Geographic patterns and socioeconomic differences in the nutritional quality of household food purchases in the United States

Abstract

Background:

It is unclear whether dietary quality varies by geography in the US.

Purpose:

Assess patterns in packaged food purchases (PFPs)

Methods:

We characterized variation in PFP quality from 2008–2018 by 1) examining geographic clustering and 2) using regression analysis to control for household characteristics.

Results:

Lower quality purchases clustered in the Southeast and Appalachia, whereas higher quality purchases clustered in the West and Northeast. Spatial patterns were similar for low socioeconomic households but not high socioeconomic households. Geographic differences in quality remained after controlling for demographic composition.

Conclusion:

This analysis should inform research into systemic drivers of PFP quality.

1. Introduction

A poor diet is a risk factor for poor health.¹ Many diet-related diseases in the U.S. exhibit distinct spatial patterns, with above average rates of stroke,² obesity^3,4 and type 2 diabetes⁵ clustering in the Southeast and Appalachia (the cultural region of the U.S. following the Appalachian Mountains from southern New York to northern Mississippi⁶). However, although it is well-established that food culture varies throughout the country,⁷ it is unclear whether dietary quality also follows a spatial pattern.⁸ Furthermore, it is unclear whether national estimates of socioeconomic disparities in dietary quality^9–12 differ throughout the US.

Associations between diet quality and census region have been described;^13–15 however, census regions are large and may obscure differences between smaller, more meaningful geographic units. Examples of smaller geographic areas include media markets, where diet quality may differ due to food-related advertising, and US states, where differences in diet quality may be associated with social welfare policies. Analysis of spatial patterns in diet quality across the US has previously not been possible due to the lack of a sufficiently large national sample that captures diet-related data throughout the year. Data from Nielsen Homescan on household packaged food purchases (PFPs) provides a unique opportunity to examine spatial patterns because data is collected from about 60,000 households per year across 93% of counties in the contiguous US (93% coverage from Authors own calculations).

Packaged foods (or foods with a universal barcode, e.g., a bag of onions, frozen entrees, etc) are a subset of the overall diet, which also includes unpackaged foods (e.g., loose onions, meat from a butcher) as well as food eaten away from home (e.g., from schools, restaurants). Although not a complete picture of diet, PFPs contribute significantly to overall intake. Food purchased from stores constitute about 70% of calories consumed¹⁶ and over 50% of a household’s food budget.¹⁷ Of this food purchased from stores, an estimated 70–80% of store calories come from PFPs.¹⁸

The objective of this research was to determine whether the nutritional quality of households’ PFPs varies across the US and to identify regional disparities in food purchases among socioeconomic subpopulations. Our first aim was to describe spatial patterns in the quality of PFPs. Clusters of geographic units (e.g., Homescan markets, states) with above or below average quality PFPs would suggest there may be regional factors associated with the quality of PFPs. Our second aim was to determine whether variation in the quality of PFPs remained after controlling for household characteristics. In this analysis, differences between Homescan markets or states after controlling for demographic composition would indicate that systemic factors at the market or state level are associated with the quality of PFPs. For both aims, we first analyzed our full sample and then stratified by high and low socioeconomic groups to understand variation in disparities across the US.

Nielsen Homescan markets were chosen as the geographic unit for our primary analysis because they are more numerous than states and are delineated relative to metro centers or rural areas. A sensitivity analysis was conducted using states as the unit of analysis since state-level social policies, such as eligibility for the Supplemental Nutrition Assistance Program (SNAP), may impact food purchasing decisions.

2. Methods

2.1. Household Packaged Food Purchase Data

We used household panel data from the Nielsen U.S. Homescan Consumer Panel from 2008–2018 (n=673,422 household-year observations). Participating households were instructed to scan barcodes on all purchased items, including products purchased from all food retailers. For inclusion in the panel, households needed to report at least ten months of purchases. Demographic data, including county of residence, was collected by questionnaire. Nielsen Homescan used direct mailing (targeting low-income and racial/ethnic minority groups) and the Internet to recruit households. The panel’s open cohort study design allowed households to drop out at any time; new households were enrolled to replace dropouts based on demographic and geographic targets. Households were sampled from 52 metropolitan and 24 non-metropolitan “remaining” areas (i.e., 76 Homescan markets) and were weighted to be nationally representative (the term “remaining” was chosen by Nielsen to refer to non-metropolitan areas).

2.2. Nielsen Market Boundaries

There is no available GIS boundary file for the 76 Homescan market areas, so we derived geographic boundaries by merging counties located within the same Homescan market. To match counties to markets, we used the county FIPS code and market identifier that Nielsen provided for each household-year observation. For each county, we created a list of all household-year observations from 2008–2018 and their corresponding markets. If all household-year observations located in a county corresponded to the same market, then the county was matched to that market. However, if the household-year observations located in a county corresponded to more than one market, then the county was matched to the market that had the most observations.

