Results

In Fig. 1, we map the spatial distribution of key violence data together with temperature anomalies [6-mo temperature index (TI6)] and 6-mo precipitation anomalies [standard precipitation index (SPI6)] for the study area. The sum of violent events by grid over the 33-y study period, 1980–2012, is shown in Fig. 1A with the increasing size of the cartographic breaks reflecting the highly skewed distribution of these data. We use the Armed Conflict Location and Event Dataset (ACLED), based on media reports (32), for our analysis. In SI Text, we also compare our results in a robustness check to the Uppsala Conflict Data Program Georeferenced Event Data (33) for the limited period (1989–2010) where these data are available. Fig. 1 B and C show the spatial distribution of hot and dry grid-months using a value of 1 SD to define these thresholds. Our climatological data are from the University of East Anglia Climate Research Unit (34). Assuming a normal data distribution, we normalize these observed counts by the expected number of dry and hot grid-months. Thus, grids with an observed-to-expected ratio of 2 had twice as many hot or dry months than expected. The maps clearly show that, during the last 33 y, most of sub-Saharan Africa has been much warmer than expected compared with long-term trends (Fig. 1B) and many parts (e.g., the Sahel and southern Africa) have been dryer than usual (Fig. 1C).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

Descriptive maps of (A) ACLED violent events, (B) hot temperature anomalies, and (C) dry precipitation anomalies, 1980–2012. For temperature (TI6) and precipitation (SPI6) data, the number of observed grid-months beyond 1 SD is compared with the expected number from a standard normal distribution. Distribution plots of the TI6 and SPI6 values and the grid means are shown in Figs. S1 and S2.

To isolate and comparatively assess the effects of climate variability on violence, all of our models incorporate important social, political, and economic indicators that have been associated with violence in previous conflict studies research (see Table S1 for data sources, spatial and temporal resolutions, and aggregation rules; see Table S2 for summary statistics). All data are measured with a temporal lag so that the reported results are not biased by endogeneity between key indicators and conflict as an outcome.

Using a quasiexperimental matching analysis, we examine temperature and rainfall anomalies in relation to the specific political–economic settings and we can determine whether climate conditions are associated with variation in the level of violence observed across the continent. We use coarsened exact matching (CEM) (35) to drop outlying observations for key social indicators in the quasiexperimental analysis data. Our goal is to evaluate a treatment effect (TE) of climate variability after controlling for alternative economic and political explanations of conflict occurrence. We also present the simple differences in the average conflict event counts for each treatment condition (Table S3). In Fig. 2, we report the TE estimates and balance improvement statistics (E, Bottom). For each temperature and precipitation TE estimate, we control for the possible influences of the other climate indicator (e.g., the extremely dry precipitation TE is measured after controlling for temperature). We define extreme values as ±2 SD of the long-term values (Fig. S3 shows the maps for extremely dry, SPI6 ≤ −2, and extremely hot grid months, TI6 ≥ 2), but we also report smaller extremes with the effects for ±1 SD indicated.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Coarsened exact matching treatment effects (CEM TE) results for the effect of climate variability on conflict in sub-Saharan Africa, 1980–2012. See Table S1 for listing of controls and maps (Fig. S3) of binary climatological conditions. The relationship is only statistically significant when the 95% confidence interval (gray horizontal line) does not overlap the vertical dashed line representing no effect. (A) Climate variability effects for all types of violence. (B–D) Effects for the subsets of violence, battles, riots and protests, and violence against civilians, respectively.

The TE estimates indicate that extremely high temperatures increase the level of conflict. TE is interpreted as the change in number of events per grid-month that can be attributed to the corresponding climate indicator (for extremely high temperatures, the value is 0.008 greater events). Colder anomalies are associated with dramatically reduced levels of conflict (0.028 fewer events). Our precipitation result suggests that dry conditions reduce the level of conflict (0.008 fewer events). In Fig. 2 B–D, we compare the effects of these indicators across different types of conflict because climate anomalies are not expected to influence all forms of violence equally. The additional models show that the main high temperature result (Fig. 2A) is driven primarily by the effect for violence against civilians (Fig. 2D), and the relationship does not emerge for any of the other forms of violence in our database. This is a noteworthy result considering that existing studies [e.g., Hsiang et al. (3)] have been based on a problematic conflation of dramatically varying forms of violence. Violence against civilians may drive the high temperature effect because, at the fine spatial and temporal resolution that we use, responses to weather stress are likely to be immediate rather than long term in their scope. Such conflict would be characterized by loose organizational structure (not interactions between two armed actors, or “battles”) and may involve land seizures or cattle-raiding activity rather than cohesive (e.g., rebel group) actor formation and organized warfare (e.g., against a government), qualities that characterize civil-war definitions of conflict used in other studies (as in ref. 7).

Dry precipitation anomalies have the effect of reducing the number of violent events between two armed actors (Fig. 2B) and in the main model (Fig. 2A), but extremely dry conditions increase the observed level of rioting (Fig. 2C) and violence against civilians (Fig. 2D). In a similarly inconsistent conclusion that further illustrates the importance of disaggregating violence, the combined effect of extremely dry and extremely hot conditions reduces the number of observed battle events (Fig. 2B) but elevates the prevalence of rioting and protesting (Fig. 2C).

