Errors in climate datasets and their effects on statistical crop models

Although weather data are widely acknowledged to contain measurement errors, the implications of these errors for models that relate weather to yields have not been adequately examined. From statistical theory and applications in many other fields, it is clear that measurement error in a single predictor variable can lead to bias in estimating the effects of that variable, as well as any other correlated predictors. Of particular concern for statistical crop models is that errors in measuring precipitation can lead to bias in inferences about yield responses to both temperature and precipitation. In this study, simulation extrapolation (SIMEX) is used to gauge the importance of measurement error for two recent studies that employed statistical crop models. In both cases, estimates of yield responses to temperature were only slightly changed when considering measurement errors. However, yield responses to precipitation were significantly larger when assuming that precipitation is measured with 30% error, compared to the common assumption of error-free measurements. Thus, results indicate that studies that ignore measurement errors are unlikely to be biased for estimating T sensitivity of yields, but can easily underestimate P sensitivity by a factor of two or more. More work is needed to test effects of measurement errors in other cases, as well as to better quantify the magnitudes of errors in weather measurements for cropped regions. As a rough substitute for detailed measurement error analysis, sensitivity tests that double the yield response to precipitation are advised when applying statistical crop models to projections from climate ensembles. Depending on the magnitude of precipitation projections, which in turn depend on the spatial and temporal scale of analysis, the conclusions of a study may or may not be altered by considering the effects of measurement errors.