Conference Paper International Sociological Association Research Committee on Logic & Methodology
Measurement Error in Event History Data from Panel Interviews
02 Sep 2008
Measurement error in event history data leads to biases in estimated spell durations and estimated coefficients and duration dependence from hazard models. Statistical methods of adjusting for measurement error in spell durations depend on specific assumptions about the properties of errors, most notably the assumption that errors are independent of true durations. Little is however known about the properties of errors in event history data.
This paper examines errors in event history data derived from panel interviews and their impact on estimates. The study uses survey reports on the receipt of State benefits, which were collected in the context of the British Household Panel Survey, and matched to government administrative records. These validation data are rare, in that they cover monthly receipt status from annual interviews over a four-year period.
The data are used to test assumptions typically made about the nature of measurement error. The evaluation is based on an individual-level validation of the survey responses, and examines under-reporting, mis-classification and misdating, all of which contribute to errors in reported spell durations. The results suggest that errors in reported durations are negatively related to true durations. The implications of measurement errors for estimates from event history models are illustrated by comparing estimates based on the survey and administrative data.