GIS-Based Home Neighborhood Food Outlet Counts, Street Connectivity, and Frequency of Use of Neighborhood Restaurants and Food Stores

Abstract

Researchers have linked neighborhood food availability to the overall frequency of using food outlets without noting if those outlets were within or outside of participants’ neighborhoods. We aimed to examine the association of neighborhood restaurant and food store availability with frequency of use of neighborhood food outlets, and whether such an association was modified by neighborhood street connectivity using a large and diverse population-based cohort of middle-aged U.S. adults. We used self-reported frequency of use of fast food restaurants, sit-down restaurants, and grocery stores in respondents’ home neighborhoods using data from the Coronary Artery Risk Development in Young Adults study Year 20 exam in 2005–2006 (n = 2860; Birmingham, AL; Chicago, IL; Minneapolis, MN; and Oakland, CA) and geographically matched GIS-measured neighborhood-level food resource, street, and U.S. Census data. We used mixed-effects logistic regression to examine the associations of the GIS-measured count of neighborhood fast food restaurants, sit-down restaurants, and grocery stores with self-reported frequency of using neighborhood restaurants and food stores and whether such associations differed by GIS-measured neighborhood street connectivity among those who perceived at least one such food outlet. In multivariate analyses, we observed a positive association between the GIS-measured count of neighborhood sit-down restaurants (OR = 1.02, 95% CI 1.00–1.04) and the self-reported frequency of using neighborhood sit-down restaurants. We observed no statistically significant association between GIS-measured count of neighborhood fast food restaurants and self-reported frequency of using neighborhood fast food restaurants, nor did we observe a statistically significant association between GIS-measured count of neighborhood grocery stores and self-reported frequency of using neighborhood grocery stores. We observed inverse associations between GIS-measured neighborhood street connectivity and the self-reported frequencies of using neighborhood fast food restaurants (OR = 0.42, 95% CI 0.26–0.68) and grocery stores (OR = − 2.26, 95% CI − 4.52 to − 0.01). Neighborhood street connectivity did not modify the association between GIS-measured neighborhood restaurant and food store count and the self-reported frequency of using neighborhood restaurants and food stores. Our findings suggest that, for those who perceived at least one sit-down restaurant in their neighborhood, individuals who have more GIS-measured sit-down restaurants in their neighborhoods reported more frequent use of sit-down restaurants than those whose neighborhoods contain fewer such restaurants. Our results also suggest that, for those who perceived at least one fast food restaurant in their neighborhood, individuals who live in neighborhoods with greater GIS-measured street connectivity reported less use of neighborhood fast food restaurants than those who live in neighborhoods with less street connectivity. The count of neighborhood sit-down restaurants and the connectivity of neighborhood street networks appear important in understanding the use of neighborhood food resources.

Electronic supplementary material

The online version of this article (10.1007/s11524-019-00412-x) contains supplementary material, which is available to authorized users.

Background

A growing body of literature is devoted to how individuals interact with the food environment [1], including a focus on the relationship between neighborhood food availability and frequency of use of restaurants and food stores. How frequently people use food outlets provides information about how people engage with their surrounding food environments [2]. Frequent away-from-home eating has been suggested to associate with lower quality diet and higher likelihood of obesity [3, 4]. Studies of neighborhood restaurant/food store availability in relation to frequency of use have produced mixed findings. Several have provided support for a positive association [3, 5, 6], whereas others have not [7–10]. However, previous studies have focused merely on the frequency of restaurant or food store use, without measuring whether such restaurants or food stores were within individuals’ perceived neighborhoods. Thus, there are fundamental gaps in understanding how people interact with their surrounding neighborhood food environments. Neighborhood food availability might be weakly associated with overall frequency of use if most participants used restaurants/food stores in other settings, such as school or workplace neighborhoods [7, 11]. Thus, identifying how neighborhood food outlets are perceived and used by residents is an important step toward a better understanding of complex dietary behaviors and in developing neighborhood-based environmental strategies to increase healthy diet behaviors. In addition, few population-based studies have explicitly addressed whether street connectivity plays a role in the association between neighborhood restaurant/food store availability and use of restaurants/food stores [12]. Previous studies have indicated that people in highly connected (or walkable) neighborhoods were more likely to participate in transport walking (e.g., walking to work or a grocery store) than those in poorly connected neighborhoods [13, 14], as connectivity affects the directness of travel and the number of alternative routes [15]. Street connectivity, reflecting the convenience of accessing food resources, might modify the strength of the relationship between neighborhood restaurant/food store availability and frequency of using those food resources. To address these gaps in the literature, we used data on the self-reported frequency of using fast food restaurants, sit-down restaurants, and grocery stores in participants’ home neighborhoods from the Coronary Artery Risk Development in Young Adults (CARDIA) study Year 20 exam in 2005–2006. Using geographically matched restaurant and food store locations and street information, we estimated the associations between the geographic information system (GIS)-measured numbers of neighborhood fast food restaurants, sit-down restaurants, and grocery stores and self-reported frequencies of using fast food restaurants, sit-down restaurants, and grocery stores in respondents’ neighborhoods among those who reported that they had at least one type of restaurant or food store in their neighborhood. To address the potential role of neighborhood fast food restaurants, sit-down restaurants, and grocery stores in differences in street connectivity, we also examined how such associations differed by GIS-measured neighborhood street connectivity.

