Identifying Dietary Consumption Patterns from Survey Data: A Bayesian Nonparametric Latent Class Model

Abstract

SummaryDietary intake is one of the largest contributing factors to cardiovascular health in the United States. Amongst low-income adults, the impact is even more devastating. Dietary assessments, such as 24-hour recalls, provide snapshots of dietary habits in a study population. Questions remain on how generalizable those snapshots are in nationally representative survey data, where certain subgroups are sampled disproportionately to comprehensively examine the population. Many of the models that derive dietary patterns account for study design by incorporating the sampling weights to the derived model parameter estimates post hoc. We propose a Bayesian overfitted latent class model that accounts for survey design and sampling variability to derive dietary patterns in adults aged 20 and older. We compare these results with a subset of the population, adults considered low-income (at or below the 130% poverty income threshold) to understand if and how these patterns generalize in a smaller subpopulation. Using dietary intake data from the National Health and Nutrition Examination Surveys, we identified six dietary patterns in the US adult population. These differed in consumption features found in the five dietary patterns derived in low-income adults. Reproducible code/data are provided on GitHub to encourage further research and application in this area.

Racial and ethnic heterogeneity in diets of low-income adult females in the United States: results from National Health and Nutrition Examination Surveys from 2011 to 2018

“…The multilevel clustering is achieved by borrowing information across and within subgroups to better discriminate subgroup-specific differences. Compared with the standard LCM, the RPC improves identifiability, fit to the observed data, and precision of pattern estimation from the model, compared with the standard LCM [ 15 , 19 ].…”

“…We see this in the United States, where non-Hispanic (NH) White participants and/or those living above the 130% poverty income level define the majority of the population. Modeling dietary behaviors from studies sharing this makeup will yield mischaracterized behaviors of the general population and prevent us from better understanding the factors affecting the subgroups at the greatest risk for health outcomes (racial and ethnic minority and low income) [ [15] , [16] , [17] ]…”

Racial/Ethnic Heterogeneity in Diet of Low-income Adult Women in the United States: Results from National Health and Nutrition Examination Surveys 2011-2018

“…For example, in many national surveys and cohorts of adults in the United States, the study population is often a majority of non-Hispanic white participants living above the 130% poverty income level. Modeling dietary behaviors from study populations sharing this makeup will yield behaviors reflective of this demographic, and prevent us from better understanding factors impacting the populations at greatest risk for health outcomes (racial/ethnic minority and low-income) (Stephenson, 2021; Field, 2007; Gavin, 2011)…”

“…For example, in many national surveys and cohorts of adults in the United States, the study population is often a majority of non-Hispanic white participants living above the 130% poverty income level. Modeling dietary behaviors from study populations sharing this makeup will yield behaviors reflective of this demographic, and prevent us from better understanding factors impacting the populations at greatest risk for health outcomes (racial/ethnic minority and low-income) (Stephenson, 2021;Field, 2007;Gavin, 2011) More flexible model approaches have recently been introduced to improve the way we consider demographic features that may drive dietary behaviors. Robust Profile Clustering (RPC) is an extension to the latent class model that distinguishes consumption patterns that may be shared across the study population, and those that may be specific to a defined demographic.…”