A Characterization of Optimal Designs for Observational Studies

摘要: SUMMARY An empirical investigation of the effects a treatment is an observational study if it involves comparison treated and control groups that were not formed by randomization. In such studies, may differ systematically with respect to pretreatment measures or covariates, addressing these pre- differences central concern. Matching subclassification are two standard methods adjusting for observed differences; they be used alone (e.g. Cochran (1968)) in conjunction analytical model-based adjustments Rubin (1973, 1979), Holford (1978), Rosenbaum (1984), section 3.3, (1987, 1988a)). particular, Rubin's simulation studies suggest applied matched subclassified samples more robust than unmatched samples. The purpose this paper identify structure optimal subclassification, i.e. one which subjects same subclass are, on average, as similar possible covariates. It turns out simple intuitive. While imbalances cannot