Abstract:
|
Monte Carlo methods were used to investigate the impact of covariate measurement error on the efficacy of propensity score (PS) methods. Seven factors were crossed in the simulation design: number of covariates (3, 9, 15, 30), population treatment effect (0, .2, .5, .8), covariate relationship to treatment (.1, .2, .4), covariate relationship to outcome (.1, .2, .4), correlation among covariates (0, .2, .5), sample size (50, 100, 250, 500, 1000), and covariate reliability (.4, .6, .8, 1.0). Each sample (5000 replications) was analyzed using seven PS methods (matching with and without a caliper, ignoring covariates, ANCOVA, PS as covariate, stratification, and PS weighting). Outcome measures included treatment effect bias, SE, 95% CI coverage and width. Results indicate that even low levels of measurement error lead to substantial statistical bias in treatment effect estimates and reduction in CI coverage. Such effects were evident across conditioning methods and effects increased with greater amounts of measurement error, larger numbers of covariates, and greater strength of relationship between the covariates and both the treatment assignment and the outcome variable.
|
Copyright © American Statistical Association.