Relevance & Research Question: Web surveys of the general population are suspected to yield less representative samples than traditional survey modes for two major reasons. First, population coverage with internet access is incomplete in most countries. Second, response rates to Web surveys are generally lower than in traditional modes. Sequential mixed-mode strategies using personal (CAPI) or telephone (CATI) interviews have been suggested to cope with the presumed representativity problem. In this study we assess the underlying assumptions of this reasoning, i.e., first, whether a probability sample using Web for data collection is less representative than other modes and, second, whether CAPI or CATI possess properties that can compensate for bias when added in sequence to Web.
Methods & Data: We conducted an experiment based on a national probability sample of 8,800 persons in the context of the Dutch ‘Security Monitor’ survey. Persons were assigned to one out of four modes (Web, Mail, CATI, CAPI). Mailed invitations contained a link (Web), a questionnaire (Mail), or announced the contact by an interviewer (CAPI/CATI). After 4-8 weeks we approached every person again in a second wave, either by CAPI or CATI. We included questions on internet coverage in this wave. To assess representativity, background data was available from national registries.
Results: Though The Netherlands are a country of high internet coverage, the impact of under-coverage on representativity was strong (assessed by R-indicators). However, many of the coverage deficits were mitigated in the process of contact and cooperation. The final Web response sample was as representative as CAPI and even more representative than CATI and Mail. Mixing modes in sequence (i.e. Web followed by CAPI or CAPI+CATI) was not capable of compensating the deficits of Web beyond well-known net increases in response rates.
Added Value: Our results question two basic assumptions about Web surveys and mixed-mode designs. We suggest that Web surveys possess equally good or better response properties as traditional modes, if they are based on probability samples and population coverage with internet access is high. Also, from a representativity perspective, we did not find evidence in favour of using mixed-mode designs.