Propensity Score Adjustment Sample Clauses

Propensity Score Adjustment. Traditional methods to adjust for confounding include regression adjustment (often re- ferred to as epidemiologic models that include confounders in the final regression model) and matching treatment and control individuals based on identical (or at least similar) confounder values. However, both methods may be problematic in real world applications. Xxxxx has repeatedly voiced concerns about the possibility for traditional regression adjust- ment to actually add rather than reduce bias when response surfaces are nonlinear [1979], and stated that vastly differing groups were problematic for regression methods: “[t]he statistical literature has, for many years, warned that regression analysis cannot reliably adjust for differences in observed covariates when there are substantial differences in the distribution of these covariates in the two groups [2001].” As demonstrated by Xxxxxxx as early as 1957, in cases where the two comparison groups of interest differ substantially, regression adjustment merely pulls each group’s covariates toward a common mean, which may not accurately represent either group’s covariate distribution. As for matching tech- niques, Xxxxxxx noted [1972] that as the number of covariates on which it is desirable to match increases, the number of subclassifications within which a sample must contain both treatment and control observations increases exponentially and rapidly becomes untenable. Therefore one of the primary advantages of the propensity score method is that it provides a summary measure for a potentially large number of covariates, without the subclassification difficulties of direct matching on covariate values. Depending on which adjustment method is used, propensity score analyses may suffer from similar bias problems in the presence of non-linear response surfaces as traditional methods. Xxxxxxxxx and Xxxxx have shown [1983a] that propensity score analyses can be used to calculate an unbiased estimate of average treatment effect in observational studies. In their initial paper [1983a], only three methods of adjustment were outlined - pair matching, sub- classification, and covariance (regression) adjustment. Since then, these three methods have been expanded and added to - observations may be weighted based on inverse probability of treatment assignment (i.e., 1/propensity score) [Robins et al., 2000] or inverse odds ra- tios [Hirano et al., 2003], and matching may be 1 : 1 or 1 : k (i.e., matching one treated individua...