Methods

CARDIA is a prospective cohort study examining the development of cardiovascular disease and its risk factors in 5115 white or black U.S. adults aged 18 to 30 years at baseline in 1985–1986 who were recruited to attain an approximately balanced representation of age (18–24 or 25–30 years at baseline), race (white or black), gender, and education (≤ high school versus > high school) from four metropolitan study centers (Birmingham, AL; Chicago, IL; Minneapolis, MN; and Oakland, CA) [16]. We used data on 3549 participants, 72% of the surviving cohort, at the Year 20 exam in 2005–2006, in which CARDIA surveyed use of neighborhood food outlets through a neighborhood environment questionnaire. We included only the participants who completed the neighborhood environment questionnaire (n = 3539, Fig. 1). We excluded n = 576 participants who reported that there were none of the three types of food outlets (fast food restaurants, sit-down restaurants, or grocery stores) in their neighborhood, since our focus was on frequency of use of these facilities. In addition, we excluded n = 42 participants due to invalid response patterns (i.e., participant reported no grocery store in his neighborhood in the first question, but stated use of a neighborhood grocery store in the second question). In addition, we excluded n = 61 participants due to missing covariate information, resulting in a final sample of n = 2860. We used a GIS to link the three types of food outlets, street network data, and U.S. Census data to CARDIA participants’ residential addresses.

Fig. 1.

Inclusion and exclusion criteria used for the analyses of the self-reported frequencies of use of fast food restaurants, sit-down restaurants, and grocery stores among those who perceived the food establishments, CARDIA, 2005–2006

Outcome: Self-Reported Frequencies of Using Neighborhood Fast Food Restaurants, Sit-Down Restaurants, and Grocery Stores among Participants Who Perceived at Least One Type of Restaurant or Food Store in Their Neighborhood

We used the CARDIA neighborhood environment questionnaire, which elicited information on the frequency of using the food outlets in each participant’s neighborhood with a series of questions for each type (fast food restaurants, sit-down restaurants, or grocery stores). These sets of survey questions are summarized in Table 1. If the participant answered “yes” to the first question (“Is there a fast food restaurant/sit-down restaurant/grocery store in your neighborhood (q1)?”), they were then asked a second question (“In the past year, did you use a fast food restaurant/sit-down restaurant/grocery store in your neighborhood (q2)?”; if the answer to this question was “yes,” then the participant was asked a third question (“How often did you use a fast food restaurant/sit-down restaurant/grocery store (q3)?”), with response options of “more than once per week,” “weekly,” “monthly,” and “yearly.” The final analytic samples comprised 2007, 2122, and 2191 participants who perceived at least one type of restaurant or food store in their neighborhood and explicitly reported the frequency of using such outlets, respectively.

Table 1.

