The Role of the Built Environment, Food Prices and Neighborhood Poverty in Fruit and Vegetable Consumption: An instrumental variable analysis of the Moving to Opportunity Experiment

Abstract

The food environment has been associated with fruit and vegetable consumption, however many studies utilize cross-sectional research designs. This study examined 3,473 participants in the Moving to Opportunity experiment, who were randomized into groups that affected where they lived. The relationship between the built environment, food prices and neighborhood poverty, assessed over four to seven years, on fruit or vegetable consumption was examined using instrumental variable analysis. Higher food prices and neighborhood poverty were associated with lower fruit or vegetable consumption. Policies and programs that address food prices should be implemented and evaluated for their effects on fruit and vegetable consumption.

BACKGROUND

Inadequate intake of fruits and vegetables has been associated with a number of chronic disease outcomes including increased risk of hypertension (1), coronary heart disease (1), stroke (1) and several cancers.(2) A diet low in fruit consumption was ranked as the fourth leading risk factor for global disease burden in 2010.(3) Given these significant health impacts, increasing fruit and vegetable consumption is an important public health goal.(4, 5) Yet consumption of fruits and vegetables remains very low, with only 12% and 9% of adults in the United States (US) meeting fruit and vegetable recommendations, respectively.(6)

In the past decade, efforts to improve healthy eating, including fruit and vegetable consumption, have often focused on modifying the built environment and enacting supportive policies.(7–13) These changes typically address food availability and food prices, the latter through implementation of taxes and subsidies. Both “food deserts,” areas of low socio-economic status with no or a low density of establishments that sell healthy food (14), and “food swamps,” areas of low socio-economic status with a high density of establishments selling unhealthy foods and drinks (15, 16), are associated with healthy eating.(17–19) The price of food is also a primary driver of food purchasing and consumption.(20, 21) Food prices can be affected by taxes, which result in price increases (e.g., taxes on sugar-sweetened beverages),(11) and subsidies, which result in price decreases (e.g., United States Department of Agriculture’s Food Insecurity Nutrition Incentive grant program (FINI), now the Gus Schumacher Nutrition Incentive Program (GusNIP)). GusNIP supports projects that provide incentives aimed at increasing the purchase and consumption of fruits and vegetables among Supplemental Nutrition Assistance Program (SNAP) participants at the point of purchase.(22)

Gathering evidence of the association between the food environment and healthy eating based on strong research designs is critical and has important policy implications. The US federal government launched the $400 million Healthy Food Financing Initiative, which provides one-time grants or loans for creating or supporting places like farmers markets and supermarkets in areas that are underserved. Agricultural subsidies, which largely support corn, wheat, cotton and soybean, have been supported by billions of dollars from the US federal government over the past two decades. Soda taxes, implemented in several cities, may be enacted in other cities across the US.(11) Thus, continuing to examine these issues is imperative to move the field forward and to guide policy decisions.(23)

The vast majority of the research that examines the relationship between food availability and healthy eating has been based on cross-sectional studies.(17, 23) The results of these studies are mixed.(17, 23) A number of recent evaluations of food availability have been completed as part of natural experiments, where assignment of exposed vs. unexposed is due to factors outside the investigator’s control. These studies have found that providing access in a food desert does not increase healthy eating or reduce BMI (24) and did not improve fruit and vegetable consumption.(25) Studies of food pricing and healthy eating include a large body of non-experimental designs (26), as well as a reasonable body of literature based on intervention studies.(27) The evidence for an effect between food prices and healthy eating is much more consistent, with the majority of the studies supporting the implementation of taxes and/or subsidies to promote healthy eating, including fruit and vegetable consumption.(12, 13, 26–29)

In this paper, we take advantage of a unique research experiment to examine the association between the built environment, including the food environment, food price, neighborhood poverty and fruit or vegetable consumption. Specifically, we utilize the Moving to Opportunity (MTO) experiment that randomized over 4,000 people who lived in public housing in high poverty neighborhoods into three different groups whereby the experimental group received assistance to move to census tracts with low poverty rates (i.e., less than 10% of the population in the census tract had incomes below the federal poverty threshold). The MTO experiment examined how neighborhood poverty affected economic, educational and health outcomes and the findings have greatly contributed to the field.(30, 31) However, few studies have examined what factors in the neighborhoods of the MTO participants were associated with the health outcomes. A notable exception was a study that found neighborhood poverty was more associated with long-term well-being than racial residential segregation using instrumental variable analysis.(32)

Instrumental variable (IV) analysis can help address confounding and selection issues that are often found in neighborhood effects research. Residential self-selection bias is particularly problematic in observational studies of the food environment and healthy eating. Residential self-selection bias, in this case, would occur when one’s attitudes or socio-demographic traits leads them to choose a residential location that conforms with their food-related needs or preferences, potentially leading to a spurious relationship between neighborhood factors related to the food environment and healthy eating.(33) Outside of MTO, several studies have used IV analysis to examine the effect of the food environment on healthy eating and obesity.(34, 35) One study found no relationship between availability of restaurants and obesity (34) whereas an association was found for availability of fast food restaurants and obesity in non-white rural residents but not white rural residents.(35)

