Neighborhood price of healthier food relative to unhealthy food and its association with type 2 diabetes and insulin resistance: The Multi-Ethnic Study of Atherosclerosis

Abstract

This study examined the association between the price of healthier food relative to unhealthy food and type 2 diabetes prevalence, incidence and insulin resistance (IR). Data came from the Multi-Ethnic Study of Atherosclerosis exam 5 administered 2010–2012 (exam 4, five years prior, was used only for diabetes incidence) and supermarket food/beverage prices derived from Information Resources Inc. For each individual, average price of a selection of healthier foods, unhealthy foods and their ratio was computed for supermarkets within 3 miles of the person’s residential address. Diabetes status was confirmed at each exam and IR was assessed via the homeostasis model assessment index. Multivariable-adjusted logistic, modified Poisson and linear regression models were used to model diabetes prevalence, incidence and IR, respectively as a function of price and covariates; 2,353 to 3,408 participants were included in analyses (depending on the outcome). A higher ratio of healthy-to-unhealthy neighborhood food price was associated with greater IR (4.8% higher HOMA-IR score for each standard deviation higher price ratio [95% CI -0.2% to 10.1%]) after adjusting for region, age, gender, race/ethnicity, family history of diabetes, income/wealth index, education, smoking status, physical activity, and neighborhood socioeconomic status. No association with diabetes incidence (relative risk=1.11, 95% CI 0.85 to 1.44) or prevalence (odds ratio=0.95, 95% CI 0.81 to 1.11) was observed. Higher neighborhood prices of healthier food relative to unhealthy food were positively associated with IR, but not with either diabetes outcome. This study provides new insight into the relationship between food prices with IR and diabetes.

INTRODUCTION

An estimated 29 million individuals are living with diabetes, accounting for more than 9% of the population (Centers for Disease Control and Prevention, 2014), and the prevalence of diagnosed diabetes has grown sharply since 2000 (Centers for Disease Control and Prevention, 2016). It is vital that we identify potential areas of intervention to reverse this trend. Type 2 diabetes is mainly driven by obesity, and diet quality plays an important role in its development. Diets high in unhealthy, sugary, energy-dense and nutrient-poor foods are associated with an increased risk of obesity and diabetes (Brunner et al., 2008; Malik et al., 2010), while consumption of healthier foods – such as fruits and vegetables and dairy – has shown protective effects (Choi et al., 2005; Margolis et al., 2011; Mursu et al., 2014). Unfortunately, foods that are energy dense and generally considered unhealthy tend to be the most affordable (Andrieu et al., 2006; Drewnowski, 2010; Drewnowski and Darmon, 2005; Kern et al., 2016; Monsivais et al., 2010). The price of food is associated with food purchasing decisions and consumption (Aggarwal et al., 2016; Andreyeva et al., 2010; Connors et al., 2001; Glanz et al., 1998), and thus price differences between healthy and unhealthy foods are expected to be associated with downstream effects of diet, including diabetes and insulin resistance.

Prior work that has examined prices of healthy and unhealthy foods and health outcomes has largely focused on the relationship between food prices and body weight. Most (Chaloupka and Powell, 2009; Chou et al., 2004; Finkelstein et al., 2014; Powell et al., 2007; Powell and Bao, 2009) but not all (Beydoun et al., 2008; Han and Powell, 2011) of these studies have shown higher healthier food prices are associated with higher BMI/obesity and/or have found that higher unhealthier prices are associated with lower BMI/obesity. However, more research is needed on the association between food prices and downstream health effects. No studies have examined food price association with type 2 diabetes and only a few studies have linked food prices directly to insulin resistance. Those studies focused primarily on metropolitan-area fast-food prices (hamburger, pizza, fried chicken) (Duffey et al., 2010; Meyer et al., 2014; Rummo et al., 2015). Duffey et al (2010) found metropolitan area prices inversely associated with insulin resistance (Duffey et al., 2010), while Rummo et al (2015) found no association (Rummo et al., 2015) and Meyer et al (2014) found associations only among less advantaged residents (Meyer et al., 2014). No work to date has examined food prices at stores nearby residents and examined associations with type 2 diabetes and insulin resistance.

