Statistically bias-corrected and downscaled climate models underestimate the adverse effects of extreme heat on U.S. maize yields

The study investigates how uncertainties introduced by statistical bias-correction and downscaling of global climate model (GCM) outputs propagate into modeled agricultural outcomes, focusing on U.S. maize yields. While GCMs provide valuable global-scale insights, their coarse resolution and systematic biases limit utility at regional scales, leading to widespread use of bias-corrected and downscaled products (e.g., NEX-GDDP). However, these products carry their own uncertainties (e.g., stationarity assumptions, representation of variability and extremes). The authors aim to quantify how these uncertainties affect a sector-specific outcome—maize yields—by comparing hindcasts and projections driven by CMIP5 parent models versus their bias-corrected, statistically downscaled counterparts, with emphasis on the role of temperature extremes. The purpose is to inform decision-makers about trade-offs among resolution, historical fidelity, and projection confidence when selecting climate information for agricultural risk management.

The paper situates its analysis within literature noting that bias-corrected and downscaled climate products add value over raw GCM outputs but may violate assumptions (e.g., stationarity), misrepresent variability and extremes, and face challenges in changing climates. Many studies develop/evaluate bias-correction and downscaling techniques, yet fewer conduct end-to-end impact assessments with sector-specific models. Prior agricultural impact studies document negative effects of extreme heat on crops and commonly model yields using degree days, with GDDs benefiting and EDDs harming yields; precipitation (often a quadratic term) is a standard proxy for water availability. Prior evaluations show CMIP5 captures a wide range of extremes and that bias correction can improve process-based model inputs, but outcome-focused evaluations (like maize yield responses to temperatures >29 °C) are rarer. The authors build on this gap by evaluating how bias-corrected, statistically downscaled data affect a statistical panel yield model, particularly for extremes.

Climate data: The NASA NEX-GDDP dataset provides statistically bias-corrected and spatially downscaled daily maximum/minimum temperature and precipitation from 21 CMIP5 models at 0.25° resolution, using the Bias-Correction Spatial Disaggregation (BCSD) algorithm applied at daily scale with GMFD observations. Historical period: 1950–2005; projections: 2006–2100 under RCP4.5 and RCP8.5. Corresponding parent CMIP5 simulations are also used. Yield model: A panel econometric model for county-level U.S. rainfed maize yields (counties east of the 100th meridian) with a quadratic, state-specific time trend and county fixed effects: log Y_it = f(t) + c_i + β1 GDD_it + β2 EDD_it + β3 P_it + β4 P_it^2 + ε_it. GDD and EDD are computed from sinusoidally interpolated diurnal cycles (from daily Tmax/Tmin), with maize-specific thresholds: 10 °C < T ≤ 29 °C for GDD, and T > 29 °C for EDD; aggregated to county level via area averaging. P is season-total precipitation (March–August). The model is trained on 1956–2005 for counties with ≥50% USDA data coverage. Model fit: overall R^2 = 0.81; Within R^2 = 0.25. Standard errors clustered at state level. For evaluation, the ‘weather only’ component log Y_B = β1 GDD + β2 EDD + β3 P + β4 P^2 is analyzed to isolate climate effects and ensure stationarity. Evaluation framework: Hindcasts from each CMIP5 and NEX-GDDP model are compared against yields simulated using the same observational product used for bias correction and downscaling (to avoid confounding with USDA data). The comparison is distributional (insensitive to year-by-year phasing) using empirical distributions of national and county-level log yields. Summary statistics include mean, standard deviation, 10th percentile (or 90th for EDD), and minimum (or maximum for EDD). County-level analyses identify spatial patterns. Additional evaluations compare underlying predictor distributions (GDD, EDD, precipitation) and assess precipitation–temperature correlations. Additional checks: (1) Examine correlations between precipitation and temperature variables (Spearman in observations: precipitation with GDD = −0.11 (P=0.43), with EDD = −0.23 (P=0.11)); assess model representation. (2) Compare alternative yield specifications using growing season averages of daily max/min/mean temperature (with/without precipitation), evaluate fit (AIC/BIC) and distributional similarity via two-sample Kolmogorov–Smirnov tests. (3) Analyze representative time series of Tmax/Tmin and cumulative EDD to illustrate integration-driven amplification of temperature biases into EDD errors. Projections and risk metrics: Under RCP8.5, mid-century (2030–2059) national-level yield distributions are generated for each ensemble. Return periods (1-, 5-, 10-, 20-year events) for large national-level yield drops are estimated to illustrate implications for risk management. National aggregation uses mean 1980–2020 production weights due to lack of future county production shares.