After each county was linked to a single market, the boundaries of all counties in the same market were dissolved together to create a single market region. Using this approach, we identified 11 markets that had multiple noncontiguous regions that fell on opposite sides of an urban market. Since these noncontiguous markets would cause problems for spatial cluster analysis, which relies on identifying neighboring regions, we split them into distinct market areas (for example, the noncontiguous market “Remaining” Omaha was split into two regions: East “Remaining” Omaha and West “Remaining” Omaha). This process increased the number of distinct geographic market areas from 76 to 88 (see Supplementary Figure 1). Lastly, 203 US counties (6.8% of U.S. counties) were not represented in the Nielsen Homescan panel data; for visualization purposes only, we merged these counties with their adjacent Homescan market area (n=88).

2.3. Outcome Measures

Household packaged food and beverage purchases (PFPs) were linked to Nutrition Facts Panel (NFP) data, which includes data on calories, saturated fat and sugar content. These matches were updated annually to account for product reformulation as well as product entry and exit in the market.¹⁶

A range of outcomes were used to capture the nutritional quality of PFPs. In comparison, studies examining overall dietary quality often use a singular, summary measure like the Healthy Eating Index.¹⁹ In the case of PFPs, tools exist to measure the healthfulness of individual products (e.g., Guiding Stars²⁰), yet no validated summary index exists to assess diet quality across all PFPs. In addition, using a range of outcomes had more practical public health policy implications, including targeting specific nutrients for product reformulation or incentivizing / taxing specific food groups.

Outcomes included nutrient measures: total sugar (% total calories from PFPs (purchased)), saturated fat (% total calories purchased), sugar (g per capita per day), saturated fat (g per capita per day), sodium (milligrams per capita per day), and total calories (per capita per day). In addition, we measured calories (per capita per day) from PFP food groups of public health interest, including those groups that tend to be unhealthy (processed meats, mixed dishes,²¹ sugar-sweetened beverages (SSBs) and junk foods) and those that tend to be healthy food groups (fruits, non-starchy vegetables). Mixed dishes include foods like canned soups and frozen entrees. Junk foods include salty snacks, desserts, sweeteners, toppings, and candy. Examples of the types of products included each food group are detailed in Supplementary Table 1.

All outcomes were continuous measures, and annual household purchases were normalized to daily per capita values for interpretability. We described the nutritional quality of PFPs relative to dietary recommendations (Table 1). We considered PFPs to be of low nutritional quality if nutrients of concern (sugar, saturated fat and sodium) surpassed dietary recommendations, if the calories per capita per day from processed meats was greater than zero, if servings of mixed dishes, SSBs or junk foods reflected multiple servings per day, or if the calories per capita per day from fruits or non-starchy vegetables reflected fewer than the recommended 5 servings per day. PFPs were of high nutritional quality if the opposite conditions were true. We note that while there are no dietary recommendations specific for mixed dishes, prior research has found most pre-packaged mixed dishes are highly processed and exceed recommendations for saturated fat, sodium and sugar,²¹ so greater purchases of these foods reflect a lower quality purchase basket.

Table 1:

Average nutrient profile of household packaged food purchases, Nielsen Homescan Panel 2008–2018

Nutrients from packaged food purchases, per capita per day
Outcome	National Average1 (± SE)	Context: Relative to health guidelines37
Total calories (per capita per day)	1,281 ± 3.1	2,000 kcal/day per average adult
Sugar (% total calories)	25.1% ± 0.03%	Limit: less than 10% total daily intake from added sugar2
Sugar (g per capita/day)	80.6 ± 0.2	10% of a 2,000 calorie diet = 50g
Saturated fat (% total calories)	12.2% ± 0.01%	Limit: less than 10% total daily intake
Saturated fat (g per capita/day)	17.5 ± 0.05	10% of a 2,000 calorie diet = 22g
Sodium (mg per capita/day)	2,462 ± 7.2	Upper limit: 2,300 mg/day
Calories from packaged food purchases, per capita/day
(kcal per capita/day)	National Average (± SE)	Context: Relative to health guidelines; Examples of common serving sizes
Fruit	20 ± 0.1	2–3 servings/day; Ex: 1 apple: 80 calories
Non-starchy vegetables	17 ± 0.1	2–3 servings/day; Ex: 1 cup broccoli: 40 calories
Processed meats	52 ± 0.2	Not recommended3; Ex: 1 Nathan’s hot dog: 140 calories
Mixed dishes	93 ± 0.4	Recommendation: n/a; Ex: 1 slice DiGiorno cheese pizza: 330 calories
Sugar-sweetened beverages	60 ± 0.4	Recommendation: limit; Ex: 1 can of soda: 160 calories
Junk foods	330 ± 1.0	Recommendation: limit; 1 Oreo: 80 calories

^1:National averages are derived from Nielsen PFP data using survey-weighted averages

^2:While guidelines exist for added sugar, no guidelines exist for the consumption (and therefore purchase) of total sugar. However, much of the sugar in processed and packaged foods is added during product formulation. As evidenced by the low average calories from fruits, it is unlikely that consumption of total sugar in PFPs is driven by naturally-occurring sugars.

^3:The International Agency for Research on Cancer and World Health Organization categorized processed meats as a Group 1 carcinogen in 2015.⁵¹ Therefore, recommended consumption is zero.

