Model-Based Comparative Effectiveness Estimation for Observational Time-to-Event Data Using Instrumental Variables

• Activity Number: 517
• Type: Contributed
• Date/Time: Wednesday, August 1, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Purushottam Laud*+ and Rodney Sparapani and Jessica Pruszynski and Robert McCulloch
• Companies: Medical College of Wisconsin and Medical College of Wisconsin and Medical College of Wisconsin and The University of Texas
• Keywords: comparative effectiveness research ; observational data ; instrumental variables ; Bayesian methodology ; time-to-event data ; selection bias

Abstract

In estimating comparative effectiveness of alternative medical treatments from observational data, it should be recognized that patients are not randomized to treatment. Self-selection of treatment, if taken as randomized assignment, leads to systematic error in estimation. One approach to avoiding this error utilizes instrumental variables (IVs) that predict treatment choice but, conditioned on the actual treatment chosen, do not affect the outcome variable. With this approach traditional methods employ moment-based IV estimates that are asymptotically unbiased. These methods, working mainly with linear cases and linear approximations to others, are not directly suitable for time-to-event data. In this talk we present a complete probability model and an accompanying Bayesian inference procedure. The method employs separate models for treatment choice and patient outcome. The former is a probit model, the latter a Cox model. The two are linked via joint modeling of the two error terms. Inference is implemented using Markov chain Monte Carlo techniques. The talk outlines the method development, then illustrates its use and need via real data and simulation examples.