- CMIP5 hindcasts: Parent GCMs show a wide spread and tend to overestimate historical yield variability and tail statistics across many maize-growing counties; many summary statistics deviate by more than 100% from observations. - NEX-GDDP hindcasts: Bias-corrected/downscaled models align better with observations for several bulk metrics but overestimate average yields and underestimate interannual variability and the magnitude of the largest weather-induced yield declines. Nationally, the lowest observed yield (historical minimum) is underestimated by 63%–233% across models, with a median underestimation of 162%, while the 10th percentile is represented more accurately. - County-level patterns: In the Corn Belt (e.g., Illinois, Iowa), NEX-GDDP underestimates yield standard deviation by up to ~50% and substantially underestimates the severity of largest yield drops; a latitudinal gradient is evident with better performance in warmer southern counties. - Source of bias: The primary driver is poor representation of EDDs (integrated heat above 29 °C). NEX-GDDP captures GDDs and season-total precipitation comparatively well but underestimates EDD means, upper tails (90th percentile, maximum), and variability, especially in northern counties. CMIP5 tends to overestimate EDD variability and tail statistics, while generally representing GDDs and precipitation reasonably, aside from variability. - Correlational structure: Models from both ensembles overstate the magnitude of negative correlations between precipitation and both temperature variables compared to observations (observed Spearman: P–GDD = −0.11; P–EDD = −0.23). Despite making hot years relatively drier (which should lower yields), NEX-GDDP still underestimates the severity of yield losses, reinforcing EDD bias as the dominant issue. - Alternative temperature models: Yield models using seasonal mean/max/min temperature (rather than degree days) fit USDA data worse than degree-day models, but, when coupled with NEX-GDDP, more often pass distributional similarity (K–S tests) and better replicate the mean, minimum, and 10th percentile of observed-simulated yields; no meaningful improvement for variability. Degree-day models fail K–S tests (P<0.01) for all climate models. - Time-series illustration: For a representative grid point in McLean County, IL (1980 season), observed Tmax/Tmin excursions exceed the NEX-GDDP ensemble range, leading to under-accumulation of EDDs from which models never recover within the season, producing overestimated yields. - Projections (RCP8.5, 2030–2059): All models project increased yield variability and lower yields. NEX-GDDP projections show reduced variability and narrower tails relative to CMIP5. Return period analysis indicates systematically larger magnitude rare-event yield drops in CMIP5 than NEX-GDDP. Example 20-year event: median NEX-GDDP ~0.25 (ensemble 0.1–0.4) contribution to log-yield; median CMIP5 ~−0.29 (ensemble −2.4 to 0.54); historical ~0.36. Differences imply materially different risk assessments for stocking, insurance, and adaptation decisions.

The findings confirm that while bias correction and statistical downscaling improve agreement with observations for many bulk metrics, they can produce overconfident, overly narrow yield distributions that underestimate extreme, rare yield losses when using degree-day-based econometric models. The central mechanism is the underrepresentation of temperature extremes (EDD) in NEX-GDDP, particularly in cooler northern regions, whereas CMIP5 often overestimates EDD variability and extremes. This divergence leads to substantially different projections and risk metrics (e.g., return periods of large yield drops), creating non-trivial decisions for stakeholders about which climate dataset to use. Outcome-specific evaluation (e.g., yields’ sensitivity to temperatures above 29 °C) can yield different conclusions about dataset suitability than generic climate metrics, underscoring the value of sector-specific validation. Ensemble weighting/selection approaches could potentially reduce overconfidence or underconfidence but introduce additional assumptions and complexities.

Using a transparent, skillful econometric yield model, the study shows that statistically bias-corrected and downscaled climate products (NEX-GDDP) underestimate the largest weather-driven historical maize yield declines in the U.S., primarily due to underrepresentation of extreme heat (EDD). Conversely, parent CMIP5 models tend to overestimate variability and extremes. These biases propagate into projections, with NEX-GDDP yielding reduced variability and smaller tail risks compared to CMIP5, affecting decision-relevant metrics such as return periods. The work highlights trade-offs among resolution, historical fidelity for bulk metrics versus extremes, and projection confidence, and argues for holistic, outcome-specific hindcast evaluations before applying bias-corrected/downscaled products in impact assessments. Future research should evaluate more advanced bias-correction/downscaling techniques, consider alternative econometric structures and additional drivers (e.g., soil moisture), and assess generalizability across crops and regions.

- Bias-correction/downscaling method: Results pertain to the BCSD approach used in NEX-GDDP; more advanced methods may better represent variability and extremes. - Model structure: Conclusions may differ for econometric models using seasonal temperature averages rather than degree days, for other spatial/temporal scales, or for process-based crop models. - Yield model simplicity: Omits soil moisture dynamics, intra-seasonal timing effects, radiation, wind, humidity, management practices, and assumes fixed coefficients (no adaptation), no CO2 fertilization effects, and ignores air pollution changes. - Evaluation design: Observational dataset used both for bias correction and as reference for simulated yields; quantile estimation uncertainty is neglected but argued small relative to inter-model variation. - Spatial/temporal aggregation: National aggregation uses fixed 1980–2020 production weights due to lack of future shares; residuals exhibit spatial correlation not explicitly modeled. - Attribution: Because NEX-GDDP is only available post bias-correction and downscaling, isolating the exact contribution of each step to EDD biases is difficult.