Population Size and Civil Conflict Risk:

Is There A Causal Link?

by

Markus Brckner*

15 October 2008

Abstract: Does an expansion of the population size expose nation states to a higher
risk of civil conflict? Obtaining empirical evidence for a causal relationship is difficult
due to reverse effects and omitted variable bias. This paper exploits the significantly
negative impact of randomly occurring drought on Sub-Saharan population size as an
instrumental variable. The IV-estimates yield that a one percentage point population
increase raises African civil conflict incidence by over 6.6 percentage points and
increases the risk of an intra-state conflict onset by over 5.1 percentage points.

Key words: Population Size, Civil Conflict, Reverse Causality

JEL codes: O10, O55, P0, Q0

* CAEPS, Universitat de Barcelona and Universitat Pompeu Fabra. Contact e-mail: markus.bruckner@upf.edu. I
thank Antonio Ciccone and participants at the Second Riccardo Faini Doctoral Conference for comments and
discussion.

"The prodigious waste of human life occasioned by this perpetual struggle for room
and food was more than supplied by the mighty power of population. ...An Alaric, an
Attila, or a Zingis Khan, and the chiefs around them, might fight for glory, for the fame
of extensive conquests...but the true cause was a scarcity of food, a population extended
beyond the means of supporting it." Thomas R. Malthus (1798) in: An Essay on the
Principal of Population (p. 23)

1. Introduction
The immense humanitarian and economic costs associated with civil conflict are now well
recognized (e.g. World Bank, 2003; Copenhagen Consensus, 2004). It is estimated that since the
end of World War II total battle casualties due to civil conflict sum to at least 16.2 million (Fearon
and Laitin, 2003), with many more killed or disabled due to violence against civilians and the
spread of lethal diseases (Ghobarah, Huth and Russeth, 2003). Highly concerning is also evidence
by Bellow and Miguel (2006) that shows that the negative impact of civil conflict on economic
development may have long-run effects translating into permanently lower levels of per capita
income due to the adverse effects of civil conflict on social networks and political institutions.
At least since Malthus (1798) it has been argued that a larger population size increases a
country's likelihood of suffering from civil conflict. This claim bears substantial implications for
public policy and theories of economic development as it raises concern that population increases
may not only have a potentially negative effect on countries' per capita income levels but also on
aggregate output and social and political stability. A key question of course is whether the claim is
empirically validated by the data. In particular, is there evidence for a causal relationship?
Obtaining a clean estimate for the impact that population exhibits on civil conflict likelihood
is not an easy task. The majority of empirical studies investigating this link has relied on the
analysis of panel data, exploiting both cross-country as well as within country time series variation
of civil conflict (Gleditsch and Urdal, 2002; Fearon and Laitin, 2003; Collier and Hoeffler, 2004;
1

Urdal, 2005; Hegre and Sambanis, 2006).1 It is questionable though to what extend empirical
research has achieved in testing for the existence of a causal effect as the employed econometric
framework does not allow to address biases due to reverse causality and omitted variables (e.g.
Blattman and Miguel, 2008).

On the one hand side, large downward bias on the estimated

population coefficient may arise due to civil conflicts killing people and being associated with
waves of mass migration (Davenport, Moore, and Poe, 2003; Montalvo and Reynal-Querol, 2007).
On the other hand side, there are many difficult to measure variables proxying for economic
conditions and domestic policy that are related to both civil conflict and population size, potentially
pushing total bias in arbitrary directions.
This paper estimates the impact that increases of population size exhibit on civil conflict
likelihood for a Panel of 37 Sub-Saharan countries over the period 1981-2004, using an
instrumental variable framework. The basis for the IV approach is a highly significantly negative
association between randomly occurring drought and changes in the size of African countries'
population size. Exploiting this link in the second stage regressions the paper finds statistically
significant and quantitatively strong evidence for a causal effect of population on civil conflict:
increasing Sub-Saharan population size by one percentage point raises the region's average
incidence of civil conflict by over 6.6 percentage points and increases the risk of a conflict onset by
over 5.1 percentage points.
An important part to the paper's estimation approach is that it controls for unobservables
related to both civil conflict and population size. Country fixed effects are used to capture difficult
to measure cross-country heterogeneity such as differences in institutional quality, ethnic and
religious fragmentation, or time-invariant geographic conditions such as the share of mountainous
terrain, for instance. Additional unobservable heterogeneity arising from an upward sloping trend

1 A different approach to investigate the link between population size and civil conflict is by means of case study
analysis (see for instance Homer-Dixon, 1994, 1999). Despite its potential for providing a rich understanding of the
underlying forces that generate civil conflict case study analysis has been heavily criticized on the grounds of
suffering from selectivity bias and lack of generality (Gleditsch and Urdal 2002; Urdal, 2005).

in population as well as global climate change is taken into account by implementing country
specific time trends and year fixed effects. Moreover, endogeneity of per capita income to both
civil conflict and population size is addressed by building on the work of Miguel, Satyanath, and
Sergenti (2004) and Brckner and Ciccone (2007) that uses smooth variations in rainfall and
international commodity prices to study the effect of economic growth shocks on civil conflict risk.
The findings of the paper bear a strong message regarding an optimal design of the size of
nations. A larger population size effectively puts African states at a higher risk of suffering from
civil conflict, pointing towards the benefits of smaller country size (e.g. Alesina and Spolaore,
1997, 2005, 2006). The instrumental variable approach makes it credible that the identified link
between population and civil conflict is causal. It is worth noting that least squares (probit)
regression produce insignificant point estimates on the population coefficient. A Hausman test
strongly rejects the exogeneity of population size to civil conflict. Tests from overidentification
restrictions yield no evidence that instruments are correlated to second stage error terms.
The remainder of the paper is organized as follows: Section 2 discusses the estimation
methodology; the data is described in Section 3; Section 4 presents the main results and Section 5
concludes.