Characteristics of the CARDIA study sample at Year 20 exam in 2005–2006

Characteristics	N = 2860
Perceived presence of at least one outcome food outlet
Is there a fast food restaurant in your neighborhood, %
Yes	69.8
Is there a sit-down restaurant in your neighborhood, %
Yes	74.1
Is there a grocery store in your neighborhood, %
Yes	76.5
Self-reported frequency of use of neighborhood food outlets
Fast food restaurant, %
Never use	13.2
Once per year	5.9
Monthly	28.5
Weekly	15.1
More than once per week	7.5
Did not perceive a fast food restaurant in the neighborhooda	29.8
Sit-down restaurant, %
Never use	15.0
Once per year	10.7
Monthly	32.6
Weekly	11.5
More than once per week	4.3
Did not perceive a sit-down restaurant in the neighborhooda	25.8
Grocery store, %
Never use	1.9
Once per year	1.7
Monthly	12.3
Weekly	35.3
More than once per week	25.4
Did not perceive a grocery store in the neighborhooda	23.4
GIS-measured neighborhood food outlet count measures
Number of food outlets within the 1-km buffer, count, mean (SD)
Fast food restaurantb	6.2 (13.1)
Sit-down restaurantc	5.6 (15.9)
Grocery stored	3.9 (8.3)
Supermarkets and convenience storese	4.0 (5.2)
GIS-measured neighborhood street connectivity measure
Link-to-node ratiof within the 1-km buffer, mean (SD)	1.6 (0.2)
Other GIS-measured neighborhood environmental measures
Population density within the 1-km buffer: 1000 person/km2, mean (SD)	2.4 (3.1)
Neighborhood SES deprivation based on participants’ home census tract, mean (SD)	0.4 (1.1)
Density of vacant housing units in participants’ home census block group: 100 housing units/km2, mean (SD)	1.3 (3.8)
Measures of self-reported individual-level sociodemographic and reasons to moving to/staying in the neighborhood
Education > high school, %	62.4
Family income, 1000$, mean (SD)	74.3 (41.0)
Black, %	44.3
Female, %	56.5
Household size, person, mean (SD)	3.0 (1.5)
Age, years, mean (SD)	45.2 (3.6)
Employed, %	82.4
Married, %	56.8
Neighborhood food environment (grocery stores, restaurants, corner stores) is one of the most important reasons of moving to/staying in the neighborhood, %	25.2
Study center
Birmingham	24.1
Chicago	23.3
Minneapolis	24.4
Oakland	28.3

SD standard deviation, GIS geographic information system, km kilometer

^aParticipants who were ineligible to answer frequency of use questions because they did not report fast food or sit-down restaurant or grocery store in their neighborhood

^bFast food restaurants defined by SIC codes 58120300–58120315 and 58120600–58120602

^cSit-down restaurants defined by SIC codes 58120100–58120111, 58120113–58120114, 58120115–58120117, 58120500–58120502, 58120800–58120802, 58120700–58120702, and 58129904

^dGrocery stores defined by SIC codes 54110000, 54119900, 54119903–54119905, and 54990202

^eSupermarkets and convenience stores defined by SIC codes 53310000, 54110100–54110105, 54110200–54110202, 55410000, and 55419900–55419901

^fLink-to-node ratio equaled to the number of links divided by the number of nodes within the 1-km buffer

Exposures: GIS-Measured Neighborhood Fast Food Restaurants, Sit-Down Restaurants, and Grocery Stores and Neighborhood Street Connectivity

The CARDIA neighborhood environment questionnaire defined neighborhood as an area within a 10- to 15-min walk from the participant’s home. We therefore used the 1-km Euclidean buffer around each participant’s geocoded residential location to operationalize neighborhood because the distance that can be traveled from home by walking in 10–15 min is 0.7–1.1 km (given a walking speed of approximately 4.4 km per hour) [17]. We defined number of neighborhood fast food restaurants, sit-down restaurants, and grocery stores within the 1-km buffer. We calculated the numbers of fast food restaurants, sit-down restaurants, and grocery stores by geocoding the food outlet records retrieved from Dun and Bradstreet (D&B) in 2006, a commercial dataset of U.S. business records. We defined food outlets according to their primary 8-digit Standard Industrial Classification (SIC) codes. We used the link-to-node ratio (number of links divided by the number of nodes) within the 1-km buffer to describe neighborhood street connectivity. A higher ratio indicates higher connectivity [18]. We obtained the road network maps (interstate highways and access ramps excluded) from ESRI Data and Maps StreetMap North America for 2010. Additional information on our food outlet classification and the detailed method we used to produce exposures can be found in Appendices 1 and 2 in the Electronic Supplementary Material.

