Examining the Impact of Bias Correction Configurations on a Multivariate Meteorological Index: A Case Study of Heat Index Analysis in the Northeast United States Journal of Applied Meteorology and Climatology

Bias correction (BC) is commonly used to refine dynamical climate model results for use in applied research. However, uncertainties linked to the BC of multivariate meteorological indices, crucial for many applied research applications, are often underappreciated. Focusing on daily maximum values of the multivariate heat index (HI, a function of temperature and relative humidity), this study investigates uncertainties across four BC workflow configurations available to users with varying expertise levels and data availability. Univariate quantile delta mapping (QDM) and multivariate BC (MBC) were used to bias correct nine global climate models trained on 30 years of data from 13 Northeast U.S. weather stations. Two workflows use QDM to directly BC HI variables. Two workflows adopt different componentwise approaches to BC temperature and humidity, computing HI from those bias-corrected outputs. Workflow accuracy is assessed by comparing extreme HI day occurrences against reference station data. The best results are produced by the QDM-direct workflow, when hourly historical station data provide accurate daily maximum HI for training. Conversely, the common componentwise QDM (QDMc) workflow that independently adjusts temperature and humidity yields errors up to 147%, due to lost intervariable dependence. The MBC-N-dimensional (MBCn) componentwise approach, which jointly adjusts temperature and humidity, better preserves correlations than QDMc, yielding more reliable results. These findings underscore that BC methodological choices, including the order of operations, training data accuracy, and whether variable dependence is preserved, markedly impact accuracy. This offers practical guidance for researchers and practitioners quantifying climate risks using bias-corrected climate model data. Significance Statement This research quantitatively evaluates the uncertainties introduced by different bias correction methods and analysis configurations for multivariate meteorological indices, using extreme heat index days as a case study. The findings show that, depending on analytical choices and data availability, certain bias correction setups can result in errors as high as 150% in the Northeast U.S. study region. These results highlight the importance for applied researchers, data developers, and end users to critically assess bias correction products and account for their uncertainties in quantitative analyses, especially when considering multivariate stressors. Grant no. NA18OAR4320123