In this study we utilize IV analysis to investigate the association between aspects of the built environment, including the food environment, food prices, and neighborhood poverty, on the consumption of fruits and vegetables in participants from the MTO experiment. Using the MTO randomization as an instrument leverages the experimental design to help address residential self-selection and confounding issues.(36, 37)

METHODS

Study Sample

The MTO experiment was originated by the United States Department of Housing and Urban Development (HUD) in five cities: Boston, MA, Baltimore, MD, Chicago, IL, Los Angeles, CA, and New York, NY. In order to be included in the program, participants needed to be living in public housing developments located centrally in one of the five cities in a census tract with high poverty rates (i.e., 40% or more persons in the census tract had incomes below the federal poverty threshold). Further families needed to have very low incomes and have at least one child under the age of 18.(30) All participants who entered the study agreed to be randomized into one of three groups: (1) Experimental group in which participants were offered housing vouchers that could only be used in neighborhoods where less than 10% of the population had incomes below the federal poverty threshold. Local counselors worked with participants to help them find available units in qualified neighborhoods; (2) Section 8 group in which participants were offered housing vouchers that could be used in housing where the owner of the property agrees to rent within the rules of the Section 8 program at that time. There were no geographic restrictions nor was any special assistance provided to help these participants find qualified housing; (3) The control group in which participants were not offered vouchers to move. A total of 4,608 participants (nearly all women) were randomized over a five-year period.

A survey was conducted at baseline to document characteristics of the participants. From the baseline data, 19 variables were used as covariates in the models in order to improve the precision of the models and to maintain consistency across studies (these covariates are described in the measures section below). The residential location of MTO participants, including any additional residential moves during the follow-up period, were tracked using a variety of data sources including the National Change of Address Database, credit databases and tracking data available through HUD.(38) The location data were used in the current study to generate participants’ environmental exposures. In 2002, four to seven years post-randomization, an in-person follow-up survey (N=3,526) was conducted to assess the participants’ economic, educational and health outcomes. The effective response rate of the follow-up survey was 90%. The response rate was documented as part of the main study and was derived from the response rate in the main sample (MRR) combined with the weighted response from a hard-to-reach subsample (SRR) (i.e., MRR + SRR*(1-MRR)).(30) Data were obtained from HUD and Inter-university Consortium for Political and Social Research.(39)

Measures

The main outcome for this study, daily fruit or vegetable consumption, was calculated based on self-report (i.e., “In a typical week, how many days do you eat at least some green vegetables or fruit”). The response (i.e., number of days) was dichotomized to daily fruit or vegetable consumption versus less than daily fruit or vegetable consumption, based on dietary recommendations of daily fruits and vegetables (2 and 3 servings daily, respectively).(5) Nineteen baseline covariates were included in the models including age in 5 categories (18–24; 25–34; 35–44; 45–54; 55+), race/ethnicity (black, Hispanic, other, white), gender (male or female), randomization site (Baltimore, Boston, Chicago, Los Angeles, New York), education (less than high school (HS), GED, HS graduate, enrolled in school, and missing education), marital status (never married vs. married, separated, divorced or widowed), employment status (currently working vs. not working or working only for benefits), whether or not a recipient of Aid to Families with Dependent Children/Temporary Assistance for Needy Families (AFDC/TANF), whether or not has access to a working car, whether or not a victim of crime within past six months, past moving behavior (living in neighborhood for at least 5 years, moved more than 3 times in past 5 years), reasons for moving (moving to get away from gangs or drugs, moving to have better schools) and five neighborhood perception measures (streets unsafe at night near home, chatted with neighbors at least once a week, very dissatisfied with neighborhood, no family living in neighborhood, no friends in neighborhood). For a small proportion of variables (less than 5% of all observations), missing values were imputed using conditional mean imputation. For education, which was missing more than 5%, a dummy variable was used for participants with missing education status.

Characteristics of the environment were calculated using data obtained from secondary data sources. Environmental exposures were calculated for each documented residential location using the centroid of the participant’s census block group as a proxy for residential location. Environmental attributes within various spatial buffers were calculated for each participant. Different buffer sizes were utilized for the different exposures in part because of data availability but also because the geographic area that is salient likely varies based on the relationship being examined.(40) Time duration-weighted environmental variables were created by weighting each environmental exposure by the time spent at each address. Thus, if a participant was in the study for five years and had three different addresses during the study period with different durations (e.g., three years, one year and one year) the environmental exposure at the first address would be weighted with a multiplier of 0.6 and exposures at the second and third addresses would each be weighted with a multiplier of 0.2. In addition, exposures at the participant’s last location (e.g., location at time of survey) were calculated.

