Statistical Issues in Comparative Effectiveness Research

The goal of this research is to provide empirical results that can be used to guide decisions regarding treatments/interventions. This work focuses on two different problems of interest in comparative effectiveness research. The first problem is to understand if the proportion of an event changes over time, when the populations are nested within each other. This often happens in the health care system and is illustrated here by a study of hospital readmissions within 48 hours of a first visit to an emergency department (ED). The nested structure of the data must be taken into account at the analysis stage and there are no standard statistical methods for doing this. We propose a likelihood ratio test based on the product of conditional probabilities in the form of generalized mixed model. This test accommodates conditionality, within subject dependence and between hospital cluster effects. Simulations show that it preserves the type-I error level given no difference, and provides estimates that are less biased in the presence of a large cluster effect. This approach can be implemented using SAS PROC NLMIXED making it easy to apply in this setting. The second problem focuses on the identification of subgroups within a clinical trial, with the goal being the identification of subjects who benefit from the treatment of interest. The focus is on the use of interaction trees which are an extension of the classification and regression trees (CART). The use of interaction trees overcomes both the subjectivity and multiple comparisons that plague a conventional subgroup analysis. However, the method is greedy in finding each local node by exhausting every predictor and its available values. We propose a greediness reduction interaction tree (GRIT) algorithm that integrates random forests and the evolutionary algorithm into the interaction trees. Simulations show that this proposed method outperforms the interaction trees without accessing every predictor given the interaction. The strengths of the proposed method are demonstrated through a real data example from the Biological Markers for Recovery of Kidney (BioMaRK) study. Public Health Significance: Two methodologies proposed provide less bias and more accurate information under certain circumstances. One is for medical and public policy decisions based on administrative datasets and the other is for finding subgroups and generating hypotheses for future clinical trials.