A comparison by simulation of different methods to estimate the usual intake distribution for episodically consumed foods

Abstract

The purpose of the ETUI simulation study is to compare different methods to estimate the usual intake distribution for episodically consumed foods. An important criterion is the bias and precision with which the upper tail of the usual intake distribution is estimated. It is further investigated whether it is beneficial to include Food Frequency Questionnaire (FFQ) information in the models as a covariate. The data are simulated with the logistic-normal model for the frequencies and a two-way random effects model for the log-transformed amounts. Four different frequency models are employed with low, moderate and high consumption frequencies, as well as a so-called bathtub model. This is combined with two ratios (1 and 4) for the variance components in the amount model as well as three values (0, -0.5 and 0.5) for the correlation between frequency and amount. This gives a total number of 24 different scenarios. The answer to a hypothetical FFQ with seven response categories is derived by discretizing the simulated consumption. Three datasets were simulated for each scenario with 6250 individuals and two recall days. The following methods were used: OIM, ISUF, MSM, SPADE, BBN, LNN0 and LNN. In an additional simulation, with 50% never-consumers and four scenarios, it was tested whether it is advantageous to include information on which individuals are never-consumers and which are episodic consumers which happen to have zero consumption on the recall days. Also, for three practically relevant scenarios with two foods, the Add-Then-Model approach is compared with the Model-Then-Add approach. The main conclusions are that a practical approach for single foods would be to fit the LNN model and to revert to the LNN0 model when the estimated correlation is low, or when the frequency of consumption is large. For data similar to those in this simulation study, when the right model is used, inclusion of FFQ information is not beneficial when interest is in the upper percentiles only. For multiple foods the Model-Then-Add approach seems to be quite promising. Correlations between foods can then be accommodated by using the model assisted approach.