Covariates

We used self-reported individual-level sociodemographic information collected at the baseline exam in 1985–1986 using the CARDIA standardized questionnaire, including race (black, white), gender, and age. We used self-reported individual-level sociodemographic and other information collected at the Year 20 exam in 2005–2006 using the CARDIA sociodemographic and neighborhood environment questionnaires, including current educational attainment, family income, household size, employment status, marital status, and reasons for moving to or staying in the current neighborhood. We used the information on the highest degree obtained to create a dichotomous indicator of whether the participant received a degree beyond high school (≤ high school, > high school). Participants reported their combined family income as falling into one of nine categories (e.g., $5000–11,999/year), and we created a measure in U.S. dollars as the midpoint of the selected income category. We defined household size as the number of individuals living in the family. Employment status consisted of two categories: working full- or part-time and unemployed. We created a dichotomous indicator of marital status (married, not married), classifying participants as married if they reported being currently married or living with someone in a marriage-like relationship. We used the information on reasons for moving to or staying in the participant’s current neighborhood to create a dichotomous indicator of whether the neighborhood food environment (grocery stores, restaurants, corner stores) was one of the most important reasons for moving to or staying in the neighborhood. Additionally, we used five GIS-measured neighborhood-level variables to account for the contextual influences of other neighborhood-level built environment and sociodemographic characteristics, namely population density within the 1-km buffer, socioeconomic status (SES) deprivation factor score (defined as the first factor score from a principal components analysis of four census indicators of socioeconomic status, higher value indicated greater neighborhood deprivation) in a participant’s home census tract, density of vacant housing units in a participant’s home census block group, total number of neighborhood fast food restaurants and sit-down restaurants within the 1-km buffer, and total number of neighborhood supermarkets and convenience stores within the 1-km buffer. The detailed methods we used to produce these covariates can be found in Appendix 2 in the Electronic Supplementary Material.

Statistical Analysis

We used a separate set of models for each type of outlet (fast food restaurants, sit-down restaurants, grocery stores) to predict the self-reported frequency of use of each. Not all participants perceived all three types of neighborhood food outlets; therefore, reported frequency of use for outlets is missing where a given participant did not report presence of a given outlet. However, it is likely that participants reporting that they did not have a neighborhood fast food restaurant may be different from participants who reported at least a neighborhood fast food restaurant, leading to bias in the estimated coefficients of frequency of use. To address this, we used propensity scoring for two steps: first, we estimated the probability of whether or not a participant reported having at least one neighborhood fast food restaurant as a function of GIS-measured count of neighborhood fast food restaurants, neighborhood street connectivity, and covariates; and second, we estimated frequency of use of fast food restaurants in the neighborhood as a function of GIS-measured count of neighborhood fast food restaurants, neighborhood street connectivity, and interaction term between GIS-measured count of neighborhood fast food restaurants and neighborhood street connectivity, including the estimated step 1 probability as a covariate. The perceived presence of at least one neighborhood fast food restaurant helped us to obtain the final study sample in the frequency of use model, while the GIS-measured measure of neighborhood fast food restaurants served as the exposure in the frequency of use model. We developed sit-down restaurant and grocery store models in the same manner.

For each food outlet type, the first equation is a random intercept mixed-effects logistic regression that estimates the probability of perceiving at least one neighborhood food outlet across the full sample (n = 2860). The second equation is a random intercept mixed-effects generalized ordered logistic regression, also known as a mixed-effects proportional odds model that estimated the participant’s self-reported frequency of using the food outlet for participants who reported perceiving at least one outcome food outlet in their neighborhood. We adjusted the coefficients from the second equation by adding the predicted probability of perceiving at least one outcome food outlet from the first equation into the model as a covariate in the second equation.

We further adjusted for the following self-reported individual-level covariates in all models, as suggested by previous studies: family income [5–7, 19], race [5–7, 19], gender [6, 10, 20], age [5, 20], employment status [7, 21], and whether neighborhood food environment (grocery stores, restaurants, corner stores) was one of the most important reasons for moving to/staying in the neighborhood. We also adjusted for the following GIS-measured neighborhood-level covariates in all models: population density within the 1-km buffer [5, 20, 21], SES deprivation factor score in a participant’s home census tract [20], and density of vacant housing units in a participant’s home census block group. In addition to the predicted probability of perceiving at least one food outlet, we adjusted for three self-reported individual-level covariates (educational attainment [5], household size [5], and marital status [5]) in the frequency of use equations since we assumed that these covariates were associated with the self-reported frequency of use but not necessarily associated with perceiving at least one neighborhood food outlet. To adjust for the presence of other complementary neighborhood food outlets, we controlled for other restaurants and other food shopping sources in each of the models (e.g., we controlled for the number of GIS-measured sit-down restaurants in the fast food restaurant model).

For the possible modification effect of neighborhood street connectivity, we included an interaction term between the GIS-measured number of each type of food outlet and the link-to-node ratio measure. We ran the collinearity diagnostics and mean-centered exposures to address potential collinearity concerns. We used the collin command in STATA 14 (StataCorp, College Station, TX) to run the collinearity diagnostics. Additional information about collinearity diagnostics can be found in Appendix 3 in the Electronic Supplementary Material.

