Speaker:Jingchao Sun, Ph.D Candidate, Department of Bioinformatics and Biostatistics, University Of Louisville

"Directed Acyclic Graph Assisted Methods For Estimating Average Treatment Effect"

Observational data, such as electronic clinical record and claims data, can be very useful to examine average treatment effects (ATE) and help decision making, if used appropriately. Propensity score based inverse probability weighting (IPW) method has been very powerful to estimate ATE if the assumptions on exchangeability and positivity hold. Directed acyclic graph (DAG) provides a feasible approach to examine the exchangeability assumption, that is, the treatment and the potential outcome are independent given the variables which block the back-door path from treatment to the potential outcome. That is, we do not need to adjust all confounding variables, instead we only need to adjust the variables which block the back-door paths from treatment to outcome. There could be several sets of such variables. We consider a minimal set of the variables which block the backdoor from treatment to the potential outcome, and we also consider a maximum set of such variables. We carry out extensive simulations to examine the performance of the propensity score based IPW method in estimating ATE when different sets of confounding variables are included in the propensity score estimation. The simulation results indicate that the performance of ATE estimation based on the minimal set of confounding variables is comparable with the maximal set of confounding variables. We applied the method to examine if tracheostomy (a medical procedure involving creating an opening in the neck in order to place a tube into a person's windpipe) is a cause of death for infants based on the 2016 health-care cost and utilization project (HCUP) kids inpatient database (KID).

Stay connected TwitterFacebookLinkedInYouTubeInstagram