On the Use of Statistical Models to Predict Crop Yield Responses to Climate Change
On the Use of Statistical Models to Predict Crop Yield Responses to Climate Change
Predicting the potential effects of climate change on crop yields requires a model of how crops respond to weather. As predictions from different models often disagree, understanding the sources of this divergence is central to building a more robust picture of climate change's likely impacts. A common approach is to use statistical models trained on historical yields and some simplified measurements of weather, such as growing season average temperature and precipitation. Although the general strengths and weaknesses of statistical models are widely understood, there has been little systematic evaluation of their performance relative to other methods. Here we use a perfect model approach to examine the ability of statistical models to predict yield responses to changes in mean temperature and precipitation, as simulated by a process-based crop model. The CERES-Maize model was first used to simulate historical maize yield variability at nearly 200 sites in Sub-Saharan Africa, as well as the impacts of hypothetical future scenarios of 2◦C warming and 20% precipitation reduction. Statistical models of three types (time series, panel, and cross-sectional models) were then trained on the simulated historical variability and used to predict the responses to the future climate changes. The agreement between the process-based and statistical models' predictions was then assessed as a measure of how well statistical models can capture crop responses to warming or precipitation changes. The performance of statistical models differed by climate variable and spatial scale, with time-series statistical models ably reproducing site-specific yield response to precipitation change, but performing less well for temperature responses. In contrast, statistical models that relied on information from multiple sites, namely panel and cross-sectional models, were better at predicting responses to temperature change than precipitation change. The models based on multiple sites were also much less sensitive to the length of historical period used for training. For all three statistical approaches, the performance improved when individual sites were first aggregated to country-level averages. Results suggest that statistical models, as compared to CERES-Maize, represent a useful if imperfect tool for projecting future yield responses, with their usefulness higher at broader spatial scales. It is also at these broader scales that climate projections are most available and reliable, and therefore statistical models are likely to continue to play an important role in anticipating future impacts of climate change.