Separating Selection Bias and Non-coverage in Internet Panels using Propensity Matching.

G. J. L. M. Lensvelt-Mulders, P.J. Lugtig, M. Hubregtse

Research output: Contribution to journalArticleAcademic

Abstract

Many internet-panels consist of self-selected respondents and hence cover a relatively small part of the population. Estimates based on Internet-panels therefore may suffer from non-coverage and self-selection bias. One way to correct for these biases is to use adjustment weighting(Lee, 2006). However, when Internet-panel respondents are intrinsically different from the general population, previous studies showed that weighting may result in an increase in bias (for example, see Loosveldt and Sonck, 2008). How can we show that panel-members are intrinsically different from respondents that take part in a conventional random-sample survey? To answer this question we compared the results of a volunteer Internet-panel to the results of a web-interview (WI) based on a random sample of the same population. First, differences in population coverage are studied. Secondly, we test if significant differences in coverage predict differences on dependent variables. Finally, we use propensity matching to test for self-selection bias. This contribution sheds light on the extent of coverage bias relative to self-selection bias in random- and volunteer opt-in Internet surveys. We use propensity score matching to answer our question. Propensity scores summarize the conditional probability of a respondent to be member of either the random or volunteer sample based on a set of covariates. When the propensity score includes relevant covariates, respondents with the same propensity scores can be matched. Remaining differences between dependent variables after matching cannot be caused by coverage errors, and are indicative for the size of self-selection bias
Original languageAmerican English
Number of pages10
JournalSurvey Practice
Volumeaugust
Publication statusPublished - 1 Jan 2009

Cite this