|
Abstract:
|
Recent methodological developments in precision medicine allow estimating heterogeneous treatment effects from observational data. Two novel approaches, called twin and contrast regressions, use cross-validation to construct an individualized treatment score by solving doubly-robust estimating equations that adjust for baseline covariate imbalances between the treatment arms. However, it does not take into account post-treatment information such as suboptimal treatment adherence or treatment discontinuation. Therefore, the estimated individualized treatment score may be biased if the patients’ adherence and discontinuation profiles are not balanced over time between the treatment arms. We propose an extension of twin and contrast regressions that uses artificial censoring with a weighting argument to account for suboptimal adherence and discontinuation. This presentation is a proof-of-concept about the importance of such adjustment, with a demonstration motivated by the comparison of treatments for multiple sclerosis.
|
