Background Surveys commonly form the bases of population estimation of health measures and health-related determinants. The validity of such inference depends on study generalisability which can be threatened by low participation levels. Classically, efforts to reduce resultant selection bias are limited to weights based on socio-demographic characteristics which typically do not capture health differences within population sub-groups. We propose the use of administrative record-linkage and novel methodology to overcome non-response bias, and illustrate this using the 2003 Scottish Health Survey measures of alcohol consumption.
Methods Of the 54% of individuals who responded to the survey, 91% consented to record linkage of their responses to routine data; this study used data on alcohol-related harms (hospital admission or mortality) to the end of 2011. Contemporaneous census information and (unlinked) admission/mortality data on the general population were also available. We compared directly age-standardised survey-weighted estimates of alcohol-related harm rates in the 2353 male and 3028 female participants aged 20–65 years at interview with rates for the whole population. These differences and inferred characteristics of those missing from the survey were used to derive probabilities of alcohol-related harm in non-responders by age, sex, area deprivation and health board region. Observations were simulated for non-responders with corresponding alcohol-related harm probabilities and their unknown alcohol consumption estimates were multiply-imputed. Corrected results were obtained from the proxy representative sample comprising responders and simulated non-responders.
Results Overall mean weekly unit alcohol consumption estimates were 10.4% higher among men [uncorrected = 21.8 (95% CI: 20.8–22.8), corrected = 24.0 (21.8–26.3)] and 5.0% higher among women [uncorrected = 10.5 (9.6–11.4), corrected = 11.0 (9.9–12.1)]. For those living in the most deprived quintile areas, the uncorrected means were corrected from 23.1 (19.9–26.4) to 27.4 (19.9–34.9) units (18.6% increase) among men, and from 9.0 (7.5–10.4) to 10.1 (8.5–11.8) units (12.8% increase) among women. These compared with rises in the least deprived quintile areas from 22.9 (20.0–25.7) to 24.6 (20.0–29.2) units (7.5% increase) among men, and from 12.2 (10.9–13.6) to 13.0 (11.3–14.7) units (6.1% increase) among women.
Conclusion Corrected estimates suggest non-response bias leads to an underestimation of overall alcohol consumption as well as of disparities between men living in deprived and non-deprived areas. There is potential for wider application to other survey-derived estimates (including behaviours like smoking) and to other studies which have capacity to record-link to administrative health records. This methodology offers a promising route for advancing efforts to resolve non-response bias.