We tested the proportional odds assumption for the random intercept mixed-effects ordinal frequencies of use of neighborhood fast food restaurants, sit-down restaurants, and grocery stores models to ensure the relationship between all pairs of outcome groups was the same and that there was only one set of coefficients. The results of the likelihood ratio chi-squared test suggested some variables in the ordinal frequency of use models violated the assumption, which were shown as bolded text in Tables A3-4, A3-5, and A3-6 in Appendix 3 in the Electronic Supplementary Material. We relaxed the proportional odds assumptions for the variables that violated the assumption. For example, in the 1-km buffer models, educational attainment and household size in the ordinal frequency of use of neighborhood fast food restaurants model violated the assumption; therefore, we relaxed the assumption for educational attainment and household size, which obtained four coefficients which indicated the effects of switching from no use to use yearly, from use yearly to use monthly, from use monthly to use weekly, and from use weekly to use more than once per week, respectively. We used Stata 14.0 for regression analyses (xtmelogit for mixed-effects logistic regression, and gllamm, threshold, and test for testing the proportional odds assumption and mixed-effects generalized ordered logistic regression).

Sensitivity Testing

Since participants may misestimate the geographic boundaries achievable by a 10- to 15-min walk from home [22], we tested whether our results were sensitive with respect to distance by also testing a 3-km buffer. To investigate a potential source of selection bias, we examined the similarity of individual-level and neighborhood-level characteristics between individuals who perceived at least one outcome food outlet versus those who did not perceive any, using kernel density plots of each to compare similarity between the two groups. The overlap of the plotted curves between these two groups reflected the balance of neighborhood-level and individual-level characteristics between these two groups. Additional information about the balance tests can be found in Appendix 3 in the Electronic Supplementary Material.

Results

More than two thirds of the participants reported that they perceived at least one neighborhood fast food restaurant, sit-down restaurant, or grocery store. More than half of the participants who perceived that they had these outlets in their neighborhoods reported that they used the neighborhood fast food restaurants and sit-down restaurants while around three fourths reported that they used the neighborhood grocery stores (Table 1).

Associations between GIS-Measured Neighborhood Food Count, Neighborhood Street Connectivity, and Perceiving at Least One Neighborhood Food Outlet Type

After adjusting for covariates, we observed that (1) the probability of perceiving at least one neighborhood fast food restaurant was positively associated with the GIS-measured number of fast food restaurants within the 1-km buffer and inversely associated with the link-to-node ratio within the same buffer (Table 2); (2) the probability of perceiving at least one neighborhood sit-down restaurant was positively associated with the GIS-measured number of sit-down restaurants within the same buffer (Table 3); and (3) there was no association between the probability of perceiving at least one neighborhood grocery store and the GIS-measured number of grocery stores within the same buffer (Table 4).

Table 2.

Associations between GIS-measured neighborhood fast food restaurant count, neighborhood street connectivity, and self-reported frequency of use of neighborhood fast food restaurants among those who perceived at least one neighborhood fast food restaurant, CARDIA study sample at the Year 20 exam in 2005–2006

GIS-measured exposure	First-step model: perceiving at least one neighborhood fast food restauranta (full sample)	Second-step model: self-reported frequency of use of neighborhood fast food restaurantsb (restricted sample)
	OR (95% CI) (n = 2860)	OR (95% CI) (n = 2007)
Number of fast food restaurantsc	1.05 (1.03, 1.07)	1.01 (0.99, 1.03)
Link-to-node ratiod	0.61 (0.38, 0.97)	0.42 (0.26, 0.68)

Italics indicates statistically significant at p < 0.05

OR odds ratio, CI confidence intervals, GIS geographic information system

^aEstimated coefficients of perceiving at least one neighborhood fast food restaurant were derived from the random intercept mixed-effects logistic regression model of GIS-measured number of fast food restaurants and link-to-node ratio within the 1-km buffer across the full sample, while adjusting for neighborhood SES deprivation, neighborhood population density, density of vacant housing units in the neighborhood, family income, race, gender, age, employment status, and whether neighborhood food environment is one of the most important reasons for moving to/staying in the neighborhood

^bEstimated coefficients of self-reported frequency of use of neighborhood fast food restaurants within the respondents who perceived at least one neighborhood fast food restaurant were derived from the random intercept mixed-effects generalized ordered logistic regression model of GIS-measured number of fast food restaurant and link-to-node ratio within the 1-km buffer, while adjusting for total sit-down restaurants, neighborhood SES deprivation, neighborhood population density, density of vacant housing units in the neighborhood, family income, race, gender, age, educational attainment, employment status, household size, marriage status, and whether neighborhood food environment is one of the most important reasons for moving to/staying in the neighborhood, and the estimated probability of perceiving at least one neighborhood fast food restaurant from the first-step model

^cFast food restaurants defined by SIC codes 58120300–58120315 and 58120600–58120602; calculated within the 1-km buffer

^dLink-to-node ratio equaled to the number of links divided by the number of nodes within the 1-km buffer

Table 3.

