Economic Shocks and Civil Conflict: A Comment

TLDR: In this article, the authors argue that conflict is unrelated to rainfall in Sub-Saharan Africa and that this finding is driven by a positive correlation between conflict in t and rainfall levels in t-2.

Abstract: Miguel, Satyanath, and Sergenti (2004) argue that lower rainfall levels and negative rainfall shocks increase conflict risk in Sub-Saharan Africa. This conclusion rests on their finding of a negative correlation between conflict in t and rainfall growth between t-1 and t-2. I argue that this finding is driven by a positive correlation between conflict in t and rainfall levels in t-2. If lower rainfall levels or negative rainfall shocks increased conflict, one might have expected MSS’s finding to reflect a negative correlation between conflict in t and rainfall levels in t-1. In the latest data, conflict is unrelated to rainfall.

Figures

Table 1. Rainfall and civil conflict onset

Table 3. Rainfall and civil war onset or incidence 1979-2009

Table 2. Rainfall and civil conflict incidence

Full text

Economic Shocks and Civil Conflict:
A Comment
by
Antonio Ciccone*
February 2011
Abstract
Miguel, Satyanath, and Sergenti (2004) argue that lower
rainfall levels and negative rainfall shocks increase conflict
risk in Sub-Saharan Africa. This conclusion rests on their
finding of a negative correlation between conflict in t and
rainfall growth between t-1 and t-2. I argue that this finding
is driven by a positive correlation between conflict in t and
rainfall levels in t-2. If lower rainfall levels or negative
rainfall shocks increased conflict, one might have expected
MSS’s finding to reflect a negative correlation between
conflict in t and rainfall levels in t-1. In the latest data,
conflict is unrelated to rainfall.
Key words: Transitory shocks, mean reversion, civil conflict
JEL codes: O0, P0, Q0
_____________________________________________________________________________________
* ICREA-Universitat Pompeu Fabra and Barcelona GSE, antonio.ciccone@upf.edu
. This paper was
previously entitled “Transitory Economic Shocks and Civil Conflict.” I am grateful to Natalie Chen,
Esther Duflo, Edward Miguel, Shanker Satyanath, Ernest Sergenti, Kurt Schmidheiny, two referees, and
participants at the “Conflicts, Globalization, and Development” CEPR/PSE Workshop for comments. I
also thank Halvard Buhaug for sharing his temperature data, Philipp Ager and Walter Garcia-Fontes for
help with the rainfall data, and Andrea Tesei for excellent research assistance. Research support from
CREI, FEDEA-BBVA, and Spanish Ministry of Science grants SEJ2007-64340 and ECO2008-02779 is
gratefully acknowledged.

1
Does poor economic performance cause violent civil conflict? Collier and Hoeffler’s (1998,
2004) and Fearon and Laitin’s (2003) empirical work suggests this is the case. Their findings
are not based on exogenous changes in the economic environment however, and could reflect
feedback from conflict to economic performance or omitted social and political factors. To
address these concerns, Miguel, Satyanath, and Sergenti (MSS, 2004) examine the link
between (exogenous) rainfall and civil conflict in Sub-Saharan Africa 1979-1999. Their
empirics lead them to the conclusion that higher levels of rainfall are associated with
significantly less conflict (e.g. MSS, page 737). Or, equivalently, lower rainfall levels are
associated with significantly more conflict. MSS explain this association by negative rainfall
shocks reducing incomes and thereby increasing conflict risk. Their focus on exogenous
rainfall shocks is an important step forward. MSS’s study has advanced quickly to one of the
most cited articles on civil conflict, and their conclusion has become a cornerstone of the
literature on the economics of civil conflict (e.g. Collier and Hoeffler, 2005; Collier, Hoeffler,
and Rohner, 2009; Hegre and Sambanis, 2006; Fisman and Miguel, 2009).
1
MSS’s (2004) interpretation of the Sub-Saharan African rainfall and civil conflict data
rests on their finding of a negative correlation between civil conflict in year t and the year-on-
year rainfall growth rate between t-1 and t-2. I argue that this finding is driven by a positive
correlation between conflict in year t and rainfall levels in year t-2. If conflict was triggered by
lower rainfall levels or negative rainfall shocks, one would have expected the negative
correlation found by MSS to reflect a negative correlation between conflict in t and rainfall
1
MSS is the 7th most cited article on the topics civil conflict or civil war in history,
economics, political sciences, and sociology according to the ISI Web of Knowledge
http://isiwebofknowledge.com/.

