Evaluation of the Performance of the High-Dimensional Propensity Score Algorithm to Adjust for Confounding of Treatment Effects Estimated in Healthcare Claims Data Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 22, 2019
  • Le, Hoa Van
    • Affiliation: Gillings School of Global Public Health, Department of Epidemiology
  • The aim of the high-dimensional propensity score (hd-PS) algorithm is to select and adjust for baseline confounders in pharmacoepidemiologic studies based on healthcare claims data. It is not well understood how the performance of the hd-PS is affected by 1) the channelling of drugs at specific calendar time periods and differences in administrative claims databases; 2) low outcome incidence or exposure prevalence in medium sized or large cohorts; and 3) aggregation of medical diagnoses and medications in cohorts with small size, low outcome incidence and low exposure prevalence. We estimated risk ratios for upper gastrointestinal complication in patients with rheumatoid arthritis or osteoarthritis after initiating oral celecoxib versus ibuprofen or diclofenac in two large longitudinal healthcare claims databases. We conducted separate analyses for subcohorts before and after withdrawal of rofecoxib, a drug in the same class as celecoxib. We applied the hd-PS algorithm using a combination of demographic, predefined and hd-PS covariates with either PS deciles or 1:1 greedy matching for each cohort. In addition, we conducted pooled analyses for two combined databases stratified by data source and adjusted by either deciles of separate PSs or 1:1 greedy matching within the data source. The different methods of propensity score confounder selection inconsistently reduced confounding by indication across calendar time periods and administrative data sources. To evaluate the effects of aggregation of medical diagnoses and medications on the performance of the hd-PS, we resampled studies to assess the influence of size, outcome incidence, and exposure prevalence. For each sample, baseline covariates were identified with and without the hd-PS algorithm to estimate the treatment effect using propensity score deciles. In an empirical pharmacoepidemiologic study using claims data, aggregations of medications into chemical, pharmacological or therapeutic subgroups (level 4) of the Anatomical Therapeutic Chemical classification alone or in combination of aggregation of diagnoses into largest groups (level 1) of the Clinical Classification Software improved the hd-PS adjustment for confounding in most scenarios including ones with small cohort size, rare outcome incidence, and low exposure prevalence.
Date of publication
Resource type
Rights statement
  • In Copyright
  • Stürmer, Til
  • Doctor of Philosophy
Graduation year
  • 2012

This work has no parents.