Targeted Maximum Likelihood Estimation (TMLE) of Causal Effects of Single Time Point Interventions in Non-IID Data: A Simulation Study

• Activity Number: 353
• Type: Contributed
• Date/Time: Tuesday, August 5, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Oleg Sofrygin*+ and Mark J. van der Laan
• Companies: and University of California, Berkeley
• Keywords: networks ; non-IID data ; interference ; semi-parametric estimation ; targeted maximum likelihood estimation ; simulation study

Abstract

We conduct a simulation study to examine the performance of the novel doubly-robust TMLE of the causal effect of treatment in the presence of dependencies among observations, often referred to as interference or spillover. We simulate several networks of dependent units, where each unit's treatment is generated as a function of the baseline covariates in unit's network. Similarly, each unit's outcome is generated as a function of the baseline covariates and treatments in unit's network. Our causal parameter of interest is the expectation of the average counterfactual outcome over all units under point-treatment intervention. We propose a nonparametric structural equation model, formally define the causal parameter of interest and establish identifiability of the causal parameter from the observed data, under appropriate conditions. In our simulations we demonstrate that TMLE is doubly-robust and asymptotically efficient. We investigate TMLE's finite sample properties in comparison to other estimators, such as parametric maximum likelihood and inverse probability weighted (IPW). Finally, we propose a conservative estimator of the TMLE variance and construct confidence intervals.