2
levels in t-1. As civil conflict risk in MSS’s data is not higher following low rainfall levels or
negative rainfall shocks, I argue that MSS’s interpretation is an artifact of their empirical
approach. The latest available datasets on rainfall and civil conflict in Sub-Saharan Africa
have been extended to 2009. In these data, civil conflict risk is unrelated to year-on-year
rainfall growth, rainfall levels, or rainfall shocks. This suggests that uncovering an effect of
rainfall on civil conflict will require using more disaggregated data.
The main difference between my empirical approach and the empirical approach of MSS is
that I focus on the correlation between civil conflict and current as well as lagged rainfall
levels. To see why, it is useful to start with the question whether lower rainfall levels are
associated with more or with less civil conflict. MSS answer this question by examining the
correlation between civil conflict and current as well as past year-on-year rainfall growth,
while I address this question by examining the correlation between civil conflict and current as
well as past rainfall levels.
2
These two empirical approaches (growth versus levels) need not
yield the same answer as years with low year-on-year rainfall growth need not be years where
rainfall levels are low. For example, low year-on-year rainfall growth may reflect that rainfall
levels fell from a high level to the historical mean. To understand why MSS and I also reach
different conclusions regarding the link between rainfall shocks and civil conflict, it is
important to note that MSS’s empirical approach uses year-on-year rainfall growth as a
measure of rainfall shocks (MSS, page 733). As rainfall levels are strongly mean reverting,
rainfall growth between two years t and t-1 may be low because of a negative rainfall shock in
t or because of a positive rainfall shock in t-1. Put differently, because rainfall shocks are
transitory, low rainfall growth may reflect negative shocks or mean reversion following
2
Burke, Miguel, Satyanath, Dykema, and Lobell’s (2009) and Buhaug’s (2010) investigation
of the effect of global warming on civil war risk in Africa also focuses on rainfall levels rather
than rainfall growth rates.

3
positive shocks. Inferring the effect of rainfall (transitory) shocks on civil conflict from the
effect of year-on-year rainfall growth may therefore be misleading.
To see these points more precisely, it is useful to consider MSS’s linear-probability model
linking civil conflict to rainfall. Their model predicts the probability of civil conflict
t
PPconflict in year t based on current and lagged year-on-year rainfall growth,
(1)
1
LS LS
ttt
PPconflict a RGr b RGr
−
=+ ,
where
t
R
Gr is rainfall growth between year t and t-1, and
LS
a ,
LS
b are least-squares estimates.
MSS find a statistically insignificant value for
LS
a and a statistically significant, negative
value for
LS
b . They then use this finding to make inferences about the effect of rainfall levels
and rainfall shocks on conflict. To examine whether such inferences are feasible, suppose that
rainfall levels are distributed identically and independently over time. This implies that rainfall
levels are strongly mean reverting and that rainfall shocks are transitory.
3
Suppose also that the
true probability of conflict
t
Pconflict depends on current and lagged log rainfall levels
(log
R
),
(2)
01122
log log log
ttt t
Pconflict R R R
α
αα
−
−
=
++ .
If 0
i
α
> for 0,1,2i = , the probability of conflict is lower following low rainfall levels and
negative rainfall shocks (lower than expected rainfall levels).
4
Now imagine running a least-
squares regression to predict the probability of conflict based on current and lagged year-on-
3
Empirically, rainfall levels are strongly mean reverting. For example, regressing log rainfall
levels on lagged log rainfall levels using MSS’s data and controlling for country fixed effects,
yields a system-GMM coefficient on lagged log rain of 0.17 with a standard error of 0.04.
Accounting for the empirical persistence of rainfall does not affect the conclusion but
complicates the coefficient formulas in (3).
4
For an insightful theoretical analysis of the link between transitory economic shocks and
civil conflict see Chassang and Padró i Miquel (2009).