Environmental variables including food prices, a built environment factor, and neighborhood poverty were created. Food prices for a variety of food categories were obtained from Information Resources, Inc. (IRI). Through agreements with retail establishments, IRI obtains store-level scanner-based food price data for hundreds of food products at grocery stores, drug stores, convenience stores, mass merchant outlets, dollar stores and club stores. A variety of foods with higher price elasticity were selected including bakery snacks, lunch meats, soda, salty snacks, cold cereal, spaghetti and sauce, packaged vegetables, frozen pizza, ice cream, processed cheese, frankfurters, oil, and pies/cakes. In order to represent the costs to purchase a variety of foods, the price per unit was averaged across these 13 categories based on food prices in 2002. Food prices within a 1.5-mile Euclidean buffer around the participant’s home locations were calculated. Since the IRI data were obtained at the zip code level, if more than one zip code was part of a single participant’s buffer, the average price within the buffer was interpolated by weighting the food price by the proportion of the buffer that fell within each zip code. Average food price was then log transformed.

A built environment factor was also utilized. This factor was previously developed from a factor analysis of seven correlated constructs.(41) These constructs included open space and high-density development based on data from the 2001 National Land Cover Database. Open space would include mostly vegetation such as lawn grass, and high-density development would include commercial and industrial land uses as well as uses such as apartment complexes. The proportion of each of these land uses were determined for the participant’s census tract. The built environment factor also included food availability (i.e., supermarkets and fast food restaurants) and commercial physical activity availability from Dun & Bradstreet (years 1994–2002). Dun & Bradstreet data are derived using multiple sources and contain millions of business records. Using SIC codes to identify supermarkets, fast food restaurants, and commercial physical activity establishments, the count of supermarkets, fast food restaurants, and commercial physical activity establishments were generated within a 2 mile (3.2 km) Euclidean buffer of each participant. Tract-level population density (from the 2000 Census) and block density, which measures the number of census blocks in a given area (from 2000 Census TIGER/Line files) were also included. The built environment factor was a weighted average of the items in the factor where weights were defined by the factor score. All variables loaded high (above .5) on the factor.(41) Items representing commercial presence loaded the highest and all these factor loadings were above .7 (i.e., supermarkets, fast food restaurants, recreational facilities, and high density development). Population density also loaded above .7. Factor loadings for open space and block density were −.60 and .53, respectively. Higher scores on the built environment factor indicate a more developed, dense area with more retail establishments. Finally, neighborhood poverty rate at the census tract level was obtained from the 2000 Census and maintained as an independent variable. The full methods for exposure calculation are described elsewhere.(41) All environmental variables were standardized to the control group defined as the deviation from the control group mean, in standard deviation units of the control group for the full follow-up sample. This study was approved by the University of Michigan Institutional Review Board.

Statistical methods

Differences between randomization groups for baseline covariates, environmental exposures and the outcome were estimated from weighted linear regressions. These comparisons were made between the assigned randomization groups regardless of compliance. Instrumental variable probit models were used to estimate the effects of the experimentally-induced variation in the exposures on the dependent variable, daily fruit or vegetable consumption. The ivprobit command in Stata with maximum likelihood estimation was used. Similar to the methods used by Ludwig et al. (32), randomization group, site, and the interaction of randomization group and site were used as instruments to address the limited precision from using randomization alone as an instrument. Thus, in the first stage, the MTO randomization group, site indicators, interactions between MTO randomization group and site indicators, and covariates were used to predict the three environmental exposures in three separate models. The predicted values of each environmental variable were then used one at a time in the second stage of analysis along with covariates to obtain associations between these variables and daily fruit or vegetable consumption (hereafter called single environmental exposure models). Environmental exposures that were significant in single environmental variable models were tested together (hereafter referred to as combined environmental exposure models).

We also estimated the models using limited information maximum likelihood estimation (LIML) and examined the Angrist-Pischke F statistic and the corresponding partial R-squared from the first stage regression in order to investigate the possibility of weak instruments for each of the environmental exposures using the ivreg2 command in Stata. Inverse probability of sampling weights were used in all analyses.(42) The analyses were conducted using Stata version 14.1.(43)

As a sensitivity analysis, these analyses were repeated for exposures calculated at the last location for the participant (i.e., location contemporaneous with the in-person survey). Also, the sample was restricted to those who had data for all of the environmental measures for at least 80% of their residential locations; thus anyone who was missing environmental exposure data on any variable for more than 20% of residential history was excluded (N=157).

Instrumental variable analyses have a number of important assumptions and three critical assumptions were considered and tested when possible.(44) Specifically, the independence or exchangeability assumption states that the instruments are independent of the factors that would affect fruit and vegetable consumption. The relevance assumption states that the instruments should have an effect on the exposures. Finally, the exclusion assumption requires that the instruments affect the outcome only through the exposures examined.

