Which drought indicators predict impact? A cross-country test

The ten countries where OCHA has an endorsed or in-development drought anticipatory-action framework: Afghanistan, Burkina Faso, Chad, El Salvador, Ethiopia, Guatemala, Honduras, Kenya, Mauritania and Niger. Every indicator behind our triggers, plus growing-season temperature, tested against six ground truths: national staple production (FAOSTAT), CERF and EM-DAT drought seasons, the bad years each framework documents, official province and district yields (FEWS NET Data Warehouse), the same restricted to framework areas, and a gridded yield dataset (GDHY).

Summary: indicators against ground truths

The primary output. Rows are indicators (and the combinations tested), columns are ground truths. Continuous targets get an out-of-sample R² (leave-one-out for the national series, leave-one-year-out for the subnational panels); binary targets get an AUC computed from within-country (positive, negative) season pairs, 0.5 being chance. All pooled across countries with every series standardised within its country or unit, so only year-to-year variation is compared. Green is better; the colour scale differs by column.

Columns: National staple production (FAOSTAT): 240 country-seasons, 10 countries; each series standardised within its country; leave-one-season-out. Impact seasons: CERF drought allocation or EM-DAT event: 1276 within-country (positive, negative) season pairs from AFG, BFA, ETH, GTM, HND, KEN, MRT, NER, SLV, TCD; positives: 83 seasons. Framework-documented bad years: 496 within-country (positive, negative) season pairs from BFA, GTM, HND, MRT, SLV; positives: 26 seasons. Official subnational yield, all units (FEWS NET DW): ~4,800 unit-years in 7 countries (Guatemala has no yield); within-unit detrended; leave-one-year-out. Official subnational yield, framework-area units only: 1173 unit-years in AFG, BFA, ETH, GTM, KEN, MRT. GDHY gridded maize/wheat yield, admin 1 (secondary): 2001–2016, maize and wheat only, partly circular with NDVI.

What was tested

• Indicators, all from the JRC ASAP per-admin export (one consistent route, 2001 onward): temperature (ECMWF reanalysis, cropland), rainfall (CHIRPS), SPI-3, water satisfaction index (WSI), cumulative-FPAR anomaly (zFPARc). Plus ERA5 monthly rainfall from the team database and FAO GIEWS ASIS at admin 1: the Agricultural Stress Index (share of cropland under stress) and mean VHI. Temperature and biomass are detrended per unit, so "hot" means hotter than that year would be expected to be.
• Season and area. Each country's main growing season; indicators averaged over the framework area and over all ASAP units (country pages show both); biomass, WSI and SPI-3 over the last 60 percent of the season, where the trigger windows sit.
• Ground truths. (1) National production of the country's staples, FAOSTAT, percent departure from a 2001 to 2024 trend. (2) Seasons with a CERF drought allocation dated to that season by the CERF supplement's rainfall-deficit period, or an EM-DAT drought event that follows it. (3) The bad years each framework documents (five countries). (4) Official province or district yields from the FEWS NET Data Warehouse, eight countries, within-unit. (5) The same restricted to framework-area units. (6) GDHY gridded maize and wheat yields at admin 1, 2001 to 2016, as a secondary check.

By country: national production and the binary records

Pooled regressions on national production (each series standardised within its country)

By country: official subnational yields (FEWS NET Data Warehouse)

Official government production statistics compiled by FEWS NET, at the finest level with a usable series. Each unit's log yield and each indicator (ASAP unit statistics, FAO ASI and VHI at admin 1) are detrended and standardised within the unit. Leave-one-year-out R² holds out all units of a year together, the honest test for a trigger that has to call a new year. Green means the indicator predicts something out of sample.

Only Afghanistan (temperature 0.14, SPI-3 and ASI 0.10) and Niger (water balance 0.12, temperature 0.10) show out-of-sample skill at the unit level; Burkina Faso, where the national temperature signal is clear, shows none at province level with any indicator. Pooled, every coefficient has the expected sign and is significant (n above 4,300), but explained variance is small. The ranking matches the national study: VHI and ASI first, then water balance and temperature, the NDVI-only biomass anomaly last.

Framework-area units only

Restricting to units inside each framework's area changes the picture where that area is small and drought-prone: Burkina Faso's four trigger provinces and Afghanistan's five framework provinces show real unit-level skill, led by water balance and ASI with temperature a step behind. The large Ethiopian and Kenyan areas show nothing, and dominate the pooled framework-area column of the summary table.