Associations between GIS-measured neighborhood sit-down restaurant count, neighborhood street connectivity, and self-reported frequency of use of neighborhood sit-down restaurants among those who perceived at least one neighborhood sit-down restaurant, CARDIA study sample at the Year 20 exam in 2005–2006

GIS-measured exposure	First-step model: perceiving at least one neighborhood sit-down restauranta (full sample)	Second-step model: self-reported frequency of use of neighborhood sit-down restaurantsb (restricted sample)
	OR (95% CI) (n = 2860)	OR (95% CI) (n = 2122)
Number of sit-down restaurantsc	1.05 (1.02, 1.07)	1.02 (1.00, 1.04)
Link-to-node ratiod	1.53 (0.93, 2.53)	0.84 (0.52, 1.35)

Italics indicates statistically significant at p < 0.05

OR odds ratio, CI confidence intervals, GIS geographic information system

^aEstimated coefficients of perceiving at least one neighborhood sit-down restaurant were derived from the random intercept mixed-effects logistic regression model of GIS-measured number of sit-down restaurants and link-to-node ratio within the 1-km buffer across the full sample, while adjusting for neighborhood SES deprivation, neighborhood population density, density of vacant housing units in the neighborhood, family income, race, gender, age, employment status, and whether neighborhood food environment is one of the most important reasons for moving to/staying in the neighborhood

^bEstimated coefficients of self-reported frequency of use of neighborhood sit-down restaurants within the respondents who perceived at least one neighborhood sit-down restaurant were derived from the random intercept mixed-effects generalized ordered logistic regression model of GIS-measured number of sit-down restaurant and link-to-node ratio within the 1-km buffer, while adjusting for total fast food restaurants, neighborhood SES deprivation, neighborhood population density, density of vacant housing units in the neighborhood, family income, race, gender, age, educational attainment, employment status, household size, marriage status, and whether neighborhood food environment is one of the most important reasons for moving to/staying in the neighborhood, and the estimated probability of perceiving at least one neighborhood sit-down restaurant from the first-step model

^cSit-down restaurants defined by SIC codes 58120100–58120111, 58120113–58120114, 58120115–58120117, 58120500–58120502, 58120800–58120802, 58120700–58120702, and 58129904; calculated within the 1-km buffer

^dLink-to-node ratio equaled to the number of links divided by the number of nodes within the 1-km buffer

Table 4.

Associations between GIS-measured neighborhood grocery store count, neighborhood street connectivity, and self-reported frequency of use of neighborhood grocery stores among those who perceived at least one neighborhood grocery store, CARDIA study sample at the Year 20 exam in 2005–2006

GIS-measured exposure	First-step model: perceiving at least one neighborhood grocery storea (full sample)	Second-step model: self-reported frequency of use of neighborhood grocery storesb (restricted sample)
	OR (95% CI) (n = 2860)	OR (95% CI) (n = 2191)
Number of grocery storesc	0.98 (0.95, 1.01)	− 0.13 (− 0.32, 0.06)d
		0.03 (− 0.02, 0.09)d
		0.01 (− 0.01, 0.04)d
		0.02 (− 0.00, 0.04)d
Link-to-node ratioe	0.81 (0.49, 1.35)	− 2.26 (− 4.52, − 0.01)d
		0.59 (− 0.54, 1.72)d
		0.14 (− 0.44, 0.72)d
		− 0.03 (− 0.52, 0.47)d

Italics indicates statistically significant at p < 0.05

OR odds ratio, CI confidence intervals, GIS geographic information system

^aEstimated coefficients of perceiving at least one neighborhood grocery store were derived from the random intercept mixed-effects logistic regression model of GIS-measured number of grocery store and link-to-node ratio within the 1-km buffer across the full sample, while adjusting for neighborhood SES deprivation, neighborhood population density, density of vacant housing units in the neighborhood, family income, race, gender, age, employment status, and whether neighborhood food environment is one of the most important reasons for moving to/staying in the neighborhood

^bEstimated coefficients of self-reported frequency of use of neighborhood grocery stores within the respondents who perceived at least one neighborhood grocery store were derived from the random intercept mixed-effects generalized ordered logistic regression model of GIS-measured number of grocery stores and link-to-node ratio within the 1-km buffer, while adjusting for total supermarkets and convenience stores, neighborhood SES deprivation, neighborhood population density, density of vacant housing units in the neighborhood, family income, race, gender, age, educational attainment, employment status, household size, marriage status, and whether neighborhood food environment is one of the most important reasons for moving to/staying in the neighborhood, and the estimated probability of perceiving at least one neighborhood grocery store from the first-step model

