Evaluating the ability of statistical models to capture crop yield responses to climate change

Abstract:

Statistical models trained on historical yield-weather relationships are a commonly used tool for estimating the impacts of climate change on crop yields. Accurately predicting future climate impacts on yields requires choosing regressors that capture fundamental biophysical processes, since historical interannual temperature variability differs in its distributional characteristics from future climate shifts. In this work we assess the ability of commonly-used statistical models (Schlenker & Roberts, 2009) to predict future yields in a synthetic dataset from the Global Gridded Crop Model Intercomparison (GGCMI) Phase 2 experiment: simulated potential yields of 5 crops from 7 process-based models, each driven by AgMERRA reanalysis weather and a systematic suite of climate perturbations. We train the statistical models on simulated historical yield-weather relationships, and then evaluate their predictions under uniform-warming scenarios. Results suggest that the choice of regressors may not allow capturing future yield changes. For U.S. maize, historically trained statistical models have high in-sample fits (R2>0.9), but underpredict yields in warmer scenarios because they exaggerate losses caused by severe heat exposure. Performance is better if statistical models are trained on warmer scenarios and used to predict historical yields, suggesting that historical data does not provide adequate sampling variation in severe heat exposure. Ongoing work seeks to evaluate whether prediction errors are driven primarily by limited sampling variation, or because model regressors are imperfect proxies for real-world factors affecting crop yields. Additional results include that prediction error can show geographic patterns, and that the use of location-specific fixed effects in the statistical model improves in-sample fits but can lead to increased error in predictions under warming. Concerns about the use of statistical models for impacts assessment have previously been raised because they cannot capture adaptations to new climate conditions, e.g. changing cultivars or planting dates. This work raises an additional concern, that statistical models may lead to systematic errors even in scenarios without adaptation.