RESULTS

The sample used for the analysis included 3,473 participants with outcome data, environmental exposure data and covariates (98.5% of follow-up participants). The analytic sample (N=3,473) was not significantly different from the excluded sample (N=53) across most of the covariates. However, relative to the excluded sample, the analytic sample was slightly younger, and more likely to have graduated from high school, to have reported their educational attainment, to have never been married, and to be from the Baltimore or Los Angeles site. The outcome variable and all other covariates were not significantly different across the analytic sample and excluded sample.

Nearly the entire sample was female (98%). The majority self-reported their race/ethnicity as either black (64%) or Hispanic (29%) and were never married (62%). At baseline, three quarters of the sample were a recipient of AFDC/TANF assistance, only 27% were currently working and 16% had a working car. Almost half of the sample (46%) were between the ages of 25 and 34. More than half of the sample had graduated from high school or received a GED. About half the sample said the streets near their home were unsafe at night and 42% reported that a household member had been a victim of crime in the past six months. Almost two thirds of the sample reported having no family living in the neighborhood and 40% reported having no friends living in the neighborhood. Forty-three percent of participants reported eating fruits or vegetables daily. With the exception of education, there were no significant differences by treatment group in the baseline covariates in the analytic sample (see Table 1). The results shown in Table 1 suggest that the independence assumption is likely met (i.e., the instrumental variables are statistically independent of the factors that would influence fruit and vegetable consumption). This is expected since the vouchers were randomly assigned and the randomization was properly carried out.(32) There is also evidence that this assumption was met at the site level (the level at which randomization occurred) based on findings from site specific studies (e.g., (45, 46)).

Table 1.

Baseline Covariates and Outcome Overall and Differences by Randomization Group¹

		Overall		Treatment		Section 8		Control
		(N=3473)		(N=1428)		(N=982)		(N=1063)
Age2		M	SE	M	SE	M	SE	M	SE	F	p
	18–24	0.16	0.007	0.16	0.012	0.15	0.014	0.16	0.013	0.179	0.84
	25–34	0.46	0.010	0.45	0.015	0.46	0.018	0.46	0.017	0.111	0.90
	35–44	0.27	0.009	0.27	0.013	0.26	0.016	0.27	0.015	0.231	0.79
	45–54	0.09	0.006	0.08	0.009	0.10	0.011	0.08	0.009	0.708	0.49
Race/ethnicity2
	Hispanic	0.29	0.009	0.28	0.014	0.30	0.016	0.29	0.015	0.234	0.79
	Black	0.64	0.009	0.65	0.015	0.64	0.017	0.64	0.016	0.155	0.86
	Other	0.04	0.004	0.04	0.007	0.04	0.007	0.04	0.007	0.116	0.89
Gender2
	Male	0.02	0.002	0.01	0.002	0.02	0.005	0.02	0.005	2.149	0.12
Education2
	GED	0.18	0.008	0.17	0.012	0.18	0.014	0.19	0.014	0.528	0.59
	HS graduate	0.38	0.009	0.39	0.015	0.38	0.018	0.35	0.016	1.863	0.16
	Enrolled in school	0.16	0.007	0.16	0.011	0.16	0.012	0.15	0.013	0.079	0.92
	Missing education	0.06	0.005	0.05	0.006	0.07	0.009	0.08	0.010	3.196	0.04
Site2
	Baltimore	0.15	0.007	0.15	0.011	0.15	0.012	0.15	0.012	0.074	0.93
	Boston	0.22	0.008	0.22	0.013	0.22	0.015	0.21	0.013	0.185	0.83
	Chicago	0.23	0.008	0.23	0.012	0.23	0.017	0.22	0.015	0.148	0.86
	Los Angeles	0.16	0.007	0.16	0.011	0.15	0.012	0.16	0.011	0.413	0.66
Never married2		0.62	0.009	0.62	0.015	0.62	0.018	0.63	0.016	0.032	0.97
Currently working2		0.27	0.009	0.29	0.014	0.25	0.016	0.26	0.015	1.578	0.21
Receiving AFDC/TANF3		0.75	0.009	0.74	0.014	0.75	0.016	0.75	0.015	0.077	0.93
Has working car		0.16	0.007	0.17	0.012	0.17	0.014	0.14	0.011	2.396	0.09
Household member victim of crime in past 6 months		0.42	0.009	0.42	0.015	0.43	0.018	0.41	0.017	0.160	0.85
Living in neighborhood for at least 5 years		0.62	0.009	0.61	0.014	0.63	0.017	0.62	0.016	0.404	0.67
Chatted with neighbors at least once a week		0.53	0.010	0.53	0.015	0.50	0.018	0.55	0.017	1.839	0.16
Very dissatisfied with neighborhood		0.46	0.010	0.46	0.015	0.47	0.018	0.46	0.017	0.081	0.92
No family living in neighborhood		0.64	0.009	0.65	0.014	0.61	0.018	0.65	0.016	1.586	0.20
No friends in neighborhood		0.40	0.009	0.40	0.015	0.38	0.018	0.41	0.017	0.752	0.47
Streets near home unsafe at night		0.49	0.010	0.48	0.015	0.49	0.018	0.49	0.017	0.097	0.91
Moved more than 3 times in past 5 years		0.09	0.005	0.08	0.008	0.09	0.011	0.11	0.010	2.532	0.08
Moving to get away from gangs or drugs		0.77	0.008	0.77	0.013	0.76	0.016	0.78	0.014	0.769	0.46
Moving to have better schools		0.49	0.010	0.48	0.015	0.52	0.018	0.47	0.017	2.241	0.11
Daily Fruit or Vegetable Intake		0.43	0.010	0.45	0.015	0.44	0.018	0.41	0.017	1.450	0.23