This study spatially linked participants from a large multi-ethnic sample to nearby supermarkets to examine associations between neighborhood food price and type 2 diabetes prevalence, type 2 diabetes incidence and insulin resistance.

METHODS

MESA data

This study utilized data gathered by the Multi-Ethnic Study of Atherosclerosis (MESA). The MESA is a population-based longitudinal cohort study of ethnically diverse adults aged 45–84 years (Bild et al., 2002). Individuals were recruited from six sites across the United States: Bronx/Upper Manhattan, NY; Baltimore City and Baltimore County, Maryland; Forsyth County, North Carolina; Chicago, Illinois; St. Paul, Minnesota; and Los Angeles County, California. The first examination was conducted in 2000–2002 and four examinations occurred during 10 years of follow-up. Food prices were only available concurrent to exam 5, thus this study only included participants who completed exam 5 (April 2010–January 2012). Participant characteristics for the incident diabetes analysis were gathered from exam 4 (which occurred five years earlier: 2005–07, considered the “baseline” in this analysis) while exam 5 was used for all other analyses.

Outcomes - Diabetes and insulin resistance

Type 2 diabetes status was based on the 2003 American Diabetes Association criteria, defined as fasting glucose ≥126 mg/dl and/or use of antidiabetic medications (ADA Expert Committee, 2003). Prevalent type 2 diabetes at exam 5 was defined as meeting the diabetes criteria at any of the five examinations. Incident type 2 diabetes at exam 5 was defined as meeting the diabetes criteria at the time of exam 5 with no evidence of diabetes at any prior exam.

Fasting glucose was measured by rate reflectance spectrophotometry using thin-film adaptation of the glucose oxidase method on the Vitros analyzer (Johnson & Johnson Clinical Diagnostics, Rochester, NY). Insulin resistance at exam 5 was measured according to the homeostasis model assessment index of insulin resistance (HOMA-IR). This index is well correlated with measures from the gold-standard hyperinsulinemic clamp and is calculated as (Matthews et al., 1985):

$$\mathit{HOMAIR}=\frac{\mathit{Fasting}\mathit{glucose}\left(\frac{\mathit{mmol}}{L}\right)\times \mathit{Fasting}\mathit{insulin}(in\mathit{microunit}s\mathit{per}\mathit{liter})}{22.5}$$

A log transformation was used to reduce skewness of the distribution and created a normal distribution of values for use in multivariable analyses.

Covariates

Other MESA person-level data included in this study were: diet (see details below), age (continuous), sex, race/ethnicity (Caucasian, Chinese-American, African-American, Hispanic), family history of diabetes, smoking status (never, former, current), education level (high school diploma, GED, or less; some college, technical or associates degree; bachelor’s degree or higher), income/wealth index (an ordinal measure ranging from 0 to 8 based on the combination of income level and ownership of four assets: car, home, land, and investments), and physical activity (measured as total hours of physical activity per day; operationalized into tertiles for modeling).

Price data

Data on food prices were obtained from Information Resources Inc. (IRI) (Bronnenberg et al., 2008; IRI, 2014, 2015; Symphony IRI, 2015), a market research group that monitors prices of 299 consumer packaged goods sold in large chain supermarkets and superstores across the U.S.(Bronnenberg et al., 2008; IRI, 2014, 2015; Symphony IRI, 2015). Twenty-nine parent companies owned 100 supermarket companies in our study area. Among those, 8 parent companies (that owned 28 companies) did not agree to release their stores’ data and no further information was available for those companies. Data used in this study were from 794 stores located in 11 states (including Washington DC), 72 counties, and 757 census block groups. Data years were 2009–2012.

Nine food/beverage product categories were selected to serve as proxies for either healthier or unhealthier foods. Because data for fresh fruits and vegetables were not available, refrigerated products were selected in order to roughly approximate costs of fresh fruit and vegetable spoilage and storage/distribution, and proxy fresh produce. Healthier food was represented by dairy (refrigerated milk, yogurt, cottage cheese), fruits and vegetables (frozen vegetables, and fresh orange juice). While the healthfulness of orange juice is questionable, the product was chosen to represent healthier foods due to its high correlation with the price of fresh oranges (Morris, 2011) not for its own nutritional value. Unhealthier food was represented by packaged, highly processed, long-shelf life products: soda, sweets (chocolate candy, cookies), and salty snacks.

