Measurement Error in Retrospective Work Histories
Measurement error in retrospective reports of work status has been difficult to quantify in the past. Issues of confidentiality have made access to datasets linking survey responses to a valid administrative source problematic. This study uses a Swedish register of unemployment as a benchmark against which responses from a survey question are compared and hence the presence of measurement error elucidated. We carry out separate analyses for the different forms that measurement error in retrospective reports of unemployment can take. These are misdates of ends of spells, misclassifications of work status, miscounts of the number of spells of unemployment, misreports of total durations in unemployment, and mismatches of work status in person-day observations. The prevalence of measurement error for different social categories and interview formats is also examined, leading to a better understanding of the error-generating mechanisms that arise when interviewees are asked to produce retrospective reports of work status. We are able to confirm some previously hypothesised error mechanisms – such as interference – but also identify interesting patterns – such as non-monotonic dependence of recall time on recall error.
Implications of Retrospective Measurement Error in Event History Analysis
It is commonly accepted that the use of retrospective questions in surveys makes interviewees face harder cognitive challenges and therefore leads to less precise measures than questions asking about current states. In this paper we evaluate the effect of using data derived from retrospective questions as the response variable in different event history analysis models: an accelerated life Weibull, an accelerated life exponential, a proportional hazards Cox, and a proportional odds logit. The impact of measurement error is assessed by a comparison of the estimates obtained when the models are specified using durations of unemployment derived from a retrospective question against those obtained using validation data derived from the Swedish register of unemployment. Results show large attenuation effects in all the regression coefficients. Furthermore, these effects are relatively similar across models.