Ph.D. Candidate
Department of Biostatistics
Virginia Commonwealth University
Richmond, VA

While increasingly popular because of access to massive longitudinal patient records (i.e. EHR and claims), observational studies, compared to RCTs, are more prone to bias and confounding. Reporting p-values in observational studies does not contextualize intrinsic systematic error resulting in possibly unreliable evidence. A method, introduced by Schuemie et al 2012, to identify and mitigate systematic error in observational studies is to derive an empirical null distribution using effect estimates from negative controls, exposure-outcome combinations with known “no association” (i.e. RR=1). Now accounting for systematic and random error, the empirically calibrated null distribution is used to interpret the effect estimate and p-value for the exposure-outcome combination of interest. This presentation introduces the theoretical foundation of empirical calibration, previews the framework of analysis and presents two examples from replications of observational studies: 1) a new-user cohort design by Graham et al 2016 comparing risk of gastrointestinal hemorrhage of dabigatran versus warfarin for treatment of nonvalvular atrial fibrillation in Truven Medicare Supplementary Dataset and 2) comparison of a case-control and self-controlled case series design by Tata et al 2005 for evaluating the risk of gastrointestinal bleeding in users of selective serotonin reuptake inhibitors (SSRIs) in General Electric Centricity database.

plate with fork and knife, books, microscope and test tubes
Sponsor(s)
Medicine: Biostatistics
Speaker(s)
Martin Lavallee
Audience
All ( Open to the public )