Measurement errors in retrospective reports of event histories. A validation study with Finnish register data

Abstract

It is well known that retrospective survey reports of event histories are affected by measurement errors. Yet little is known about the determinants of measurement errors in event history data or their effects on event history analysis. Making use of longitudinal register data linked at person-level with longitudinal survey data, we provide novel evidence about 1) type and magnitude of measurement errors in survey reports of event histories, 2) validity of classical assumptions about measurement errors, 3) measurement error bias and 4) effect of measurement accuracy in event history analysis. The classical assumptions about measurement errors are not supported by our measurement error models. Measurement error in both spell durations and spell outcomes are shown to be important causes of bias in an event history analysis. The effects of education and earnings-related unemployment benefit are estimated with sizeable bias. The magnitude of bias in estimated covariate effects does not depend on model type whereas the Cox model produces clearly less biased estimates of baseline hazard compared to the Weibull model. The large bias in the Weibull baseline hazard is shown to be almost entirely due to low measurement accuracy in survey data.