PDF (59,1 kb)

Background: Traditional modes of survey data collection show decreasing response rates and increasing costs. Web surveys potentially provide a cost-effective alternative. Opt-in volunteer web panels are widely used for market research or opinion polling, but less for academic or government research because of concerns about their representativeness arising from the effect of self-selection bias. Various methods attempt to make web panel surveys more representative of the population. We compared results from four UK web surveys with Natsal-3, a national probability sample survey.

Methods: The four web surveys were done by three UK market research companies, each with large volunteer web panels. A shortened Natsal-3 questionnaire was included on four web surveys: two used basic demographic quotas and two were modified with variables correlated with key outcomes as additional quotas. Panel members aged 18–44 years, who were resident in Great Britain, and who met the criteria set for the quotas were eligible. Each company was asked to provide a sample of 2000 participants, which achieved between 2000 and 2099 participants, with data collection between May and July, 2012. After weighting for age and sex, comparisons were made with Natsal-3 for demographic characteristics, key behaviours, and opinions, to examine whether modified quotas improved the results. We calculated mean absolute odds ratios (ORs) and the percentage of variables that significantly differed from Natsal-3 at the 5% significance level. We used Z tests based on bootstrap standard errors to assess whether the modified quota surveys provided estimates closer to Natsal-3 than the basic quota surveys. We used generalised estimating equations to assess whether the two modified quota surveys and likewise the two basic quota surveys provided consistent estimates. Six demographic and 35 behaviour or opinion variables were compared, including reporting of same-sex experience and attraction, vaginal sex in the past month, number of heterosexual partners, and attendance to a sexually transmitted infection clinic.

Findings: All four web surveys gave different results from Natsal-3 for most of the variables, and overall the two surveys with modified quotas did not provide estimates that were closer to Natsal-3 than the basic quota surveys. Estimates for men from the web surveys differed more from Natsal-3 than estimates for women. For men, the mean absolute OR was 1·81 for basic and 1·64 for modified surveys, and for women the results were 1·43 for basic and 1·42 for modified surveys. The differences between the web surveys and Natsal-3 were numerically greater for questions asked face-to-face than those asked in self-completion format, but no formal statistical comparison was made. The two modified quota surveys seemed to differ from each other (ie, were inconsistent, p=0·07), as did the basic quota surveys (p<0·0001). Adjustment of web panel quota controls did not lead to much improvement according to the results of Z tests, although for men improvements were more evident than for women. Moreover, meeting the modified quotas proved difficult, and the quotas had to be relaxed in both cases.

Interpretation: When measuring sensitive sexual behaviours in the Great Britain population, volunteer web panel surveys provided significantly different estimates than a probability computer-assisted personal interview or computer-assisted self-interview survey, and modified quotas did not clearly improve estimation. At present, non-probability web panels are not an appropriate means of gathering data if accurate estimates of population prevalence are the intention, regardless of whether quotas are basic or more sophisticated. This finding is consistent with findings from similar research from the USA and other European countries.

Funding: Medical Research Council and the Wellcome Trust.