This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Public Health and Surveillance, is properly cited. The complete bibliographic information, a link to the original publication on https://publichealth.jmir.org, as well as this copyright and license information must be included.
The issue of food insecurity is becoming increasingly important to public health practitioners because of the adverse health outcomes and underlying racial disparities associated with insufficient access to healthy foods. Prior research has used data sources such as surveys, geographic information systems, and food store assessments to identify regions classified as food deserts but perhaps the individuals in these regions unknowingly provide their own accounts of food consumption and food insecurity through social media. Social media data have proved useful in answering questions related to public health; therefore, these data are a rich source for identifying food deserts in the United States.
The aim of this study was to develop, from geotagged Twitter data, a predictive model for the identification of food deserts in the United States using the linguistic constructs found in food-related tweets.
Twitter’s streaming application programming interface was used to collect a random 1% sample of public geolocated tweets across 25 major cities from March 2020 to December 2020. A total of 60,174 geolocated food-related tweets were collected across the 25 cities. Each geolocated tweet was mapped to its respective census tract using point-to-polygon mapping, which allowed us to develop census tract–level features derived from the linguistic constructs found in food-related tweets, such as tweet sentiment and average nutritional value of foods mentioned in the tweets. These features were then used to examine the associations between food desert status and the food ingestion language and sentiment of tweets in a census tract and to determine whether food-related tweets can be used to infer census tract–level food desert status.
We found associations between a census tract being classified as a food desert and an increase in the number of tweets in a census tract that mentioned unhealthy foods (
Social media data have been increasingly used to answer questions related to health and well-being. Using Twitter data, we found that food-related tweets can be used to develop models for predicting census tract food desert status with high accuracy and improve over baseline models. Food ingestion language found in tweets, such as census tract–level measures of food sentiment and healthiness, are associated with census tract–level food desert status.
Healthy food is vital to everyday life. However, healthy food is not equally accessible to everyone [
Geographic location is one of the most important contributing factors to food insecurity and access to healthy foods [
The disparities in healthy food access among underserved communities have fueled the interest of public health practitioners, researchers, and community activists in not only identifying regions that are currently food deserts but also regions that are at risk for becoming food deserts in the future. The Economic Research Service at the USDA uses various indicators for the official identification of food deserts in the United States at the census tract-level. A review of the literature determined that other frequently used measures to assess food access are as follows: (1) geographic information systems (GIS) technology, where researchers use geocoding to map resources and create density maps that illustrate differences in food security and access in various locations [
Although each of these food desert identification methods have been widely used and have provided rich insights into food insecurity in the United States, each method comes with unique challenges. For example, GIS technology comes with the risk of misidentification of food stores in the GIS and mapping fails to provide information about food consumption behavior [
Researchers have increasingly looked to social media data as a means of measuring population health and well-being in a less intrusive and more scalable manner [
Prior studies have successfully extracted information from social media to address various types of health-related outcomes, relying on the naturalistic observations deduced from social media data to answer questions related to health and well-being [
As seen in this study, several other studies similarly leveraged natural language processing methods such as sentiment analysis, emotion analysis, and topic modeling to use social media to answer public health research questions. For example, some studies [
In this study, we leveraged the linguistic constructs in food-related tweets to develop a classification model for food deserts in the United States. We considered both tweet sentiment and overall nutritional values of foods found in tweets to identify associations between living in a food desert and food consumption.
To our knowledge, this is the first study to develop a model for inferring food desert status among census tracts in the United States using Twitter data. The main objective of this study was to examine the linguistic constructs found in food-related tweets to evaluate the differences in food nutritional value and food consumption behavior of individuals in food deserts versus those in non–food deserts
An overview of the entire data collection and preparation process is illustrated in
Twitter data collection process. API: application programming interface; SES: socioeconomic status; USDA: United States Department of Agriculture.