^cGrocery stores defined by SIC codes 54110000, 54119900, 54119903–54119905, and 54990202; calculated within the 1-km buffer

^dThe number of grocery stores and link-to-node ratio within the 1-km buffer had four coefficients because the variables violated the proportional odds assumption; the four coefficients indicated the effects of the exposures switching from no use to use yearly, from use yearly to use monthly, from use monthly to use weekly, and from use weekly to use more than once per week, respectively

^eLink-to-node ratio equaled to the number of links divided by the number of nodes within the 1-km buffer

Associations between GIS-Measured Neighborhood Food Count, Neighborhood Street Connectivity, and Self-Reported Frequencies of Use of Neighborhood Fast Food Restaurants, Sit-Down Restaurants, and Grocery Stores among Those Who Perceived Such Food Outlets

As none of the interaction terms was significant in the fast food restaurant, sit-down restaurant, or the grocery store models, we did not include the interaction term in the final model. After adjusting for covariates, we observed that (1) the self-reported frequency of using neighborhood fast food restaurants was inversely associated with the link-to-node ratio within the 1-km buffer (Table 2); (2) the self-reported frequency of using neighborhood sit-down restaurants was positively associated with the GIS-measured number of sit-down restaurants within the same buffer (Table 3); and (3) the self-reported frequency of using neighborhood grocery stores (never use versus use yearly) was inversely associated with the link-to-node ratio within the same buffer (Table 4).

Sensitivity Testing

After adjusting for covariates, results for the self-reported frequency of use model using the 3-km buffer were largely consistent with those using the 1-km buffer, but the results for perceived presence of food outlet models showed some inconsistencies. The probability of perceiving at least one neighborhood fast food restaurant was positively associated with the GIS-measured number of fast food restaurants within the 1-km buffer and the link-to-node ratio within the 1-km buffer but not associated with those within the 3-km buffer. The probability of perceiving at least one sit-down restaurant was associated with the GIS-measured number of sit-down restaurants within the 1-km buffer but not within the 3-km buffer. Based on these results, our use of the 1-km buffer for our measures of GIS-measured exposure appears to have been more in alignment with the participants’ perception of their neighborhoods than was the use of those measures within the 3-km buffer, especially for fast food restaurants. Regression results using the 3-km buffer can be found in Appendix 4, Tables A4-1, A4-2, and A4-3, in the Electronic Supplementary Material.

We also found that the individuals who reported perceiving at least one fast food restaurant, those who reported perceiving at least one sit-down restaurant, and those who reported perceiving at least one grocery store had similar individual-level and neighborhood-level characteristics when compared to those who did not. The plotted curves (showing overlap between these two groups) are shown in Fig. A3-1 in Appendix 3 in the Electronic Supplementary Material. This finding alleviated concerns regarding off-support inference (i.e., identification of associations through actual observations thereby excluding certain types of subgroups) [23] and allowed us to include all the participants who perceived the neighborhood food outlets and explicitly reported the frequency of use into the frequency of use models based on the evidence that they did not differ greatly from those who reported that they did not perceive any outcome food outlets in their neighborhood (and therefore did not report the frequency of use).

Discussion

We found evidence of cross-sectional associations between GIS-measured neighborhood food count, neighborhood street connectivity, and the self-reported frequency of using neighborhood food outlets among individuals who perceived at least one type of restaurant or food store in their neighborhood in a large, diverse cohort of middle-aged adults. Although some of our results highlight the association between food outlets and street connectivity in the immediate home neighborhood in relation to the use of neighborhood food outlets, they also underscore the complexity underlying the relationship between the availability of different types of neighborhood food outlets and the use of neighborhood food outlets.

GIS-measured count of neighborhood fast food restaurants was not associated with the self-reported frequency of use of such neighborhood outlets. Our findings agree with the previous literature suggesting that GIS-measured count of neighborhood fast food restaurants was not related to the overall frequency of use of fast food restaurants [7–9]. One previous study identified a significantly higher overall frequency of use of fast food restaurants among adolescent males living in neighborhoods with high numbers of such restaurants [6]. It is possible that there is a relationship between spatial distribution (e.g., clustering near home, schools, or highway off-ramps) of neighborhood fast food restaurants (beyond the simple count of such restaurants) and use of such outlets, particularly in high-density areas shown to have clustering of fast food restaurants [24], which should be examined in the future.