The primary exposure of interest was the price of healthier foods relative to unhealthy foods, which was operationalized as the ratio of the average price per serving of healthy food divided by the average price per serving of unhealthy food and referred to as the healthy-to-unhealthy price ratio. The price of healthy foods and unhealthy foods were also modeled separately as secondary exposures of interest. Average prices of healthy and unhealthy foods were calculated using weights for each product class based on national consumption estimates (i.e., food types – e.g., fruits and vegetables – that are consumed more received larger weights) and converted to z-scores. A one unit change in the z-scores for the price ratio were equivalent to a 14% change in the price of healthy food relative to unhealthy food while a one unit change in the z-scores for healthy and unhealthy foods represented differences of $0.04 and $0.03 per serving, respectively. A sensitivity analysis which calculated overall healthy and unhealthy food prices using equal weights for each product class (e.g., dairy) was also performed and model results were similar to the results using the original weighting methodology.

Other data sources

US Census data came from the American Community Survey (ACS) 2007–2011. Geographic regions (Northeast, Midwest, South, West) and population density were assigned to each participant. A neighborhood block group SES index was created using six variables representing wealth and income, housing value, education, and managerial or professional occupations, and was operationalized as a single continuous measure as described by Diez-Roux et al (Diez Roux et al., 2001).

Addresses of MESA participants and supermarket addresses from the pricing dataset were used to link individuals to the average food/beverage prices at supermarkets within a 3-mile buffer (radius) of each MESA participant residence using ArcGIS 10.0. For simplicity, we will refer to this buffer as an individual’s “neighborhood”. Further information regarding the relationship between supermarket food prices and neighborhood characteristics can be found in a previous publication(Kern et al., 2017). Median number of supermarkets per MESA participant in the analytic sample was 5 (25^th–75^th percentile 3–6). A 3-mile radius was used for consistency with other research examining neighborhood food environments, and is in line with prior research estimating the average distance individuals travel to their primary supermarket (Drewnowski et al., 2012; Fuller et al., 2013; Michimi and Wimberly, 2010; 2014).

Statistical models

The following sequence of models was examined for each outcome: unadjusted for covariates (model 0), and then adjusted for potential confounders: model 1, geographic region, age, gender, and family history of diabetes; model 2, plus income/wealth index, education, race, smoking status and physical activity; and model 3, plus neighborhood level SES.

The analysis of diabetes prevalence included all individuals completing exam 5 and who lived within 3 miles of a supermarket from our dataset. Individuals with type 1 diabetes, those missing data for covariates, and those missing type 2 diabetes status at exam 5 were excluded (n=258). For diabetes prevalence, logistic regression was used to compute odds ratios (ORs) and 95% confidence intervals (CIs).

The analysis of diabetes incidence included all individuals free of type 2 diabetes at exam 4 and who lived within 3 miles of a supermarket at any time between exam 4 and exam 5. Of those included for analysis, 97% lived within 3 miles of a supermarket for the entire duration between the two exams, and just 1.3% lived at an address near a supermarket for less than half the time. A modified Poisson regression model with robust error variance(Zou, 2004) was used to measure the relative risk of developing diabetes between exam 4 and 5 for each unit increase in the z-score of the price exposure. Covariates were measured during exam 4 and included the same variables described above. Risk ratios (RRs) and 95% CIs are reported. A sensitivity analysis was performed to account for potential clustering within the same census tract and results were identical to those without clustering thus results thus are not shown.

The analysis of insulin resistance levels included all individuals free of type 2 diabetes at exam 5 and with complete data for fasting insulin and glucose levels. HOMA-IR levels were assessed using a normal linear model. Because HOMA-IR was heavily skewed a log transformation was used and the resulting distribution was normal, and thus the log-transformed HOMA-IR values were modeled. Regression coefficients of the exposure illustrate the mean change in the log of HOMA-IR (reported as the percent change in the geometric mean of HOMA-IR for more meaningful interpretation) for every one unit change in the z-score of the price exposure (healthy price, unhealthy price, or the ratio). Log-linear effect estimates and 95% CIs are reported.

