Saturday, November 9, 2013
Thomas Boardroom (Westin Georgetown)
*Names in bold indicate Presenter
Estimation of duration models of transfer program participation are important for understanding what factors and policies determine how quickly individuals move off these programs. However, survey data are well known to suffer from measurement error, in particular in the detailed recall of dates of participation in transfer programs. While previous work has examined mismeasurement of static models of participation (simple indicators for participation at a point in time). Very few studies have examined the structure and extent of measurement error in duration measurements. Using an administrative record match between the British Household Panel Survey and administrative records of participation in three income transfer programs (Job Seekers Allowance, Income Support and Working Family Tax Credit), we present detailed estimation of the measurement error structure in reporting duration of these programs. Further, we demonstrate how to use these results to correct survey estimates for the measurement error and present corrected estimates of simple duration models for these three programs. Duration models present an interesting challenge in modeling measurement error. Like simple participation, there are both errors of omission (the failure to report a spell to the survey) and errors of commission (reporting a spell which did not actually occur). In addition to these errors, are errors in the length of spells, both shortening and lengthening spell. As we will demonstrate, the errors in reporting the length are highly correlated with the length. There are other important regularities in the misreporting of spell length such as seam effects. Correcting estimation for these requires a likelihood function which links an underlying behavioral model for the true spell with the observable likelihood for the reported spell: the links characterize the response error structure. We formally derive this link in general terms, which motivates the estimation of particular models of response error. We then estimate the underlying behavioral model.