¹Weighted using wt_totsvy; group differences were tested using weighted linear regression.

²Reference category for Age is 55+; Race is White; Gender is Female; Education is Less than High School; Site is New York; Marital Status is married, separated, divorced or widowed; and Working Status is not working or working only for benefits.

³AFDC/TANF is Aid to Families with Dependent Children/Temporary Assistance for Needy Families

The three treatment groups did experience statistically significant differences in their exposure to neighborhood poverty, built environment factor and food prices during the follow-up period. The experimental and Section 8 groups lived in areas with lower neighborhood poverty, that had less dense built environments (i.e., a lower score on the built environment factor that included food and physical activity commercial establishments, population density, block density, developed high intensity land cover, (lack of) open space) and had higher food prices relative to the control group (see Table 2). The results shown in Table 2 support the relevance assumption, namely that the instrument is associated with the exposures (further support is provided by values of the Angrist-Pischke F statistic, described below).

Table 2.

Environmental Exposures¹ Overall and Differences by Randomization Group²

	Overall		Treatment		Section 8		Control
	(N=3473)		(N=1428)		(N=982)		(N=1063)
	M	SE	M	SE	M	SE	M	SE	F	p
Built environment factor3	−0.14	0.021	−0.29	0.032	−0.13	0.040	0.05	0.036	25.28	<.001
Average food price4	0.21	0.017	0.30	0.026	0.24	0.030	0.07	0.031	16.50	<.001
Poverty rate	−0.44	0.022	−0.73	0.037	−0.57	0.034	0.06	0.037	127.38	<.001

¹Weighted using wt_totsvy and z-scored relative to the full control group

²Differences were tested using weighted linear regression

³Built environment factor is based on a factor analysis of seven exposure measures including supermarkets, fast food restaurants, commercial recreation facilities, open space, high density development, block density and population density

⁴Price per unit averaged across 13 categories (log transformed)

In single environmental exposure models, food price was significantly associated with daily fruit or vegetable consumption: those living in areas of higher food prices had a lower probability of eating fruits or vegetables every day. Neighborhood poverty was also associated with daily fruit or vegetable consumption: those living in areas with higher neighborhood poverty had a lower probability of eating fruits or vegetables every day. The built environment factor was not associated with fruit or vegetable consumption in the single environmental exposure models (see Table 3).

Table 3.

Single Environmental Exposure Instrumental Variable Regression Models, Fruit or Vegetable Intake as outcome (N=3473)

	Single environmental exposure models1						First stage statistics2
	β	SE	z	p	95% Confidence Interval		Angrist-Pischke F statistic	P R2
Built environment factor score3	0.008	0.03	0.25	0.8	−0.054	0.070	17.02	0.070
Average food price4	−0.213	0.04	−5.22	<.001	−0.293	−0.133	12.42	0.053
Neighborhood poverty	−0.158	0.05	−3.05	0.002	−0.260	−0.057	36.86	0.113

¹Results derived from IV probit models (2^nd stage results) that control for age, race/ethnicity, gender, education, site, marital status, employment, receiving AFDC/TANF, has working car, household member victim of crime, living in neighborhood for at least 5 years, chatted with neighbors at least once a week, very dissatisfied with neighborhood, no family living in neighborhood, no friends in neighborhood, streets near home unsafe at night, moved more than 3 times in past 5 years, moving to get away from gangs or drugs, and moving to have better schools.

²Results derived from IV linear regression

³Built environment factor is based on a factor analysis of seven exposure measures including supermarkets, fast food restaurant, commercial recreation facilities, open space, high density development, block density and population density

⁴Price per unit averaged across 13 categories (log transformed)

P R² = Partial R Squared

In the combined environmental exposure models both food price and neighborhood poverty remained significant in the model together: those living in areas with higher food prices and those living in areas with higher neighborhood poverty both had a lower probability of eating fruits or vegetables every day (see Table 4). A one-unit increase in food price (defined in standard deviation units of the control group) decreased the probability of fruit or vegetable consumption by 0.08. A one-unit increase in neighborhood poverty (defined in standard deviation units of the control group) decreased the probability of fruit or vegetable consumption by 0.05. To determine whether we had weak instruments for any of our analyses, a LIML analysis was run. Across all our models, the Cragg-D value far exceeded the LIML critical value, suggesting no weak instrument problems. Furthermore, the Angrist-Pischke F statistic was over 10, suggesting that randomization and site were reasonable instruments, lending further support for the relevance assumption.(47)

Table 4.

