Inverse probability of treatment weighting (IPTW) is frequently used to estimate the causal effects of treatments and interventions. The consistency of the IPTW estimator relies not only on the well-recognized assumption of no unmeasured confounders (Sequential Randomization Assumption or SRA), but also on the assumption of experimentation in the assignment of treatment (Experimental Treatment Assignment or ETA). In finite samples, violations in the ETA assumption can occur due simply to chance; certain treatments become rare or non-existent for certain strata of the population. Such practical violations of the ETA assumption occur frequently in real data, and can result in significant bias in the IPTW estimator of causal effects. This manuscript presents a diagnostic tool for assessing the bias in the IPTW estimator due to violation of the ETA assumption. The Diagnostic of ETA Bias (DEB), implemented in a public R routine, relies on parametric bootstrap sampling from an estimated data-generating distribution. The article presents results of simulations to assess the performance and applications of the diagnostic and to investigate the extent of ETA bias in a range of contexts. In addition, results are presented from two data examples drawn from the treatment of HIV infection, in which DEB is used to assess the ETA bias of the IPTW estimator.


Statistical Methodology | Statistical Theory