
ferent Climates and Localities." Journal of the Statistical Society of London 8 (3): 193-209
Cantlie, Sir Neil. 1974. A History of the Army Medical Department , Volume I. London: Churc
ingstone.
Curtin, Philip D.1968. "Epidemiology and the Slave Trade." Political Science Quarterly 83 (2): 190-216.
Curtin, vui uii^ m. Philip D. xyuy 1989. » Death * by isjr in Migration f'g i : Europe's i/^/& j Encounter with rvuit the Tropical i / World rri/f m in in the t/ tc- 19th iyvui uii^ m. xyuy » * isjr in f'g i i/^/& j rvuit i / rri/f m in t/ tc- iyvui uii^ m. xyuy » * isjr in f'g i i/^/& j rvuit i / rri/f m in t/ tc- iy
Century. New York: Cambridge University
Curtin, Philip D. 1998. Disease and Empire: th
New York: Cambridge University Press.
Curtin, Philip D., Steven Feierman, Leonar
From Earliest Times to Independence. 2nd e
Dufour, Jean-Marie. 1997. "Some Impossi
Structural and Dynamic Models." Econom
Earle, Carville V. 1979. "Environment, Dise
in the Seventeenth Century , edited by Tha
NC: University of North Carolina Press.
Easterly, William, and Ross Levine. 2003. "T
Economic Development." Journal of Mone
Fails, Matthew D., and Jonathan Krieckhau
World Income Distribution." British Jour
Froot, K. A. 1989. "Consistent Covariance Ma
eroskedasticity in Finanical Data." Journal o
Gutierrez, Hector. 1986. "La Mortalité des
Annales de Démographie Historique 1986
Hall, Robert E., and Charles I. Jones. 1999.
Per Worker than Others?" Quarterly Journ
Knack, Stephen, and Philip Keefer. 1995. "
Tests Using Alternative Measures." Econo
La Porta, Rafael, Florencio Lopez-de-Silanes
Finance." Journal of Political Economy 106
Mauro, Paulo. 1995. "Corruption and Growt
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms
3076 THE AMERICAN ECONOMIC REVIEW OCTOBER 2012
Moreira, Marcelo J. 2009. "Tests with Correct Size when
nal of Econometrics 152 (2): 131-40.
Moulton, Brent R. 1990. "An Illustration of a Pitfall in E
on Micro Units" The Review of Economics and Statist
Raynaud, Gustave A. 1898. Considérations sanitaires su
autres expéditions coloniales, françaises et anglais. Par
Rodrik, Dani. 1999. "Where Did All the Growth Go?" J
Rodrik, Dani, Arvind Subramanian, and Francesco Tre
of Institutions over Geography and Integration in Eco
Growth 9(2): 131-65.
Sachs, Jeffrey D., and John L. Gallup. 2001. "The Econom
nal of Tropical Medicine & Hygiene 64 (1-2): 85-96.
Sargan, J. D. 1958. "The Estimation of Economic Relation
metrica 29 (3): 393-415.
Staiger, Douglas O., and James H. Stock. 1997. "Instrumen
ments." Econometrica 65 (3): 557-86.
Sundbärg, G. 1905. "Tab 1: Döde efter alder och kön, 17
Tulloch, A. M. 1847. "On the Mortality of Her Majesty'
Years 1844 and 1845." Journal of the Statistical Societ
Wooldridge, Jeffrey M. 2002. Econometric Analysis o
MA: MIT Press.
This content downloaded from 195.113.13.94 on Wed, 22 Nov 2017 12:30:23 UTC
All use subject to http://about.jstor.org/terms