Evaluating the impact of instrumental variables in propensity score models using synthetic and negative control experiments

Abstract

In pharmacoepidemiology research, instrumental variables (IVs) are variables that strongly predict treatment but have no causal effect on the outcome of interest except through the treatment. There remain concerns about the inclusion of IVs in propensity score (PS) models amplifying estimation bias and reducing precision. Some PS modeling approaches attempt to address the potential effects of IVs, including selecting only covariates for the PS model that are strongly associated to the outcome of interest, thus screening out IVs. We conduct a study utilizing simulations and negative control experiments to evaluate the effect of IVs on PS model performance and to uncover best PS practices for real-world studies. We find that simulated IVs have a weak effect on bias and precision in both simulations and negative control experiments based on real-world data. In simulation experiments, PS methods that utilize outcome data, including the high-dimensional propensity score, produce the least estimation bias. However, in real-world settings underlying causal structures are unknown, and negative control experiments can illustrate a PS model's ability to minimize systematic bias. We find that large-scale, regularized regression based PS models in this case provide the most centered negative control distributions, suggesting superior performance in real-world scenarios.