Methodology of Longitudinal Surveys Conference
July 14, 2006
There are at least two main factors affecting the interviewee contact and cooperation in sample surveys: the individual and fieldwork characteristics. When focusing on the attrition problem in household panel surveys it is possible to evaluate the impact of those two factors by estimating contact and co-operation probability models as in Lepkowski and Couper (2002) and Nicoletti and Peracchi (2005).
Those models are useful to predict the impact of changes in the data collection process, to compute weights to correct for a potential nonresponse bias or to provide the base for other sample correction estimation methods. Moreover, as suggested by Nicoletti and Buck (2004) they can help in explaining differences in contact and co-operation rates between two different surveys. In particular, it is possible to decompose the difference in the contact (co-operation) rate between two surveys in two components: a part due to differences in the household, personal and fieldwork characteristics distribution and a residual part due to a different impact of those characteristics. The residual part reflects a genuine difference between contact (co-operation) rates between two surveys everything else being equal. In other words, the residual part reflects differences which would persist even if all fieldwork, household and personal characteristics were the same. But what happens if there is unobserved heterogeneity, that is, if there are unobserved characteristics which differ between surveys?
When the unobserved heterogeneity is due to unobserved area or interviewer characteristics then is possible to control for it by considering random effects. The aim of this work is then to clarify how to decompose response rate differences in presence of area and/or interviewer random effects. We show that it is possible to decompose the rate differences in three components: a component due to differences in the distribution of the observed characteristics, a component due to differences in the distribution of the random effects, and a component due to differences in the impact of those variables. If we assume that the random effects follow the same distribution in the two surveys to be compared, then the component due to differences in random effects becomes zero. This seems to suggest that random effects may be omitted in the decomposition analysis.
Focusing attention on 4 household panel surveys as in Nicoletti and Buck (2004) - the European Community Household Panel in Germany and the UK, the German Socio Economic Panel Survey and the British Household Panel Survey - we apply the decomposition analysis to contact and co-operation rate differences between pairs of surveys and compare the results obtained considering and neglecting the random effects.