Differences are likely between people that respond to public health surveys and those that do not. Among non-respondents, there is commonly a disproportionate number of young,(1-8) male,(2-10) and unmarried people,(1, 2, 5, 8, 11-14) as well as those with lower education,(1, 2, 5-8, 10, 12-15) and lower socioeconomic status.(5, 6, 8, 11, 13, 16) Non-respondents are also more likely to be smokers, (1, 4, 10, 14, 17-19) and to have different patterns of alcohol consumption,(10, 16, 20-22) poorer physical and/or mental health,(5, 7, 9, 10, 23) and higher rates of mortality and morbidity.(20, 24-29) If researchers fail to account for nonresponse bias, prevalence estimates (in particular)(9) and analyses of associations between variables will likely be incorrect.
As nonresponse bias is less of a problem as survey response rates approach 100%, researchers often use a variety of strategies to encourage participation, such as sending multiple reminders to non-respondents, encouraging them to complete the survey. Nevertheless, few large-scale public health surveys achieve a response rate adequate to avoid substantial nonresponse bias, which – depending on how much respondents and non-respondents differ in the outcomes of interest – is estimated to be between 70% and 90%.(30) As participation rates in epidemiologic studies have been declining over time,(2, 31) it is increasingly important for researchers to identify and account for nonresponse bias when summarising and analysing data. Obviously, this is a major challenge because information on non-respondents is often unavailable, particularly for the outcomes of interest.
Researchers have sought methods to account for nonresponse bias for decades. In 1939, Pace proposed that the existence and direction of nonresponse bias in a given survey could be detected by comparing the responses of people who respond quickly to those who only respond after repeated contact attempts (delayed respondents).(32) This approach, often referred to as the “continuum of resistance” model,(33) is based on the presumption that people who are slow or reluctant (i.e. resistant) to complete a questionnaire are more similar to non-respondents than early respondents are.
The continuum of resistance model has resurfaced periodically in the literature since its proposal, despite having performed inconsistently under empirical testing. Some early studies supported the existence of a continuum of resistance for outcomes of interest;(34, 35) however, others have found that early and delayed responders do not differ at all,(36, 37) that associations between delayed responders and non-responders are weak,(33) or that a continuum of resistance exists for demographic variables but not outcomes of interest.(38) Recently, the model has been applied in a number of surveys of alcohol consumption and other health behaviours,(3, 13, 39-42) which have demonstrated significant and consistent differences between early and late respondents, for both demographic variables and outcomes of interest. In several of these surveys, investigators subsequently used weightings based on delayed respondents’ data to adjust their prevalence estimates, in an effort to account for nonresponse bias.(3, 13, 39, 41, 42)
Given the apparent value of the continuum of resistance model in these recent studies, and because we are unaware of any investigations of the utility of the model in large public health surveys involving digital data collection, we compared early and delayed respondents of the internet-based Norwegian Counties Public Health Survey. We hypothesized that there would be a relatively higher proportion of male, young and less-educated people among late responders, and that a continuum of resistance would exist for health outcomes and behaviours related to these demographic differences.