Instrumental variable methods under dependent censoring

Instrumental variable methods are a popular set of econometric techniques to identify and estimate causal effects. They can however fail when the outcome variable of interest is a duration. For example, consider evaluating the causal effect of a job training service on unemployment duration, where job seekers often stop answering follow-up surveys before they find new employment. In that case, the unemployment duration is said to be right censored by some censoring time. To solve this issue, it is customary to assume that the duration and censoring time are statistically independent of each other. However, this assumption is likely to be violated in the example on job seekers, that is there is dependent censoring. The goal of this fellowship is to develop new instrumental variable methods that are valid under dependent censoring. The work that follows from this research proposal aims to (i) suggest an estimator for the complier causal hazard ratio under a semiparametric copula model for dependent censoring; (ii) develop a model to recover a quantile treatment effect under dependent censoring; (iii) provide a review paper on instrumental variable methods in duration analysis; and (iv) explore other avenues such as double machine learning or local distribution regression. This research will fill an important gap in the literature on instrumental variable methods for duration outcomes, where the independent censoring assumption is often taken as a given without proper reasoning.