Performance of the High-dimensional Propensity Score in a Nordic Healthcare Model

The high-dimensional propensity score (hdPS) is increasingly used as a tool to adjust for confounding in observational studies of drug effects. It was developed within very rich data sources, for example the American claims databases. Thus, it is unknown whether it can be applied in settings that provide little more than primary care prescriptions and diagnoses from hospital contacts, as in the Nordic data sources. Our objective was to evaluate the performance of hdPS under such circumstances. As our case, we chose the association between use of selective cyclooxygenase-2 inhibitors (coxibs) and traditional NSAIDs (tNSAIDs) and the risk of upper GI bleeding. Using Danish health registries, we identified 110,285 incident users of coxibs and 575,980 incident users of tNSAIDs and followed them for 90 days with respect to the occurrence of serious upper GI bleeding. Data were analysed using Cox regression, estimating the coxib/tNSAID hazard ratio (HR). Values below 1.00 indicate a lower estimated hazard with coxibs. We build hdPS models with inclusion of up to 500 diagnosis and 500 prescription drug covariates. The crude HR was 1.76 (95% confidence interval: 1.57-1.97), decreasing to 1.12 (1.00-1.26) and 0.99 (0.88-1.12) after adjustment for age and sex and 11 pre-selected confounders, respectively. A hdPS with inclusion of 500 most prevalent diagnoses and 500 most prevalent prescription drugs resulted in a HR of 0.89 (0.77-1.02). These estimates were consistently lower when the analysis was restricted to non-users of low-dose aspirin. The estimate based on 500 diagnoses alone was higher than an estimate based on 500 prescription drugs alone (0.99 versus 0.91). We conclude that hdPS does work within a Nordic setting that prescription data are more effective than diagnosis data in achieving confounder adjustment and that hdPS seems more effective than simple confounder adjustment by variables selected on the basis of clinical reasoning.