Open Access

The availability of spatially-explicit time-series estimates of rainfall and other weather outcomes has expanded rapidly over the past decade and a half. This proliferation of publicly-accessible rainfall data has changed how empirical analysis of farm production and productivity takes place: fortifying survey-derived data (e.g., on farm management and production outcomes) with spatial estimates of seasonal rainfall outcomes is now standard practice, much to the benefit of applied agricultural economic analysis. Yet guidance on which dataset to use, among the many available alternatives, has largely been lacking: while there have been some studies comparing rainfall data quality across alternative datasets, relative to some accuracy benchmark, there has not yet been systematic analysis of whether the choice of rainfall dataset may affect econometric model performance, or the estimation of other (non-rainfall) model parameters. Such a concern would arise from cases where any mismeasurement of rainfall is correlated with model error terms or other model covariates. To investigate this, we use panel data on plot-level maize yield outcomes in Ethiopia from 2018 and 2021, along with fifteen alternative spatio-temporal rainfall data products, including gauge-based, satellite-derived, and reanalysis datasets. We estimate yield response models, in which our primary interest is on the estimate of nitrogen use efficiency (NUE). Estimation results from alternative specifications and panel estimators indicate that, while coefficient estimates for rainfall vary considerably with alternative rainfall estimates, the coefficient estimates on nitrogen fertilizer (and consequently, our estimates of NUE) do not vary in a statistically significant way. Our results suggest that the choice of rainfall dataset does not significantly affect analytical conclusions about fertilizer response.

Text in English