Does heat predict drought impact, beyond the biomass proxy?

The heat-and-drought page showed that growing-season temperature explains part of the ASAP biomass anomaly that Trigger 2 keys on. That still uses biomass as the stand-in for drought. This page steps outside the proxy and asks whether temperature, alone and combined with biomass, explains drought impact: national cereal production shortfalls, and the seasons that produced EM-DAT drought events and CERF allocations, 2001 to 2024.

1 · Building an impact record #

Two targets, as agreed. A continuous one: FAOSTAT national millet plus sorghum production, expressed as the percentage departure from a 2001 to 2024 linear trend (the trend removes area expansion and intensification). And a binary one: growing seasons that produced a recorded humanitarian drought impact, assembled from EM-DAT drought events, CERF drought allocations, and the framework's own target years.

Impact records are dated to the lean season in which people were affected or funds were allocated, which is usually the year after the failed harvest. Each record was assigned to a growing season on the evidence below. Where the evidence is ambiguous the framework's own labelling is carried as an alternative and every result is reported under both.

The main binary target is therefore 2011, 2013, 2017, 2019, 2021; the alternative is 2011, 2014, 2017, 2019, 2022. Everything after 2018 is also contaminated by the security crisis, so results are shown for 2001 to 2018 as well. Cadre Harmonisé and FEWS NET classifications were examined but do not help: the national phase-3-plus series starts in 2020 and is conflict-driven, and FEWS NET's northern pastoral zone reached phase 3 in the 2012 lean season and otherwise stayed at 2 or below until the conflict years.

2 · What explains production shortfalls #

Predictors are the ASAP province series aggregated two ways: the four framework provinces (what the trigger sees) and the unweighted mean of all 45 provinces (the scale of the production target). Temperature is the July to September anomaly from each province's own trend line; biomass is the cumulative-FPAR anomaly over the Trigger 2 window, detrended as on the companion page. All predictors are standardised; the leave-one-out R² is the honest measure with 24 seasons.

• Temperature is the only predictor with out-of-sample skill. Every other single predictor has a negative leave-one-out R², meaning it predicts production worse than the mean does. The temperature coefficient is −5.5 to −5.9 percent of trend per standard deviation, which is −16 percent per degree in the framework provinces and −26 percent per degree nationally (the national anomaly has a smaller spread).
• Adding biomass to temperature adds nothing, and the interaction term is zero (p 0.99 for the framework provinces). The hypothesis that heat makes a biomass deficit worse cannot be seen because the biomass deficit itself carries no signal for production. Adding water balance and biomass together to temperature raises the in-sample R² by three points and lowers the out-of-sample one to zero.
• This is not an artefact of the window. Biomass over the end of season (dekads 27 to 30) does a little better (r 0.28 to 0.33, not significant), the full season no better than the window. Temperature over August to September, the grain-filling weeks, does as well as the July to September mean (r −0.54); pre-season heat does nothing (r −0.01).

Hot with normal biomass: the seasons the trigger cannot see

The quadrants say it plainly. In the framework provinces, the cool-and-low-biomass seasons averaged +3.9 percent production with no impact season among them; the hot-and-low-biomass seasons averaged −4.4 percent with two impact seasons; and the hot-and-high-biomass seasons were the worst of all, −11.4 percent, containing 2017 and 2021. Those two are the deepest production shortfalls in the record and the two most recent CERF drought responses, and in both the framework-province biomass was above its detrended median. A biomass-only trigger, however calibrated, would not have activated in either. The national split shows the same ordering.

3 · Ranking the impact seasons #

With five positive seasons a fitted model is not meaningful, so the binary target is scored by AUC: the probability that a randomly chosen impact season scores as more drought-like than a randomly chosen other season. 0.5 is chance, 1.0 is a perfect ranking. Production itself is included as a reference for how well a physical outcome ranks the humanitarian one.

• Temperature ranks the impact seasons best in every variant except one, and that one is the alternative dating over the full record, where framework-province biomass reaches 0.77 because 2014 and 2022 were poor-biomass seasons in the provinces. Restrict the same dating to the pre-conflict years and temperature leads again.
• Water balance ranks below chance throughout: the impact seasons had slightly better water balance than the others. This matches the earlier finding that ASAP's water-balance index is rainfall-total-driven and misses what matters.
• The combined score does not beat temperature alone. Adding biomass to temperature lowers the AUC in five of seven variants.
• The chronic underfunded allocations are not droughts. Adding 2005 to 2008 pushes every score toward 0.5, which is the expected result if those seasons were not meteorologically unusual, and the CERF supplement's note on 2008 says as much.
• 2013 is probably mis-dated. Under the main dating it is an impact season, yet production was 11 percent above trend, the framework provinces were cool and biomass was normal. The evidence points to the May 2014 EM-DAT event reflecting the 2014 season after all, as the framework assumed. The alternative-dating rows are therefore the more credible ones for that record, which strengthens biomass slightly and leaves temperature where it is.

4 · What this means #

• For the question asked. Heat combined with biomass does not lead to worse outcomes than biomass alone in any measurable way, because biomass in the Trigger 2 window barely relates to outcomes. Heat on its own does: the hot seasons are the poor-harvest seasons and the CERF seasons, whether or not the satellite saw a vegetation deficit.
• For the trigger. The observational arm is keyed on the indicator with the least outcome skill in this record. A temperature condition, whether as a second arm or as a replacement for the biomass condition, is the first thing to backtest in the 2027 revision. On this record, a July to September anomaly of more than +0.3 °C above trend in the framework provinces selects 2002, 2004, 2009, 2011, 2014, 2019, 2021 and 2023: six of the eight are impact seasons or ASAP activation seasons (seven under the alternative dating), 2021 is caught where biomass missed it, and the one clear false alarm, 2023, was a below-trend production season (−4 percent), not a good one. It misses 2017, which was only +0.1 °C in the provinces though hot nationally.
• For 2026. The framework provinces are running about +1 °C above trend for July to September so far. The fitted relationship would put this harvest 15 to 20 percent below trend, the worst in the record. ASAP's biomass says otherwise. One of them is wrong, and the harvest assessments in November will say which. That is the single most useful check on this whole line of analysis.
• For the data request. The weak link is the target. National production is dominated by the southern cereal belt, while the framework provinces are pastoral and millet-growing. Regional production and yield series from the agricultural statistics directorate (DGESS/MAERAH), which the SAP-GTP campaign evaluations draw on, would allow this test at the scale the framework operates. That request should go alongside the one to ANAM for station rainfall.