Aggregated back to the national year

Averaging the within-unit standardised series over all units of a year gives a national series built from the official subnational statistics. The second column is its agreement with the FAOSTAT anomaly used above; the rest is its correlation with the national mean of each indicator, green when the sign is the expected one. Yield first, then production.

The two sources agree closely (r 0.83 to 0.90) in Afghanistan, Burkina Faso, Niger and Mauritania, moderately in Ethiopia and Kenya, poorly in El Salvador, and Guatemala has only ten years. Where they agree the national ranking reappears: Afghanistan temperature −0.73 with aggregated yield and VHI, ASI, water balance and temperature all around 0.7 with aggregated production; Niger water balance 0.64, temperature −0.59, VHI 0.59; Mauritania VHI 0.70 and temperature −0.61 on production; Burkina Faso temperature −0.48 on production against −0.56 with FAOSTAT. The biomass anomaly is never above 0.6 and often has the wrong sign. Ethiopia (temperature with the wrong sign), Kenya, Guatemala and El Salvador show nothing consistent.

A second yield dataset: GDHY

The Global Dataset of Historical Yields (Iizumi and Sakai 2020, doi:10.1038/s41597-020-0433-7) gives 0.5-degree maize, wheat, rice and soybean yields for 1981 to 2016 and is the dataset several of the global studies below use. Two limits here: no millet or sorghum, so the Sahel staples are not covered; and it blends national statistics with satellite NDVI, so agreement with vegetation indicators is partly circular. National and admin-1 means, detrended 1996 to 2016, analysed 2001 to 2016.

GDHY agrees only weakly with FAOSTAT in these countries (r 0.5 to 0.6 in Kenya, Guatemala and Mauritania, near zero or negative elsewhere) and Honduras and El Salvador have five cells each. Where usable it points the same way (Mauritanian maize follows temperature −0.66 and rainfall and water balance 0.7; Ethiopian wheat temperature −0.70, ASI and VHI 0.7; Kenyan maize and wheat ASI and VHI 0.45 to 0.65). At admin 1 the pooled panel explains almost nothing, and its best indicator is the circular one. GDHY neither contradicts nor strengthens the result.

Reading the results

• Production and the binary records disagree on the leader, and both are informative. Temperature, ASI and VHI predict how bad the harvest is; rainfall and SPI-3 predict whether the season was recorded as a drought. The two are not the same event: a hot, patchy-rain season with a poor harvest may never become a CERF allocation, and the frameworks' documented bad years were chosen by people looking at rainfall. The framework-bad-year column should be read as "agrees with how the framework already thinks", not as an independent outcome.
• Where the weather explains the harvest, temperature is at the top. Mauritania (leave-one-out R² 0.54 nationally), Afghanistan (0.38), Burkina Faso (0.23) and Niger (0.11). In these four the hot seasons are the poor harvests, and the official subnational statistics, aggregated to the year, say the same (temperature r −0.73 with aggregated yield in Afghanistan, −0.59 in Niger, −0.61 with production in Mauritania, −0.48 in Burkina Faso). Where the impact record is sparse enough to test, temperature also ranks the CERF seasons well there: AUC 0.72 to 0.80 in Burkina Faso, Niger, Chad and Afghanistan.
• ASAP's biomass anomaly is the weakest indicator against every ground truth. Pooled leave-one-out R² 0.02 on production, negative in eight of ten countries; the lowest AUC on both binary targets; nothing within provinces. FAO's ASI and VHI do much better, and the reason is in their construction: VHI is an equal blend of a vegetation condition index and a land-surface-temperature index, so it already carries half the heat signal that zFPARc lacks.
• Rainfall is the weaker predictor of production everywhere except Kenya, where ERA5 rainfall over the long rains is the best indicator and temperature adds little; Kenya's maize is grown in the highlands, not the arid counties the framework targets.
• Nothing works in Ethiopia or the Dry Corridor, in any dataset. Ethiopia's national cereal production is a smooth growth curve with no relation to any indicator, CERF or EM-DAT record drought in 15 of 24 seasons (CERF alone in 12) so the binary target is saturated, and the official zone-level Meher yields show nothing either. Guatemala, Honduras and El Salvador produce most of their maize outside the Dry Corridor departments, and neither the national nor the departmental series follow the framework-area weather.
• Within a province, no indicator predicts the year. Pooled over about 4,800 unit-years of official yields, every coefficient has the expected sign and is significant, but the best leave-one-year-out R² is 0.04. Official subnational statistics are noisy, and indicators averaged over a unit and a season miss timing and location within it. Skill reappears inside small, exposed framework areas (Burkina Faso's four trigger provinces: water balance 0.25, temperature 0.12; Afghanistan's five: ASI 0.33, temperature 0.19), but the pooled framework-area column is dominated by the large Ethiopian and Kenyan areas and shows nothing.
• Heat and vegetation stress are complementary. Temperature plus ASI is the best two-indicator combination on production and both coefficients stay significant; temperature plus biomass or plus rainfall adds nothing to temperature alone.

