Integrating the Evidence from Evidence Factors in Observational Studies

• Activity Number: 72 - Methods for Causal and Integrative Analysis in Health Studies
• Type: Contributed
• Date/Time: Sunday, July 30, 2017 : 4:00 PM to 5:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Bikram Karmakar* and Benjamin French and Dylan Small
• Companies: University of Pennsylvania and Radiation Effects Research Foundation and University of Pennsylvania
• Keywords: Causal inference ; Design sensitivity ; Evidence factors ; Observational study ; Sen­sitivity analysis

Abstract

In an observational study, it is typically assumed that treatment is effectively randomly assigned conditional on covariates, but there is often uncertainty about whether this assumption holds. A sensitivity analysis determines how much the assumption can be violated, i.e., how much bias in treatment assignment there could be, without refuting the conclusion. Evidence factors allow for testing a hypothesized treatment effect in more than one way using the same observational study where tests are based on different assumptions. Inferences based on each test could be affected by different biases. This paper develops methods for combining the evidence from the evidence factors. In combining the evidence factors, we need to be cautious about the overall type I error incurred. We develop a method for controlling the family wise error rate for sensitivity analysis with evidence factors. We show that by combining the evidence factors appropriately, the Bahadur efficiency of sensitivity analysis is greater than for either evidence factor alone. We illustrate our methods in a study of the effect of radiation exposure on cancer.