Join us on Feb. 22 for a great new presentation by Dr. Eloise Kaizar. This event is a joint effort between the Research Methodology Center and the OSU Department of Statistics. Attendees are welcome to attend one or both parts of this presentation.
Introduction to Cross-Design Synthesis and Generalizability
Eloise Kaizar, Department of Statistics, The Ohio State University
First Part (Feb. 22, 2018, from 2:00 – 3:00 pm)
“Cross-Design Synthesis: Estimating causal effects in target populations by using multiple study designs to correct for two types of selection effect”
Abstract: Collections of studies on similar topics not only provide more evidence about interventions or exposures, but their design heterogeneity may hold the key to answering difficult scientific questions. Researchers often favor randomized trials, since treatment randomization typically provides strong internal validity. However, convenient or constrained participant recruitment may limit the general usefulness of the study results. Studies with an observational design often have strengths and weaknesses opposite from their randomized counterparts, with weaker internal validity but broader applicability. I will overview methods that jointly analyze data from both types of study, and highlight cross-design approaches that exploit the heterogeneity of different designs.
Second Part (Feb. 22, 2018, from 3:00 – 4:00 pm)
“Generalizability: Reweighting approaches to estimating average treatment effect in a target population”
Abstract: When treatment effects are heterogeneous, average treatment effects that are estimated from typical randomized trials rarely directly correspond to such averages in target populations of interest that are defined by observational datasets. The mismatch between estimator and estimand arises from different distributions of treatment modifiers in the study sample and target population. Recently, researchers across several fields have been interested in statistical methods that adjust trial results to reflect the different distribution in the target population. Such approaches naturally arise from methods historically popular in both survey sampling and causal inference. However, many decisions needed to translate these historical methods to this new application area still require careful consideration. I will overview two such topics: the precise formulation of the estimator, and the estimation of the variance. I make recommendations for practice based the statistical properties that result from these choices.