Consider a setting where a treatment that starts at some point during a spell (e.g. in unemployment) may impact on the hazard rate of the spell duration, and where the impact may be heterogeneous across subjects. We provide Monte Carlo evidence on the feasibility of estimating the distribution of treatment effects from duration data with selectivity, by means of a nonparametric maximum likelihood estimator with unrestricted numbers of mass points for the heterogeneity distribution. We find that specifying the treatment effect as homogenous may yield misleading average results if the true effects are heterogeneous, even when the sorting into treatment is appropriately accounted for. Specifying the treatment effect as a random coefficient allows for precise estimation of informative average treatment effects including the program’s overall impact on the mean duration.