There was no evidence of non-linearity in the association between prices and the outcome variables, thus prices were analyzed as continuous exposures, operationalized as z-scores, in all analytic models. Tertiles of healthy-to-unhealthy price ratio are used in tables 1 and 2 for descriptive purposes.

Table 1.

MESA participant characteristics by healthy-to-unhealthy price ratio at the time of exam 5 (N=3,408)

	All participants		Lowest healthy-to- unhealthy ratio (1.55 – 1.88)		Middle healthy- to-unhealthy ratio (1.88 – 2.01)		Highest healthy- to-unhealthy ratio (2.01 – 2.39)
	N /Mean	Column %/ SD	N /Mean	Row %/ SD	N /Mean	Row %/ SD	N /Mean	Row %/ SD
Number of participants (N)	3408		1148	33.7%	1141	33.5%	1119	32.8%
MESA location (n, %)
Forsyth County, NC	645	18.9%	295	45.7%	349	54.1%	1	0.2%
New York, NY	665	19.5%	322	48.4%	273	41.1%	70	10.5%
Baltimore, MD	570	16.7%	420	73.7%	147	25.8%	3	0.5%
St. Paul, MN	7	0.2%	6	85.7%	0	0.0%	1	14.3%
Chicago, IL	814	23.9%	102	12.5%	333	40.9%	379	46.6%
Los Angeles, CA	707	20.7%	3	0.4%	39	5.5%	665	94.1%
Region of residence (n, %)
Northeast	660	19.4%	318	48.2%	273	41.4%	69	10.5%
Midwest	797	23.4%	95	11.9%	331	41.5%	371	46.5%
South	1236	36.3%	732	59.2%	496	40.1%	8	0.6%
West	715	21.0%	3	0.4%	41	5.7%	671	93.8%
Age (mean, SD)	70.2	9.4	70.4	9.2	70.4	9.4	69.8	9.8
Female (n, %)	1809	53.1%	603	33.3%	636	35.2%	570	31.5%
Race/ethnicity (n, %)
White	1274	37.4%	475	37.3%	563	44.2%	236	18.5%
Chinese American	481	14.1%	10	2.1%	89	18.5%	382	79.4%
Black	1072	31.5%	549	51.2%	321	29.9%	202	18.8%
Hispanic	581	17.0%	114	19.6%	168	28.9%	299	51.5%
Family history of diabetes (n, %)	1529	44.9%	546	35.7%	490	32.0%	493	32.2%
BMI (mean, SD)	28.2	5.6	29.1	5.6	28.3	5.8	27.1	5.2
<25 (n, %)	1040	30.5%	277	26.6%	337	32.4%	426	41.0%
25–29.9 (n, %)	1283	37.6%	422	32.9%	446	34.8%	415	32.3%
≥30 (n, %)	1085	31.8%	449	41.4%	358	33.0%	278	25.6%
Education level (n, %)
HS/GED or less	1022	30.0%	310	30.3%	331	32.4%	381	37.3%
Some college, Technical or
Associate degree	945	27.7%	328	34.7%	306	32.4%	311	32.9%
Bachelor’s degree or higher	1441	42.3%	510	35.4%	504	35.0%	427	29.6%
Per capita income (in $10k)	2.6	1.9	2.7	1.8	2.8	2.0	2.2	1.8
Wealth index	2.6	1.2	2.6	1.2	2.6	1.2	2.5	1.2
Income/wealth index	5.0	2.2	5.2	2.1	5.2	2.2	4.7	2.4
Marital status (n, %)
Not married or living with partner	1395	40.9%	533	38.2%	474	34.0%	388	27.8%
Married/Living with partner	2013	59.1%	615	30.6%	667	33.1%	731	36.3%
Smoking status (n, %)
Never smoked	1575	46.2%	444	28.2%	507	32.2%	624	39.6%
Former smoker	1585	46.5%	598	37.7%	556	35.1%	431	27.2%
Current smoker	248	7.3%	106	42.7%	78	31.5%	64	25.8%
Physical activity, hours per day (mean, SD)	8.8	5.4	9.5	5.8	9.1	5.3	7.7	4.8
Daily calorie intake (mean, SD)	1643	782	1668	793	1641	765	1619	789

Table 2.

