Understanding Algorithmic Bias in Clinical Prediction Models

• Activity Number: 286 - Missing Data Methods
• Type: Contributed
• Date/Time: Wednesday, August 11, 2021 : 1:30 PM to 3:20 PM
• Sponsor: Biometrics Section
• Author(s): Mengying Yan* and Michael Pencina and Benjamin A Goldstein
• Companies: Department of Biostatistics & Bioinformatics, Duke University and Department of Biostatistics & Bioinformatics, Duke University and Duke University
• Keywords: Racial Disparity; Clinical Prediction Models; Electronic Health Records Data; Missing Data

Abstract

Informative visit process can induce biases. In this paper we extend this work to clinical prediction models (CPMs) and discuss how including a grouping variable such as race in CPMs could induce algorithmic bias. We define such bias as differential miscalibration and frame the creation of it via a differential missing data process, where conditional on having the outcome of interest the outcome is differentially observed. Simulations are conducted based on a two-stage process where the grouping variable may impact either the risk of the true outcome (stage 1) or probability of observing the outcome (stage 2). Calibration plots show that if the grouping variable only impacts the true outcome, then including it mitigate the algorithmic bias; if the grouping variable only impacts the observation process, including it induce algorithmic bias. For more complex cases, the algorithmic bias is determined by the direction and magnitude of the effects of that variable on both stages. Furthermore, the potential correlation between the grouping variable of interest and other risk factors, or the grouping variable as a proxy don’t meaningfully affect the results.