2. Estimation Methodology
The econometric framework employed in this paper follows a two-stage instrumental variable
approach that treats both population size and per capita GDP as endogenous variables to the
incidence and onset of civil conflict. In a first stage regression population size and per capita GDP
are regressed on the set of instruments and control variables. Equations (1a) and (1b) show
formally the functional specification,
log POP c , t=1, c 1,ct1,t1,1 Drought c ,t 1,2 log Rainc ,t 1,3 log Indexc ,t 1,c ,t

(1a)

log GDP c ,t =2, c2,ct2,t 2,1 Drought c , t 2,2 log Rain c ,t 2,3 log Indexc ,t 2,c , t (1b)

where POP stands for population size, GDP is the level of (real) per capita GDP, Drought is a
dummy variable indicating episodes of drought, Rain is the amount of rainfall observed in a given
country-year, and Index is an index of international prices for exported commodity goods (see
Section 3 for a description of these variables). The control variables are: (i) country fixed effects c,
(ii) country specific time trends c*t, and (iii) year fixed effects t. The first sub-indice "1" refers to
equation number "1" while the second sub-index identifies the coefficient on the regressor; log
stands for the natural logarithm. The error terms c,t are clustered at the country level to allow for
serial-correlation within countries across time.
The second stage estimates the impact that population expansions exhibit on Sub-Saharan
civil conflict. Formally it is represented by equation (2),
Conflict c ,t = 3,c 3,ct3, t3,2 log Popc , t13,3 logGDPc , t13,c , t

(2)

where Conflict is an indicator function that is one in the event of civil conflict and zero else. Note
that despite the presence of a binary dependent variable equation (2) is specified as a linear
probability model, which is in two-stage instrumental variable estimation the preferred method
(Angrist and Krueger, 2001; Wooldridge, 2003). Although a linear model may potentially violate
Kolmogorov axioms it provides usually a good approximation of the average effect, while a
nonlinear specification would require strong identification assumptions.
Note also that population and per capita GDP are introduced in the second stage explicitly in
levels, rather than growth rates. This guarantees that all possible information contained in the levels
of these variables is exploited for specific country-years as testing the effect of population
expansions on civil conflict requires controlling for the average amount of food output available to
the population at a given point in time and not for changes relative to previous periods. Moreover, a
level specification is immune to producing potentially confounding effects associated with a
corresponding growth specification that arise due to rapid reversion of rainfall to its mean (see here
Ciccone, 2008).

3. Data and Measurement

The data on civil conflict is taken from the 2007 Armed Conflict Dataset of the Uppsala Conflict
Data Program (UCDP) and the Centre for the Study of Civil War at the International Peace
Research Institute, Oslo (PRIO). The UCDP/PRIO Armed Conflict Database defines civil conflict
as a "contested incompatibility which concerns government and/or territory where the use of armed
force between two parties, of which at least one is the government of a state, results in at least 25
battle deaths." Statistically, civil conflict incidence is captured by an indicator variable that is one
in the event of civil conflict and zero else, while civil conflict onset is an indicator variable that
takes on the value of one each time a new conflict has started. For some summary statistics on civil
conflict see Table I, Panel A.
Data on population size and real per capita GDP is taken from the Penn World Tables 6.2
(Heston et al., 2006). Other measures capturing forces of population pressure - working age
population (ages 15-64), youth bulges (ages 0-14), male population, and population density - are
obtained from the World Development Indicators (2007). Number of refugees by country of origin
are obtained from the United Nations Higher Commission for Refugees (UNHCR). Summary
statistics for these variables (in log points) can be found in Panel B.
Following Miguel, Satyanath, and Sergenti (2004) observations on rainfall for the SubSaharan region come from the NASA Global Precipitation Climatology Project, Version 2 (Adler et
al, 2003). Based on this data, droughts are identified by an indicator function that is one for the 5%
largest drops in the level of rainfall over two consecutive years. Selecting such a biennial drop
rather than just a decrease in rainfall over one year follows the notion that it are episodes of
repeated drought rather than short time periods of low rainfall that lead to severe famines in SubSaharan Africa (Sen, 1981; Fafchamps, 2003). The choice of the 5% quantile in turn is motivated
by a trade-off between capturing those episodes of drought that are very unexpected and particularly

severe and the number of observations that are coded as ones. In the sample the 5% quantile
corresponds to a more than 36% drop in the level of rainfall between a given year t and t-2.
Drawing on data provided by the Universite Catholique de Louvain's EM-DAT (2008) natural
disaster database, Appendix Table I shows that these rainfall identified droughts affected in nearly
all cases at least hundreds of thousands of people, and in many instances the numbers were even in
the millions. Section 4.3.3 elaborates on the motivation behind this rainfall based drought indicator,
considering also the use of alternative quantiles (10%, 15%) and other sources for coding drought.
The index of international prices for exported commodities is taken from Brckner and
Ciccone (2007) who employ this instrument to study the impact that commodity driven growth
shocks exhibit on Sub-Saharan civil war risk. The index is constructed using fixed export shares to
ensure that the instrument's time-series variation stems entirely from fluctuations contained in
international commodity prices.

For some summary statistics on rainfall, drought, and the

international commodity price index see Table I, Panel C.

4. Empirical Results
4.1 Rainfall, Commodity Prices, Population Size, and Per Capita GDP
The first stage estimates are reported in Table II, where the included control variables are country
fixed effects, country specific time trends, and common year fixed effects (all jointly significant at
the 1 percent level). Column (1) shows that the rainfall identified droughts lead to an average
decrease in the total population size of a typical Sub-Saharan country of over 1.6 percentage points,
significant at the 1 percent level. Minor changes in the level of rainfall, as captured by the linear
rainfall term, had no significant effect.
insignificant.

The international commodity price index is also

Columns (2) and (3) investigate whether this relationship is maintained when

restricting attention to working age population, or youth bulges. This yields virtually the same
point estimates as in column (1), with the drought dummy being highly significantly negative.

Column (4) shows that droughts had also an equally significantly negative effect on male
population.
One possible explanation why drought exhibits a negative impact on Sub-Saharan countries'
population size is a higher death rate or lower number of new-borns (FAO, 2005; WHO, 2006).
Unfortunately, within the panel data framework of this paper it is impossible to meaningfully test
this mechanism due to lack of annual data. Another explanation would be that drought leads people
to migrate and become refugees in neighboring countries (Sen, 1981; Fafchamps, 2003).

