Abstract:
|
In comparative effectiveness research using observational data sources, control for measured confounding can be achieved through numerous methods. With rare binary outcomes (e.g., 1% of sample experiences event), over-fitting is a concern, which is used as rationale for preferring propensity score methods over regression. In a simulation study, we evaluate six methods for treatment effect estimation when the outcome is rare: 1) covariate-adjusted logistic regression, 2) standardization, 3) 1:1 matching with propensity score, 4) inverse probability of treatment weighting with propensity score, 5) propensity score-adjusted logistic regression, and 6) overlap weights with propensity score. We vary the sample size, outcome prevalence, and treatment proportion to assess the bias and efficiency of each method. Results are presented relative to target estimands: marginal or average treatment effect and the conditional treatment effect.
|