An important question facing economists and policymakers is how long individuals would collect unemployment insurance (UI) if it were made available for a longer period of time. This is a difficult task for two reasons. First, even though distributional assumptions generally make little
difference for the non-censored portion of a survival curve (e.g., before UI benefits are exhausted), they can have a large impact past the censoring point. Second, there may be a behavioral response to any change in the maximum allowed benefit duration. To estimate the survival function past the censoring point, I adopt a semiparametric approach which builds on Chen, Dahl, and Khan (2005). I exibly model the location and scale parameters of an accelerated failure
time (AFT) model, without making distributional assumptions about the error term. Using administrative-level data from New Jersey's UI system, I find the commonly-used Weibull model significantly biases UI exit rates upward compared to my semiparametric approach (which includes the Weibull model as a special case). By incorrectly predicting a quick
dropo in UI claims, the Weibull model greatly underestimates the costs of extending benefits, even in the absence of a behavioral response. I also take advantage of a unique policy experiment, which exogenously increased the maximum duration from 26 to 39 weeks. This allows estimation of the latent versus behavioral response to an extension in UI benefits. I find that most of the behavioral response occurs in the choice of whether to enter the extended benefits program at 26 weeks, and not in the decision to exit the UI system between weeks 26 to 39.