Dealing with day-to-day variance in dietary intake: regression calibration for diet-disease risk models

INTRODUCTION: (In a diet-disease risk model we need the information about usual dietary intake, defined as the average of several days of intake. In the epidemiological practice it is usually not feasible to administer more than two or three 24-hour recalls (24hr) per subject. Day-to-day variation in dietary intake leads to attenuation of the association with diseases and loss of statistical power. There are methods to estimate usual intake by adjusting for within-person variance. The objective is to assess the performance of methods to adjust by within-person variance of intake in a risk model.) 

METHODS: (We presented the preliminary results from a longitudinal study with 82 individuals in whom each of them was invited to answer 20 non-consecutive 24hr collection days and one food frequency questionnaire. As a methodological example we modelled body mass index (BMI) by soft-drink consumption, age and sex. In the first model we used only the first 24hr of each individual. In the second model we used the observed usual soft-drink intake (average of the 20 24hr for each individual). In the third model we performed a regression calibration using a 2-part nonlinear mixed model (National Cancer Institute method) using the first two 24hr of each individual and the same set of covariates used in the risk model in addition to the frequency of consumption. We compared the coefficients through the three models.) 

RESULTS: (The coefficients (standard-error) for the model 1, 2 e 3 were 0.0044(0.002), 0.016(0.004), e 0.016(0.006) [soft-drink]; 0.277(1.25), 0.461(1.14), e 0.477(1.23) [sex] and 0.195(0.038), 0.235(0.037), e 0.219(0.038) [age], respectively.) 

CONCLUSIONS: (The regression calibration with only two 24hr per individual could estimate a very similar coefficients to those obtained when using usual dietary intake. The statistical power was partially recovered.)