Econometric analysis of duration data raises difficult issues of statistical inference. Building on the work of Prentice (1974) and Farewell and Prentice (1977), Addison and Portugal (1987; henceforth, AP) use the extended generalized gamma (EGG) distribution to test the sensitivity of regression estimates to more restrictive distributional assumptions which frequently have been imposed by other studies. AP find that these restrictions are rejected by data from the 1984 Displaced Worker Survey (DWS) and that imposing them creates large and variable biases in the estimated effects of the independent variables. AP also find that very different distributional forms provide the best fit for two subsamples of displaced workers. Both results suggest that many findings from earlier studies using simple parametric hazard models of unemployment duration are not robust. In this comment, we show that the sensitivity of AP's estimation results to distributional assumptions is a consequence of using an endogenous criteria to stratify the estimation sample, namely, whether or not workers collect Ul. This is an especially serious problem for a sample of displaced workers, because (a) many workers report jobless spells of zero or very few weeks; and (b) most displaced workers are Ul eligible. We also identify several other limitations of the DWS data for estimating unemployment durations and perform sensitivity analysis.