Estimation of high-dimensional propensity scores with multiple exposure levels.

PURPOSE:

Little information is available on the performance of high-dimensional propensity scores (HDPS) in settings with more than two exposure levels. Our objective was to adapt the HDPS algorithm to allow for the inclusion of multilevel treatments and compare estimates obtained via this approach with those obtained via pairwise comparisons in a case study using real-world data.

METHODS:

We conducted a retrospective cohort study of cardiovascular events associated with three smoking cessation drugs (varenicline, bupropion, nicotine replacement therapy [NRT]) using the Clinical Practice Research Datalink. We applied the binary HDPS algorithm adjusted for pre-specified and empirically-selected covariates to cohorts formed by each treatment pair. We then constructed multinomial HDPS models on a cohort of new users of any of the three drugs, adjusting for predefined covariates and different combinations of empirically-selected covariates. After trimming the area of non-overlap of the HDPS distributions, the effects of the study drugs on cardiovascular events were estimated with the Cox proportional hazards models adjusted for propensity score category.

RESULTS:

Outcome models adjusted for multinomial HDPS estimated treatment effects that were slightly more protective than those estimated in pairwise comparisons (varenicline vs

NRT:

HRMultinomial = 0.60-0.62, HRPairwise = 0.64; bupropion vs

NRT:

HRMultinomial = 0.70-0.72, HRPairwise = 0.76). Trimming rates were similar between the two approaches.

CONCLUSIONS:

The extension of HDPS to multilevel exposures is a valid and practical approach to confounder control that may be useful when comparing different classes of drugs prescribed for the same indication or different molecules within a given drug class.