Disclaimer: Calculations are based in part on data reported by Nielsen through its Homescan Services for all food categories, including beverages and alcohol for 2008–2018 across the U.S. market. The Nielsen Company, 2015. Nielsen is not responsible for and had no role in preparing the results reported herein

2.4. Spatial Clustering Analysis: Overview

To evaluate the geographic patterns in the quality of food purchases we used spatial cluster analysis, an exploratory tool used to identify whether there is complete randomness or if there is clustering in the distribution of some attribute over space. Clustering may result if there are areas where high values are surrounded by high values and/or low values are surrounded by low values. If clustering exists, their location and size can be identified. Spatial cluster analysis is exploratory because it is a data-drive approach without any a priori expectation of the location or size of high and low clusters.²² The identification of clusters of our PFP outcomes would suggest there are systemic factors that drive the quality of PFPs that are not defined by boundaries.

To conduct cluster analysis, we calculated mean values for each outcome for each geographic unit of analysis (i.e., Homescan markets and states). Means were survey-weighted to ensure the household sample was representative of the target population. However, we did not further adjust means to control for differences in the sociodemographic characteristics of households between markets or states. We used unadjusted means so that patterns in PFPs could be compared to patterns in the unadjusted prevalences of diet-related chronic diseases.

We calculated Moran’s I to determine whether the overall geographic pattern of each PFP outcome was clustered, random or dispersed across markets.²³ We defined spatial neighbors using 1^st order Queen’s contiguity (adjacent units sharing a boundary line or point). We conducted a sensitivity analysis using 1^st order Rook’s contiguity (adjacent units sharing a boundary line only). Results were highly clustered if Moran’s I was greater than 0.5, moderately clustered if Moran’s I was between 0.3 and 0.5, slightly clustered if Moran’s I was between 0.1 and 0.3, and random if Moran’s I was less than 0.1.

We identified the locations of geographic clusters of regions with high or low values for each outcome using the Local Indicator of Spatial Association (LISA). This metric identifies high-high clusters (high value adjacent to high values), low-low clusters (low value adjacent to low values), high outliers (high value adjacent to low values) and low outliers (low value adjacent to high values).²⁴ Whether these clusters reflect relatively low or high quality purchases depends on the outcome being examined (e.g., high-high clusters of fruits are indicative of regions with relatively higher quality PFPs whereas high-high clusters of SSBs are indicative of regions with relatively lower quality PFPs).

As an additional exploratory tool, we compared whether low (or high) clusters of PFP outcomes followed similar patterns. We overlaid clusters of select outcomes that showed significant clustering and were the most health relevant (e.g., clusters defined by low fruit purchases or high junk food purchases both reflect lower quality PFPs, but low or high total calories from PFPs is health-neutral). We assigned each market one score for the total number of clusters it pertained to reflecting relatively lower quality PFPs relative to other markets and one score for the total number of clusters reflecting higher quality PFPs relative to other markets. These scores were mapped to identify patterns and outliers.

We repeated all steps (the calculation of survey-weighted means, spatial cluster analysis and overlays) in a sensitivity analysis using states as our geographic unit of interest. All spatial analysis was conducted using GeoDa software,²⁵ and maps were created using ArcGIS.²⁶

2.5. Spatial Clustering Analysis: Stratification by Socioeconomic Status

We repeated all steps of the spatial clustering analysis for stratified socioeconomic (SES) subpopulations to investigate whether disparities found at the national level^9,10,12,27 are uniform throughout the country. It is unlikely that systemic factors underlying spatial clustering patterns affect households with different resources in the same way.

We defined SES groups by cost-of-living (COLI) adjusted income and by educational attainment. Income was chosen as one dimension of SES because the cost of food is associated with dietary quality.^28–31 We used COLI-adjusted income to account for differences in the cost of living across the country. We first adjusted self-reported household income using Regional Price Parities data from the Bureau of Economic Analysis.³² Next, we recalculated COLI-adjusted income as a percent of the Federal Poverty Level (FPL) to account for household size. These calculations were made for each household-year observation to reflect annual changes in the cost of living, household composition, and changes to the FPL. This measure of COLI- and FPL-adjusting income was then divided into tertiles.

We chose educational attainment as a second dimension of SES because it has also been independently associated with dietary quality.^27,33,34 We defined education as the highest level of self-reported educational attainment by either the female or male head of household and was then categorized as high school or less, some college, college graduate or post college graduate.

Finally, we calculated differences between low and high SES groups to explore whether there was spatial clustering to either income or education disparities. Differences were considered disparities when the low SES group had a relatively lower quality PFP outcome compared to the highest SES group (e.g., fewer calories from fruits or vegetables, higher percent total calories from sugar, etc.). These differences were calculated at the market and state level and steps for calculating Moran’s I and LISA were repeated.

We excluded markets or states from subpopulation analysis if the sample size of unique households within an area was less than 50 between 2008–2018. While all markets and states had at least 50 unique households by income tertile, this threshold resulted in the exclusion of three market areas from the high education (post college graduate) cluster analysis and one state from the low education (high school or less) cluster analysis.

