|
Abstract:
|
Stratification, covariance analysis, matching and propensity analysis are some common methods of bias reduction when randomized balanced treatment assignment is absent. Propensity analysis is a conditionally specified bias reduction method in casual analysis. We consider the bias reduction in treatment comparisons when a large number of imbalance baseline predictors are available. Based on the marginal distribution for the covariates, dimension reduction methods in Regression Graphics will be used as an alternative to propensity analysis for reducing the bias in treatment comparisons. We show how sliced inverse regression (SIR) and other dimension reduction estimation methods such as SAVE, p.H.d and SIR-II for constructing effective directions, in addition to dimension reduction, can be used to reduce bias and, under ignorability condition, perform casual analysis.
|