From March 2020 to December 2020, the Twitter streaming application programming interface (API), which provides access to a random sample of 1% of publicly available tweets, was used to collect tweets (including retweets and quoted tweets) from 25 of the most populated cities in the United States (
When a location-based search is specified, the Twitter API extracts tweets tied to a certain location based on two criteria that are not mutually exclusive: (1) the user has their location enabled for all tweets, in which case these tweets will have specific GPS coordinates, or (2) the user has location information in their profile, such as the city and state they live in, in which case all tweets associated with this user will be tied to this location but without specific geocoordinates. In both cases, these location-tagged tweets are eligible for selection by the Twitter API when a location-based search is specified [
As this analysis sought to assign individual tweets to their respective census tracts, all tweets in our sample were required to have specific geolocation information (latitude and longitude GPS coordinates). A parsing module was created to filter out tweets without specific geolocation information. Next, to extract tweets related to food ingestion, tweets were further filtered by a list of 1787 food-related words from the USDA FoodData Central Database (examples are presented in
Tweets related to job postings and advertisements were filtered out by excluding tweets with hashtags and keywords such as “#jobs,” “#hiring,” and “#ad.” For the purposes of this research, we assumed that the tweets in our sample, which, at minimum, contained at least one of 1787 food-related keywords, were related to food consumption, as was done in the study by Nguyen et al [
Albuquerque, New Mexico
Dallas, Texas
Atlanta, Georgia
Baltimore, Maryland
Colorado Springs, Colorado
Fresno, California
Kansas City, Missouri
Las Vegas, Nevada
Long Beach, California
Louisville, Kentucky
Mesa, Arizona
Miami, Florida
Milwaukee, Wisconsin
Minneapolis, Minnesota
New Orleans, Louisiana
Oakland, California
Oklahoma City, Oklahoma
Omaha, Nebraska
Portland, Oregon
Raleigh, North Carolina
Sacramento, California
Tucson, Arizona
Tulsa, Oklahoma
Virginia Beach, Virginia
Wichita, Kansas
Healthy
Acai
Apple
Apricot
Avocado
Banana
Blackberries
Blueberries
Cantaloupe
Cherries
Clementine
Unhealthy
Cheesecake
Cupcake
Donut
Pepsi
Sprite
Sunkist
Red velvet cake
Chicken McNuggets
Double cheeseburger
Zinger burger
Fast-food restaurants
Jack in the Box
Chick-fil-A
Burger King
Dairy Queen
Del Taco
Taco Bell
Bojangles
Checkers
Popeyes
Whataburger
We referred to similar work conducted by Nguyen et al [
To measure the healthiness of foods mentioned in tweets, the overall nutritional values of the foods mentioned in each tweet were calculated. To calculate the nutritional values of foods mentioned in each tweet, regular expression matching was used to compare the words in each tweet to the items described in the aforementioned food list (
Next, the respective nutritional values for each matched food word were then calculated for the corresponding tweet. For tweets having >1 match to food names in the food list, the assigned nutritional value was equal to the average of the nutritional values for all matched food items in the tweet.