Proportion of participants with type 2 diabetes, incident type 2 diabetes, and insulin resistance levels by tertile of the healthy-to-unhealthy price ratio and the average serving price of healthy and unhealthy foods

	All participants			Lowest ratio (1.55 – 1.88)			Middle ratio (1.88 – 2.01)			Highest ratio (2.01 – 2.39)
Number of participants (N)	3,408			1,148			1,141			1,119
Number with diabetes at exam 5 (n, %)	790	23.2%		275	24.0%		247	21.6%		268	23.9%
Number with incident diabetes at exam 5 (n, %) a	154	5.4%		45	4.8%		49	5.2%		60	6.4%
Insulin resistance level at exam 5 (mean, SD, median) b	2.32	0.67	2.33	2.31	0.66	2.32	2.32	0.68	2.33	2.31	0.69	2.34
Average food prices per serving c
Healthy food price per serving (mean, SD, median)	$0.61	$0.04	$0.60	$0.60	$0.06	$0.57	$0.59	$0.03	$0.58	$0.62	$0.03	$0.62
Unhealthy food price per serving (mean, SD, median)	$0.31	$0.03	$0.30	$0.33	$0.04	$0.31	$0.30	$0.01	$0.30	$0.29	$0.01	$0.29
Ratio of Healthy-to-Unhealthy (mean, SD, median)	1.97	0.14	1.94	1.83	0.05	1.85	1.95	0.05	1.94	2.14	0.09	2.13

^apercentages are within the analytic population for the incidence analysis, n=2,829

^bvalues are from within the analytic population for the insulin resistance analysis, n=2,353; insulin resistance is measured as the log of HOMA-IR

^cWe refer readers interested in price variation by supermarkets, regions, and sociodemographic variables to a previously published paper (Kern et al., 2017)

The multiplicative interaction between individual SES – measured separately by the income/wealth index and education level – with the price ratio was examined, adjusting for all covariates from model 3, for each outcome of interest. Our hypothesis was that the effect of price on the outcomes would be stronger within lower levels of SES. To illustrate how the effect differed by level of SES, stratified analyses were performed within each tertile of income/wealth (low, medium, and high) and education (high school degree or less, some college, or bachelor’s degree or more). Similarly, the analysis tested for an interaction between race with the price ratio and a stratified analysis was conducted within each race category (White, Chinese, Black, and Hispanic).

All analyses were performed using SAS software version 9.3 (SAS Institute, Cary, NC).

Sample

Detailed participant selection criteria for each of the analyses (type 2 diabetes prevalence, type 2 diabetes incidence, and HOMA-IR) are shown in Figure 1. Of the 4,716 individuals completing exam 5 there were 3,666 living within 3 miles of a supermarket in our dataset. Participants retained for analyses of each outcome were: 3,408 for diabetes prevalence (excluded type 1 diabetes n=7, missing data for covariates n=181, and unknown diabetes status at exam 5 n=70); 2,829 for diabetes incidence (excluded prevalent or unknown diabetes status at exam 4); 2,353 for HOMA-IR (excluded diabetes n=680 and those missing glucose or insulin data n=336). More than half of the 1,308 excluded from prevalence analyses were from the St. Paul site (a site with many Hispanics), as supermarket data was not available for this area. This led to a higher proportion of excluded participants being Hispanic (see Supplemental Table 1 for comparison of excluded and included). In addition, those excluded had slightly higher BMI and lower education; but otherwise excluded and included were generally similar.

Figure 1.

Participant selection criteria for each of the three outcomes of interest

* Approximately 60% were from St Paul MN where supermarket chains were unwilling to release their data

