Size and the Nature of Measurement Error in Gridded Weather Datasets and its Consequential Estimation Bias in Regression Model: An Application to PRISM Datasets for the US Midwest Regions

Gridded weather datasets, such as PRISM, have been widely used in econometric analysis to study the impact of weather on economic outcomes. Yet, concerns persist regarding the measurement errors in these datasets and their consequential estimation bias. This study systematically quantifies and characterizes the measurement errors in growing-season total precipitation and extreme degree days (EDD) derived from PRISM. Using exact spatial matches between PRISM grid cells and ground weather stations, we find that PRISM weather variables exhibit nontrivial measurement errors that are negatively correlated with true weather outcomes, especially for EDD. Moreover, the variance of these errors increases with the extremity of actual weather outcomes. We conduct Monte Carlo simulations to evaluate the resulting estimation bias in weather impact on farm-level corn yields. The simulation results show that the average bias is moderate when estimating weather impacts on corn yield across the entire Corn Belt region due to large weather variations. However, the bias becomes more substantial at smaller spatial scales, such as individual states, where limited variation amplifies the relative influence of measurement error. We conclude that researchers should exercise caution when using PRISM data in econometric models, especially in subregional analyses.