2.6. Multiple Regression Analysis

We used multiple linear regression to examine the association between geographic area and the nutritional quality of PFP while controlling for the sociodemographic characteristics of households. Whereas our spatial cluster analysis explored whether regional patterns in the quality of PFPs exist, the purpose of our regression analysis was to assess whether differences exist between distinct geographic units independent of the composition of their population. Moreover, while significant spatial clusters suggest systemic drivers of PFP quality that act irrespective of geographic boundaries, significant differences between regression-adjusted averages suggest systemic factors that operate specifically at the market or state level (e.g., marketing strategies or state policies, respectively).

To control for demographic composition, we adjusted all models for household COLI/FPL-adjusted income tertile, highest educational attainment of a head of household, head of household race and ethnicity, household age composition and year. Like income and education, race and household composition were self-reported and updated every year via questionnaire. We combined race and ethnicity into a covariate with five mutually exclusive categories: Hispanic, non-Hispanic White, non-Hispanic Black, non-Hispanic Asian, and non-Hispanic other race. To control for household age composition, we included a series of count variables on the number of people in the household as follows: 0–1 years, 2–5, 6–11, 12–18, 19–64 and 65 and older.

As with the spatial cluster analysis, we stratified regression models by income tertile and by education to understand variation in disparities. The income-stratified models controlled for household race/ethnicity, education, composition, and year. The education-stratified models controlled for household race/ethnicity, income, composition, and year.

Statistical analysis and data preparation was conducted using STATA version 15.³⁵ We used multilevel generalized linear models (GLM; Stata: meglm) to control for clustering of multiple years of observations at the household level. A gamma family and a log link were used for all per capita per day outcomes, while a gaussian distribution and identity link were used for the percent of total calories from saturated fat and from sugar. Multilevel models allowed for the incorporation of household survey weights that are generated on an annual basis to produce representative estimates. We rescaled annual weights to generate weights at level one (year) and level two (household) following Heeringa, et al.³⁶ All households in the sample were retained in each model for the correct calculation of standard errors, and analytic samples were specified using subpopulation commands (Stata: svyset, subpop()). We used Stata’s margins and contrast commands to compare and test differences in PFP outcomes between geographic units. By using a GLM with a log link, margins were generated in the original units of each outcome (code available upon request). For interpretability, we use margins to present our results as predicted PFP outcomes for each geographic unit after controlling for household demographics.

3. Results

3.1. Survey-weighted means for all households

From 2008 – 2018, the average household purchased 1,281 calories per person per day from PFPs, equivalent to about 64% of total recommended daily calories for the average adult. Overall, the nutrient profile of packaged food purchases (PFPs) does not meet dietary recommendations. Sugar and saturated fat as a percent of total calories exceeded dietary guidelines which recommend that Americans consume less than 10 percent of calories per day from added sugars and less than 10 percent from saturated fats³⁷ (Table 1). While PFPs are only a component of the diet, daily grams of sugar per person from PFPs alone exceeded dietary recommendations, while daily grams of saturated fat were slightly lower than recommendations, leaving little room to consume these nutrients in the remainder of the diet. Daily purchases of packaged fruits and non-starchy vegetables reflected less than one serving per day, while purchases of junk foods reflected multiple servings per person per day.

3.2. Spatial clustering: All households

The degree of clustering (as measured by Moran’s /) and the locations of high and low clusters varied by nutritional outcomes of PFPs (Table 2). Grams of saturated fat, grams of sugar, percent energy from sugar, and calories from processed meats were found to be highly clustered. Similar overall patterns were observed when states were used as the geographic unit of analysis, although the degree of clustering for each outcome was lower (see Supplementary Table 2). Defining neighbors using first order Rook’s contiguity did not change results for any outcomes.

Table 2:

Spatial clustering patterns of nutritional outcomes of PFPs across Nielsen markets

Outcome	Moran’s Index (degree of clustering)	Clusters of above average outcome1	Clusters of below average outcome	Notable outliers1
Nutrients from packaged food purchases, per capita per day
Total calories	0.44** (moderate)	Midwest	Southwest; Tristate metro area	Remaining North California (high); Atlanta and Charlotte (low)
Sugar (% total calories)	0.60** (high)	Southeast and Ohio Valley	West coast	Salt Lake City (high); Cincinnati (low)
Sugar (g)	0.49** (high)	Southeast and Ohio Valley	Southwest	Remaining North California (high); Atlanta (low)
Saturated fat (% total calories)	0.59** (high)	Northwest to the Great Lakes	Southeast	Remaining Greenville (high); Omaha (low)
Saturated fat (g)	0.43** (moderate)	Great Plains; Ohio/Indiana	Southern CA; Tristate metro area (PA/NJ/NYC/CT)	Chicago (low)
Sodium (mg)	0.48** (moderate)	Southeast	Southwest; Tristate metro area (PA/NJ/NYC/CT)	Remaining North California (high); Atlanta and Charlotte (low)
Calories from packaged food purchases, per capita per day
Fruit	0.38** (moderate)	Montana/Wyoming; Great Lakes	South	Little Rock (high); Remaining Seattle (low)
Non-starchy vegetables	0.44** (moderate)	New England; Florida	South
Processed meats	0.64** (high)	Midwest/South	Northeast; Southern CA	Atlanta (low)
Mixed dishes	0.36** (moderate)		Tristate metro area
Sugar-sweetened beverages	0.57** (high)	Southeast	Northwest; New England	Atlanta and Columbus, OH (low)
Junk foods	0.55** (high)	Midwest	Southwest; Tristate metro area	Atlanta (low)

^1:Clusters are described by the authors and do not reflect formally defined geographic areas.

