No comment yet

Marshall B. Burke et al. Warming increases the risk of civil war in Africa

Halvard Buhaug, Climate not to blame for African civil wars


Burke et al. (2009)’s study concentrated on the correlation between the intensity of warfare in Africa and the temperature. Burke et al. (2009) proposed an empirical model to find such correlation. Based on their model, Burke et al. (2009) were able to extrapolate data into 2030 and their finding suggested that in longer term, climate increase will outweigh other offsetting effects such as continued democratization and intensify armed conflicts across Africa, therefore urged a reform of governments and foreign aid policies to deal with rising temperature.

Burke et al. (2009) used a linear empirical model to reflect the relevance between climate change and the number of armed conflicts per country. The model defines armed conflicts as conflicts that directly results > 1000 casualties. A compensate factor ci were used to adjust bias towards some characteristics of certain country. Yearly correlated factor are adjusted with diyeari. The overall model can be represented simply as

〖war〗_it=f(x_it )+c_i+d_i 〖year〗_i+ε_it

The temperature variable xit is a computed average over whole year and a whole country.

Burke et al. (2009) spent large part of their article on reflecting the implications of their finding. Particularly, they used public data about prediction of future climate change to reason about future number of armed conflicts. Their calculation suggested of a 54% rise in average likelihood of conflict across Africa.

In later discussion, Burke et al. (2009) argued that comparing to previous studies, temperature served as a better indicator than precipitation. They went further and argued that since temperature and precipitation are negatively correlated, previous finding of increased conflict during drier years might be partly capturing the effect of hotter year. On speculation of why temperature caused increased conflict, Burke et al. (2009) provided the context of how temperature interacted with agriculture. Temperature affects agricultural yields both through increases in crop evapotraspiration and through accelerated crop development which accounts for 10% - 30% crop yields per ˚C of Africa. Because the poorest households in rural area Africa derive 60% - 100% of their income from agricultural activities, temperature will have strong economic consequence to them. From previous studies we knew that economic welfare is the single most significant factor for conflict incidence, Burke et al. (2009) were able to establish the link between temperature and conflict incidence.

Burke et al. (2009) believed so strongly in this link that in their conclusion, they proposed two solutions which directly addressed the agricultural part. First, aid donors and governments should improve Africa agriculture to deal with extreme heat. Second, implementing an insurance scheme to protect poor societies from climate shocks therefore reduce the risk of civil war in Afirca.


Buhaug (2010) examined the computation model and several arguments about how climate change would intensify the conflicts in Africa. His investigation argued that the climate change is a poor predictor of armed conflict. Instead, widespread ethno-political exclusion, poor national economy, and the collapse of the Cold War system are sufficient to explain Africa civil wars.

Buhaug (2010) started by examining assumptions Burke et al. (2009) made in their computation model. As we noted in the previous section, Burke et al. (2009) only marked war as conflict with > 1000 casualties. Moreover, conflicts are discretized across year. For example, the Sierra Leone civil war is widely accepted as lasting from March 1991 to late 2000, but in Burke et al. (2009)’s model, the war is only accounted for year 1997-1998. Buhaug (2010) argued that it would make little sense to account the causality of a war lasting 8 years to the climate change in the middle of 2 years. In Burke et al. (2009)’s finding, Buhaug (2010) also questioned the ignorant on the data of conflict after 2002 because the civil war incidence and severity have decreased while the global warming effect continues.

In later chapter, Buhaug (2010) presented a more comprehensive and arguably more thoughtful analysis of the relevance between civil war risk in Sub-Saharan Africa and short-term climate variations. The analysis takes into account of conflicts that far lower than 1000 annual casualties (minimum of 25 annual battle deaths) and distinguished the outbreak of war with the incidence of war. A particular attention was paid to address the climate change before the initiation of conflict.

Surprisingly, Buhaug (2010) found that the preliminary inspection supported the climate-driving conflict claim. However, with sensitivity analysis, which is common to examine the robustness of a model, Buhaug (2010) pointed out how fragile such model is. With more dependent variables have been introduced to the model, Buhaug (2010) argued that the collapse of the Cold War system was the major contributor to civil war in Africa with significant margin on statistic certainty (p < 5%). Such effect can also be explained from political viewpoint due to the increasing concern on national security. In the following discussion, Buhaug (2010) attacked the frangibility of such model for which, by slight modification, the coefficients jumped back and forth from positive to negative all with uncertainty below 5%. Even more, the data directly shows that an unusually wet period followed by more conflicts (> 1000 causalities) which is contradict to the notion of scarcity-induced conflicts. In addition, Buhaug (2010) addressed some pitfalls in his own modeling. First, he argued that in all empirical studies on this subject, they applied country-level averaged climate data which may mask out local anomalies. If it turns out such local variance is large enough, a different conclusion may be drawn. Second, his own study is only focusing on short-term climate change but long-term environment change may have more security implications, and that may change the political environment accordingly. Third, the intensity of global warming impact may be undervalued since the famous hockey stick graph suggested a dramatically change in temperature thus may trigger some major tipping point events. If that happens, this presented analysis would be invalid.