Combined Environmental Exposure Instrumental Variable Regression Models, Fruit or Vegetable Intake as outcome (N=3473)

	Model1						First Stage Statistics2
	β	Std. Err.	z	p	95 % Confidence Interval		Angrist-Pischke F stat	P R2
Average food price3	−0.201	0.041	−4.85	<0.001	−0.282	−0.120	24.1	0.05
Neighborhood poverty	−0.131	0.050	−2.62	0.009	−0.230	−0.033	17.8	0.11
Age
18–24	−0.086	0.159	−0.54	0.587	−0.398	0.225
25–34	−0.124	0.144	−0.86	0.390	−0.407	0.159
35–44	−0.046	0.145	−0.32	0.751	−0.330	0.238
45–54	−0.113	0.159	−0.71	0.478	−0.425	0.199
Race/ethnicity
Hispanic	0.010	0.172	0.06	0.953	−0.327	0.348
Black	0.249	0.170	1.47	0.143	−0.084	0.582
Other	0.432	0.208	2.08	0.038	0.025	0.839
Male	−0.200	0.204	−0.98	0.327	−0.600	0.200
Education
GED	0.018	0.074	0.25	0.804	−0.126	0.162
HS graduate	0.052	0.059	0.87	0.383	−0.064	0.168
Missing education	0.032	0.104	0.31	0.759	−0.172	0.236
Enrolled in school	−0.126	0.070	−1.81	0.070	−0.263	0.011
Never Married	−0.132	0.057	−2.31	0.021	−0.245	−0.020
Currently Working	0.012	0.064	0.18	0.854	−0.113	0.136
Receiving AFDC/TANF	0.019	0.066	0.29	0.772	−0.110	0.148
Has working car	0.142	0.070	2.01	0.044	0.004	0.280
Household member victim of crime in past 6 months	0.101	0.052	1.94	0.052	−0.001	0.203
Living in neighborhood for at least 5 years	−0.055	0.057	−0.96	0.335	−0.168	0.057
Chatted with neighbors at least once a week	0.043	0.053	0.82	0.414	−0.061	0.147
Very dissatisfied with neighborhood	0.066	0.058	1.15	0.249	−0.046	0.179
No family living in neighborhood	0.024	0.056	0.43	0.665	−0.085	0.133
No friends in neighborhood	0.006	0.056	0.11	0.913	−0.103	0.115
Streets near home unsafe at night	0.014	0.057	0.25	0.799	−0.096	0.125
Moved more than 3 times in past 5 years	0.016	0.087	0.18	0.854	−0.155	0.187
Moving to get away from gangs or drugs	−0.072	0.060	−1.18	0.237	−0.190	0.047
Moving to have better schools	−0.016	0.052	−0.31	0.760	−0.118	0.086
_cons	−0.255	0.241	−1.06	0.291	−0.727	0.218

¹Results derived from IV probit models (2^nd stage results)

²Results derived from IV linear regression

³Price per unit averaged across 13 categories (log transformed)

Reference category for Age is 55+; Race is White; Gender is Female; Education is Less than High School; Marital Status is married, separated, divorced or widowed; and Working Status is not working or working only for benefits.

P R² = Partial R Squared

In sensitivity analyses that examined exposures based on residential location at follow-up (versus duration-weighted), the results were consistent (i.e., both food price and neighborhood poverty remained significantly associated with daily fruit or vegetable consumption) when examined separately. In the combined environmental exposure models, food price remained significantly associated with daily fruit or vegetable consumption although neighborhood poverty became non-significant. However, the Angrist-Pischke F statistic was under 10, suggesting that randomization and site in the last location analyses were weak instruments.(47) In sensitivity analyses that included only those who had all the environmental exposures data at residential locations for at least 80% of their follow-up duration (N=3,308), the findings for the combined environmental exposure models were consistent (analysis not shown).

When the individual items of the built environment factor (i.e., supermarkets, fast food restaurants, commercial physical activity, open space, high density development, block density, population density) were examined separately along with neighborhood poverty, only block density was significantly associated (beta= −0.13, se 0.05, p=0.003), with those living in higher block density having a lower probability of eating fruit or vegetables every day. Supermarkets, fast food restaurants, commercial physical activity, open space, high development, and population density were not significantly associated with daily fruit or vegetable consumption. Neighborhood poverty was negatively associated with daily fruit or vegetable consumption in all the models.