This

channel is in fact testable by using refugee data from the United Nations Higher Commission for
Refugees (UNHCR) statistics. The UNHCR refugee data is publicly available on an annual basis
from 1991 onwards.
Column (5) regresses the (log) number of refugees by origin of country on the drought
indicator, the linear rainfall measure, and the international commodity price index. The result is a
positive and statistically significant point estimate on the drought dummy (p-value of 0.062),
yielding that on average severe drought in a typical African country was associated with nearly a
two-fold increase of its refugee stock. In turn, the linear rainfall measure and the international
commodity price index are insignificant, as expected.
Column (6) reports the first stage estimates for the per capita GDP regression. As in
Miguel, Satyanath, and Sergenti (2004) the paper finds a positive effect of rainfall on economic
output: a ten percentage point increase in the amount of rainfall is associated with an average
increase in per capita GDP in that same year by 0.75 percentage points, significant at the 2 percent
level. Also, windfalls from commodity prices were a blessing for Sub-Saharan countries: a ten
percentage point increase in the Brckner and Ciccone (2007) commodity price index was
associated with an increase in per capita GDP by over 0.91 percentage points, significant at the 5
percent level.
To outrule that statistical inference is invalid due to the presence of unit roots in per capita

GDP, international commodity prices, or population size the Hadri (2000) Lagrange-Multiplier
(LM) test is computed for the residuals of the first stage regressions. The test statistic is based on
the null of stationarity around a country-specific deterministic trend and takes into account serial
dependence in the disturbance term by using a Newey-West kernel estimate of the long-run
variance. As can be seen from the p-values listed in Table II, there is no evidence that first-stage
residuals follow unit-root processes.

4.2 Population Pressures, Per Capita GDP, and Civil Conflict

4.2.1 Civil Conflict Incidence
Table III presents estimates on the effect that population expansions exhibit on the incidence of
Sub-Saharan civil conflict. In columns (1) and (2) conflict incidence is regressed on a set of crosssectional control variables, the log of population size lagged one period, and the log of per capita
GDP, which is also lagged one period.2 The probit regression of column (1) is estimated using
maximum likelihood and reported coefficients are marginal effects evaluated at sample means. It
turns out that the only variable significant at conventional confidence levels is the log share of
mountainous terrain. Population size and per capita GDP are both insignificant. Also, crosssectional differences in ethnic fractionalization and the level of primary school attainment are not
significantly associated with a higher or lower average incidence of civil conflict. In column (2)
exactly the same set of variables are included but the functional form is changed to a linear
probability model. Comparing these point estimates with the marginal effects of column (1) the
least squares and probit regressions yield quite similar coefficients. Moreover, the only variable
that is significant is a country's share of mountainous terrain.
The regressions in columns (1) and (2) should be interpreted with care since there remain
many difficult to measure variables that reflect cross-sectional differences in social fragmentation,
2 The cross-sectional variables on ethnic fractionalization, the share of mountainous terrain, and colonial origin are
obtained from Miguel et al. (2004). The level of primary education is from WDI (2007).

institutional quality or geographic conditions that have only been captured imperfectly by the set of
included control variables. Also, some of these variables such as ethnic fractionalization and the
level of primary schooling are likely to be endogenous to the incidence of civil conflict. To address
these issues column (3) implements country fixed effects to account for unobserved cross-country
heterogeneity. This still results in insignificant coefficients on the population variable. And, when
additionally controlling in column (4) for country specific time trends and year fixed effects the
point estimates have even the wrong sign. In particular, the negative coefficient on the population
variable may be indicative of the presence of large downward bias that arises from civil conflict
reducing a country's population size.
Column (5) estimates the impact that population expansions exhibit on civil conflict
incidence by employing a two-stage least squares country fixed effects approach that treats both
population size and per capita GDP as endogenous regressors. Using the drought dummy, rainfall,
and the international commodity price index as excluded instruments, the IV estimate yields a
coefficient on the population variable that is highly statistically significant (p-value of 0.026) and
nearly ten times larger than the corresponding least squares estimate of column (3).3 A one
percentage point increase of past year's population size increases the likelihood of civil conflict in
the following year by over 1.2 percentage points. Per capita GDP enters this second stage country
fixed effects regression with the correct sign, although insignificant at conventional confidence
levels (p-value of 0.119).
In column (6) country trends and common year effects are included as additional control
variables.

This produces an even larger coefficient on the population variable: for each one

percentage point increase in an African country's population size the likelihood for it to suffer in the
following year from a civil conflict increases by over 6.6 percentage points. And, despite the
substantial increase in estimation uncertainty (including time controls leads to an eight-fold increase
of the second stage standard error) this point estimate is still significantly different from zero at
3 A Hausman test rejects similarity of the IV and least square estimate at over 95 percent confidence.

conventional confidence levels (p-value of 0.095). Furthermore, once the time controls are included
the per capita GDP variable enters also as highly statistically significant (p-value 0.014): increasing
a typical African country's per capita income by one percentage point decreases the likelihood of
civil conflict incidence by over 1.9 percentage points.
Essential for the testing framework is that: (i) the effect of a population expansion on civil
conflict is estimated conditional on the average amount of output available to the total population at
a given point in time (proxied for by the level of per capita GDP), and (ii) that both population size
and per capita GDP are treated as endogenous regressors. Column (7) and (8) demonstrates this
point. In column (7) the control variables are country fixed effects, country specific time trends,
and common year fixed effects, but population size, which continues to be instrumented by the
drought dummy, is regressed on civil conflict incidence without controlling for the level of per
capita GDP. The consequence of not controlling for the income proxy is an insignificant 2SLS
estimate on the population variable that is only one-third of the size of the corresponding estimate
of column (6). In column (8) per capita GDP is included in the second stage, but without being
instrumented by rainfall or the international commodity price index. Thus, column (8) tests the
effect that population expansions exhibit on civil conflict incidence by conditioning on the average
amount of output available to the population at a given point in time, but without taking into
account that the per capita GDP variable is highly endogenous to civil conflict (e.g. Blattman and
Miguel, 2008), and likely to be contaminated with severe measurement error (Heston, 1994;
Deaton, 2005). Just as in column (7), where per capita GDP was not included in the regression, the
point estimate on the population variable is insignificant and cut down to about one-third of the
effect obtained in column (6).
Returning to the double instrumentation framework of column (6), column (9) of Table III
addresses another potential pitfall in the analysis that population expansions exhibit on civil conflict
likelihood by elevating the battle death benchmark to 1000. These remaining civil conflicts are

