A Marginal Structural Cox Model Based Analysis Of The Comparative Effectiveness Of Two Dialysis Therapies

Abstract

The Institute of Medicine identified comparing the effectiveness of renal replacement therapies as the only kidney-disease related topic among the top 100 initial national priorities for comparative effectiveness research. The median life expectancy of patients starting renal replacement therapy is only a little over three years and they spend on average 12 days in the hospital annually. The overwhelming majority of patients are treated with thrice-weekly in-center hemodialysis(TWICHD). However, an increasingly larger number of patients are being treated alternate dialysis modalities, including nocturnal in-center hemodialysis (NICHD), which is a modified hemodialysis regimen that include significantly longer treatment times. As such, it is critically important to perform a rigorous assessment of the benefit, if any, conferred by NICHD. In this thesis, we present a careful analysis of a large cohort from a retrospective study of dialysis modalities in the US, using marginal structural Cox models. The data set consists of 208,820 patients who initiated dialysis between January 1, 2007- December 31, 2011 at facilities owned by Davita corporation. Prior to presenting the data analysis, we provide a overview of the relevant concepts from the theory of the counterfactual model and causal inference, and discuss marginal structural models for binary as well as time to event outcomes. For the primary analysis, we estimated the causal hazard ratio comparing the risk of death if patients received NICHD as opposed to TWICHD through their follow-up, contrary to their observed treatment history using a pre-specified marginal structural Cox model. We found that those who were treated with NICHD had a substantial reduction in their risk of death. We also performed a series of sensitivity analyses to assess the effect of potential sources of bias which were not accounted for in the primary analysis. We found that our main result was robust to these potential sources of bias. In addition, using a bootstrap procedure, we also investigated the effect of the parametric assumptions we made on the precision of our estimates. We found that the model we used provided a substantial gain in precision of our estimate of the causal effect of NICHD on survival.