Improving power with repeated measures: diet and serum lipids

The inability to detect associations between diet and serum cholesterol in cross-sectional population studies has been attributed to measurement error in diet assessments and between-subject variability in lipid concentrations. Current statistical methods can reduce the effects of measurement error and allow within-subject comparisons when replicate measures on individuals are available, even if the time between replicates is as long as 4 y and replicate data are not available for all subjects. Data from 928 nondiabetic participants of the San Luis Valley Diabetes Study with measures of 24-h dietary intake and fasting lipid concentrations at baseline, at a 4-y follow-up visit, or both were analyzed in a random-effects model that allowed for an unbalanced design. Sex was included as a non-time-varying covariate and age, body mass index, and energy intake were included as time-varying covariates. The findings when LDL cholesterol (mmol/L) was regressed on saturated fat intake (20 g/d) with all observations in a random-effects model (beta = 0.14, P = 0.0016) were compared with results with observations restricted to the first visit only (beta = 0.05, P = 0.52), a balanced design using averages across visits (beta = -0.12, P = 0.28), and a balanced design with random effects obtained by excluding subjects without two observations (beta = 0.12, P = 0.0092). Study power was greatest in the random-effects model using all observations and time-varying covariates. These findings highlight the importance of even a single replicate observation on a subsample of subjects. We recommend analyzing all data rather than averaging measures across visits or omitting observations to create a balanced design.