Performance of statistical methods to correct food intake distribution: comparison between observed and estimated usual intake.

There are statistical methods that remove the within-person random error and estimate the usual intake when there is a second 24-h recall (24HR) for at least a subsample of the study population. We aimed to compare the distribution of usual food intake estimated by statistical models with the distribution of observed usual intake. A total of 302 individuals from Rio de Janeiro (Brazil) answered twenty, non-consecutive 24HR; the average length of follow-up was 3 months. The usual food intake was considered as the average of the 20 collection days of food intake. Using data sets with a pair of 2 collection days, usual percentiles of intake of the selected foods using two methods were estimated (National Cancer Institute (NCI) method and Multiple Source Method (MSM)). These estimates were compared with the percentiles of the observed usual intake. Selected foods comprised a range of parameter distributions: skewness, percentage of zero intakes and within- and between-person intakes. Both methods performed well but failed in some situations. In most cases, NCI and MSM produced similar percentiles between each other and values very close to the true intake, and they better represented the usual intake compared with 2-d mean. The smallest precision was observed in the upper tail of the distribution. In spite of the underestimation and overestimation of percentiles of intake, from a public health standpoint, these biases appear not to be of major concern.