What the literature says

Temperature outperforming rainfall and satellite vegetation as a predictor of staple-production shortfalls is the mainstream finding of the statistical crop-climate literature for the tropics and Africa. The closest studies:

• Schlenker and Lobell (2010), Robust negative impacts of climate change on African agriculture, Environmental Research Letters 5, 014010. Panel regressions of FAO national yields on growing-season weather for maize, sorghum, millet, groundnut and cassava across sub-Saharan Africa. Temperature, not precipitation, is the dominant driver of yield variation and of projected losses; the precipitation coefficients are small and often insignificant once temperature is in the model. doi:10.1088/1748-9326/5/1/014010
• Lobell, Bänziger, Magorokosho and Vivek (2011), Nonlinear heat effects on African maize as evidenced by historical yield trials, Nature Climate Change 1, 42–45. Twenty thousand maize trials: each degree-day above 30 °C cut yield by about 1 percent under good rainfall and 1.7 percent under drought. Heat hurts most when the crop is also water-stressed, which is the interaction our hot-and-low-biomass quadrant picks up. doi:10.1038/nclimate1043
• Sultan et al. (2013), Assessing climate change impacts on sorghum and millet yields in the Sudanian and Sahelian savannas of West Africa, Environmental Research Letters 8, 014040. Crop-model simulations across 35 stations: the yield response to warming is negative for all varieties regardless of the direction of rainfall change, so temperature dominates the Sahel signal. Directly relevant to Burkina Faso, Niger, Mauritania and Chad. doi:10.1088/1748-9326/8/1/014040
• Lobell and Burke (2008), Why are agricultural impacts of climate change so uncertain? The importance of temperature relative to precipitation, Environmental Research Letters 3, 034007. Shows that for most crops and regions, uncertainty in the temperature response matters more than uncertainty in rainfall, and that statistical models attribute more yield variance to temperature. doi:10.1088/1748-9326/3/3/034007
• Ray, Gerber, MacDonald and West (2015), Climate variation explains a third of global crop yield variability, Nature Communications 6, 5989. Gridded analysis: climate explains roughly a third of global yield variability, with large regional differences and both temperature and precipitation mattering. Consistent with our finding that a large share of the variance is not climatic at all, and that national production diverges from the framework area's weather in some countries. doi:10.1038/ncomms6989
• Vogel et al. (2019), The effects of climate extremes on global agricultural yields, Environmental Research Letters 14, 054010. Random-forest models on GDHY yields: climate extremes explain 18 to 43 percent of the interannual variance of maize, soybean, rice and spring-wheat yields, with temperature extremes carrying more weight than precipitation extremes in most regions, and hot-dry compound years the most damaging. doi:10.1088/1748-9326/ab154b
• Lesk, Rowhani and Ramankutty (2016), Influence of extreme weather disasters on global crop production, Nature 529, 84–87. Using EM-DAT disasters against FAO production: droughts and extreme heat each cut national cereal production by about 9 to 10 percent; floods and cold had no measurable effect. Supports treating heat as an impact driver in its own right rather than a covariate of drought. doi:10.1038/nature16467
• Rojas, Vrieling and Rembold (2011), Assessing drought probability for agricultural areas in Africa with coarse resolution remote sensing imagery, Remote Sensing of Environment 115, 343–352, and the FAO ASIS methodology built on it. The Agricultural Stress Index is derived from the Vegetation Health Index, which is an equal-weight blend of a vegetation condition index (NDVI) and a temperature condition index (land-surface temperature, after Kogan 1995). This is why ASI and VHI track temperature so closely in our results and outperform the NDVI-only zFPARc: they already contain half of the heat signal. doi:10.1016/j.rse.2010.09.006
• Rembold et al. (2019), ASAP: a new global early warning system to detect anomaly hot spots of agricultural production for food security analysis, Agricultural Systems 168, 247–257. Describes the warning system two of our frameworks trigger on; its indicators are rainfall, water balance and FPAR-based biomass, and temperature is provided as context rather than used in the warning classification. doi:10.1016/j.agsy.2018.07.002

Two literature caveats apply to us as much as to those studies. First, temperature and rainfall are negatively correlated in the Sahel growing season (dry years are hot years because of reduced cloud and evaporative cooling), so part of what "temperature" captures is the rainfall deficit measured more precisely, as the Burkina Faso water-balance comparison suggested. Second, the literature works with national or gridded yields, where aggregation averages out local noise; the subnational check below shows how much that matters.

