This repository contains the source code for our analysis of where and to what extent flooding impacts food security in Sub-Saharan Africa. We make use of econometric time series analysis techniques—including panel Granger causality and static panel regression—to quantify the effects of floods on the integrated food security phase classification (IPC) metric at the place of their occurrence up to roughly one year after their onset.
The data used in this analysis include:
- IPC over sub-Saharan Africa at a seasonal timestep from 2009–2020 produced by FEWS NET available here
- Flood data from shapefiles produced by the Dartmouth Flood Observatory as part of the Global Active Archive of Large Flood Events, 1985–Present available here
- Population data as of 2020 produced by WorldPop available here
All of the data was harmonized to a spatial scale at the intersection of administrative level 2 units and FEWS NET Livelihood Zones (a 'panel') and a time scale corresponding to the reporting period of the IPC data. Furthermore, all time series were first-differenced to enforce stationarity. Code is available in scripts/.
We used panel Granger causality analysis as a way to identify where there exists a significant Granger-causal relationship between any of the derived flood variables (occurrences, area as proportion of panel area, and duration) and the IPC using a lag of up to four seasons. The code for this analysis is available in analysis/modeling/granger-causality.R. A global test for homogeneous non-causality (i.e., a global null hypothesis of field significance) which accounts for the cross-sectional dependence of the data was conducted in Stata.
The figure below highlights the panels for which the indicated flood variables significantly Granger-caused changes in food security at a .00579 level. This level was calculated to constrain the false discovery rate to 0.10 given the large number of hypothesis tests conducted and to help prevent overinterpretation of potential Type I errors. Some of the panels did not experience enough variance in flooding over the study period to fit a linear model needed to test for Granger causality, and so were filtered out from this analysis as shown in the figure.
(left or top) Map of where first-differenced flood variables Granger cause changes in food security based on a significance level determined using the false discovery rate method. (right or bottom) Total population (as of 2020) living in panels indicated as experiencing a Granger-causal relationship between flooding and food security.
We used static random effects panel models at multiple spatial scopes to quantify the relationship between each of the contemporaneous and lagged flood variables and IPC. Additionally, we fit these models to the full data and the Granger-filtered data at each spatial scope. Random effects were specified as the correct model type for each sub-dataset with the code in analysis/modeling/panel-model-specification.R and the models were subsequenty fit in analysis/modeling/panel-modeling.R. Significant coefficients are plotted below.
Significant (p-value < .05) coefficient estimates and 95% confidence intervals of All-Africa panel models for both the full and Granger-filtered datasets. Positive coefficients represent negative impacts on food security.
Significant coefficient estimates and 95% confidence intervals for regional and country-specific panel models with normalized data. Solid lines indicate coefficient p-value < p*FDR = .0124, representing the coefficients which are most confidently nonzero. Dashed lines indicate coefficient p-value < .05, representing coefficients that are confidently nonzero yet which warrant more targeted analysis to confirm. (top) West Africa and Chad, (middle) East Africa, and (bottom) Southeast Africa. Countries are indicated by colors in corresponding legends.
In short, these models illuminate how floods heterogeneously impact food security across space and time over Sub-Saharan Africa, showing both positive and negative impacts at different spatial scopes and time lags. Food security reflects social dynamics as much as environmental ones, and so we decided to nuance our analysis by interpreting these model results in the context of qualitative case studies from humanitarian reports such as those produced by FEWS NET. Full discussion is available in our paper, which will be linked here once publicly available.