Accurately predicting the effects of climate change on agriculture is crucial for effective adaptation strategies. Global climate models (GCMs) provide valuable large-scale insights, but their coarse resolution and systematic biases limit their applicability to regional and local scales. Bias-corrected and downscaled climate products are frequently used to address these limitations, enhancing their suitability for various applications such as economic impact assessments, infrastructure planning, and adaptation strategy development. However, these processed products introduce their own uncertainties, including the validity of stationarity assumptions in bias-correction methods, the physical plausibility of results, and the representation of atmospheric and hydrologic variability. Understanding the propagation of these uncertainties through impact models is critical. While many studies focus on developing or evaluating bias-correction and downscaling techniques, fewer investigate their end-to-end effects on sector-specific impact models. This study addresses this gap by focusing on the agricultural sector, specifically analyzing how uncertainties from one statistical bias-correction and downscaling technique (applied to the NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP) dataset) affect simulated U.S. maize yields. The researchers chose a well-established statistical panel model that incorporates season-wide temperature and precipitation measures to capture the impact of climate on yields. This model, driven by both the NEX-GDDP and CMIP5 model ensembles, allows for a direct comparison of yield hindcasts and future projections, quantifying the uncertainty introduced by the bias-correction and downscaling process.

A substantial body of literature exists on the impacts of climate change on agriculture, highlighting the sector's vulnerability and the need for reliable climate projections for informed decision-making. Studies have shown the negative impacts of climate change on agricultural productivity, and the expectation of continued challenges throughout the century. The agricultural sector relies heavily on accurate climate projections for both national policy and local farming practices. Previous research has explored the use of bias-correction for crop modeling, primarily focusing on process-based agricultural models and dynamic downscaling techniques. However, fewer studies have directly addressed the effects of statistical bias-correction and downscaling on yields simulated by statistical panel models. This study contributes to this gap by investigating the impact of bias-correction and downscaling on the outputs of an econometric model of agricultural yields, a frequently used approach for agricultural impact studies.

The researchers used county-level rainfed maize yield data from the USDA, coupled with climate data from the NEX-GDDP and CMIP5 datasets. NEX-GDDP data consists of statistically bias-corrected and downscaled outputs from 21 CMIP5 models, offering daily maximum and minimum temperatures and mean precipitation. The bias-correction and downscaling were performed using the Bias-Correction Spatial Disaggregation algorithm. The yield model incorporated growing degree days (GDDs), which are beneficial for yields, and extreme degree days (EDDs), which represent heat stress and negatively impact yields. A quadratic function of season-total precipitation was also included. The model's performance was evaluated using the coefficient of determination (R²). To isolate the effects of climate and weather, the researchers removed state-specific time trends and county-level fixed effects, focusing solely on weather's contribution to yield variability. The yield model was fitted using data from 1956–2005 for counties with at least 50% data coverage. The analysis compared the distribution of simulated yields from each model to observed distributions, using summary statistics including the mean, standard deviation, 10th percentile, and minimum to assess the models' performance in capturing both average yields and yield extremes. The researchers further investigated the sources of biases by analyzing the models' representation of individual climate variables (GDDs, EDDs, and precipitation). Additional models were created to verify the key results by modifying input variables, and the analysis was expanded to include a series of checks that corroborate the principal findings regarding biases and their sources. Yield projections under a high-emissions scenario (RCP 8.5) were then generated and compared to assess potential differences in future yield variability and return periods.

The study found significant discrepancies between the yield simulations from the CMIP5 and NEX-GDDP models. CMIP5 models considerably overestimated historical yield variability, while NEX-GDDP models, although closer to observations, underestimated the largest historically observed weather-induced yield declines. At the national level, the NEX-GDDP models underestimated the lowest observed yield (representing the most extreme weather-induced yield drop) by 63% to 233%, with a median underestimation of 162%. This underestimation also held true at the county level, particularly in high-yield counties. The analysis revealed that the NEX-GDDP ensemble struggled to reproduce the observed distribution of EDDs, particularly in the Corn Belt region, while performing better with GDDs and total precipitation. The biases in NEX-GDDP yield hindcasts were primarily attributed to the poor representation of EDDs. Additional analyses confirmed that these biases persisted even when using alternative yield models that did not employ degree-day variables. The study also projected yields under the RCP 8.5 scenario, revealing substantial differences between NEX-GDDP and CMIP5 projections. NEX-GDDP projected lower variability in future yields compared to CMIP5, potentially leading to overly optimistic assessments of climate risk. This difference is evident when examining yield return periods, illustrating the challenge for decision-makers in choosing between models with differing levels of confidence and the potential for substantial implications when assessing the likelihood and magnitude of adverse events. The analysis also highlighted that biases in EDD representation resulted from biases in daily temperature time series, amplified by the integration process used to derive EDDs. These biases were not due to spatial aggregation, as they persisted at the model grid scale. Analysis of representative time series further illustrated the underestimation of extreme temperatures and consequent EDDs by NEX-GDDP, in contrast to the CMIP5 ensemble which often overestimated them.

The study's findings align with previous research showing that observational measures of interannual temperature variability fall within the CMIP5 ensemble range, and that CMIP5 models simulate a wide range of historical and future meteorological extremes. The study adds to this literature by offering a comprehensive evaluation of NEX-GDDP and its impacts on a specific agricultural outcome (U.S. maize yields). The results emphasize how framing the evaluation around a particular outcome can significantly change the interpretation of downscaling and bias-correction methods, particularly regarding extremes. The study acknowledges several limitations: the use of a relatively simple statistical downscaling and bias-correction method; the possibility that conclusions may not apply universally to all econometric models or spatial/temporal scales; and the relative simplicity of the yield model, which excluded factors such as soil moisture and intra-seasonal variations that could be included in more complex process-based models. The study also assumed no future adaptation and didn't account for other factors that affect yields (like CO2 levels or air pollution). Nonetheless, the results highlight the need to carefully consider the potential for bias in downscaled and bias-corrected climate projections, particularly concerning temperature extremes.

The study demonstrates that statistically bias-corrected and downscaled climate models, while often considered an improvement, can underestimate the severity of climate impacts, especially on agricultural yields sensitive to temperature extremes. The underestimation stems primarily from a flawed representation of extreme temperature events. This underscores the importance of carefully evaluating climate datasets and choosing models appropriate to the specific application. The findings highlight the need for more sophisticated bias-correction and downscaling techniques, considering the sector-specific impacts, particularly when temperature extremes influence the outcome. Future research should focus on developing and evaluating more advanced methods that accurately capture extreme events and their downstream consequences. This holistic approach is crucial to mitigating potential biases in climate impact assessments and informing effective decision-making.

The study's limitations primarily involve the methodology employed. The relatively simple statistical downscaling and bias-correction method used might not generalize to more advanced techniques that may yield different results. The model's simplicity, focusing primarily on temperature and precipitation while neglecting other potentially influential factors like soil moisture and intra-seasonal variability, could limit the generalizability of findings. The absence of future adaptation measures and other yield-affecting factors such as CO2 levels and air pollution may also limit the accuracy of future yield projections. The study’s focus on maize yields in the US might not fully translate to other crops, regions, or econometric model specifications. Despite these limitations, the study provides valuable insights and highlights the need for more comprehensive evaluations of bias-corrected and downscaled climate data.