Participants in neighborhoods with a (GIS-measured) greater number of sit-down restaurants tended to report that they used such restaurants more frequently. A previous study similarly indicated such an association for the overall frequency of eating at all restaurants other than fast food restaurants (defined as those selling quick service burger, roast beef, and pizza parlor) as the outcome [8]. Our results suggest a greater concern with respect to the use of sit-down restaurants because such restaurants are a heterogeneous group and many of such restaurants sell food options with high calories. The small proportion of our study population who used such restaurants could be greater in neighborhoods with a greater number of sit-down restaurants, which should be examined in the future.

We found an inverse association between neighborhood street connectivity and self-reported use of neighborhood fast food restaurants and grocery stores. We measured street connectivity as the directness of possible routes and the number of optional routes from which study participants could choose in their home neighborhood. Thus, higher connectivity connotes higher density of intersections. Our findings contrast with those of other investigators who have observed a positive association between neighborhood street connectivity and frequency of trips to destinations such as grocery stores and restaurants [25, 26]. Of these studies, the Khan et al. [25] study was similar to ours in age of respondents and size of neighborhood, though this study considered walking only and not biking, whereas Chudyk et al. [26] was in an older cohort. In contrast, we observed lower frequencies of neighborhood fast food restaurants and grocery stores with higher connectivity. It is possible that high connectivity facilitated access to outlets outside of the neighborhood by decreasing the travel cost from within the neighborhood.

Consistent with a previous study [20], we found that higher numbers of neighborhood fast food and sit-down restaurants increased participants’ awareness of such neighborhood food outlets. However, we did not find such an association for grocery stores. This finding suggests that increasing the number of grocery stores alone is unlikely to change resident’s awareness of grocery store availability. It might be that related interventions that assist residents in recognizing, mapping, and sharing information about the types of stores in residential neighborhoods can help with awareness [27]. It is also possible that there is measurement error in the definition of food outlets, for example if participants had different ideas about how to define a grocery store than researchers [28].

We found that approximately 16% (576 out of 3549) of participants reported that they did not have any of the three types of food outlets in their neighborhood. Of these, only 83 had none of the three types of food outlets in their 1-km buffer using electronic databases on food records via GIS. This was consistent with a report that the self-reported presence of food outlets was not strongly correlated with the objective, GIS-based presence of outlets in a South Carolina sample, the majority being female (79%) and non-Hispanic white (67%) [21].

Our study had several limitations. Our assumption that people use food outlets partly based on their awareness of such neighborhood outlets may be incorrect. Thus, perception (“I’m aware of the food outlet”) may not translate to intention (“since I’m aware of the presence of the food outlet, and other conditions (e.g., time, money) are also met, I will eat there”). It is possible that participants had difficulty distinguishing neighborhood fast food and sit-down restaurants by memory and the questionnaire did not define a fast food restaurant nor did the questionnaire provide examples (names) of typical fast food restaurants (such as McDonald’s). If so, we might underestimate the number of participants who perceived fast food restaurants in their neighborhoods. Our measures of frequency of using food outlets were derived from self-reports, which are prone to recall bias and other reporting errors. In the case of grocery stores, people might forget to report small purchases (for example, beverages or a bag of chips) and therefore underestimate frequency of their use. People might incorrectly include the use of food outlets outside of the 10- to 15-min walk area. We addressed this concern by using 3-km buffers in our sensitivity analysis. Our analysis may have omitted important factors that explain residential selection and use of food outlets. We noted that CARDIA participants were recruited from urban environments, which may limit the generalizability of our findings to more geographically diverse samples [7]. Also, the Dun & Bradstreet food business record data may have contained location or categorization errors, but these errors were probably random in nature and small. By relying on SIC code alone as compared with alternative approaches using a name search strategy, we may have misclassified food outlets [29]. In addition, we used only cross-sectional data, so we are unable to study changes in the built environment and temporality in these associations.

Conclusion

Our findings suggest that, for those who reported perceiving at least one type of restaurant or food store in their neighborhood, individuals who have more GIS-measured sit-down restaurants in their neighborhoods reported more frequent use of such restaurants than those whose neighborhood contains fewer such restaurants. Our findings also suggest that, for those who reported perceiving at least one type of restaurant or food store in their neighborhood, individuals who live in neighborhoods with greater GIS-measured neighborhood street connectivity reported less use of neighborhood fast food restaurants and grocery stores than those who live in neighborhoods with less neighborhood street connectivity. The inverse associations between street connectivity with fast food restaurant use and grocery store use are intriguing and point to future research to disentangle these causal pathways.