The NA-CORDEX data archive provides climate change scenario data over North America for use in impacts, decision-making, and climate science. This data comes from RCM simulations dynamically downscaling CMIP5 GCMs run as part of the WCRP Coordinated Regional Climate Downscaling Experiment (CORDEX).

These climate model outputs often contain bias that hinders their use in impacts and adaptation analysis. To improve their usability, the NA-CORDEX data archive provides both raw simulation outputs and outputs with statistical adjustments to reduce the bias. We use Cannon's MBCn method for multivariate bias correction, a quantile-mapping approach that corrects the joint distributions of the variables as well as their marginal distributions.

The selection of observational datasets suitable for use as the basis of these bias-corrections is limited, and even though the climatologies may be very similar, the daily values at a given time and location can vary considerably from one another. To evaluate the uncertainty associated with the choice of observational dataset, we bias-correct outputs from the NA-CORDEX data archive using two different observational datasets and compare the results.

The gridMET dataset from the Univeristy of Idaho's Climatology Lab uses climatically aided interpolation to blend 4-km monthly data from PRISM with 12-km daily data from NLDAS-2 to generate 4-km daily data for a suite of impacts variables over CONUS. The Daymet dataset from NASA's ORNL DAAC processes GHCND station data to generate 1-km daily data for a slightly different suite of impacts variables over the entire United States, Canada, and Mexico.

We average these datasets to a common quarter-degree lat-lon grid and evaluate the differences between them, as well as the differences in bias-corrected model outputs that result. We find that in both cases the differences can be significant and pervasive.