^2:Outliers are those markets with a value notably higher or lower than their geographic neighbors.

^**Results statistically significant at p<0.001

Overlays of clusters representing low and high quality PFPs are presented in Figures 1 and 2 (see Supplementary Figures 2 and 3 for results by state). In general, we observed that markets in the Southeast, Appalachia and Midwest belonged to more clusters that reflected relatively lower quality PFPs compared with the national average, while markets in the Northeast and West Coast belonged to more clusters reflecting relatively higher quality PFPs. The Atlanta market emerged as a notable outlier across several outcomes (Figure 1).

Figure 1:

Overlay of clusters of relatively lower quality PFPs, by market.

Cluster analysis was conducted using Nielsen market areas (n=88). All state averages were calculated using the survey-weighted average of households located within each state. Grams of sugar and of saturated fat were excluded from this summary map due to redundancy with percent calories from sugar and from saturated fat.

Figure 2:

Overlay of clusters of relatively higher quality PFPs, by market.

Cluster analysis was conducted using Nielsen market areas (n=88). All state averages were calculated using the survey-weighted average of households located within each state. Grams of sugar and of saturated fat were excluded from this summary map due to redundancy with percent calories from sugar and from saturated fat.

3.3. Spatial clustering: Socioeconomic stratification

In the stratified analysis, we found some differences when comparing Moran’s / values for all households to Moran’s / values for sociodemographic subpopulations (Table 3). However, the locations of clusters by socioeconomic group were largely consistent with patterns found among all households. High and low clusters were the most similar among low income households, middle income households and households with lower educational attainment (high school or less; some college). Clusters tended to decrease in size and slightly shift for high income households and highly education households (post college graduates) (Figures 3A and 4A). Results for states can be found in Supplementary Table 2.

Table 3:

Market-level Cluster Analysis – Moran’s I for each nutritional outcome, by sociodemographic subpopulations

>0 to 0.1: random			0.3–0.5: moderately clustered
0.1 – 0.3: slightly clustered			>0.5: highly clustered
	All Households	Income1				Education2
Outcome		Low income	Mid income	High income	High - Low disparity	<HS	Some college	College grad	Post college	Post college - HS disparity
Saturated fat %	0.59**	0.55**	0.57**	0.37**	0.00	0.51**	0.54**	0.44**	0.22**	0.20**
Saturated fat g	0.43**	0.45**	0.49**	0.27**	0.22**	0.43**	0.39**	0.38**	0.25**	0.09
Sugar %	0.60**	0.48**	0.56**	0.50**	0.07	0.44**	0.57**	0.43**	0.15*	0.18**
Sugar g	0.49**	0.52**	0.53**	0.30**	0.15*	0.50**	0.45**	0.40**	0.22**	0.24**
Sodium	0.48**	0.42**	0.50**	0.24**	0.06	0.47**	0.40**	0.31**	0.30**	0.17*
Total Calories	0.44**	0.42**	0.49**	0.29**	0.23**	0.44**	0.37**	0.36**	0.27**	0.12
Fruits, kcal	0.38**	0.21**	0.31**	0.30**	0.11*	0.18*	0.34**	0.23**	0.21**	0.07
NS veg, kcal	0.44**	0.32**	0.38**	0.43**	0.28**	0.32**	0.40**	0.24**	0.27**	0.18*
Proc’d meats, kcal	0.64**	0.54**	0.72**	0.45**	0.12*	0.62**	0.60**	0.53**	0.37**	0.40**
Mixed dishes, kcal	0.36**	0.24**	0.31**	0.25**	0.05	0.29**	0.25**	0.24**	0.21**	0.14*
SSBs, kcal	0.57**	0.47**	0.51**	0.32**	0.09	0.47**	0.40**	0.27**	0.23**	0.25**
Junk foods, kcal	0.55**	0.57**	0.54**	0.38**	0.22**	0.54**	0.54**	0.41**	0.40**	0.24**

^*p<0.05

^**p <0.01

^1:Household income is adjusted for the local cost of living and then categorized into tertiles

^2:Highest educational attainment of either the female or male head of household

HS: high school; NS: non-starchy; Proc’d meats: Processed meats; SSBs: Sugar-sweetened beverages

Figure 3:

Spatial clustering and disparities by socioeconomic subgroups: percent of calories from sugar in packaged foods and beverages.

