Identifying and Characterizing Shared and Ethnic Background Site-Specific Dietary Patterns by Hispanic/Latino Background and Site: The Use of Bayesian Multi-Study Factor Analysis in The Hispanic Community Health Study/Study of Latinos

Abstract

Introduction

Dietary patterns (DPs) are combinations of dietary components intended to summarize key aspects of diet, while taking advantage of synergies between single components. A posteriori DPs are defined from the application of multivariate statistics, including principal component and factor analyses. New statistical methods like multi-study factor analysis have been recently used to distinguish subpopulation-specific DPs (i.e., study/country-specific features within an international consortium or subpopulation-specific features within a single study), as well as those shared among all groups in a population [1].

The Hispanic Community Health Study/Study of Latinos (HCHS/SOL), the most extensive and ongoing community-based cohort of Hispanic/Latino adults from 4 US sites to date, provides a unique opportunity to identify shared and subpopulation-specific a posteriori DPs.

Aims

The present work aims to: 1. identify shared and ethnic background-site (EBS)-specific (nutrient-based) DPs within the HCHS/SOL study and 2. characterize the identified DPs in terms of food-group consumption, an overall measure of diet quality, socio-demographic and lifestyle characteristics.

Methods

The HCHS/SOL

The HCHS/SOL is a population-based cohort study designed to identify disease prevalence rates and risk factors of Hispanic/Latino populations residing within 4 urban US communities (Bronx, Chicago, Miami, and San Diego) and representing individuals with 7 ethnicity backgrounds (Cuban, Dominican, Mexican, Puerto Rican, Central and South American, and mixed). Participants were selected using a probability sampling design [2].

Dietary habits at baseline (16,415 subjects from 2008 to 2011) were assessed using two 24-hr recalls, the first conducted in person and the second via telephone <=30 days after. The Nutrition Data System for Research software allowed for nutrient intake estimation [3].

Selection of subjects and variables

We excluded Hispanics and Latinos from other/mixed backgrounds, with unreliable dietary recalls, or providing extreme energy intake. We also excluded subpopulations <200 participants after previous exclusions. This gave a final sample size of 15,021 participants.

We selected 42 nutrients that well represent the overall diet for Hispanics/Latinos. For each participant, nutrient intakes were derived from either one available reliable recall or the mean of the two available reliable recalls.

Statistical analysis

Bayesian multi-study factor analysis (BMSFA) was carried out on the correlation matrices of the log-transformed nutrient intakes. The total number of factors to retain was selected using the spectral decomposition of the factors. After the singular value decomposition method used in the BMSFA for identifiability, the varimax rotation was applied to the shared factor-loading matrix to achieve a better-defined loading structure [4]. Characterization of DPs against selected food groups, a measure of diet quality, selected socio-demographic and lifestyle factors was based on survey-weighted regression models. Calculations were carried out using the R software [5].

Results

The selected model included 4 shared (62.5% total variance explained) and 12 EBS-specific DPs (variance around 10%), one for each of the 12 EBS combinations (Figure 1). Among shared DPs, the first, named Plant-based foods, loaded highly on vegetable protein, several minerals, vitamin B1, niacin, natural folate, soluble and insoluble fiber, the second, named Dairy products, loaded highly on short- and medium-chain saturated fatty acids and calcium, vitamins B2, B12, D, and retinol; the third shared factor, named Seafood, loaded highly on EPA, DPA, and DHA and the fourth, named Processed foods, loaded highly on several fats, including long-chain saturated and monounsaturated fatty acids, linoleic and linolenic acids, total trans fatty acids, and natural alpha-tocopherol. Most EBS-specific DPs were further grouped into overarching profiles: Animal vs. vegetable source, Animal source only, and Poultry vs. dairy products, to capture nuances within animal-based DPs. Puerto Rican background participants from Chicago expressed a strikingly different DP.

When interpreted in terms of food groups, the identified DPs confirmed the names based on nutrients. Higher overall diet quality was observed with increasing categories of Plant-based foods, Seafood, and the “Puerto Rican background–Chicago” EBS-specific DP, whereas increasing categories of Dairy products, Processed foods, and the remaining EBS-specific DPs were related to lower diet quality. Compared to non-US-born participants, US-born individuals exhibited lower adherence to the Plant-based foods and Dairy products DPs but higher adherence to Processed foods, Seafood, and 6 EBS-specific DPs.

Conclusions

In its first application in nutritional epidemiology, BMSFA succeeded in simultaneously estimating well-interpretable shared and EBS-specific DPs within 12 combinations of background and site.