10

usually referred to as civil war, which have been the focus of many empirical studies investigating
the Malthusian hypothesis (e.g. Hegre and Sambanis, 2006). However, by analyzing only those
civil conflicts that incur more than 1000 battle deaths per annum one systematical excludes those
conflicts that may be minor for countries with large populations but potentially major for smaller
African countries. To illustrate the implications, consider for instance Sierra Leone that had an
average sample population size of 4.4 million. During the period 1981-2004 more than 41% of all
observations were coded as civil conflict, but only 8% were coded as civil war. In turn, for a large
country like Sudan that had an average sample population of over 28 million this coding difference
mattered little as 92% of all observation were marked as civil conflict, and 83% percent were
marked as civil war. The choice of the battle death benchmark hence not only affects the coding of
civil conflict across countries but also across time within countries.

Column (9) shows that

excluding those conflicts that incur less than 1000 battle deaths (which constitute more than 50
percent of all the civil conflicts in the sample) implies that the point estimate on the population
variable turns insignificant, although the estimate on per capita GDP remains highly statistically
significant.

4.2.2 Civil Conflict Onset.

The dependent variable in Table III was the incidence of civil conflict, which incorporates elements
of conflict onset as well as elements of conflict duration. One possible interpretation of the
obtained results is hence that increases in population and lower per capita GDP act as fuel to an
already ongoing civil conflict. But an equally plausible interpretation would be that an increase in
an African country's population size increases the likelihood for this country to become befallen in
the following year by a new or recurrent civil conflict.
In order to explore specifically the impact that past year's population size exhibits on current
year's likelihood of civil conflict outbreak Table IV presents second stage estimates for civil

11

conflict onset.

The first two columns report the estimates of the least squares regressions that

include as control variables country fixed effects (column (1)) as well as country specific time
trends and year fixed effects (column (2)). In both of the least squares regressions is the population
variable insignificant. The coefficient is negative, pointing towards strong downward bias that
arises from an onset of civil conflict reducing population size.
In columns (3)-(4) population size and per capita GDP are instrumented by a drought
dummy, rainfall, and the index of international commodity prices. Now the effect of population
expansions is positive and highly statistically significant. The point estimate of column (4), where
the control variables are country fixed effects, country specific time trends, as well as year fixed
effects, indicates that a one percentage point increase in past year's population size increases the
likelihood for an African country to experience an onset of civil conflict by over 5.1 percentage
points, significant at the 5 percent level.4 In turn, an increase of past year's per capita GDP due to
better rainfall conditions or higher international commodity prices significantly reduces the risk for
Sub-Saharan countries to become befallen by civil conflict.

Thus, increases in Sub-Saharan

population size and decreases in per capita income raise the region's rate of conflict incidence, and,
in particular, increase the risk for African countries to become befallen by a new or recurrent civil
conflict.

4.3 Robustness
4.3.1 Overidentification
On the statistical side the results of Tables III and IV bear a clear message: not addressing the bias
that results from reverse causality and possibly severe measurement error would lead to the
conclusion that the Malthusian prediction of population pressures increasing the likelihood of war
failed in significantly explaining Sub-Saharan civil conflict. In turn, the instrumental variable
estimates provided stark evidence that population expansions not accompanied by significant
4 A Hausman test rejects equality of the IV and least square estimate at over 95 percent confidence.

12

increases in per capita income place Sub-Saharan countries at a substantially higher risk of suffering
from civil conflict.
An important assumption underlying the validity of the IV-estimates is that the drought
dummy, rainfall, and the index of international commodity prices have no effect on civil conflict
other than through per capita GDP or population size. For instance, droughts may stimulate an
influx of food aid. The instantaneous average effect would be captured by the per GDP variable,
but aid could have effects on income distribution, generating a more egalitarian society. If a
reduction in inequality is associated with less conflict then this implies that both the per capita GDP
and population coefficient are over-estimated, e.g. the true impact would be larger (smaller) for
increases in per capita GDP (population size) than shown in Tables III and IV. On the empirical
side, there is not much evidence though pointing towards a systematic effect of inequality on civil
conflict (see for example Hegre and Sambanis, 2006, or Blattman and Miguel, 2008). A somewhat
more concerning issue is that rainfall could affect warfare strategies. One such factor is for instance
troop mobility, which may be higher during times of low rainfall. Another factor may be that
heatwaves associated with episodes of drought make soldiers and civilians more aggressive,
increasing potential for violent action. Note however that second stage variables (and instruments)
are all lagged one period. Hence, it would have to be the case that rainfall of the past year
significantly affects warfare possibilities of the following year.
Regarding the international commodity price index a concern may be that the instrument is
not entirely exogenous to Sub-Saharan civil conflict since large producers of world commodities
may have an impact on international commodity prices. Note that if this were the case then the
second stage estimate of the per capita GDP variable would be an upper bound since civil conflict
will act as a supply shock to the international commodity market, increasing prices and hence per
capita GDP. Brckner and Ciccone (2007) addressed this concern by excluding large commodity
producers, running a reduced form and second stage regression without those commodities where

13

an African country's exports constitute more than 3% of world production.

This did not

significantly change point estimates. Furthermore, in their second stage regressions international
commodity price growth conditional on per capita GDP growth was always insignificant.
The instrumental variable regressions employ three instruments for two endogenous
variables, implying that the system is overidentified.

