Is Cohort Representativeness Passé? Poststratified Associations of Lifestyle Risk Factors with Mortality in the UK Biobank.

Abstract

BACKGROUND: The UK Biobank (UKB) has been used widely to examine associations between lifestyle risk factors and mortality outcomes. It is unknown whether the extremely low UKB response rate (5.5%) and lack of representativeness materially affects the magnitude and direction of effect estimates. METHODS: We used poststratification to match the UKB sample to the target population in terms of sociodemographic characteristics and prevalence of lifestyle risk factors (physical inactivity, alcohol intake, smoking, and poor diet). We compared unweighted and poststratified associations between each lifestyle risk factor and a lifestyle index score with all-cause, cardiovascular disease (CVD), and cancer mortality. We also calculated the unweighted to poststratified ratio of HR (RHR) and 95% confidence interval as a marker of effect-size difference. RESULTS: Of 371,974 UKB participants with no missing data, 302,009 had no history of CVD or cancer, corresponding to 3,298,958 person years of follow-up. Protective associations between alcohol use and CVD mortality observed in the unweighted UKB were substantially altered after poststratification, for example, from a hazard ratio (HR) of 0.63 (0.45-0.87) unweighted to 0.99 (0.65-1.50) poststratified for drinking ≥5 times/week versus never drinking. The magnitude of the poststratified all-cause mortality hazard ratio comparing least healthy with healthiest tertile of lifestyle risk factor index was 9% higher (95% confidence interval: 4%, 14%) than the unweighted estimates. CONCLUSIONS: Lack of representativeness may distort the associations of alcohol with CVD mortality, and may underestimate health hazards among those with cumulatively the least healthy lifestyles.