Sensitivity analysis

Although both propensity-score matching and inverse-probability-weighted regression adjustment approaches can estimate average treatment effect and potential outcome mean, the principles of the two approaches are different. One of the drawbacks of the PSM approach is that biased estimates may be obtained if the propensity score model is mis-specified and that cases without appropriate matched controls are dropped from the analysis. Unlike PSM, the IPWRA approach provides efficient estimates by allowing the modelling of both the outcome and the treatment equations and requires that only one of the two models are correctly specified to consistently estimate the impact. It also uses the full sample. However, the PSM approach is less sensitive small sample/cell sizes and will more reliably produce estimated coefficients. We therefore use PSM as our baseline approach but test the robustness of the results using the IPWRA approach, when possible.