Bounds on the conditional and average treatment effect in the presence of unobserved confounders
share
The causal effect of an intervention can not be consistently estimated when the treatment assignment is influenced by unknown confounding factors. However, we can still study the causal effect when the dependence of treatment assignment on unobserved confounding factors is bounded by performing a sensitivity analysis. In such a case, the treatment effect is partially identifiable in that the bound of the treatment effect is still estimable based on the observed data. Here, we propose a sensitivity analysis approach to bound the conditional average treatment effect over observed covariates under bounded selection on unobservables. Additionally, we propose a semi-parametric method to estimate bounds on the average treatment effect and derive confidence intervals for these bounds. Combining the confidence intervals of the lower and upper bound gives a confidence region that includes the average treatment effect when the bounded selection on unobservables holds. This method scales to settings where the dimension of observed covariates is too high to apply a traditional sensitivity analysis based on covariate matching. Finally, we provide evidence from simulations and real data to illustrate the accuracy of the confidence intervals and value of our approach in practical finite sample regimes.
READ FULL TEXT