Hence, it is possible to formally test

instrument validity by running the Hansen J-test. This Lagrange-Multiplier test is based on the null
that instruments are jointly uncorrelated to second stage error terms. If the obtained test-statistic is
significantly different from zero then this would cast serious doubt on instrument validity. But as
can be readily seen from the computed p-values at the bottom of Tables III and IV the Hansen J-test
statistic is always insignificant. Thus, the test does not provide evidence against the assumption
that the instruments fulfill the exclusion restriction and are exogenous to the presence of SubSaharan civil conflicts.

4.3.2 Fuller LIML

The other important criteria in instrumental variable analysis is the instrument strength in providing
a sufficiently precise first stage fit. Staiger and Stock (1997) suggested as a rule of thumb criteria a
first stage F-stat of around 10, which for the population (per capita GDP) regression was 10.79
(8.75). The tabulations in Stock and Yogo (2002) show however that in the presence of multiple
endogenous regressors the first stage F-stat should increasingly exceed this rule-of-thumb value as
the number of instruments increases.
To address this issue Table V presents second stage estimates using Fuller limited
information maximum-likelihood estimators. These estimators have been shown to be more robust
to weak instruments than two-stage least squares (e.g. Stock, Wright, and Yogo, 2002; Hahn and
Hausman, 2003). The two Fuller limited-information maximum likelihood estimates are calculated
for Fuller constants 4 and 1. The Fuller 1 estimator yields the most unbiased estimator and is

14

recommended when one wants to test hypotheses; the Fuller 4 estimator minimizes the mean
squared error of the estimator (Fuller, 1977).
In columns (1)-(4) the dependent variable is civil conflict incidence, and in columns (5)-(8)
civil conflict onset. In all cases is the population coefficient positive and statistically significant.
The Fuller second stage conflict incidence country fixed effects estimates yield that a one
percentage point population increase raises the likelihood of civil conflict in the following year by
over one percentage point, with a p-value of 0.03 for the Fuller 1 estimator and 0.008 for the Fuller
4 estimator (columns (1) and (2)). Quantitatively these point estimates increase when including
country specific time trends and Africa wide year controls (columns (3) and (4)).

Now the

coefficient is 5.98 for the Fuller 1 estimator and 4.13 for the Fuller 4 estimator. The Fuller second
stage estimates also highlight the quantitatively strong and statistically significant impact that
Malthusian forces exhibited on Sub-Saharan civil conflict onset: a one percentage point increase of
past year's population size increased the likelihood of a civil conflict outbreak in the following year
between 3.4 (Fuller 4) and 4.8 percentage points (Fuller 1), significant at least at the 4 percent level.

4.3.3 Alternative Drought Data

The so far presented instrumental variable estimates relied on the use of a drought indicator that was
defined as taking on the value of one if the drop over two consecutive years in the level of rainfall
was more than 36%, which corresponded to the 5% quantile. As a robustness check Table VI,
Panel A, columns (1)-(3) show second stage fixed effects estimates that are based on extending the
the drought dummy to include also those drops in the level of rainfall that are in the 5% to 10%
percent quantile range (e.g. a drop between 26% and 36%). Within the country fixed effects
framework of Panel A this still results in a significantly positive point estimate on the population
variable with coefficients ranging from 1.27 to 2.188 (the corresponding p-values are 0.002 and
0.083 respectively). In Panel B however where in addition to the country fixed effects country

15

specific time trends and year fixed effects are included coefficients are very imprecisely estimated
and not statistically significant at conventional confidence levels (although coefficients are
quantitatively quite large, ranging from 3.30 to 6.53). In general, extending the cut-off much
beyond the 5 percent quantile, say, to the 15 percent quantile, would not yield significant first nor
second stage point estimates (regressions not shown). This negative result should not be too
surprising though given that it are severe and repeated droughts, rather than episodes of low rainfall,
that lead to the famines associated with people dying in Africa due to starvation or large streams of
migration in search for food and better living conditions (Sen, 1981; Fafchamp, 2003; FAO, 2005;
WHO, 2006).
An alternative to the rainfall based drought indicator would have been to draw directly on
drought data provided by the Universite Catholique de Louvain's International Emergency Disasters
Database (EM-DAT, 2008).5

In this database droughts are reported if any of the following

minimum criteria are fulfilled: (i) ten or more people reported killed; (ii) hundred or more people
affected; (iii) a declaration of a state of emergency; or (iv) a call for international assistance. Table
VI, Panel A, columns (4)-(6) compute second stage country fixed effect estimates that use instead
of the GPCP rainfall based drought dummy the EM-DAT drought indicator as an instrumental
variable. This drought indicator is defined in complete analogy to the rainfall based drought
indicator, taking on the value of one for the 5 percent harshest droughts, as measured by number of
people affected according to EM-DAT. The result is a positive and significant effect on the lagged
population variable with coefficients ranging between 1.27 to 2.15 (the p-values are 0.008 and
0.093 respectively). Finally, columns (7)-(9) combine EM-DAT drought data with the GPCP
rainfall based indicator generating a dummy variable that is one if both EM-DAT and GPCP agree
jointly on the event of drought. This yields somewhat smaller point estimates on the population
variable (coefficients range between 1.61 and 1.09), which are significant at the 5 percent level.
For the purpose of the instrumental variable analysis the rainfall based drought indicator has
5 The data is publicly available at www.emdat.be.

16

at least three important advantages over the EM-DAT data.

First, rainfall drought data is

interpretable as triggering famine, which in Sub-Saharan Africa has been associated with waves of
mass migration and people dying due to starvation. Second, since rainfall is random the drought
indicator is completely exogenous to the presence of Sub-Saharan civil conflict. This stands in
stark contrast to the EM-DAT drought indicator, which is an outcome variable potentially
endogenous to the conditions of the local environment. Moreover, generating a drought indicator
based on number of people affected may produce confounding effects due to cross-country
differences in the size and the trend of the population variable. Finally, as pointed by EM-DAT
itself obtaining a concise estimate of the number of people affected by drought is a difficult task due
to compilation dilemmas and misreporting. Hence, in so far as the GPCP satellite based rainfall
data collected by Adler et al. (2003) is correct, the rainfall indicator is more robust to measurement
error than the EM-DAT drought data.