DISCUSSION

This study considered a number of neighborhood factors that have been shown to be associated with fruit and vegetable consumption in the literature, but only food prices and neighborhood poverty were significantly associated with daily fruit or vegetable consumption. Supermarket availability, when considered as part of a built environment factor with other correlated variables or considered by itself, was not associated with daily fruit or vegetable consumption. The important influence of food prices on fruit and vegetable consumption in our study reflects the findings from a large body of literature including studies of food subsidies.(12, 13, 26, 27, 29) The finding that food availability is not associated with daily fruit or vegetable consumption is consistent with more recent studies that examine the association between food availability and healthy eating, including fruit and vegetable consumption, through natural experiments.(24, 25, 48)

These findings must be considered in light of the exclusion assumption in IV analyses, namely that the instruments affect the outcome only through the exposures examined. This assumption is not directly testable. When participants moved within the MTO experiment, many neighborhood factors changed and non-neighborhood factors may have changed as well. Factors across the socio-ecological model have been proposed to influence fruit and vegetable consumption in low-income households, including factors at the intrapersonal, interpersonal, organizational, community and policy levels.(49) We measured a comprehensive set of built environment exposures, as well as neighborhood poverty and food prices. When examined separately, the built environment exposure was not significantly related to fruit or vegetable consumption. However, it cannot be ruled out that the absence of an effect related to the built environment may be due to offsetting influences of other factors correlated with the built environment. For neighborhood poverty and food prices, which were significantly associated with fruit or vegetable consumption, an analysis was conducted that simultaneously included both variables. Estimates from this joint model were very similar to those obtained in separate analyses. While not definitive, this suggests that neighborhood poverty and food prices are not strongly correlated. It also suggests that potentially omitted factors correlated with neighborhood poverty and food prices are not strongly correlated with the other exposure. For example, if omitted factors correlated with neighborhood poverty, which may partly explain the neighborhood poverty effect, were correlated with food prices, then adding neighborhood poverty to the model should change the estimated effect of food prices on fruit and vegetable consumption. This is not the case in these analyses. Similarly, the estimated effect of neighborhood poverty does not change much when food prices are added to the model, suggesting that omitted factors correlated with food prices, which may partly explain the food price effect, are not correlated with neighborhood poverty. Thus, there is some evidence that neighborhood poverty and food prices have some substantial independent effects on fruit and vegetable consumption. However, this is a heuristic argument, which while supported by some evidence, is not definitive. In the end, the IV estimates represent the effect of the exposure and other omitted factors correlated with the exposure (conditional on inclusion of the other exposures). Thus systematic changes in other exposures across the socio-ecological model cannot be ruled out as the reason for the differences seen in the results that are attributed to changes in neighborhood poverty and food prices.

This study suggests that housing mobility programs may have the ability to affect health behavior through changes in exposures that result from the move. The health effects of the MTO experiment have been well documented.(30, 31) However few studies have examined the potential mechanisms that were associated with these health effects, perhaps because MTO was not designed to examine the pathways through which health was affected.(50) This study found significant effects for some proposed pathways and null associations for others. The findings are consistent with a proposed model for how mobility programs can affect health, namely through housing stability, providing safety and quality of the home, alleviating financial burden of high-cost housing, and impacts from the social and built environment where people live.(51)

While significant efforts have been undertaken to increase food availability, recent natural experiments suggest that living in a food desert may not be a primary driver of healthy eating.(24, 25) Alternatively it has been suggested that food availability “may be necessary but not sufficient” for behavioral change.(52) Importantly, this study did not assess areas that had limited food availability as most areas had access to food retailers.(41) Further, our IV analysis examined a built environment factor and not food availability alone, but a complimentary analysis did not find a relationship between supermarket availability and daily fruit or vegetable consumption when examined by itself. Future interventions directed at increasing food availability through federal grants and loans should consider including requirements for addressing the affordability of healthy foods as well.

The Farm Bill, which authorizes major food and agricultural programs provides an infrastructure to address food affordability, but, relative to its budget (~$86 billion a year), its potential to improve health is not maximized.(10, 53, 54) For example, GusNIP, formerly known as FINI, provides grants to help low-income people access and purchase fruits and vegetables through subsidies. This program is effective in increasing fruit and vegetable consumption (12, 13, 29) and in the 2018 Farm Bill, the financial commitment to GusNIP program was increased to 250 million (from 100 million) over the five years. But given that there are 42 million SNAP eligible recipients, the benefit equates to just over a dollar a year per person.(53) Addressing the purchase price of fruits and vegetables is important as these products have seen larger price increases than other commodities (e.g., corn, soybeans, sugar).(55)