RESULTS

Demographics and other characteristics of the MESA participants are reported in Table 1. Overall the serving price of healthy food was nearly twice as expensive as unhealthy food (mean±SD of healthy-to-unhealthy price ratio = 1.97±0.14; $0.61±$0.04 vs. $0.31±$0.03 average price per serving for healthy and unhealthy food, respectively). Large difference across the healthy-to-unhealthy price ratio tertiles was only apparent for region (highest ratio in west which was primarily the Los Angeles area and lowest in south which was primarily North Carolina and Baltimore) and for race/ethnicity (Chinese resided in highest ratio areas and Black in lowest ratio areas) (Table 1). The relatively higher ratios for individuals from the Chicago and LA sites were due to higher costs of healthy food ($0.62±$0.02 and $0.64±$0.03, respectively vs. $0.59±$0.05 elsewhere). Unhealthy foods were also relatively lower in those two areas compared to others ($0.31±$0.02 and $0.29±$0.01, respectively vs. $0.32±$0.03 elsewhere). Participants with the lowest income (primarily Hispanic and Chinese) lived in areas with the highest ratio. There were no apparent differences in the proportion of individuals with prevalent type 2 diabetes or levels of insulin resistance across tertiles of the price ratio (Table 2). There appeared to be a slight increase in the rate of incident diabetes as the price ratio increased (4.8%, 5.2%, and 6.4% from lowest to highest).

Multivariable models

Results from the analytic models adjusting for covariates are shown in Table 3. Our hypotheses were that the ratio of healthy-to-unhealthy price would be positively associated with insulin resistance and diabetes. The healthy-to-unhealthy price ratio was modestly associated with the level of insulin resistance: a one unit increase in the price ratio z-score was associated with a 5.2% higher geometric mean HOMA-IR score (95% CI=[0.2% to 10.5%], adjusted for age, gender, race/ethnicity, family history of diabetes, income/wealth index, education level, smoking status, and physical activity). The first adjusted model (Model 1 in Table 3) showed a significant association while the unadjusted model (Model 0) did not, mainly due to the inclusion of region and race, which were negative confounders, and thus the exposure estimate became stronger after they were added. After additional adjustment for neighborhood SES, the estimate remained largely the same though confidence intervals included the null (4.8%, 95% CI=[−0.2% to 10.1%]). For diabetes prevalence and incidence, there was no evidence of an association with healthy-to-unhealthy ratio (OR model 3 =0.95, 95% CI=[0.81 to 1.11] and adjusted RR=1.11, 95% CI=[0.85 to 1.44], respectively).

Table 3.

Multivariable models for each of the three outcomes of interest: prevalence of type 2 diabetes at exam 5, incidence of type 2 diabetes from exam 4 to exam 5, and level of insulin resistance at exam 5 ^a Per editor comment E-1, p-values have been removed from this table.

	Exposure of interest
	Healthy-to-unhealthy ratio			Healthy food price			Unhealthy food price
		95% CI			95% CI			95% CI
A. Type 2 diabetes prevalence (n=3,408)	OR	Lower	Upper	OR	Lower	Upper	OR	Lower	Upper
Model 0: No covariates	1.003	0.927	1.085	0.943	0.869	1.022	0.949	0.874	1.031
Model 1: Region, age, gender, race,	0.959	0.824	1.116	0.888	0.784	1.005	0.948	0.855	1.051
family history of diabetes
Model 2: SES, smoking, PA	0.968	0.831	1.129	0.896	0.791	1.016	0.950	0.856	1.055
Model 3: Neighborhood SES	0.948	0.812	1.106	0.931	0.816	1.061	0.981	0.879	1.093
B. Type 2 diabetes incidence (n=2,829)	RR	Lower	Upper	RR	Lower	Upper	RR	Lower	Upper
Model 0: No covariates	1.179	1.025	1.355	0.985	0.855	1.134	0.828	0.684	1.002
Model 1: Region, age, gender, race,	1.146	0.882	1.490	0.864	0.698	1.071	0.857	0.695	1.057
family history of diabetes
Model 2: SES, smoking, PA	1.138	0.874	1.481	0.868	0.700	1.077	0.862	0.698	1.064
Model 3: Neighborhood SES	1.107	0.850	1.441	0.941	0.754	1.174	0.912	0.734	1.134
C. Insulin resistance (n=2,353)	Estimate	Lower	Upper	Estimate	Lower	Upper	Estimate	Lower	Upper
Model 0: No covariates	−0.007	−0.034	0.020	−0.025	−0.052	0.003	−0.014	−0.041	0.013
Model 1: Region, age, gender, race, family history of diabetes	0.051	0.002	0.100	−0.052	−0.091	−0.012	−0.046	−0.079	−0.014
Model 2: SES, smoking, PA	0.051	0.002	0.100	−0.051	−0.090	−0.011	−0.046	−0.078	−0.014
Model 3: Neighborhood SES	0.047	−0.002	0.097	−0.046	−0.087	−0.004	−0.042	−0.076	−0.008

CI, confidence interval; OR, odds ratio; RR, risk ratio; SES, socioeconomic status; PA, physical activity.

