Causal mediation analysis under the presence of unmeasured confounder
Causal mediation analysis under the presence of unmeasured confounder
Date: | Tuesday, April 08 |
Time: | 4:10 am -- 5:00 pm |
Place: | Morrill Hall 2019 |
Speaker: | Cheng Zheng, University of Washington, Seattle |
Abstract:
Causal mediation analysis is an important tool in medical and behavioral sciences as it helps understand why and how an intervention works. Knowledge on the causal relationships can further assist the design of future interventions. Traditional regression based methods do not guarantee causal interpretation. Rank preserving model has been proposed to handle the unmeasured confounder for binary intervention and single mediator cases. The count type outcomes are frequently observed in the studies of risk behavior. In this presentation, we propose a general method that can handle multi-arm treatment, multiple mediators for both continuous, count type outcomes. We illustrate our methods with two college drinking studies. We found that the effect of a dialectical behavior therapy on depression is not fully mediated by its effect on emotion regulation. We also confirmed that the normative perception of peer drinking is causal related to one’s drinking behavior. Finally we show how the censoring, measurement error and interference among subjects can be handled and discuss some future directions.