Vrijheid M, Richardson L, Armstrong BK, et al (2008) Quantifying the Impact of Selection Bias Caused by Nonparticipation in a Case–Control Study of Mobile Phone Use. Ann Epidemiol 19:33–42
A common issue in epidemiologic studies is that of selection bias due to high non-participation rates. In studies of mobile phone use and brain cancer, there is a possibility that reports of “preventive” odds ratios are in fact biased by low participation rates and non-representative respondents. The study examined study participation in the INTERPHONE study and its relation to mobile phone use, allowing estimation of the extent of selection bias.
Data was collected from the INTERPHONE case-control study, in which cases were diagnosed with primary glioma, meningioma, or acoustic neuroma, and they were resident in regions of 13 participating countries. Controls were matched on age, sex, and region. Data was collected by a computer-assisted personal interview (CAPI). Reasons for non-participation were collected when possible, and in some regions individuals were asked to fill a short “non-respondent” questionnaire, which included short questions on mobile phone use. Analyses were based on comparisons of respondents and non-respondents. By assuming a number of different scenarios, an assessment was also made of bias introduced to risk estimates of regular mobile phone use.
A total of 14,354 controls (6696 non-respondents) and 8776 cases (2465 non-respondents) were approached. Non-respondent questionnaires were completed by 1,704 controls and 215 cases. Regular mobile phone use was reported less commonly by non-respondents than by interviewed study subjects. Non-respondents were also less likely to have reported starting using mobile phones in earlier years. With the exception of one region, respondents were more highly educated. Depending on assumptions of phone use among non-respondents who did not fill the non-respondent questionnaire, bias selection factors ranged from 0.87 to 0.92 in the glioma portion of the study when the same assumptions held for cases and control, and the bias factors ranged from 0.56 to 1.48 when different assumptions held (with factors below or above 1.0 indicating downward or upward bias in the true odds ratio for regular phone use).
Discussion and Conclusion
The authors identify that their results may be limited because the “questionnaire non-respondents” may not have been representative of the total group of refusers. The lower phone-use among non-respondents can have many explanations. For controls, the most common reason for not participating was inability to be traced, and for cases it was illness and the physician refusing contact, all of which may be related to mobile phone use. In the most plausible scenarios in the study, downward bias factors were indicated in the odds ratio of regular mobile phone use. After applying the average downward correction factor of 10% to previously published results from the INTERPHONE study, odds ratio estimates were still largely below 1.0, indicating an implausible preventive effect of mobile phone use on tumours; the authors discuss various other explanations for these results.