Online Program

Return to main conference page
Friday, October 19
Fri, Oct 19, 11:45 AM - 1:15 PM
Caprice 3-4
Speed Session 3

Inverse Probability Weighting Approach for Modeling Longitudinal Change in Disease Biomarkers in the Presence of Disease Treatment (304824)

*Nicole M Butera, The University of North Carolina at Chapel Hill 
Jianwen Cai, The University of North Carolina at Chapel Hill 

Keywords: longitudinal change, biomarkers, inverse probability weighting, doubly robust, Cox regression

In observational public health studies, there is often interest in identifying factors associated with longitudinal change in disease biomarkers. However, some people may start treatment during follow-up, resulting in healthier biomarker values than would be observed if untreated. If the biomarker is related to the disease being treated, then treatment initiation would be informative for the biomarker value that would have been observed at the follow-up visit if treatment had not been received, because less healthy people are more likely to start treatment. We propose inverse probability weighting to address this problem. We model time between baseline and treatment initiation using Cox proportional hazard regression to estimate the probability of not starting treatment before the follow-up visit. Then a weighted regression model is fit for longitudinal change in the biomarker, where weights are calculated based on this estimated probability. Doubly robust estimators are also introduced. We derive asymptotic properties of these estimators and present a simulation study evaluating finite sample properties. We apply these methods to the Atherosclerosis Risk in Communities study.