Conference Paper European Survey Research Association Conference
The Impact of Dependent Interviewing on Measurement Error
01 Jul 2009
Repeated measures data are used, amongst other things, to estimate and predict the extent of change experienced by individuals, of time spent in particular states, or to control for unobserved individual effects in explanatory models. Repeated measures data from panel surveys, where sample units are interviewed at regular intervals, are affected by measurement error at each interview. The causes of measurement error are similar to those in cross-sectional surveys. The impact on estimates may however be worse. To improve the quality of panel data, survey organisations are increasingly using dependent interviewing (DI) techniques, whereby information about the sample member from prior interviews is incorporated into the question wording and routing.
This article assesses in which way, and to what extent, DI alters the distributional properties of measurement error in repeated measures data. The study examines reports on the timing and amount of various income sources, collected in the context of the British Household Panel Survey. The survey included a randomized experiment comparing different methods of independent and dependent interviewing and obtained validation data for respondents from government administrative records.
Previous validation studies have focused on the impact of DI on measurement error in a single wave. The results indicate that DI reduces under-reporting and hence reduces bias in cross-sectional estimates of prevalence. The impact on time series properties of errors across panel waves have however been ignored so far. Previous studies have also shown that DI increases the consistency of reports across waves, thereby reducing spurious change. Whether this is achieved by reducing errors, or merely by replacing random error at each wave with systematic (correlated) error, is however unknown.
This study firstly tests assumptions about the distributional properties of measurement error, typically made in the analysis of repeated measures data. Independent and dependent interviewing data are compared, to assess how and to what extent DI changes the properties of errors. Secondly, to illustrate the implications for actual estimates, the biases in exemplary analyses of repeated measures data from independent and dependent interviewing data are compared.