3A: Moran’s I and spatial clusters of the percent calories from sugar

3B: Average market-level values and disparities for the percent of calories from sugar Differences are also considered disparities when the lower SES group also has a less healthful purchasing pattern (i.e., higher percent calories from sugar). In Figure 3B, differences are represented in yellow, while disparities are in green.

Figure 4:

Spatial clustering and disparities by socioeconomic subgroups: calories per person per day from packaged processed meats.

4A: Moran’s I and spatial clusters of calories from processed meats

4B: Average market-level values and disparities for calories from processed meats Differences are also considered disparities when the lower SES group also has a less healthful purchasing pattern (i.e., more calories per capita per day from processed meats). In Figure 4B, differences are represented in yellow, while disparities are in green

For each outcome, there were markets where differences reflected disparities (i.e., low- income or low-education households had a lower quality outcome) and markets where they did not (i.e., wealthier or more educated households had a lower quality outcome) (Figures 3B and 4B). However, the magnitude of differences/disparities showed no significant clustering pattern across almost all outcomes.

The lack of spatial clustering in income and education differences were consistent with visual examination of clustering patterns stratified by subpopulation. Although sociodemographic groups differ in their average purchase outcomes and in the degree of overall clustering, the locations of high and low clusters were largely consistent. Therefore, there is no clustering pattern in differences between groups. Visual examples for percent calories from sugar and calories per capita per day from processed meats are presented in Figures 3B and 4B. For cluster maps of other outcomes by markets, see Supplementary Figure 4. For cluster maps of all outcomes by states, see Supplementary Figure 5.

3.4. Multivariate Regression

Significant differences were found between markets using multilevel, multivariate regression models, indicating that differences exist even after controlling for household demographic characteristics. Overall, markets belonging to high (or low) clusters also had comparatively high (or low) averages in adjusted regression models, although there are some examples where this is not the case. Figures 5 and 6 show regression results for percent calories from sugar and for calories from processed meats, which show substantial variation. For example, for processed meats, the five lowest ranked markets purchased less than 40 calories per person per day, while the five highest ranked markets purchased more than 65 calories per person per day. Remaining regression results can be found in Supplementary Figure 6 for results by market and Supplementary Figure 7 for results by state.

Figure 5:

Comparing average purchases by market after adjusting for household sociodemographic characteristics¹ to markets found in high/low outcome cluster areas² in spatial analysis

5A: Percent of calories from sugar

5B: Calories from processed meats

1: Market-level values are predicted values using coefficients from multilevel models that control for household income, education, race/ethnicity, age composition and year and are adjusted for complex survey design.

2: For high/low clusters for the percent calories from sugar, refer to Figure 3A, and for high/low clusters of calories from processed meats, refer to Figure 4A.

Rem: Remaining (term provided by Nielsen)

Figure 6:

Disparities¹ in the average percent of total calories from sugar in PFPs after controlling for household demographic characteristics,² by Nielsen market

6A: Disparities by educational attainment – alternative option

6B: Disparities by household income tertile

1: Markets are ordered from least difference (left) to greatest difference (right) between low SES and high SES groups. Differences are also considered disparities when the difference is statistically significant and the low SES group has the worse outcome (i.e., higher percent of calories from sugar). Therefore, the greatest disparities are found on the right side of the figure.

2: Market-level values are predicted values using coefficients from multilevel models that control for household income, education, race/ethnicity, age composition and year and are adjusted for complex survey design.

Rem: Remaining (term provided by Nielsen)

When models were stratified by education and income groups, we found significant differences between markets after adjusting for other household characteristics. Figure 6 illustrates stratified results for the percent calories from sugar and Figure 7 for the calories from processed meats (see Supplementary Figures 8 and 9 for these results by state). We also found that outcomes were significantly different between high and low education households within most geographic areas, that these differences largely reflect disparities, and that there is substantial variation in the magnitude of within-market disparities. For example, for processed meats, we found disparities by education in 69 of 88 markets areas, ranging from 13.5 kcal/person/day in Boston to 46.1 in the “Remaining” Oklahoma City market area. While there is variation in differences by income group, these differences are not always significant and do not always reflect disparities. With processed meats, disparities by income were found for only 2 market areas, Houston and Columbus, and both were less than 1 kcal/person/day.

Figure 7:

Differences¹ in the average calories from packaged processed meats after controlling for household demographic characteristics,² by Nielsen market

7A: Disparities by educational attainment

7B: Disparities by household income tertile

1: Markets are ordered from least difference (left) to greatest difference (right) between low SES and high SES groups. Differences are also considered disparities when the difference is statistically significant and the low SES group has the worse outcome (i.e., more calories from processed meats). Therefore, the greatest disparities are found on the right side of the figure.

2: Market-level values are predicted values using coefficients from multilevel models that control for household income, education, race/ethnicity, age composition and year and are adjusted for complex survey design.

Rem: Remaining (term provided by Nielsen)

4. Discussion

We found geographic patterns and significant variation in the quality of PFPs, both between Nielsen market areas and states, suggesting there may be systemic factors and multilevel processes that influence the nutritional quality of PFPs in the US. Spatial clustering suggests there are factors that operate at a regional level, while statistical differences between areas after controlling for household demographics indicate some factors operate at the market level and state level. Characterizing these patterns is an important first step in identifying factors that drive low quality of PFPs and disparities by income and education. We present further discussion of the context and implications of these findings for each aim.