^aModel 0 includes only the price exposure; Model 1 covariates include region, age, gender, race/ethnicity, and family history of diabetes; Model 2 covariates include those in Model 1 plus income/wealth index, education level, smoking status, and physical activity; Model 3 covariates include those in Model 2 plus neighborhood SES

Separate models for healthy and unhealthy food prices yielded similar results to those reported above: significant negative associations with insulin resistance were detected, while no significant associations were detected with either diabetes outcome. As expected, higher unhealthier food price was negatively associated with log HOMA-IR; but contrary to expectation, higher healthier food price was also negatively associated with log HOMA-IR.

There was no evidence of effect modification -- by income/wealth, education or race -- of the healthy-to-unhealthy price ratio association with any of the three outcome measures (p for interactions ranged 0.07 to 0.82). See Supplemental Table 2 for results.

DISCUSSION

This study of healthy-to-unhealthy food price ratios at local supermarkets found positive associations with insulin resistance, indicating that residents living in areas with larger price differentials between healthier and unhealthy food have greater insulin resistance. No association was found between healthy-to-unhealthy food prices and prevalence or incidence of type 2 diabetes.

No previous studies have examined the price of unhealthy foods at nearby grocery stores and insulin resistance, but the association detected between local food prices and insulin resistance in this study is consistent with prior research linking insulin resistance to other neighborhood factors including the availability of healthy food and physical activity resources (Auchincloss et al., 2008), neighborhood deprivation (Andersen et al., 2008), and neighborhood SES (Diez Roux et al., 2002). Prior work has examined fast food price and insulin resistance in a cohort of young adults (CARDIA) and found mixed results: one study found expected results with soda and pizza prices and insulin resistance (Duffey et al., 2010) but results were null in another study (Rummo et al., 2015) or significant only within relatively disadvantaged groups (middle income range, lower education, Black (Meyer et al., 2014)).

In our study, no significant association between the price of foods at neighborhood supermarkets and diabetes was observed. This may not be surprising given the chronic nature of diabetes. The lack of association with prevalence is especially unsurprising given that diabetes may have been diagnosed in the far past and an individual’s most recent price exposure may not be reflective of the exposures faced many years prior. However, an association with diabetes incidence was more likely to be observed given the concurrent timing of diagnosis and the price exposure. In fact, the association trended in the expected direction, but was not significant due to limited power and wide confidence intervals. And, while the price exposure was measured synchronously with the diagnosis of diabetes, risk factors of diabetes, such as diet and weight, tend to be chronic exposures, and not limited to the time period during which the disease was diagnosed. Thus, while we measured the price of food around the time of first diagnosis of diabetes in the incidence analysis, it would have been beneficial to know what the long-term exposures to food prices were, but these data were not available. There is little prior literature with which to compare to our results. Studies examining NHANES data suggested that higher prices of healthy and low glycemic foods were associated with higher blood sugar levels in those with diabetes (Anekwe and Rahkovsky, 2014) and without (Rashad, 2007). Compared to our dataset, the NHANES population was younger, less educated and had a higher proportion of White individuals. Market region food prices from 10 distinct food groups obtained from Nielsen Homescan consumer surveys were used in one study (Anekwe and Rahkovsky, 2014) while the other used aggregate national food prices for four high glycemic index foods from the Bureau of Labor and Statistics (Rashad, 2007). Thus, our study differentiates itself from prior work by using prices found at stores within the immediate neighborhood of individuals.

Food price is just one piece of a complex food environment, and the lack of association or weak association between food/beverage prices ratio and health may be due to the influence of other factors such as the neighborhood built environment (Auchincloss et al., 2008; Christine et al., 2015; Diez Roux and Mair, 2010), food marketing (Zimmerman, 2011) and nutrition information in retail environments (Campos et al., 2011), among others, which were not captured in this study. More research is needed to understand exactly what the causes are for varying food prices in each of the neighborhoods, including infrastructure and transportation networks; government policies such as taxes, subsidies and price floors; local food availability; and cultural influences and demand of specific products.

