Surveys using general practice registers: Who are the non-responders?

Abstract

Background: General practice population registers are important sampling frames for surveys of disease prevalence, lifestyles and health needs. In any survey there is a problem of non-response in the target population which can reduce precision and may bias any estimates derived from the study. Bias related to those who choose not to respond has been studied, but the problems posed by those who no longer live at the address on the register have received less attention. The specific objectives of this study were to estimate the number of non-responders in a registered general practice population who had not received a questionnaire (‘ghosts’) as distinct from those who chose not to respond to it (‘refusers’) and to determine whether ghosts and refusers were different from non-responders in ways which might affect the purpose of the survey. Methods: A total of 500 people aged 18–75 years were selected from a general practice register in Stockport for a postal survey of shoulder pain in the adult population. Forty-one were identified as ‘ghosts’ during the survey. After three mailings of a baseline screening questionnaire for shoulder pain and a further mailing of a final reminder, there were 120 further non-responders. They were classified as ghosts or refusers on the basis of a comparison of their current address on the practice register with that on the electoral roll. A matched sample of 120 survey responders was identified and was used to check the accuracy of the electoral roll in classifying non-responders. Ghosts and refusers were compared with the sample of responders using consultation information obtained from medical records at the practice, if the latter were available. Results: In total, 108 of the 500 people in the sample (22 per cent) were classified as ‘ghosts’ - 41 on the basis of information obtained during the survey and 67 on the basis of a comparison with electoral roll details. In addition, 48 were classified as refusers on the basis of the electoral roll comparison. This compared with 8 (7 per cent) of the matched sample of responders who would have been classified as ‘ghosts’ using the electoral roll. People who had addresses outside the electoral district were excluded from classification. Medical records were available for 105 (90 per cent) of the responders, 31 (46 per cent) of the ghosts and 40 (83 per cent) of the refusers. Of those who had consulted at least once for any reason in the previous two years, a similar proportion of responders and refusers (48 per cent and 46 per cent) had attended because of musculoskeletal problems, but responders were more likely to have consulted specifically about shoulder pain (10 per cent compared with 3 per cent, but based on small numbers). Ghosts by contrast had a lower proportion of consultations for both all musculoskeletal (33 per cent) and shoulder-specific problems (none), but were more likely to have consulted with a psychological problem than responders. Conclusions: We have estimated the number of non-responders to a survey of shoulder pain who were likely not to have received the baseline questionnaire (ghosts) to have been 22 per cent of a sample drawn from a practicebased Family Health Services Authority register. The error of using the electoral roll to make such a classification was found to be small, with only 8 (7 per cent) of the sample of responders being misclassified as ghosts. Those ghosts who still had records at the practice appeared to differ from refusers with respect to musculoskeletal and psychological morbidity. It may be inappropriate to exclude ghosts from the denominator in population surveys. © 1996, Oxford University Press.