5. Conclusion
This paper addressed core issues at the heart of the debate of whether population expansions cause
civil conflict by exploiting randomness of drought as an instrumental variable for Sub-Saharan
population size. Taking the Malthusian claim seriously, the second stage regressions explicitly
accounted for year-specific cross-sectional differences in the average amount of (food) output
available to the population at a given point in time by using rainfall and international commodity
prices as an instrumental variable for per capita GDP.
Quantitatively large and statistically significant empirical support is found for population
size affecting civil conflict likelihood: a one percentage point population increase raises the chance
of observing in the following year a civil conflict of up to 6.6 percentage points and increases the
risk of a conflict onset by over 5.1 percentage points.

Comparing the instrumental variable

estimates to the insignificant and quantitatively very small least square estimates highlights the need

17

to employ econometric methods that are able to convincingly deal with the biases that arise when
testing the the impact that population size exhibits on civil conflict likelihood.
From the policy perspective the paper bears a clear message: population expansions in SubSaharan Africa increase these countries' likelihood of suffering from civil conflict. One way to
combat the higher risk of intra-state conflict is to ensure that population increases are accompanied
by a substantial improvement in the economic environment. As income rises the opportunity cost
of war increases and this reduces the incentives to engage in combat. Ultimately then, the future
role of population in affecting African civil conflict likelihood will depend on whether increases in
prosperity are associated with equally large percentage increases in population size or whether
these countries follow the path of western societies where higher wages lead optimal household
decision making to endogenously solve the challenges arising from food scarcity and the threat of
overpopulation.

18

6. References
Adler, R.F., G.J. Huffman, A. Chang, R. Ferraro, P. Xie, J. Janowiak, B. Rudolf, U. Schneider, S.
Curtis, D. Bolvin, A. Gruber, J. Susskind, P. Arkin, and E. Nelkin (2003). The Version 2 Global
Precipitation Climatology Project (GPCP) Monthly Precipitation Analysis (1979-Present).
Journal of Hydrometeorology 4: 1147-1167.
Alesina, A. and E. Spolaore (2006). "Conflict, Defense Spending and the Number of Nations."
European Economic Review 50:91-120.
Alesina, A. and E. Spolaore (2005). "War, Peace and the Size of Countries." Journal of Public
Economics 89:1333-1354.
Alesina, A. and E. Spolaore (1997). "On the Number and Size of Nations." Quarterly Journal of
Economics 112: 1027-1056.
Angrist, J. and A. Krueger (2001). Instrumental Variables and the Search for Identification: From
Supply and Demand to Natural Experiments. Journal of Economic Perspectives 15: 69-85.
Bellow, J. and E. Miguel (2006). "War and Institutions: New Evidence from Sierra Leone."
American Economic Association Papers and Proceedings 96 (2):394-399.
Blattman, C. and E. Miguel (2008). "Civil War." Journal of the Economic Literature, forthcoming.
Brckner, M. and A. Ciccone (2007). "Growth, Democracy, and Civil War." CEPR Discussion
Paper No. 6568.
Ciccone, A. (2008). "Transitory Economics Shocks and Civil Conflict." Working Paper UPF,
Barcelona.
Collier, P. and A. Hoeffler (2004a). "Greed and Grievance in Civil War. Oxford Economic Papers
56 (4): 563-596.
Collier, Paul and Anke Hoeffler (2004b). 'The challenge of Reducing the Global Incidence of Civil
War.' presented at Copenhagen Consensus 2004.
Davenport, C., W. Moore, S. Poe (2003). "Sometimes You Just Have to Leave: Domestic Threats
19

and Refugee Movements: 1964-1989." International Interactions 29: 27-55.

Deaton, A. (2005). "Measuring Poverty in a Growing World (Or Measuring Growth in a Poor
World)." Review of Economics and Statistics 87 (1):1-19.
EM-DAT (2008). "EM-DATA: The OFDA/CRED International Disasters Database." Universite
Catholique de Louvain, Brussels, Belgium, available at www.emdat.net.
Fafchamps, M. (2003). Rural Poverty, Risk, and Development. Cheltenham: Edward Elgar.
FAO (2005). "Food Supply Situation and Crop Prospects in Sub-Saharan Africa." Food and
Agriculture Organization of the United Nations, Africa Report, April 2005.
Fearon, J. and D. Laitin (2003). "Ethnicity, Insurgency and Civil War." American Political Science
Review 97 (1): 75-90.
Fuller, W. (1977). "Some Properties of a Modification of the Limited Information Estimator."
Econometrica 45: 939-954.
Ghobarah, H., P. Huth, and B. Russet (2003). "Civil Wars Kill and Maim People Long after the
Shooting Stops." American Political Science Review 97 (2):189-202.
Gleditsch, N. and H. Urdal (2002). "Ecoviolence? Links Between Population Growth,
Environmental Scarcity and Violent Conflict in Thomas Homer-Dixon's Work." Journal of
International Affairs 56 (1): 283302.
Hegre, H. and N. Sambanis (2006). "Sensitivity Analysis of Empirical Results of Civil War Onset."
Journal of Conflict Resolution 50 (6):937-961.
Heston, A. (1994). "A Brief Review of Some Problems in Using National Accounts Data in Level
of Output Comparisons and Growth Studies." Journal of Development Economics 44: 29-52.
Heston, A., Summers, R., and B. Aten (2006). Penn World Table Version 6.2. Center for
International Comparisons of Production, Income and Prices, University of Pennsylvania,
September 2006.
Homer-Dixon, T. (1994). Environmental Scarcities and Violence Conflict: Evidence from Cases."

20

International Security 19 (1): 5-40.

