Confounding is usually controlled by either cross-stratification or multivariate modeling. The first approach is simple and intuitive, but it is not practical for controlling many factors. The second approach, although less intuitive, may provide a more efficient means for controlling many confounders, but its ability to control confounding depends on the appropriateness of the chosen model. Hybrid methods based on a multivariate confounder score or a propensity score combine the favorable characteristics of both methods and may be better suited for controlling many confounders. However, the resulting strata are defined by subranges of a multivariate model, and, therefore, may possess little intrinsic meaning. The authors propose the principle of asymmetric stratification to control efficiently a number of confounders in cohort studies while retaining the intuitive appeal and general framework of cross-stratification. The proposed method resembles a propensity score analysis but does not use a multivariate model to define the strata. Instead, strata are defined by the categories of only a subset of the original potential confounders. The authors also demonstrate how our proposed method can be implemented by an application of classification and regression trees (CART) (recursive partitioning), as outlined by Breiman et al. (Classification and Regression Trees. Belmont, CA: Wadsworth, 1984). Computer simulations and an actual example suggest that the proposed method is a potentially simpler alternative to the standard propensity score analysis. Specific recommendations on how the proposed method can be improved are also presented.
Cook et al. (1988) studied this question.