Population-level estimation of comparative risks of Celecoxib versus non-selective NSAIDs

Background Observational data is often used to estimate causal effects of prescription drugs on health outcomes. One popular study design that is used for this purpose is the cohort method, where we compare the risk of the outcome of interest in two groups (cohorts) of subjects. Typically the two groups represent two treatment options, where one option could be no treatment. In many ways the cohort method mimics a randomized controlled trial, except that the assignment of subjects to one of the two groups is not random. This non-random assignment can easily lead to confounding, when variables associated with the treatment assignment are predictors of the outcome. Two options are often considered for reducing the risk of confounding. The first option is to fit a propensity score (PS) model, a model that tries to predict treatment assignment based on information available at the start of treatment, and use the PS to make the two groups more similar for example by stratifying the analysis based on PS. The second option is use an elaborate outcome model that tries to predict the outcome not only based on the treatment assignment, but also covariates that are potentially risk factors for the outcome. Traditionally, the covariates included in the PS model and outcome model are hand-picked based on expert assessment of potential for confounding. The CohortMethod package developed in OHDSI (Observational Health Data Science and Informatics) uses a different approach, where very large sets of covariates are created, and regularized regression is used to estimate models. In the study described here we would like to provide a proof of principle of the CohortMethod package, and more specifically of our approach to generating and incorporating covariates into the cohort method. In order to this we will focus on the well-documented example of celecoxib vs nsNSAIDs for GI bleed. We would also like to show that studies implemented using the CohortMethod package can be easily deployed in a distributed research network, and across several other outcomes.

Objective

Primary objective - Show that the OHDSI CohortMethod is capable of reproducing known findings Secondary objectives - Show effect of various forms of adjusting for confounding - Show feasibility of running CohortMethod in a distributed data network

Project Lead(s): Marc Suchard

Coordinating Institution(s): UCLA

Additional Participants: Martijn Schuemie, Patrick Ryan

Full Protocol: Word doc for the protocol

Initial Proposal Date: 7/10/2015

Launch Date: TBD

Study Closure Date: TBD

Results Submission: Email