Left-Censoring in Duration Data: Theory and Applications
In this paper, we discuss how to best exploit the information contained in spells that are in progress when an observation period begins, that is, left-censored and left-truncated duration data. We provide a survey of censoring and truncation mechanisms in event history models. We describe some approaches that have been suggested in the literature to deal with left-censoring. Our contribution is the description of ways to use additional information to obtain more efficient parameter estimates using the left-censored informations, and particularly, the derivation of the associated likelihood expressions. In order to use the information efficiently, we often resort to the stationarity assumption. Hence, we provide a Hausman test for this assumption. The second part of the paper briefly presents some empirical examples which demonstrates the efficiency gains associated with the use of the information contained in the left-censored observations. In particular, we show how the use of some additional pieces of information allows us to obtain more efficient estimates of the parameters of interest. In doing this, we use the information reported in the waves 1990-1992 of the French Labour Force Surveys on young French individuals.