4.1. Spatial Clustering

Overlaying clusters over several PFP outcomes showed consistent patterns in the locations of clusters of PFPs. Patterns of relatively lower quality PFPs were highly clustered in the Southeastern US and Appalachia, whereas higher quality PFPs were clustered in the Northeast and on the West Coast. Although these patterns are consistent with other national-level studies,^10,38,39 our results provide additional evidence by incorporating an explicit spatial cluster analysis and a subpopulation analysis that indicates that these patterns are generally consistent for high and low SES groups.

Geographic analysis of dietary behaviors for vulnerable sociodemographic groups can be combined with information about other social determinants of health, current policies, and disease outcomes that can inform the location of future health services and policy reform.⁴⁰ For example, the Southeast and Appalachia have a high prevalence of poverty and diet-related disease. Our study showed that households in these regions purchase relatively higher amounts of sugar, processed meat, and SSBs. While these clusters are not indicative of absolute measures of unhealthy purchases – indeed, the proportion of calories from sugar is high in all areas of the country – it is important to understand differences in diet at the population level in order to understand underlying causes of the incidence of diet-related disease.⁴¹

Further investigation is warranted into location-specific factors that may be associated with these trends, such as regional characteristics of the food supply. Differences in local food supply systems could also explain regional clustering. For example, store type is associated with the quality of PFPs, where households tend to buy less healthful PFPs from mass merchandisers (e.g., Walmart) and convenience stores compared to club (e.g., Costco) and grocery stores.^42,43 In our sample, we found that households in markets belonging to high clusters for purchases with a high percentage of sugar had a lower share of PFPs from club and grocery stores and a higher share of PFPs from mass merchandisers ( average of 15% of PFP calories purchased from club stores, 54% from grocery stores and 34% from mass merchandisers) compared to households in markets belonging to clusters of low sugar PFPs (25% club, 61% grocery and 19% mass merchandisers). Future research could consider the association between specific retail chains (e.g., Walmart) and the quality of PFPs across markets or states. In addition, the identification of clusters of lower quality purchases could be used to target funding for state and local-level public health marketing that focus on avoiding SSBs, processed meats, junk foods and sugar. Finally, geographic outliers may serve as useful tools for generating new hypothesis. For example, the Atlanta market emerges as a healthy outlier compared to neighboring markets in the Southeast across several outcomes. Atlanta also has a higher median household income compared to neighboring regions,⁴⁴ and further investigation is needed to understand how these are interrelated with the quality of household food purchases. Public health policies or social programs specific to heathy outliers like Atlanta could be modeled as best practices.

Geographic patterns for high socioeconomic subpopulations (high income and post-college graduates) were often less clustered than patterns found among all households and among lower SES groups. Population health research has similarly found that mortality for more highly educated individuals varies little across states compared to significant variation among individuals with low education, suggesting that high-SES individuals may be “buffered” from contextual influences on health.⁴⁵ Households with higher income or higher education often may be able to afford packaged foods that have lower sugar and saturated fat content, shop in stores that sell healthier packaged foods, or be more aware of the nutritional content of the food they are purchasing. This increased access may result in food purchasing patterns that are less dependent on the regional food supply or less influenced by local food marketing campaigns. Although significant variation in the magnitude of income and education disparities was found across markets and states, no geographic patterns in disparities were found. This suggests that the processes that result in disparities in one market or state have little influence on disparities in neighboring markets.

When comparing our sensitivity analysis by state to our analysis by market, we found clustering patterns for PFPs for the whole sample were similar whereas clustering patterns for some stratified SES groups differed. In general, when using market areas, clusters tended to occur in more places and covered a larger geographic area. For example, California is a large state which contains multiple markets. In the market area data, clusters often centered around the central and southern California markets, while the remaining area of Northern California often appeared as an outlier. In comparison, when states were used, a local cluster for the same outcome either centered around California (i.e., the remaining northern market was overpowered by the central and southern markets) or California was no longer part of a cluster (i.e., the remaining and urban areas cancelled each other out). These results suggest there are important differences in the quality of PFPs between urban and rural areas. Therefore, a limitation of state-level cluster analysis is that states have a mix of urban and rural households, which may mask each other, while Nielsen’s markets are delineated in part by level of urbanization. On the other hand, market area analysis may overstate the degree of clustering if they typically occur within large states with multiple markets, where similarities may be driven by state-specific food-related policies. Our results indicate that it is important to consider geographic units at multiple scales when attempting to understand systemic factors that influence the quality of PFPs.

