1 Rio de Janeiro State University, Brazil; 2 Fluminense Federal University, Brazil; 3 University of São Paulo, Brazil
Introduction: Day-to-day variation in dietary intake leads to attenuation in its association with diseases and loss of statistical power. The objective is to assess the performance of methods that correct the coefficient by adjusting for within-person variance in a risk model. Methods: Data of 285 individuals with 20 non-consecutive 24hour recalls (24hr) and one food frequency questionnaire (FFQ) was used. As methodological example it was simulated an outcome to have food consumption coefficient equal to 0.1 in a risk model adjusting for age and sex. To generate these outcomes it was used the average of the 20 24hr per subject (here assumed to be the true usual intake) for soft drink and also for vegetables. To overcome the sample variation due to the small sample size, it was generated 190 datasets containing all possible combinations of collection days (one with 1st and 2nd 24hr, then 1st and 3rd, 1st and 4th, and so on, until the 19th and 20th 24hr). For each combination it was performed a regression calibration using a 2-part nonlinear mixed model in order to estimate a corrected coefficient using two collection-days for each individual and the same set of covariates used in the risk model in addition to the FFQ. It was compared the estimated with the simulated true coefficient. Results: True simulated coefficients [95%CI] were 1.01[0.46-1.59] and 1.03[0.14-3.02] for soft-drink and vegetables, respectively. The mean of the 190 estimated coefficients were 1.12 [0.21 - 2.02] and 1.47[0.14-3.02], respectively. The inclusion of the FFQ as covariate shrunk 95%CI: [0.41-1.95] and [0.22-2.76], besides reducing false negatives for the statistical significance (from 25% to 2%, and from 37% to 29%, respectively). Conclusion: The regression calibration may de-attenuate coefficients and recover sample power when using only two non-consecutive 24hr for estimating diet-disease relationship.