Characterizing unemployment duration data with stock sample measures

In this paper, we describe the usefulness of stock sample measures for average unemployment durations, when the parameter of interest is the expected unemployment duration. If both job separations and job accessions follow a Poisson process which are constant over time, the stock sample measure and the parameter of interest are identical. We discuss deviations from this benchmark in a theoretical framework and show how the stock sample measure develops compared to the parameter of interest. Over time the stock sample measure is useful (ie, moves in the same direction as the parameter of interest), when (the rate of) job accessions change whereas it is not useful when (the rate of) job separations changes. Comparing groups, the stock sample measure only provides consistent rankings under certain conditions on the moments of the underlying duration distribution. We propose a new measure for unemployment durations (an inflow corrected stock sample measure of unemployment duration) that can be easily calculated from stock sample data and takes account of the fact that it is primarily the changes of the inflow rates that invalidates the stock sample measure. Using a large administrative microdataset for Germany, we assess the performance of stock sample measures in predicting changes in the actual duration distribution and show that while the levels of the stock sample measures are positively related to the parameter of interest, the differences are strongly negatively correlated and thus the stock sample measure points not in the same direction as the parameter of interest. However, we also show that changes to the previous year of the stock sample measures are stable and positively related to changes of a moving average of expected durations. Thus, empirically, we can interpret the stock sample measures in terms of parameters of interest. In terms of comparisons between groups, we demonstrate for the example of men and women that the stock sample measure predicts the sign of the difference correctly only in about 60 percent of the cases.