We also aim to advance the current understanding of how climate and conflict operate alongside other characteristics of social settings by following the multilevel modeling approach (MLM) recently published by Bell and Jones (36). For time-variant explanatory variables, this approach separates out within-unit and between-unit variance to allow a more nuanced explanation of the predictor’s effect. For instance, in considering the effect of temperature anomalies on violence, we would like to know if there is a relationship due to a grid cell being consistently hotter than other cells (1980–2012 grid mean compared with its climatological mean, 1949–2012, the between effect), and/or if the effect is due to increasing temperatures within that grid cell (the within effect). For each time-variant indicator, we calculate a grid mean to extract the between-unit variance and then subtract this mean from the original value to generate the demeaned (within) variance. Although researchers often prefer observation unit fixed effects in statistical analysis (controlling away all unknown influences upon the outcome of interest), we select a multilevel model that retains all variables of interest.

While the grid means for the temperature and precipitation anomalies are close to zero in our data, there is sufficient variation to warrant their inclusion in the model (see maps and distributions in Figs. S1 and S2). With the geographic cells as the first level, the second level for our MLM–random effects (MLM-RE) model is the country. To improve ease of model estimation, the infant mortality rates were rescaled to mean 0 and SD 1.

Our overall conclusion from the MLM-RE analysis is that climate variability effects are either absent or are quite specific to higher temperatures over time in an area (within effects) (Table 1). No between grids effect is significant but hotter periods for a given geographic cell are associated with a higher risk of violence. For every 1 SD increase in temperature, we expect an increase of 0.102 violent events, corresponding to a 10.7% increase in the level of violence. We find no statistically significant influence of SPI6 (precipitation measure) upon conflict, either as a between or within effect. Stated another way, cells that are wetter or drier than their long-term means are neither more or less likely to experience a greater number of violent events, and dry and wet periods within a grid cell are also not associated with a change in violence risk. Recent nearby political violence (space–time lag) is a strong indicator of violence in a given cell location, predicting 0.184 more violent events in an area (for each unit increase in neighboring violence in the previous month, a 20.2% increase in violence is predicted). Similarly, those areas in less democratic states (higher value for political rights indicator) experience more violence. Cells that include national capitals experience a greater risk of violence than other regions of the country. Table 1 also reports the null model, with no predictors, to show how much the country-level variance is reduced after adding the explanatory variables to the model.

We explore the robustness of these results by using a linear functional form (Fig. S4 and Table S4), by testing ACLED violent event subsets (i.e., riots/protest, violence against civilians, and battles; Table S5, models i–iii), and by testing with the independently coded Uppsala Conflict Data Program Georeferenced Event Data (UCDP-GED) violent event dataset (Table S5, models iv and v). These checks reveal that the temperature effect is not consistently robust to the alternate ACLED data specifications. To see whether any particular region is driving the relationships we observe, we divide sub-Saharan Africa into geographical regions (Fig. S5) and find that of the five regional subdivisions, the temperature effect is statistically significant only for the Sahel region (Table S6). This result suggests that temperature anomalies in the Sahel region are driving the overall continent-wide relationship. We also test an alternate governance indicator [Polity (37)] and nonlinear governance effects, with little impact to the model (Table S7).

Predictive power analysis allows us to present improvements in the in-sample predictive power of the model as we systematically exclude each indicator in turn (38). Instead of using SEs (the usual metric) to judge the importance of a given predictor, this approach shows the relative predictive capability of each variable to explain violence. We plot the level of significance (absolute z value) of each indicator on the horizontal axis against the effect that the indicator has on the ability of our model to accurately predict conflict events on the vertical axis. Indicators near the zero dashed line have little influence on our ability to predict the timing and location of violence, whereas variables above the line improve the accuracy of conflict predictions.

From Fig. 3, we conclude that existing conflict in neighboring areas (time–space lag), socioeconomic well-being (within cell effect), distance to national borders, capital city location, and political rights (within-cell effect) are all better predictors of observed conflict levels than either of the climate variability measures, temperature or precipitation. Although changing temperature (TI6, within-cell effect) has a larger z score than roughly one-half of the other variables, it lies very near the zero line, indicating that it provides little improvement in the predictive power of the model. Whether or not an area experienced precipitation anomalies (SPI6) has little value for understanding where and when conflict takes place compared with other alternative explanations of conflict. These general findings are confirmed for a linear functional form of the relationships (Table S4 and Fig. S4).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

In-sample predictive power plot of the violent event count Poisson MLM-RE model for sub-Saharan Africa, 1980–2012. The predictive contribution of each variable is quantified using the area under the (receiver operator characteristic) curve (AUC) metric, which plots the false-positive rate against the true-positive predictions of conflict for thresholds from 0.0 to 1.0, modified for our count data by truncating predicted values above 1. Figure contains the same predictors as the full model in Table 1. B, between grids; W, within grid.