Exploring regression dilution bias using repeat measurements of 2858 variables in up to 49 000 UK Biobank participants

BackgroundMeasurement error in exposures and confounders can bias exposure-outcome associations but is rarely considered. Our aim was to assess measurement error between repeat measures of all continuous variables in UK Biobank, and explore approaches to mitigate its impact on exposure-outcome associations. MethodsIntraclass correlation coefficients (ICC) were calculated for all continuous variables with repeat measures. Regression calibration was used to correct for measurement error in both exposures and confounders, using the association of C-Reactive Protein (CRP) with mortality as an illustrative example. ResultsThe 2858 continuous variables with repeat measures, varied in sample size from 109 to 49 121. They fell into three groups: (i) baseline visit measures (529 variables; median ICC=0.64, IQR= [0.57, 0.83]); (ii) online dietary measures (22 variables; median ICC=0.35, IQR=[0.30, 0.40]) and (iii) imaging measures (2307 variables; median ICC=0.85, IQR=[0.73, 0.94]). Highest ICC were for anthropometric and medical history measures, and lowest for dietary and heart magnetic resonance imaging. The ICC for CRP was 0.29 (95% CI=[0.27, 0.30]), and for body mass index and smoking pack-years (confounders), were 0.93 (95% CI=[0.92, 0.93]) and 0.85 (95% CI=[0.84, 0.86]) respectively. The association of CRP with all-cause mortality (Hazard Ratio (HR)=1.029 per mg/L, 95% CI=[1.028, 1.031]) increased when correction for RDB was applied to the exposure (HR=1.119, 95% CI=[1.095, 1.143]). Confounder correction did not influence estimates. ConclusionsMeasurement error varies widely and is often non-negligible. For UK Biobank we provide relevant statistics and adaptable code to help other researchers explore and correct for bias due to measurement error. Key MessagesO_LIRandom measurement error in the exposure and confounders can bias the association between exposure and outcome towards or away from the null. C_LIO_LISome prospective studies, including UK Biobank, provide repeat measures of variables in a sub-sample for exploring bias due to random measurement error; these are rarely used. C_LIO_LIOur results demonstrate that measurement error is often non-negligible and may bias estimates. C_LIO_LIAgreement between repeated measures varies by category. In UK Biobank, dietary measures and heart magnetic resonance imaging are least stable while anthropometric and medical history variables are most stable. C_LIO_LIWe have provided intraclass correlation coefficients for 2858 continuous variables from UK Biobank and adaptable code to support researchers to correct for bias due to random error in exposures and confounders. C_LI