Our finding that lower food prices were associated with higher fruit or vegetable consumption is consistent with other studies.(26–28) For example, a recent intervention in subsidized housing complexes found that offering discounts at fresh fruit and vegetable markets on a regular basis along with educational components was successful in increasing fruit and vegetable consumption.(56) In addition, our study found that those in the experimental group who lived in neighborhoods with less poverty faced higher food prices than the control group. This finding is consistent with a recent systematic review that investigated the price of foods by neighborhood income levels and found that prices were highest in high income areas (57), though historical research has found people in lower income neighborhoods pay slightly more.(58) These findings together may seem inconsistent with other studies showing the MTO experimental group had lower obesity levels at follow-up (30) since consumption of fruits and vegetables generally is associated with reduced risk of weight gain and consuming fruit and vegetables has a small impact on weight loss.(59, 60) However obesity is affected by multiple factors including total caloric intake, energy expenditure and genetics. In addition, smaller effect sizes and limited follow-up time could play a role in these seemingly contradictory findings. Our study also suggests that reducing neighborhood poverty could have effects on fruit and vegetable consumption. Future research is needed to determine which policy changes can improve fruit and vegetable consumption in low-income populations either by directly reducing neighborhood poverty or by addressing the mechanisms through which neighborhood poverty exerts its effect.

Our study is not without limitations. Our outcome (i.e., daily fruit or vegetable consumption) was based on a self-reported item, as has been the case with some other studies of the food environment. More comprehensive longer dietary instruments are ideal, although brief dietary measures have been found to be useful.(61) Food price data from grocery stores with sales of less than $2 million annually were not available for inclusion in the food price calculations. However nationally, almost 90% of shopper’s primary grocery store was a supermarket or supercenter regardless of income in the US.(62) Still, in a systematic review of food availability and fruit and vegetable consumption, consistent positive effects on fruit and vegetable consumption were seen from the introduction of mobile produce carts but not supermarkets.(63) Our study did not capture mobile produce carts. Additionally, data from Walmart was not available and the data were aggregated to the zip code level. Nationally, the average distance to one’s primary grocery store is almost 4 miles, thus the larger geographic area used in this study would be reasonable for most shoppers.(62)

Moreover, the food price index included a variety of food items, measuring the affordability of food, but did not explicitly include a fruit price indicator. The prices of foods that were based on weights, such as fresh fruits and vegetables, were not available to be included in the study. The overall affordability of food is postulated to affect fruit and vegetable consumption because when food budgets get restricted (due to low income, high food prices, etc.), people are less likely to purchase healthy foods such as fruits and vegetables.(64) However, this study was not able to determine whether addressing food prices generally versus focusing on specific foods such as fruits and vegetables would be more effective in increasing fruit and vegetable consumption. Other limitations include that time-varying exposure data was not available for all exposures and the study considers only the area near where the participants live although other spatial areas may be relevant. The buffer sizes used are not the same for all variables, which may make it more difficult to directly compare the importance of the various environmental factors. It is also possible the exposures have synergistic effects but interaction effects among the exposure variables were not tested.

This study is significant in that it examined both food pricing and the built environment (including food availability) simultaneously; studies doing so are limited in number.(65) Our study only examined one dimension of food access (i.e. supermarket availability), whereas multiple dimensions of food access (e.g., acceptability) have been identified in the literature.(17) In a systematic review, it was postulated that the introduction of new supermarkets may not influence fruit and vegetable consumption because it mostly addresses availability, but not necessarily other factors like accommodation and acceptability.(63) This study includes a unique yet important population, namely those who had lived in public housing (a particularly vulnerable population). Participants were very low-income and mostly self-identified their race/ethnicity as either black or Hispanic, a group for whom access to healthy foods has been an important policy concern. It is possible that our findings on fruit or vegetable consumption may not be generalizable to other populations or food categories. Notably, the MTO experiment did not include any US rural locations or sites in the Deep South. The results of this study are based on an experiment during which people moved, and moving may have had other effects in addition to a change in the participant’s environment (e.g., change in their social network). Modifying neighborhood factors among individuals that stay in place may or may not have similar effects.

This study is distinctive in its method of exploring the association between the built environment, neighborhood poverty and food prices on fruit and vegetable consumption. The study takes advantage of a large multi-site randomized control trial in which variation due to randomization and site resulted in groups of participants living in markedly different environments in terms of neighborhood poverty and built environment factors.(41, 66, 67) Randomization is considered an ideal instrument for an IV analysis.(68) Many studies have cited the need to use more robust methods to examine the relationship between the local environment and food consumption.(23) This study adds an additional and unique piece of evidence on the importance of food prices to encourage fruit and vegetable consumption.

CONCLUSION

This study utilized a unique methodology that capitalized on a randomized study across five US cities to investigate whether differential neighborhood factors were associated with daily fruit and vegetable consumption. Food prices and neighborhood poverty were inversely associated with daily fruit or vegetable consumption in this low-income, mostly female racial/ethnic minority population. The study suggests housing policy can have important effects on health, potentially through changes in neighborhood factors. Subsidies or other methods to reduce food prices should be further evaluated as potential mechanisms to improve health in low-income populations.

Using instrumental variable analysis, neighborhood features and diet were examined.

Higher neighborhood food prices were associated with lower fruit or vegetable intake.

Higher neighborhood poverty was associated with lower fruit or vegetable intake.

A factor of built environment features was not related to fruit or vegetable intake.