In this paper we present different estimates of spell durations in the Food Stamp and Aid to Families with Dependent Children programs, based on data from the 1987 panel of SIPP. We represent 4 types of duration estimates: a) estimates based on spells that started during the 28 months of the panel, b) estimates based on spells that started during observation periods of varying lengths ("subpanels" of the 1987 SIPP panel), c) point-in-time estimates, and d) estimates based on all spells, left-censored as well as non left-censored. Our purpose is to ascertain the robustness of duration estimates to various changes in sample design.

We find that estimates of spell distributions vary considerably for different types of samples. As an illustration, the median duration of food stamp participation is significantly higher for spells in progress at a point in time than for spells with observed beginnings, 8 years as compared to 6 months. However, estimates based on spells with observed beginnings did not differ greatly for different panel lengths. Thus, this study suggests that if the main point of interest is to investigate the distribution of the majority of spells into narrowly defined intervals, even panels shorter than the 1987 panel seem adequate.