4.2. Multivariate Regression

The significant variation that exists between markets and states after controlling for household demographics provides a complementary but distinct set of comparisons for formulating hypothesis about the systemic drivers of the quality of PFPs. Nielsen markets are either predominantly urban or rural and may have different food marketing campaigns, while states have policies that affect food access and social programs. Contextual factors appear to be more influential for low-education households, since disparities by education tend to be wider than disparities by SES. This aligns with previous findings that education is more strongly associated with demand for healthier purchases than income.^38,46

Differences that remain after controlling for demographic composition indicate possible systemic factors that are delineated by market or state level boundaries. For example, food-related tv ads are geographically targeted. One study found that Nielsen Designated Market Areas with a higher proportion of Black children and adolescents viewed more ads per week for SSBs and sweets while areas with a higher proportion of low-income households viewed more ads for cereal, snacks, sweets and SSBs.⁴⁷ Future studies could pair data from Nielsen Media Research with Nielsen Homescan to understand how food and beverage marketing mediates the relationship between market-level sociodemographic composition and the quality of household food purchases. While public health marketing campaigns have a lower impact on viewers than food and beverage ads, there is also significant variation in public health campaigns across Nielsen DMAs. In 2010, state and local level marketing represented a full third of obesity prevention marketing in select DMAs, largely based in California and Arizona.⁴⁸ These ads were more likely to encourage healthy eating compared to national obesity prevention ads. Future research could investigate whether these campaigns may have contributed to the Southwest frequently emerging as a low cluster area for unhealthy PFPs.

Similar to the identification of geographic outliers, comparing markets/states with the lowest and highest magnitudes of disparity can help identify systemic factors associated with disparities in the quality of food purchases. To take processed meats as an example, the market area that encompasses the suburban region south of NYC and the state of Massachusetts have a relatively low average calories from processed meats PFPs and no disparity (South Exurban NY) or a small disparity (MA) by education level.

4.3. Strengths and Limitations

This study has several limitations. While Moran’s I and LISA are used to find statistically significant patterns of geographic clusters, whether these clusters are meaningful remains to be determined. Clusters indicate relatively high and low areas of each of the PFP nutritional outcomes, but their substantive significance depends on the underlying magnitude and range of values. Although our primary focus is describing geographic patterns, we attempted to integrate this contextual information (Table 1; Figures 3B and 4B). While market areas and states are subnational spatial units, they may still be too large to capture variation in the healthfulness of household food purchasing patterns. In addition, our dietary data is limited to household purchases of packaged foods and beverages and thus does not capture foods and beverages sold without barcodes nor food eaten outside of the home. As such, we are unable to combine nutritional outcomes of PFPs together to get a singular measure of healthfulness, nor are we able to correct for the share of total dietary intake that comes from PFPs. Purchases also do not equal consumption, and we are unable to account for food waste. Our purchase data is unable to capture store inventory, so it is unknown to what extent clusters reflect regional patterns in the supply of PFPs. Lastly, our ability to assess different sociodemographic trends is limited by data availability. Nielsen provides household income data as a categorical variable with wider bins at higher income levels, capping income at $100,000/year or greater, limiting our ability to contrast low- and high-income groups. Small populations of different racial and ethnic subgroups in many areas of the country precluded analysis by race/ethnicity. In addition, since Nielsen predominantly samples urban counties, the rural sample is much smaller than our urban sample, and rural households are more likely to be sampled from counties closer to major markets. Therefore, although we apply Nielsen’s survey weights, samples in market areas and states with a larger rural population may not be fully representative of their populations. This may be particularly true for the 33 market areas derived from the 24 “remaining” non-metropolitan Nielsen markets.

Despite these limitations, our study has several important strengths. Our sample has the highest spatial resolution of any nationally-representative dataset of nutrition and sociodemographic information, allowing for the first investigation of geographic patterns and clustering in PFPs by income and education. In addition, data gathered from household purchases has several advantages in comparison to other measures of dietary intake. First, Homescan is an open, longitudinal cohort with year-round data collection, from which we captured usual purchase patterns without bias from seasonal changes in purchases and diet. Linking scanned barcodes to product-specific nutrition facts panel information in a time-specific manner allows for improved measurement of nutrients of concern, including saturated fat and sugar.¹⁶ In contrast, dietary intake based on 24-hour recall data may be weaker indicators of usual diet and lacks specificity because items are linked to the USDA Nutrition Facts Database, which only captures a small fraction of the total packaged products in the US food system and is not updated to keep pace with a rapidly changing food supply.^49,50

5. Conclusion

Using a national dataset with spatially-rich food purchase data and sociodemographic information, we found significant clustering of a variety of nutrient and food group outcomes relating to dietary quality. We observe that high income and highly educated households do not seem to be as impacted by systemic factors that may influence food purchases. Furthermore, after adjusting for household composition, we find significant variation remains between Nielsen markets and US states in the average quality of PFPs. We also find variation in the magnitude of disparities, with disparities by education often wider than disparities by income. The identification of these geographic patterns and variation between markets and states should inform future research on location-specific marketing, food culture, and health policies that may impact the healthfulness of purchase decisions, particularly for households with low education.

Highlights.

• The quality of packaged foods purchased by households differs across the U.S.

• Higher quality purchases tend to cluster on the West Coast and in the Northeast

• Lower quality purchases tend to cluster in the Southeast and Appalachia

• The size of disparities by income and by education differs across the U.S.