What this means for trigger design

• A detrended growing-season temperature anomaly belongs in the candidate set for every drought framework's observational window, alongside ASI or VHI and a rainfall measure. It is available at the same time as the vegetation indices (ASAP updates every dekad) and needs no new data agreement. In the Sahel and Afghanistan it is the single indicator most tied to harvest outcomes.
• Rainfall and SPI-3 remain the indicators most aligned with the humanitarian record. A trigger that must match when CERF and partners will call a drought should keep a rainfall arm; one that aims at harvest shortfall should add heat or use ASI/VHI.
• Where a framework's observational window rests on a biomass anomaly alone (Burkina Faso on ASAP's zFPARc, Chad on ACF/GeoSahel's DMP, a different product in the same family), the backtest should be re-run with temperature, ASI and VHI as alternatives before the next revision. The Burkina Faso case is worked through on the Burkina site.
• Triggers defined on a handful of admin-2 units should not be expected to have much skill against locally reported outcomes, whatever indicator they use. The evidence for skill is at the scale of a region or country, or of a small area chosen for its exposure.
• Temperature does not lead. On the Burkina Faso dekadal analysis heat coincides with dry spells rather than preceding them, so this is an in-season observational indicator, not a forecast-window one; a forecast temperature arm is worth a separate test.
• Next ground truth: IPC phase 3+ population by admin unit (team database, from 2017) or FEWS NET classifications would test the provincial question with less measurement noise, over a shorter record.

Data notes

• Indicators. JRC ASAP per-admin indicator statistics export (September 2026 pull; ids and levels in data/config/asap_ids.json): temperature (ECMWF, cropland), rainfall (CHIRPS), SPI-3, water satisfaction index, cumulative-FPAR anomaly (zFPARc). Temperature detrended per unit with a Theil-Sen slope over 1991 to 2025, biomass over 2001 to 2025. ERA5 monthly rainfall from the team Postgres public.era5. FAO GIEWS ASIS country csv endpoints for ASI and mean VHI at admin 1. Indicators averaged over each country's main season (dekads in data/config/countries.json); biomass, WSI and SPI-3 over its last 60 percent.
• Ground truths. FAOSTAT production bulk file (December 2025), staples per country, percent from a 2001 to 2024 linear trend. CERF drought allocations from the OneGMS mirror with the CERF supplement's rainfall-deficit periods (dev DB, schema aa), dated to the overlapping main season; EM-DAT blob snapshot via ocha-stratus, dated to the last main season before the event start; undated chronic allocations excluded. Framework bad years from the team knowledge base. FEWS NET Data Warehouse cropproductionfacts (18 September 2026): staples and main season per country at the finest level with a usable series; yield is production over harvested area, falling back to planted area and to the reported yield; "All (PS)" production-system rows used where present, other systems summed; series of at least 8 years, detrended and standardised within the unit; Chad has no series, Honduras a national one to 2009 only, Guatemala production without area, El Salvador's reporting changed in 2013 and Kenya's districts became counties in 2013 (separate series). GDHY v1.2/1.3 (PANGAEA doi:10.1594/PANGAEA.909132), cells assigned to CODAB admin 1 by centre, detrended 1996 to 2016, analysed 2001 to 2016.
• Statistics. Ordinary least squares on standardised predictors with leave-one-out R² for the national series; within-unit detrended panels with leave-one-year-out R² (all units of a year held out together); AUC from within-country (positive, negative) season pairs, indicators oriented so that higher means worse; pooled regressions with every series standardised within its country or unit.
• Code. OCHA-DAP/ds-aa-drought-indicators: build_tables.py, analyse.py, panel.py, gdhy.py, summary.py, make_site.py; processed tables and results under data/processed/. Pages by the OCHA Centre for Humanitarian Data, September 2026.

Country pages

Each page shows both scopes (framework area and national), all binary targets, the quadrant split, how impact records were dated, and the season table.