A novel use of record linkage: resolving non-representativeness in health surveys and improving alcohol consumption estimates to inform strategy evaluation

Abstract

BackgroundPopulation health estimates are commonly inferred from survey data to inform policy. However, if the survey is not representative, bias can arise, invalidating estimates (especially in relation to prevalence and quantity estimates such as population alcohol consumption and other health behaviours). A key aspect determining the extent to which surveys are representative is the level of participant response. Inverse probability weights are usually applied in an attempt to correct for non-response, but are typically based on a limited range of sociodemographic variables and make the assumption that non-responders have similar behaviours to responders in the same sociodemographic category. We aim to explore and address non-representativeness in national health survey data by exploiting health and mortality record linkage, obtaining improved estimates of population alcohol consumption (weekly; binge drinking) and related dysfunction (problem drinking). MethodsData from the cross-sectional cluster-sampled Scottish Health Surveys (SHeSs; response levels 67–81%) have been confidentially linked (91% consent) to prospective and retrospective hospital admission and mortality data (roughly 90% accurate diagnosis, 99% complete). For preliminary analyses, we investigated whether weighted estimates of all-cause mortality and mortality from alcohol-related disorders derived from the SHeSs reflect those in the population of Scotland. Directly age-standardised survey-weighted mortality rates were calculated for the 3117 men and 3980 women aged 20 years and older at interview. We calculated equivalent mortality rates for the whole of Scotland using population estimates and mortality data contemporaneous with the linked survey data. Subsequently, we shall use information about differentials between the survey sample and general population to derive probabilities of alcohol-related harm (admissions to hospital or deaths) in non-responders by age, sex, deprivation, and region group. From general population data, we shall identify the number missing from the surveys in each group and simulate observations for non-responders with corresponding alcohol-related harm probabilities. We shall multiply impute alcohol consumption in simulated non-responders and modify imputations to represent any consumption-harm association differences, finally devising a correction factor. FindingsDeaths of 201 (6%) men and 215 (5%) women were recorded in the 2003 SHeS by the end of 2008. In men, all-cause mortality was lower in the SHeS sample (928 per 100 000 person-years, 95% CI 793–1075) than in the Scottish population (1269, 1262–1276). Figures for women were also significantly different (735, 633–846, for the SHeS and 902, 897–907, for the Scottish population). Alcohol-related mortality was somewhat lower in the SHeS sample (38, 15–75, in men and 9, 1–30, in women) relative to the Scottish population (54, 53–56, in men [p=0·317] and 23, 23–24, in women [p=0·190]). InterpretationWe found that people who responded to the 2003 SHeS differed from the population they were intended to represent, with lower than expected all-cause and alcohol-related mortality, despite the application of conventional weighting and age-standardisation methods. We offer a way of obtaining added value from record linkage by gaining additional insight into representativeness. By enabling more accurate estimates of population alcohol intake, this approach should lead to improved evaluation of interventions, informing policy aimed at reducing consumption. The development of effective corrective procedures offers a valuable boost to all survey-based research. FundingMedical Research Council and Chief Scientist Office for Scotland.