Intervention studies suggest that price interventions modify purchases of the targeted foods (Harnack et al., 2016; Waterlander et al., 2012), yet nutritional quality does not necessarily improve due to substitution with other untaxed less healthy items (Epstein et al., 2012; Mursu et al., 2014; Ni Mhurchu et al., 2010). Combining subsidies for healthy foods with restrictions or increased prices of unhealthy foods can partially address negative substitutions (Duhaney et al., 2015; Epstein et al., 2012; Harnack et al., 2016; Niebylski et al., 2015; Thow et al., 2014). For this reason, our study focused mostly on the price ratio of healthy-to-unhealthier foods/beverages as the main exposure. Separately modeled estimates of the adjusted association between unhealthier and healthier prices on HOMA-IR indicated lower levels of insulin resistance for both (when prices of unhealthier soda, sweets and salty snacks were more expensive and, counter-intuitively, when price of healthier food was more expensive). Prior work suggests that incentivizing purchase and use of healthier foods may have weaker impacts relative to disincentivizing the purchase of unhealthy foods (Epstein et al., 2010; Mayne et al., 2015). Associations between availability and price of healthier foods/beverages and health are complex and potentially mediated or moderated by available time, education, and skills in food preparation (Duhaney et al., 2015; Niebylski et al., 2015).

Limitations

Across the U.S., chain supermarkets represent 75% of all supermarket retailers and thus are likely a fair representation of supermarket prices in general, although not all supermarkets are included in the IRI dataset. Results of this study are only generalizable to individuals living within 3 miles of a chain supermarket in the IRI dataset. An examination of the characteristics of those included and excluded found differences with region and race/ethnicity, which were due to those excluded being largely concentrated in a single recruitment site (that had a large Hispanic sample) where supermarket data were unavailable. While prior studies have noted that lower SES areas have fewer supermarkets than other areas(Gustafson et al., 2012; Moore and Diez Roux, 2006), the SES composition of our analytic sample was roughly similar for included vs. excluded (Supplement Table 1). It is also unknown whether individuals shopped for food at the stores within 3 miles of their residences; however, 3 miles is in-line with prior research examining the average distance individuals travel to their primary supermarkets (Drewnowski et al., 2012; Fuller et al., 2013; Michimi and Wimberly, 2010; The Capitol Forum, 2014).

Because data for fresh fruits and vegetables were not available, refrigerated products were selected in order to roughly approximate costs of fresh fruit and vegetable spoilage and storage/distribution, and proxy fresh produce. Healthier food was represented by dairy (refrigerated milk, yogurt, cottage cheese), fruits and vegetables (fresh orange juice and frozen vegetables). It is unknown how the results would have changed if we had access to data on fresh fruits and vegetables. However, it is possible that results would be roughly similar for the following reasons: the price of fresh oranges correlates well with the price of refrigerated fresh orange juice (Morris, 2011) and the price of frozen vegetables was only one component of our healthy food measurement.

Additionally, data on other healthy foods such as whole grains and legumes was not available. Lastly, data were limited to branded products in order to include products that are available in many regions of the country.

CONCLUSIONS

As the prevalence of type 2 diabetes increases we must understand what factors are contributing to the development of this disease and what steps need to be taken to reverse the trend. While a large amount has been written regarding the hypothetical effect of subsidizing and taxing foods to influence consumption and the resulting health effects, very little work has examined real world data to understand how changes in price affect rates of diabetes and levels of insulin resistance. This study was the first of its kind to examine the association between neighborhood food prices with insulin resistance and diabetes for individuals in multiple areas across the U.S. While prior work has examined food prices at the metropolitan level, current literature regarding the direct association between local neighborhood food prices and diabetes is lacking, and our study provides the only currently published data examining this relationship.

HIGHLIGHTS.

• A large multi-ethnic cohort was linked to food prices at nearby stores

• Healthy food cost relative to unhealthy food was associated with insulin resistance

• No food price association was found with diabetes incidence or prevalence.