CMStatistics 2021: Start Registration
View Submission - CMStatistics
B1052
Title: Imputing biomarkers from cognitive assessments: combating covariate shift by assuming causal stationarity Authors:  Chelsea Krantsevich - Arizona State University (United States) [presenting]
Richard Hahn - Arizona State University (United States)
Abstract: Motivated by the problem of developing accurate biomarkers to track the progression of Alzheimer's disease, the aim is to consider how incorporating a causal understanding of the underlying biology can improve the prediction of biomarker trajectories. We introduce a causal imputation method based on biologically-motivated causal graphs and compare its performance to an unconstrained supervised learning method that ignores causal relationships. We demonstrate that the causal approach is substantially more accurate in the presence of ``covariate shift'', where the test population differs in important but unforeseen ways from the training population.