Software for Survival Analysis Using Structural Models

Period of Performance: 05/01/2007 - 04/30/2008


Phase 1 SBIR

Recipient Firm

Insightful Corporation
Seattle, WA 98109
Principal Investigator


DESCRIPTION (provided by applicant): Causal inferences from observational data increasingly drive policy and medical decisions aimed at decreasing incidence rates of disease and premature death. Standard statistical models for survival analysis (e.g., Cox proportional hazards model, accelerated failure time model) are commonly used to make these inferences. Unfortunately, these methods are not designed to handle longitudinal data with time-varying treatments because they impose unrealistic assumptions regarding time-dependent confounding. Further, they cannot estimate either the joint effect of multiple treatments, or the effect of dynamic interventions (i.e., interventions that vary over time to maximize health benefits). Consequently, the most effective interventions may not be incorporated into policies for preventing morbidity and premature mortality. As an alternative to standard models for survival analysis, there exist "causal" or structural semiparametric models for estimating the causal effects of time-varying exposures on survival. These structural models overcome the limitations of the standard statistical methods listed above, and therefore deliver a better answer to key health questions. The semiparametric theory underlying these models has been developed by James Robins and his collaborators. We propose to create software that implements causal models for survival analysis: marginal structural Cox models and nested structural accelerated failure time models. Our ultimate goal is to develop a modular addition to Insightful's data analysis and data mining products called S+STRUCTURALFTA for survival analysis of complex longitudinal data with time-varying exposures and confounders. We also propose to produce detailed case studies that guide analysts in applying the proposed methods. Making these methods accessible to practicing epidemiologists will better inform scientific and policy decisions for preventing morbidity and premature mortality. S+STRUCTURALFTA will provide a set of tools to help researchers better assess the effects of health interventions from complex observational data, even when it is too expensive, too time consuming, or not ethical to conduct randomized trials. We anticipate a ready market for such software among medical research institutions, epidemiology centers, universities and in pharmaceutical and biotechnology companies. It could also be useful in other fields such as education, sociology, marketing, and econometrics that aim to estimate causal effects from observational longitudinal data.