In the closing section, Buhaug (2010) suggested that even though global warming is a real challenge, letting the debate based on some unverified and nonrobust scientific findings would be too daunting to make real progress.


In social studies, empirical data is usually hard to obtain, and the correlations between them is hard to serve as an argument for causality analysis. There are usually three reasons for a pair of correlation which have significant certainty to appear. First, one side of the equation is the cause, and the other side is the consequence. Second, both sides are the consequences of a “hidden” cause. Third, an unlikely random event that incurred the significant certainty happened due to many tries and errors. Social scientists paid a lot of attentions to distinguish the cause from consequence. For example, it usually indicates a cause if one side of the equation is hard to change. If both sides are easy to change, tracing and reasoning the link between them becomes crucial. However, much less attentions are paid to distinguish from the third reason.

Burke et al. (2009)’s analysis is arguably unconventional. They took climate change and tried to recover the association with the number of conflicts in Africa. It is obvious that since temperature is hard to change by artificial events, if such correlation does exist, we can safely attribute temperature as the cause of conflicts in Africa.

Though the conclusion is eye-catchy and Burke et al. (2009) went further to extrapolate their data into future in order to gain some insights into political choices, they failed to address some fundamental issues of their model.

The discretization method for counting war is arbitrary. Burke et al. (2009) didn’t provide any justification for the choice of 1000 as threshold. The civil war is counted as a single one instead of treating them as individual battles. Although the outbreak of a civil war took unusual momentum to pull off, the intensity of such war is only relevant to individual battles. For a linear regression model which essentially would only fit by painting intensity of warfare, such ignorant is worrisome.

To further assess their argument, Burke et al. (2009) tried to draw the actual causality chain out and their temperature-agriculture-economic-warfare chain seems convincing. However, it brings up another question: if agriculture is so important, why don’t analyze the relationship between agricultural activities and warfare at first place? The causality chain also pointed out some weaknesses in their model. For example, the variables in the model such as temperature and GDP are actually dependent on each other. Such existence of dependency would diminish their finding because now, for prediction, the cascading effect (one variable change would cause other variables to be adjusted accordingly) would not be properly captured in the existing model when plug in one variable (the linear regression model assumes independency of its variables).

The trouble of interdependence between variables gets reflected on their future projection of civil war risk in Africa. If temperature has a direct impact on economic growth as they suggested, the offsetting effect of economic growth they mentioned later will collide with the former one, it is unclear which side will win since the dependency is not properly analyzed. By having such dependency of variables, they entered the loop of reasoning and cannot get any sound conclusion out.


Buhaug (2010) attacked Burke et al. (2009) in a more constructed way than mine. The most significant contribution in his paper is to rebuild and reexamine Burke et al. (2009)’s model. The sensitivity analysis confirmed my worrying on several seeming arbitrary choices in original model.

To go beyond reexamining original model, Buhaug (2010) also presented a modified version which in principle should reveal more insights into the causality of civil war in Sub-Saharan Africa. The irrelevancy of temperature in his new model is expected based on his analysis of original model even he included more aspects of short-term climate change (first derivative of temperature etc.) in the new model. It seems like a second confirmation of how fragile Burke et al. (2009)’s model is.

It is very interesting to see a model that consists of a lot variables, however, it should be noted that, to make such a complex model robust, a large number of training data is required. This is known as overfitting problem to statisticians. However, to solve the overfitting problem, Buhaug (2010) used linear regression model rather than logistic regression model. The choice is a source of trouble. For me, it is inconsistent since he specifically addressed the problem of using linear regression model in Burke et al. (2009)’s paper (mainly because in previous step, we discretized/binarized warfare as target function).

Buhaug (2010) took a lengthy paragraph to reason why the discretization method of warfare used in Burke et al. (2009) was broken. His argument deserves some merits, but I don’t find that is as important as he argued. Though the outbreak of civil war is significant, the intensity of a certain war is never be a function of how severe the outbreak is. The World War I started as revenge to political murder but the argument for the war quickly shifted away. A civil war should be viewed as a collection of battles, thus, the severity of certain battle can be determined by other factors in that time frame. If Burke et al. (2009)’s analysis is robust enough, such discretization method should be OK for their model.


Since more and more data has been gathered thanks to the advancement of information technology, social scientists and alike started to use statistic model to reason about certain events. However, extreme caution should be taken in the process and certain mathematic competency should be required in order to avoid faulty analysis.

War is a serious business. Doing causality analysis for warfare does not only require macro-scoped statistics, also requires detailed field work. Social science doesn’t share the luxury with physics of doing controlled-variable experiment. Thus, any study consists of empirical analysis should be carefully carried out.

In short, there is no magic dose for the Africa problem.

blog comments powered by Disqus