Extreme weather impacts do not improve conflict predictions in Africa

Existing quantitative research on climate and conflict primarily focuses on temperature and precipitation, neglecting the role of biophysical impacts of extreme weather events (droughts, floods, crop production shocks, tropical cyclones) on conflict dynamics. This is significant because the frequency and intensity of such events are increasing with climate change. While research highlights the complexity of the relationship between extreme weather and conflict (the same event can have drastically different consequences depending on context), this paper investigates whether incorporating data on extreme weather impacts can enhance conflict prediction models. The study addresses this by examining if adding extreme weather data to established conflict predictors (socioeconomic factors, governance indicators, and conflict history) improves the accuracy of conflict forecasts. It uses high-resolution data from state-of-the-art climate impact models (ISIMIP2a) and the non-parametric Generalized Random Forest (GRF) algorithm to test this hypothesis. The study’s predictive approach allows for the estimation of the total predictive information contained in extreme weather impacts as a group, rather than isolating effects of individual events, thus enabling a more comprehensive assessment of their predictive value compared to previous research that often focuses on individual events or climate variables in isolation. The research also aims to assess whether extreme weather events provide unique predictive information beyond that already captured by well-established conflict predictors.

The existing literature on the relationship between climate and conflict predominantly centers on temperature and precipitation. Studies using causal analysis often examine specific extreme weather events (droughts, floods, or crop shocks) in isolation, failing to consider their cumulative effect. Several studies have attempted to integrate climate data into conflict prediction models with mixed results. Some research suggests that extreme disasters (above a certain threshold) increase armed conflict risk globally, especially in ethnically diverse countries. Others have found that the inclusion of climate variables, such as drought and flood frequency or temperature and rainfall fluctuations, either improves prediction accuracy marginally or worsens it. The overall lack of inclusion of a full suite of known conflict predictors along with extreme weather data leaves uncertainty regarding their overall contribution to prediction ability. Previous research also has limitations in sample size and modeling choices, preventing a robust assessment of extreme weather's collective impact. The current paper seeks to address these shortcomings by using a broader set of predictors along with an advanced model.

This study utilizes data from mainland Africa between 1994 and 2012 (UCDP) and 2002-2012 (ACLED) to predict conflict incidence one year ahead. The analysis leverages data from multiple sources, including the Uppsala Conflict Data Program (UCDP) and the Armed Conflict Location and Events Data Project (ACLED) for conflict events, the ISIMIP2a project for extreme weather impacts (droughts, floods, tropical cyclones, and crop production shocks), and various sources for socioeconomic and governance indicators. The spatial units of analysis include national, first sub-national, and second sub-national administrative levels. The study uses the Generalized Random Forest (GRF) algorithm, a non-parametric machine learning method, to construct predictive models. GRF is chosen for its ability to handle non-linear relationships and avoid overfitting, which is crucial given the complex interactions between climate, socioeconomic factors, governance, and conflict. The predictive performance of models is evaluated using the area under the Receiver Operator Characteristic (ROC) curve and the area under the Precision-Recall (PR) curve, standard metrics for assessing conflict prediction models. The core methodology involves comparing the predictive performance of models that include extreme weather impacts with those that do not, using all three administrative levels, to assess the added predictive value of the extreme weather variables.

The study's main finding is that the inclusion of extreme weather impacts does not significantly improve conflict forecasts when a full suite of known conflict predictors (socioeconomic indicators, governance indicators, and conflict history) are already included. This result holds across different administrative levels (national and sub-national) and conflict outcome measures (UCDP and ACLED data). While extreme weather impacts demonstrate a moderate ability to predict conflict independently, they add minimal predictive power when combined with other predictors. Analysis reveals that the information in extreme weather impacts is largely captured by the spatial patterns already present in the socioeconomic indicators. Regression analysis shows that the predictive relationship between extreme weather impacts and key socioeconomic variables disappears after controlling for place-specific fixed effects, suggesting that the apparent relationship arises from spatial co-occurrence rather than temporal causality. A further analysis of spatial and temporal variation in predicted conflict risk shows that a large portion of variation is due to spatial fixed effects for all predictor groups (including extreme weather events), highlighting the importance of spatial factors over time in conflict prediction. This analysis supports the conclusion that the predictive information in extreme weather variables is redundant given the other predictors included in the model.

The findings challenge the notion that detailed information on extreme weather impacts consistently enhances conflict prediction models, especially when other known factors are included. The lack of added predictive value suggests that, at least for the specific context of this study (mainland Africa, 1994-2012), existing socioeconomic and political factors largely account for the influence of extreme weather on conflict. While extreme weather impacts can independently predict conflict to some extent, they do not provide unique information for improved forecasting. The results align with some previous research and contrast with others, possibly due to differences in temporal scales, spatial resolutions, and methods of analysis. This suggests the possibility of conditional effects, whereby extreme weather might affect conflict onset in specific circumstances but not the overall incidence. The non-parametric modeling approach adopted in this study allows for more robust conclusions concerning the total predictive contribution of extreme weather data in comparison to prior studies.

This study demonstrates that incorporating detailed extreme weather impact data from state-of-the-art climate models does not improve conflict predictions in Africa when combined with a full suite of known conflict predictors. While extreme weather events have independent predictive power, this information is largely redundant given the inclusion of socioeconomic and political data. Future research should explicitly test the added value of extreme weather data in diverse contexts and with different methodologies before using it for conflict prediction. This study sets a benchmark for future research, emphasizing the need to demonstrate significant unique predictive power of extreme weather data, ideally using non-parametric methods to avoid biases and fully account for the complexity of factors driving conflicts.

The study's scope is limited to mainland Africa between 1994 and 2012, due to data availability. The specific climate impact models and data sources used might affect the results, and replicating the analysis with different datasets or models would strengthen the conclusions. The study focuses on annual conflict incidence, neglecting potential influences on conflict onset or duration. Although the analysis controls for many known factors, unobserved confounding variables could still influence the results. Finally, the GRF algorithm requires sufficient data; the limited number of observations at the national level might affect the robustness of those findings.