4
year rainfall growth as in (1). The coefficients of this regression would be determined by the
usual least-squares orthogonality conditions:
1
cov( , )
LS LS
tttt
Pconflict a RGr b RGr RGr
−
−− =
11
cov( , )
LS LS
tttt
Pconflict a RGr b RGr RGr
−−
−− 0
=
.
Making use of (2) and
1
log log
ttt
R
Gr R R
−
=− in these conditions yields
(3)
012 01 2
2( ) ( )2
and
33
LS LS
ab
α
αα αα α
−
++−
==.
5
Hence, the least-squares estimates for
LS
a and
LS
b in (1) are mixtures of the parameters
i
α
determining the effects of rainfall levels and rainfall shocks on the probability of civil conflict
in (2). As a result, the coefficients of the rainfall-growth specification in (1) are uninformative
about the effect of rainfall levels or shocks and using them to make inferences about the effect
of rainfall levels or shocks may be misleading. For example,
LS
b in (3) will be negative as
long as
201
2
α
αα
>+. Hence,
LS
b may be negative although lower rainfall levels and negative
rainfall shocks reduce conflict at all lags, i.e. 0
i
α
> for 0,1,2i
=
in (2). It is even possible that
both
LS
a and
LS
b are negative although lower rainfall levels and negative rainfall shocks
reduce the probability of conflict at all lags. To see this, note that both coefficients in (3) will
be negative if and only if
010
2
α
θα α θ
−< < + where
20
θ
αα
=
− .
6
As a result, if
20
0
α
α
>>,
5
The assumption that log rain is i.i.d. implies that
1
cov( , )
LS LS
tttt
Pconflict a RGr b RGr RGr
−
−−
simplifies to
01
(2)
LS LS
abV
αα
−− + and that
11
cov( , )
LS LS
tttt
Pconflict a RGr b RGr RGr
−−
−−
simplifies to
12
(2)
LS LS
abV
αα
−+ −
where V is the variance of log rain. Hence, (3) can be
obtained by solving
01
2
LS LS
ab
αα
−− + =
12
20
LS LS
ab
αα
−
+− = for
LS
a ,
LS
b . In practice
one does not observe the probability of conflict but only whether there has been a conflict or
not. This does not affect (3) however, see Wooldridge (2002), page 454.
6
Lagged rain growth will enter negatively if and only if
201
2
α
αα
>+, or equivalently
120 0
22
α
αα θα
<−=+. Current rain growth will enter negatively if and only if
012
2
α
αα
<+,
or equivalently
1020
2
α
αα αθ
>−=−. Combining inequalities yields
010
2
α
θα α θ
−< < + .

Frequently asked questions

Q1. What is the likely cause of conflict onset?

The estimate indicating that conflict onset is less likely following lower rainfall levels and negative rainfall shocks is significant at the 95% confidence level, no matter which standard error is used. 

Q2. What is the significance of the results in column (4)?

According to the results in column (4), civil war onset is either significantly less likely following low rainfall levels and negative rainfall shocks in t-1 or unrelated to rainfall levels and shocks, depending on the standard error used. 

Q3. What is the implicit assumption when using conflict incidence instead of conflict onset?

the implicit assumption when using conflict incidence instead of conflict onset as the dependent variable is that rainfall affects conflict onset and conflict continuation in the same way. 

Q4. What is the earliest civil conflict onset indicator?

(As MSS control for contemporaneous and lagged year-on-year rainfall growth and the GPCP rainfall data start in 1979, the earliest civil conflict onset observations employed correspond to 1981.) 

Q5. What is the onset indicator for a conflict?

The onset indicator in year t is 1 if there is a civil conflict in t but there was no conflict in t-1; 0 if there is no conflict in t and there was no conflict in t-1; and not defined if there was a conflict8 See Gleditsch, Wallensteen, Sollenberg, and Strand (2002). 

Q6. What is the link between civil conflict onset and rainfall?

When The authorcontrol for rainfall and temperature, there is some evidence that civil conflict onset over the 1979- 2008 period is less likely following low rainfall levels and negative rainfall shocks. 

Q7. What does the author do with the data?

Column (1) shows that a least-squares regression of war onset in year t on current and lagged rainfall growth yields a significantly negative coefficient on rainfall growthMiguel and Satyanath do not use these data and lose this observation. 

Q8. What is the correlation between civil conflict and rainfall levels in t-1?

If civil conflict was triggered by lower rainfall levels or negative rainfall shocks, one would have expected the negative correlation found by MSS to reflect a negative correlation between civil conflict in t and rainfall levels in t-1. 

Q9. What is the coefficient of year t-1 rainfall growth?

Column (1) shows that a leastsquares regression of conflict onset in year t on current and lagged year-on-year rainfall growth yields a significantly negative coefficient on year t-1 rainfall growth (year-on-year rainfall growth between t-1 and t-2). 

Q10. What is the difference between the two conclusions of the authors?

26Two of the conclusions of Miguel, Satyanath, and Sergenti’s (2004) study of civil conflict and rainfall in Sub-Saharan Africa are that lower rainfall levels and adverse rainfall shocks increase conflict risk.