Conference Paper Joint Empirical Social Science Seminar
Measurement Error in Spell Durations *from Panel Surveys: *Nature, Causes and Consequences
20 May 2009
Little is known about the nature of measurement error in panel data, or about the potential biasing effects on estimates. Methods of adjusting for errors in event history data tend to assume that errors in spell durations are random, and not related to true durations or variables predictive of spell durations. This paper uses data from four waves on a subset of the British Household Panel Survey, to examine measurement error in spell durations. Survey data on the receipt of State benefits and tax credits were linked to individual-level administrative data and used to assess 1) the distributional properties of measurement error and whether the assumptions of random error hold, 2) hypotheses about the processes leading to measurement error and whether it is possible to predict situations in which errors are more likely to occur, and 3) implications for analysis and the extent to which measurement error biases results. The findings suggest that there is considerable measurement error in spell durations derived from panel data. Errors are non-random and to some extent predictable. The resulting biases in estimates of spell durations and multivariate duration models nonetheless appear to be reassuringly small.