Abstract
Surveys are key means of obtaining policy-relevant information not available from routine sources. Bias arising from non-participation is typically handled by applying weights derived from limited socio-demographic characteristics. This approach neither captures nor adjusts for differences in health and related behaviours between participants and non-participants within categories. We addressed non-participation bias in alcohol consumption estimates using novel methodology applied to 2003 Scottish Health Survey responses record-linked to prospective administrative data. Differences were identified in socio-demographic characteristics, alcohol-related harm (hospitalisation or mortality) and all-cause mortality between survey participants and, from unlinked administrative sources, the contemporaneous general population of Scotland. These were used to infer the number of non-participants within each subgroup defined by socio-demographics and health outcomes. Synthetic observations for non-participants were then generated, missing only alcohol consumption. Weekly alcohol consumption values among synthetic non-participants were multiply imputed under missing at random and missing not at random assumptions. Relative to estimates adjusted using previously derived weights, the obtained mean weekly alcohol intake estimates were up to 59% higher among men and 16% higher among women, depending on the assumptions imposed. This work demonstrates the universal value of multiple imputation-based methodological advancement incorporating administrative health data over routine weighting procedures.
Highlights
Population health and health behaviour estimates are commonly derived from survey data to monitor trends and formulate and evaluate policies
Numerator counts of morbidity and mortality events in the population during the eight years of follow-up were combined with mid-year population estimates – by socio-demographic characteristics – to create an unlinked aggregate-level data set for the population for comparison with the record-linked survey data
Once the synthetic observations for the non-participants were created at Stage 2, the unit non-response problem had been converted into an item non-response problem with the synthetic non-participant observations having data on socio-demographic characteristics and health outcomes but missing data on alcohol consumption
Summary
Population health and health behaviour estimates are commonly derived from survey data to monitor trends and formulate and evaluate policies. Survey weights derived from inverse probability weighting[6] are usually applied in an attempt to correct for such unit non-response (as well as accounting for aspects of sampling design such as the oversampling of certain household types or geographical areas) These weights typically rely on a limited range of sociodemographic variables[7] and are based on the assumption that non-participants have equivalent behaviours to participants in the same socio-demographic category which is unlikely to be the case. Our approach involves: (1) exploitation of record-linkage to hospital discharges and mortality; (2) survey–population comparisons which inform the creation of synthetic partial observations for non-participants; and (3) MI to generate refined estimates of weekly consumption of alcohol under assumptions of MAR (weaker than when based on survey data alone) and explorations of MNAR.[18] We illustrate the application using data from the 2003 Scottish Health Survey (SHeS) individually record-linked to administrative health information from the Scottish Morbidity Records (SMR), mortality data from the National Records of Scotland (NRS) and unlinked contemporaneous data for the entire population.
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
