E

NHANES Data Analysis Methodology

DESCRIPTION OF THE NHANES DATASET

Introduction

The National Health and Nutrition Examination Survey (NHANES) is a set of studies conducted since the 1970s that regularly collects dietary data from a representative sample of the American population. The primary source of dietary intake data in NHANES comes from two 24-hour recalls administered to participants, with the first recall administered in person at a mobile examination center and the second administered by phone 3–10 days later. The design of NHANES involves data collection across all days of the week and seasons of the year. The collection of data over 2 days instead of a single day allows for the minimization of measurement error associated with day-to-day variation in individual diets using methods described below.

NHANES also includes a food frequency questionnaire (FFQ) with questions about recalled consumption of 31 seafood items in the previous 30 days. This questionnaire does not ask about quantities consumed, only about the number of meals that were consumed that included each food item of interest. Since this module collects consumption data on commonly consumed seafood items over a longer recall period (30 days) than the 24-hour recalls, it can be integrated into models to estimate usual intake as described below.

General Overview of the Approach

Seafood is known to be an episodically consumed food in the United States, presenting a challenge for the use of 24-hour recall methods to estimate usual consumption because a large proportion of the population does not consume it regularly. The National Cancer Institute (NCI) has developed a two-step method that enables the use of multiple 24-hour recalls on nonconsecutive days to estimate usual intake of foods or nutrients, implemented through statistical analysis systems (SAS) macros (Tooze et al., 2006, 2010). The NCI SAS macros have been used to analyze NHANES 24-hour recall data to help inform the Dietary Guidelines for Americans and in various research studies (HHS/USDA, 2020; Krebs-Smith et al., 2010; Shan et al., 2019). The models use mixed-effects models, containing both random and fixed effects to estimate usual intake by separating and removing the within-person variation from between-person variation (Herrick et al., 2018).

For episodically consumed foods and nutrients, first a two-part model is fit to describe the relationship between usual intake and a set of covariates to partition the variance attributable to within and between individuals (Tooze

et al., 2006). This two-part model estimates (1) usual intake as the probability of consuming a food on a given day and (2) the usual consumption day amount. Omega-3 fatty acids and certain other nutrients such as vitamin A are also known to be episodically consumed and require a two-part model. Other nutrients, such as total protein, are consumed daily and can be estimated with a one-part model. Further details about the NCI approach are available (Herrick et al., 2018; Luo et al., 2022).

DATASET PREPARATION AND ANALYSIS

The What We Eat in America (NHANES) day-1 and day-2 dietary recall dataset was analyzed for children (2–19 years old, n = 13,171) and women of childbearing age (16–50 years old, n = 7,355) in year cycles 2011–2012 to 2017–March 2020. Dietary intake and food pattern equivalent data were joined with the 30-day FFQ about seafood species meals consumed. Child age groups were defined as 2–5, 6–11, and 12–19 years old; three income-to-poverty ratio groups were used (less than 130 percent, between 130 and 499 percent, greater than 500 percent); and four race/ethnicity groups used (Hispanic, non-Hispanic Asian, non-Hispanic Black, and non-Hispanic White).

The per capita consumption of total seafood, seafood groups high and low in long-chain n-3 polyunsaturated fatty acids, total protein foods (e.g., red meat, processed meat, poultry, eggs, nuts and seeds, legumes, soy), seafood species, fatty acids, and micronutrients were calculated for the overall population of 2–19-year-old children and women and stratified by age, sex, race/ethnicity, and income. A two-part model was used for foods and nutrients consumed episodically, defined using the conventional approaches as more than 95 percent of the population having nonzero intake on the 24-hour recalls for the given food or nutrient (Herrick et al., 2018). In this model a correlation between the probability of consumption and the amount consumed was specified. All analyses of usual consumption also accounted for differences in weekend (defined as Friday–Sunday) versus weekday consumption. Covariates included in the prediction of both probabilities and amounts of seafood, protein types, and omega-3 fatty acids for children included sex, age groups, race, income quartiles, season, and consumption of any seafood meals in the seafood FFQ.

Separate modeling was conducted for each subgroup allowing for the estimation of distributions for subgroups. The same covariates were used for women’s consumption, omitting sex. Percentiles of usual total fish consumption and total omega-3 fatty acid consumption were estimated using the Distrib SAS macro, which uses pseudo-populations derived from Markov chain Monte Carlo simulation modeling to estimate the percentile of usual consumption. A default of 100 pseudo-persons was used to create this dataset.

A one-part “amount” model was used for ubiquitously consumed foods or nutrients (those consumed by nearly everyone in the sample). Covariates used in the prediction of this model included sex, race, and income quartiles.

The frequency of seafood intake by food source (retail, restaurant, etc.), meal type (breakfast, lunch, dinner), and by the top 10 species (shrimp, tuna, salmon, etc.) by age, sex, race/ethnicity, and income groups was also performed. The top 10 species’ frequencies were developed using the 30-day FFQ counts multiplied by the average seafood meal size among age-sex groups modeled using splines, and weighted by meal type (breakfast, lunch, dinner).

Separate analyses were conducted using data collected from children (n = 1,750) aged 6 months to 2 years for this same time period to examine patterns related to the timing of introduction of seafood. These analyses were primarily conducted using the 30-day FFQ data, although portion size analysis by age group used the 24-hour recall data. As seafood intake was found to be extremely infrequent in this age group the analyses were not intended to be representative, and therefore were reported as weight and unweighted estimates.

INCLUSIONS AND EXCLUSIONS

Analyses of women’s intake included all women aged 16–50 years, including pregnant and lactating women. Analyses of children’s (2–19 years) and women’s intakes relied on participants with complete covariate data.

WEIGHTING

All NHANES analyses accounted for the complex sampling design using primary sampling units, strata, and survey weights to construct nationally representative estimates of food consumption.

Sample weights were constructed for multiple year cycles using methods described by the National Center for Health Statistics (NCHS, 2021). In brief, this involved taking the original weighting variables for each round of data collection and multiplying by the share each round contributes to the total years of the study period (2011–March 2020). The 2020–2021 round was stopped in March 2020 because of the COVID-19 pandemic. NHANES merged the 2019–March 2020 round with the 2017–2018 round for a 3.2-year combined dataset, while all other rounds lasted 2 years. Day 1 sample weights were used when working with 1 day of dietary recall, and day 2 sample weights were used when working with 2 days of dietary recall.

The NCI macro requires the use of balanced repeated replication (BRR) weights to account for both the complex survey design of NHANES and differential weighting in the nonlinear mixed-effects models. This approach is described in detail elsewhere (Herrick et al., 2018). Seventy-two BRR weights were calculated for each individual in the sample following this guidance for the 9.3 years of NHANES included in our sample, using a perturbation factor of 70 percent (F = 0.3), the standard factor used in the analyses of NHANES dietary intake data (Herrick et al., 2018; Moshfegh et al., 2009; NCI, 2019).

REFERENCES

This page intentionally left blank.