Homer-Dixon, T. (1999). Environment, Scarcity and Violence. Princeton, NJ: Princeton University
Press.
Malthus, Thomas Robert (1798/1997). An Essay on the Principle of Population. London: Electronic
Book.
Miguel, E., S. Satyanath, and E. Sergenti (2004). 'Economic Shocks and Civil Conflict: An
Instrumental Variables Approach.' Journal of Political Economy 112 (41):725-753.
Montalvo, J. and M. Reynal-Querol (2007). "Fighting Against Malaria: Prevent Wars While
Waiting For The Miraculous Vaccine." Review of Economics and Statistics 89 (1): 165-177.
Urdal, H (2005). "People vs. Malthus: Population Pressures, Environmental Degradation and
Armed Conflict Revisited." Journal of Peace Research 42 (4):417-434.
UNHCR (2007). "Global Report 2006". United Nations High Commissioners for Refugees, June
2007.
Sen, A. (1981). Poverty and Famines. Oxford: Clarendon Press.
Wooldridge, J. (2002). Econometric Analysis of Cross Section and Panel Data. Cambridge, Mass.:
MIT Press.
World Bank (2003). Breaking the Conflict Trap: Civil War and Development Policy. Oxford
University Press.
WHO (2006). "WHO Responding to the Drought Crisis in the Horn of Africa Region." World
Health Organization, ICP/EHA-HAC, May 2006.

21

Table I
Descriptive Statistics
A. Measures of Conflict
Mean

Std. Dev.

Observations

Civil Conflict Incidence > 25 Battle Deaths

0.267

0.443

888

Civil Conflict Onset > 25 Battle Deaths

0.047

0.212

888

Civil Conflict Incidence >1000 Battle Deaths

0.123

0.328

888

Civil Conflict Onset >1000 Battle Deaths

0.023

0.148

888

B. Population Pressure and Per Capita Income

Total Population

15.841

1.150

888

Population Aged 15-64

15.186

1.158

888

Population Aged 0-14

15.040

1.159

888

Male Population

15.138

1.153

888

6.996

0.734

888

Number of Refugees
Real Per Capita GDP

C. Instrumental Variables
Rainfall (GPCP)

6.743

0.629

888

Commodity Price Index

4.094

0.494

888

Drought (GPCP)

0.050

0.217

888

Table II
Drought, Population Size, and Per Capita GDP
Population Size

GDP

(1)

(2)

(3)

(4)

(5)

(6)

All

Age 15-64

Age 0-14

Male

Refugees

Per Capita GDP

-0.016***
(0.005)

-0.017***
(0.005)

-0.017***
(0.006)

-0.016***
(0.005)

0.994*
(0.517)

-0.001
(0.014)

Log Rainfall, t

-0.006
(0.012)

-0.006
(0.015)

-0.007
(0.011)

-0.004
(0.013)

0.333
(0.324)

0.075**
(0.029)

Log Index, t

0.012
(0.010)

0.018
(0.013)

0.011
(0.11)

0.013
(0.010)

0.308
(0.370)

0.085**
(0.041)

Hadri Unit Root test

0.8063

0.8082

0.79

0.8092

0.7479

Country Fixed Effects

Yes

Yes

Yes

Yes

Yes

Yes

Year Effects and Trends

Yes

Yes

Yes

Yes

Yes

Yes

No Observations

888

888

888

888

462

888

Drought, t

Note: Method of estimation is least squares with Huber robust standard errors (shown in parentheses) clustered at the country level. *Significantly
different from zero at 90 percent confidence, ** 95 percent confidence, *** 99 percent confidence.

22

Table III
Population Size, Per Capita GDP, and Civil Conflict Incidence
Civil Conflict
(1)

(2)

(3)

(4)

Probit

LS

LS

LS

(5)

Civil War
(6)

2SLS
2SLS
Pop+GDP Pop+GDP

(7)

(8)

(9)

2SLS
Pop

2SLS
Pop

2SLS
Pop+GDP

2.248
(2.971)

2.000
(3.010)

2.377
(2.607)

0.173
(0.107)

-1.216**
(0.538)

Log Population, t-1

0.022
(0.43)

0.009
0.149 -0.628
(0.038) (0.144) (0.731)

1.244**
(0.560)

6.607*
(3.962)

Log per capita GDP, t-1

-0.092
(0.062)

-0.064 -0.163 0.217

(0.050) (0.103) (0.135)

-1.231
(0.791)

-1.944**
(0.787)

Mountainous Terrain

0.079** 0.083**
(0.042) (0.040)

British Colony

-0.064
(0.092)

-0.047
(0.091)

Ethnic Fractionalization

0.106
(0.249)

0.094
(0.267)

Primary Education

-0.180
(0.115)

-0.188
(0.123)

Overidentification

0.2051

0.5286

0.8222

Country Fixed Effects

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Year Effects and Trends

No

No

No

Yes

No

Yes

Yes

Yes

Yes

No Observations

888

888

888

888

888

888

888

888

888

Note: Method of estimation in columns (1)-(4) is least squares with Huber robust standard errors (listed in parentheses) clustered at the country level;
columns (5)-(9) two-stage least squares. In columns (5), (6), and (9) double instrumentation of population size and per capita GDP is implemented by
using the drought indicator, the log level of rainfall, and the log level of the international commodity price index as instrumental variables. In
columns (7) and (8) only population size is treated as an endogenous variable, instrumented by the drought indicator. The Hansen J overidentification
test is provided for all second stage regressions in form of p-values. *Significantly different from zero at 90 percent confidence, ** 95 percent
confidence, *** 99 percent confidence.

Table IV
Population Size, Per Capita GDP, and Civil Conflict Onset
Civil Conflict Onset
(1)

(2)

(3)

(4)

LS

LS

2SLS

2SLS

Log Population, t-1

-0.004
(0.056)

-0.234
(-0.78)

0.720*
(0.381)

5.179**
(2.587)

Log per capita GDP, t-1

0.001
(0.030)

0.148
(0.054)

-0.932*
(0.532)

-1.376**
(0.576)

0.5269

0.4432

Country Fixed Effects

Yes

Yes

Yes

Yes

Year Effects and Trends

No

Yes

No

Yes

No Observations

888

888

888

888

Overidentification

Note: Method of estimation in columns (1)-(2) is least squares with Huber robust standard errors (listed in parentheses) clustered at the country level;
columns (3)-(4) two-stage least squares. The excluded instruments in the 2SLS regressions are the drought indicator variable, the log level of rainfall,
and the log level of the international commodity price index, all lagged one period. The Hansen J overidentification test is provided for all second
stage regressions in form of p-values. *Significantly different from zero at 90 percent confidence, ** 95 percent confidence, *** 99 percent
confidence.

