An ensemble method based on marginal-effect models (EMM) for estimating usual food intake from single-day dietary data and internal/external two-day dietary data

Abstract

Background: With collection of repeated 24-hour recalls, there exist challenges in usual intake estimation, including infeasibility of multiple dietary assessments, and shortage of non-zero intakes for episodically consumed foods.

Objectives: We developed an ensemble method based on marginal-effect models (EMM), which estimates usual intake distribution using single-day data with internal or external two-day data.

Methods: The performance of the EMM was evaluated and compared with the National Cancer Institute (NCI) method and NCI 1-d method, via simulations with various scenarios and real data analyses of red meat, fish, and protein from Korea National Health and Nutrition Examination Survey (KNHANES).

Results: Simulations indicated the EMM (maximum bias of 1.67, 3.17, 8.57, 11.63 for average, median, 75%-tile, 95%-tile, respectively) provided more accurate estimation than the NCI method (maximum bias of 4.18, 9.43, 7.56, 37.43) across various scenarios on intake probability and within-person variation. The EMM showed robust estimation when an insufficient number of people have positive consumption on two days. In simulations with various external variance ratios, the EMM showed similar or superior performance to the NCI 1-d method. The EMM produced more stable estimates of usual intake distributions for red meat, fish, and protein than the two NCI methods.

Conclusion: The proposed EMM showed substantial improvement over the NCI methods when data contain a relatively small number of people with positive consumption on two days; is robust when food intake probability is very low or high; and estimates an external variance ratio with relatively low bias.