Ethiopia suffers more from droughts than other African countries: since 1900, over 400,000 people have died from famine following droughts. Predicting drought 1–2 years ahead is critical for preparedness, but meteorological anomalies operate on different scales. Previous studies showed correlations between sea surface temperature (SST) and rainfall in Africa, yet correlation does not imply causation. This work stands apart: neighboring studies in the archive address climate signals, El Niño, and droughts, but do not use causal inference for Ethiopia. The closest neighbor focuses on Tibetan moisture, not on drought mechanisms in Africa.
Monthly precipitation data from GPCC and sea surface temperature from NOAA ERSST V5 were used from January 1946 to December 2015. Drought was defined via the standardized precipitation index SPI: a point is considered in drought if the 12-month accumulated precipitation falls below the threshold corresponding to SPI ≤ -1. To reduce dimensionality, rotated principal component analysis (PCA) with VARIMAX rotation was applied to the 10,988 SST time series, yielding 76 non-random modes of variability. Causality was detected using the PCMCI algorithm from the Tigramite library, which removes spurious links from common drivers and autocorrelation.
The PCMCI algorithm identified 52 links to the drought variable, but after correction for multiple comparisons using the Benjamini-Hochberg method, only one remained significant: the second SST principal component with a 12-month lag (q = 0.0378). The link coefficient is negative (-0.230), meaning low SST values in the El Niño region (La Niña phase) increase the probability of drought in Ethiopia a year later. The spatial pattern of the second mode points to the ENSO region in the Pacific Ocean. Additionally, weak links with the 41st and 72nd SST modes in the Indian Ocean were found, but they did not survive correction. Adding the second SST component as a feature reduced the RMSE of forecasts for a neural network from 31.05 to 29.60 and for a random forest from 32.36 to 31.13.
This work differs from neighboring studies in that it applies causal inference to a specific region, Ethiopia, rather than correlation analysis. Neighboring works describe climate signals and correlations but do not establish causality with a time lag. Here, for the first time, it is shown that the link between La Niña and drought in Ethiopia is causal, not merely a statistical association. This strengthens the physical understanding of the mechanism: Pacific temperature anomalies influence East African rainfall through atmospheric circulation.
The method can be extended to other vulnerable regions of the world by adding more factors, such as anthropogenic influences or other meteorological anomalies. It would be promising to investigate why adding SST improves neural networks and random forests more than gradient boosting. Including nonlinear measures of conditional independence could also refine causal links.
The results could improve early warning systems for drought in Ethiopia and the Horn of Africa, providing forecasts 12 months ahead.
Test the robustness of the causal link on independent data and extend the analysis to other African regions.
This work addresses a fundamental problem in climatology: moving from correlational links to causal mechanisms in a complex climate system. It is a step toward answering which remote anomalies truly drive regional rainfall and which merely accompany them.
Since 1900, droughts in Ethiopia have affected over 77 million people and killed more than 400,000.