23

Table V
Population Size, Per Capita GDP, and Civil Conflict
Conflict Incidence

Conflict Onset

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

Fuller 1

Fuller 4

Fuller 1

Fuller 4

Fuller 1

Fuller 4

Fuller 1

Fuller 4

Log Population, t-1

1.272**
(0.586)

1.033***
(0.391)

5.980*
(3.551)

4.134*
(2.472)

0.658**
(0.333)

0.489**
(0.225)

4.773**
(2.334)

3.387**
(1.574)

Log per capita GDP, t-1

-1.273
(0.832)

-0.923*
(0.511)

-1.824**
(0.725)

-1.453**
(0.556)

-0.838*
(0.458)

-0.584**
(0.287)

-1.301**
(0.536)

-1.029**
(0.411)

Country Fixed Effects

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Year Effects and Trends

No

No

Yes

Yes

No

No

Yes

Yes

No Observations

888

888

888

888

888

888

888

888

Note: Method of estimation is the Fuller limited information maximum likelihood estimator. The excluded instruments are the drought indicator
variable, the log level of rainfall, and the log level of the international commodity price index, all lagged one period. Standard errors are provided in
parentheses. *Significantly different from zero at 90 percent confidence, ** 95 percent confidence, *** 99 percent confidence.

Table VI
Population Size, Per Capita GDP, and Civil Conflict
10% Drought Shock

EM-DAT Drought

EM-DAT & 5% Joint

Panel A: Country Fixed Effects

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

2SLS

Fuller 1

Fuller 4

2SLS

Fuller 1

Fuller 4

2SLS

Fuller 1

Fuller 4

Log Population, t-1

2.188*
(1.277)

1.801** 1.271*** 2.148*

(0.835) (0.417) (1.278)

1.805** 1.267*** 1.615**

(0.886) (0.447) (0.840)

1.511**
(0.732)

1.093***
(0.394)

Log per capita GDP,

t-1

-2.555
(1.829)

-1.990* -1.225**
(1.167) (0.523)

-2.466
(1.656)

-2.004* -1.284**
(1.127) (0.528)

-1.741
(1.159)

-1.592
(1.000)

-1.000**
(0.483)

Overidentification

0.7974

0.7966

0.7966

0.7825

0.7698

0.7698

0.3234

0.4212

0.4212

No Observations

888

888

888

888

888

888

888

888

888

Panel B: Country Fixed Effects + Time Controls

Log Population, t-1

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

(9)

2SLS

Fuller 1

Fuller 4

2SLS

Fuller 1

Fuller 4

2SLS

Fuller 1

Fuller 4

6.534
(1.25)

5.605
(4.420)

3.301
(2.714)

8.474
5.377
(10.330) (6.176)

0.984
(2.287)

8.166
(9.375)

5.275
(5.801)

1.350
(2.310)

Log per capita GDP, -1.936** -1.785** -1.375** -2.182* -1.793** -1.161**
t-1
(-2.23)
(0.767) (0.553) (1.328) (0.847) (0.457)

-2.078*
(1.130)

-1.743** -1.208**
(0.773)
(0.473)

Overidentification

0.5153

0.5819

0.5819

0.5492

0.5016

0.5016

0.4694

0.5637

0.5637

No Observations

888

888

888

888

888

888

888

888

888

Note: Method of estimation in columns (1), (4), and (7) is two-stage least squares; in columns (2), (3), (5), (6), (8), and (9) Fuller limited information
maximum likelihood. The excluded instruments are the drought indicator variable, the log level of rainfall, and the log level of the international
commodity price index, all lagged one period. Control variables in Panel A are country fixed effects. Panel B includes in addition to the country
fixed effects country specific time trends and year fixed effects. Standard errors are provided in parentheses. The Hansen J overidentification test is
provided for all second stage regressions in form of p-values. *Significantly different from zero at 90 percent confidence, ** 95 percent confidence,
*** 99 percent confidence.

24

Appendix Table I
List of Sub-Saharan Drought (1980-2004)
Country

Period of Rainfall
Shortage

Number of People Affected (EM-DAT, 2008)

Angola

1991-1992

1,900,000

Botswana

1989-1990

100,000

Botswana

2001-2003

Chad

1990-1991

300,000

Ethiopia

1983-1984

7,750,000

Gambia

1982-1983

400,000

Gambia

1989-1990

Guinea

1982-1983

Kenya

1983-1984

600,000

Kenya

1989-1999

1,600,000

Kenya

1999-2000

21,000,000

Malawi

1986-1987

1,430,000

Mali

1995-1996

Mauritania

1982-1983

1,600,000

Mauritania

1990-1992

Mauritania

1995-1996

447,000

Mozambique

1985-1986

4,750,000

Niger

1987-1988

1,000,000

Niger

1990-1991

1,630,000

Niger

1997-1998

8,500

Senegal

1983-1984

1,200,000

Senegal

1990-1991

Somalia

1980-1985

500,000

Somalia

1990-1991

Somalia

1999-2000

2,000,000

South Africa

1989-1990

1,320,000

South Africa

2003-2004

15,000,000

Sudan

1998-1999

Swaziland

1985-1986

Swaziland

1991-1992

200,000

Swaziland

2001-2003

970,000

Uganda

1983-1984

Zimbabwe

1982-1983

700,000

Zimbabwe

1991-1992

5,000,000

Zimbabwe

2001-2003

6,000,000

Note: The listed droughts are based on an indicator function that is one for the 5% largest drops in the level of rainfall over two consecutive years.
Column (3) provides an estimate of the number of people affected according to EM-DAT (2008). A dash implies that EM-DAT did not list the event
as drought; an asterix means that EM-DAT coded the event as drought without giving an estimate of the number of people affected.

25