Affiliations: [a] University of Duisburg-Essen, Forsthausweg 2, 47057 Duisburg, Germany | [b] Statistics Netherlands (CBS), CBS-weg 11, 6401 CZ Heerlen, The Netherlands
Corresponding author: Jonas Klingwort, University of Duisburg-Essen, Forsthausweg 2, 47057 Duisburg, Germany. Tel.: +49 203 379 1284; E-mail: [email protected]
Abstract: Due to their lower costs and shorter fieldwork periods, web surveys are increasingly used in official statistics. To benefit most from their advantages, it is important to determine how many reminders have to be used and when the fieldwork period has to be ended. The salience of the survey topic may affect if and when sampled persons decide to respond. Hence, timings of reminders and duration of fieldwork may be related to survey outcomes. This article presents results from a large national probability sample (Dutch Health Survey) where a web survey was used as the first response mode. The effects of number of days between invitation and the response to this request (response time) on survey estimates were studied. Since the observed differences between early, intermediate and late respondents may be correlated with demographics, a calibration approach (generalized regression) was used to control for such differences between response time groups. The results were used as indicators for the missing data mechanisms MCAR, MAR, MNAR. In this dataset, only one variable (percentage of people who visited a medical specialist) seemed to be subject to MNAR. Based on the results, a fieldwork period of at least two weeks is recommended for health web surveys.
Keywords: Time to respond, selection bias, missing-data mechanisms, calibration, generalized regression