Statistically bias-corrected and downscaled climate models underestimate the severity of U.S. maize yield shocks

Efforts to understand and quantify how a changing climate can impact agriculture often 2 rely on bias-corrected and downscaled climate information, making it important to quantify 3 potential biases of this approach. Previous studies typically focus their uncertainty 4 analyses on climatic variables and are silent on how these uncertainties propagate into 5 human systems through their subsequent incorporation into econometric models. Here, 6 we use a multi-model ensemble of statistically downscaled and bias-corrected climate 7 models, as well as the corresponding CMIP5 parent models, to analyze uncertainty 8 surrounding annual maize yield variability in the United States. We find that the CMIP5 9 models considerably overestimate historical yield variability while the bias-corrected and 10 downscaled versions underestimate the largest historically observed yield shocks. We 11 also find large differences in projected yields and other decision-relevant metrics 12 throughout this century, leaving stakeholders with modeling choices that require 13 navigating trade-offs in resolution, historical accuracy, and projection confidence. 14


18
Understanding and managing climate risk hinges on a quantitative description of the 19 human and Earth system dynamics as well as their associated uncertainties 1 . The fidelity 20 and utility of such analyses depend on using appropriate underlying climate information.

21
While global climate models (GCMs) can provide useful insights at the global scale, they 22 can face severe problems at regional to local scales. There are two main reasons for this.

23
For one, the native resolution of GCMs is too coarse to be used in fine-scale analysis 2 . In 24 addition, GCMs exhibit systematic biases relative to observations that require projections 25 to be appropriately corrected and interpreted 3,4 .

26
As a result, bias-corrected and downscaled climate products are widely employed 27 for a broad range of end-use applications, including diagnosing the impacts of climate 28 change on the economy 5 , energy demand 6 , and human populations 7 ; accounting for 29 climate change in regional infrastructure planning 8,9 ; and devising climate change 30 adaptation strategies from the national level 10 to city and urban areas 11 .

31
Although generally regarded as 'value-added' with respect to raw GCM outputs 12 , 32 bias-corrected and downscaled climate products still contain considerable uncertainties.

33
Examples include the validity of the stationarity assumption underpinning many bias-   Intercomparison Project Phase 5, or CMIP5 33 . By comparing yield hindcasts over the 57 historical period  to yields simulated by an observational dataset, we quantify 58 uncertainty due to bias-correction and downscaling as it pertains to crop yield outcomes.

59
Using a high-emissions scenario, we then project yields using both ensembles throughout 60 the remainder of this century and discuss notable differences and their implications. Our

178
for which adequate past performance is generally regarded as a necessary but insufficient 7 condition 37 . One hypothesis, based on the tendency of the NEX-GDDP ensemble to 180 underestimate historical yield variability, is that projections are similarly overconfident.

181
There are considerable differences between the yield projections of the NEX-182 GDDP and CMIP ensembles.

304
Our resulting yield model is then:

306
Here, i denotes the spatial index (for this analysis, each county in CONUS), t denotes the 307 temporal index (each year), Y is the yield in bushels/acre, GDD denotes growing degree 308 days, EDD denotes extreme degree days, and P denotes total in-season precipitation. A

341
Our yield model is able to reproduce USDA-recorded yields with good accuracy 342 ( Figure S10). In the training period, the median R 2 for all 2,371 counties is 0.35; for the   Tail-area probabilities of modeled maize yields for the NEX-GDDP ensemble measured against observationally driven yields. Results are shown for standard deviation (a), median absolute deviation (b), and the magnitude of the largest negative yield shock (c). Stippling indicates tail-area probabilities less than 0.01 or greater than 0.99.

Figure 3
Tail-area probabilities for each climate variable used in the yield model, calculated from the NEX-GDDP ensemble. The left column shows the results for standard deviation; the right column for median absolute deviation (MAD). Climate variables are organized by row: growing degree days (top), extreme degree days (middle), and precipitation (bottom). Stippling indicates tail-area probabilities less than 0.01 or greater than 0.99.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. NCCSIMaizeUncertaintyBCSD.pdf