To capture the attitudes toward foods mentioned in tweets, we conducted a sentiment analysis of all tweets using the bing lexicon from the tidytext package in R [
As this analysis examined food desert status at the census tract-level, for all census tracts in the 25 cities listed in
Percentage of tweets that mention healthy foods with positive sentiment
Percentage of tweets that mention healthy foods with negative sentiment
Percentage of tweets that mention unhealthy foods with positive sentiment
Percentage of tweets that mention unhealthy foods with negative sentiment
Percentage of tweets that mention fast-food restaurants with positive sentiment
Percentage of tweets that mention fast-food restaurants with negative sentiment
Average number of healthy food mentions
Average number of unhealthy food mentions
Average number of fast-food mentions
Average number of calories per food item (per 100 g)
Average calcium per food item (per 100 g)
Average carbohydrates per food item (per 100 g)
Average cholesterol per food item (per 100 g)
Average energy per food item (per 100 g)
Average fiber per food item (per 100 g)
Average iron per food item (per 100 g)
Average potassium per food item (per 100 g)
Average fat per food item (per 100 g)
Average protein per food item (per 100 g)
Average saturated fatty acids per food item (per 100 g)
Average sodium per food item (per 100 g)
Average sugar per food item (per 100 g)
Average trans fatty acids per food item (per 100 g)
Average unsaturated fatty acids per food item (per 100 g)
Average vitamin A per food item (per 100 g)
Average vitamin C per food item (per 100 g)
Average number of calories per healthy food item (per 100 g)
Average number of calories per unhealthy food item (per 100 g)
Once all data were collected and aggregated to the census tract-level, each census tract was classified as a food desert or not a food desert, according to the USDA Food Access Research Atlas classification of low-income and low-access tracts measured at 1 mile for urban areas and 10 miles for rural areas. The USDA classifies low-income tracts using the following criteria: (1) at least 20% of the residents live below the federal poverty level; (2) median family income is, at most, 80% of the median family income for the state in which the census tract lies; or (3) the census tract is in a metropolitan area and the median family income is, at most, 80% of the median family income for the metropolitan area in which the census tract lies [
In total, 7.52% (299/3978) of census tracts with geolocated food-related tweets were classified as low-income, low-access food deserts, measured at 1 mile for urban areas and 10 miles for rural areas.
Demographic and socioeconomic status (SES) characteristics at the census tract-level were pulled from the 2019 American Community Survey and merged onto the census tract–level tweets data set. The demographic variables used in this analysis are presented in
Percentage White and non-Hispanic
Percentage Black or African American
Percentage other race
Percentage Asian
Percentage American Indian or Alaska Native
Percentage owner-occupied housing units
Percentage of population living below the federal poverty line
Number of housing units
Number of households
Median family income (US $, 2019)
Median age (years)
Population
Analyses were performed using R software (version 3.5.1; The R Foundation for Statistical Computing) and Python (version 3.8).
To test the hypothesis that living in a food desert is associated with the food ingestion language of Twitter users, adjusted linear regression was conducted using food desert status as a treatment and the SES features listed in
where
Twitter-derived features that were found to have individual, significant associations with food desert status were later used as features in the classification model for predicting food desert status to test the hypothesis that key food ingestion language found in tweets can be used to infer census tract–level food desert status.
To test the hypothesis that food ingestion language found in tweets can be used to infer census tract–level food desert status, classification models were developed using the Twitter-derived food-related nutritional features listed in
All features were standardized using minimum-maximum normalization, a method that standardizes data by rescaling the range of individual features to (0, 1), as described in the study by Cao et al [
Classification models for predicting food desert status.
Model | Description | Features |
1 | Demographics and SESa only (baseline) | Demographics and SES features ( |
2 | Demographics and SES+nutritional values | Demographics and SES features ( |
3 | Demographics and SES+Twitter mentions sentiment | Demographics and SES features ( |
4 | Demographics and SES+nutritional values+Twitter mentions sentiment | Demographics and SES features ( |
5 | Demographics and SES+statistically significant features | Demographics and SES features ( |
aSES: socioeconomic status.
The University of Maryland College Park institutional review board has determined that this project does not meet the definition of human participant research under the purview of the institutional review board according to federal regulations.
A total of 60,174 geolocated food-related tweets were collected during the data collection period. Across the 25 cities in our sample, 3978 census tracts had at least one geolocated food-related tweet, with a median of 4 (IQR 8) geolocated food-related tweets per census tract. Long Beach, California, had the largest representation of tweets (17,303/60,174, 28.75%), as well as the largest representation of users (5189/17,978, 28.86%;
Number of tweets (N=60,174) and users (N=17,978) by city.
