Primary Submission Category: Matching
Cardinality Matching with Multiple Versions of Treatment
Authors: Lauren Liao, Amanda Ngo, Emily Wang, Rana Chehab, Yeyi Zhu, Samuel Pimentel,
Presenting Author: Lauren Liao*
Even when a study’s primary research question focuses on the effect of a binary treatment, it is common for some variation to arise in how treatment is delivered. For example, in a study comparing a medication to standard of care, different brands of that medication may be given to subjects. Effects of individual versions of treatment (e.g. one particular brand) are often of interest as secondary analyses, in addition to the effect of the overall composite treatment. However, a study that finds a balanced comparison between treated and control subjects for the overall treatment may not provide balanced comparisons for the individual versions of treatment. We propose a study design that creates a single matched comparison with balance both for the overall treatment comparison and for comparisons using individual versions of the treatment. We leverage cardinality matching to impose constraints on both overall and subgroup balance, and use a tuning parameter to encode a researcher’s relative preference between the two types of balance. We demonstrate this method using a large cohort study of pregnant subjects to evaluate the impact of categorizing subjects into stage 1 or stage 2 hypertension based on serial blood pressure measurements on adverse perinatal outcomes. Our analysis targets simultaneously the overall balance between stage 1 and stage 2 hypertension and subgroup balance, where stage 2 hypertension divides into individual versions with or without medication.