City | Number of tweets, n (%) | Number of users, n (%) |
Albuquerque, New Mexico | 839 (1.39) | 224 (1.26) |
Atlanta, Georgia | 4936 (8.2) | 1739 (9.67) |
Baltimore, Maryland | 2521 (4.19) | 684 (3.8) |
Colorado Springs, Colorado | 847 (1.41) | 268 (1.49) |
Dallas, Texas | 2472 (4.11) | 782 (4.35) |
Fresno, California | 421 (0.7) | 153 (0.85) |
Kansas City, Missouri | 1651 (2.74) | 532 (2.96) |
Las Vegas, Nevada | 2336 (3.88) | 872 (4.85) |
Long Beach, California | 17,303 (28.75) | 5189 (28.86) |
Louisville, Kentucky | 1246 (2.07) | 406 (2.26) |
Mesa, Arizona | 1888 (3.14) | 616 (3.43) |
Miami, Florida | 2576 (4.28) | 1080 (6.01) |
Milwaukee, Wisconsin | 1578 (2.62) | 388 (2.16) |
Minneapolis, Minnesota | 1282 (2.13) | 471 (2.62) |
New Orleans, Louisiana | 2144 (3.56) | 641 (3.57) |
Oakland, California | 2601 (4.32) | 614 (3.42) |
Oklahoma City, Oklahoma | 1143 (1.9) | 371 (2.06) |
Omaha, Nebraska | 742 (1.23) | 198 (1.1) |
Portland, Oregon | 5528 (9.19) | 928 (5.16) |
Raleigh, North Carolina | 1588 (2.64) | 454 (2.53) |
Sacramento, California | 1721 (2.86) | 565 (3.14) |
Tucson, Arizona | 794 (1.32) | 250 (1.39) |
Tulsa, Oklahoma | 622 (1.03) | 209 (1.16) |
Virginia Beach, Virginia | 960 (1.6) | 212 (1.18) |
Wichita, Kansas | 435 (0.72) | 132 (0.73) |
Descriptive statistics of Twitter-derived food features from geolocated food-related tweets.
Twitter-derived food features | Values, mean (SD) |
Percentage of tweets that mention healthy foods, positive sentiment | 33.8 (0.4) |
Percentage of tweets that mention healthy foods, negative sentiment | 19.8 (0.3) |
Percentage of tweets that mention unhealthy foods, positive sentiment | 33.5 (0.4) |
Percentage of tweets that mention unhealthy foods, negative sentiment | 17.1 (0.3) |
Percentage of tweets that mention fast-food restaurants, positive sentiment | 21.2 (0.3) |
Percentage of tweets that mention fast-food restaurants, negative sentiment | 11.7 (0.3) |
Average number of healthy food mentions | 0.2 (0.3) |
Average number of unhealthy food mentions | 0.4 (0.4) |
Average number of fast-food mentions | 0.1 (0.3) |
Average number of calories per food item (per 100 g) | 155.1 (96.3) |
Average calcium per food item (per 100 g) | 74 (91.3) |
Average carbohydrates per food item (per 100 g) | 23.2 (10.9) |
Average cholesterol per food item (per 100 g) | 57.3 (284.4) |
Average energy per food item (per 100 g) | 285.1 (115.7) |
Average fat per food item (per 100 g) | 10.4 (6.9) |
Average fiber per food item (per 100 g) | 1.7 (1.4) |
Average iron per food item (per 100 g) | 1.7 (8.5) |
Average potassium per food item (per 100 g) | 194.5 (93) |
Average protein per food item (per 100 g) | 7 (4.1) |
Average saturated fatty acids per food item (per 100 g) | 3.6 (2.5) |
Average sodium per food item (per 100 g) | 524.7 (962.7) |
Average sugar per food item (per 100 g) | 11.8 (8.3) |
Average trans fatty acids per food item (per 100 g) | 0.1 (0.2) |
Average unsaturated fatty acids per food item (per 100 g) | 2.6 (4) |
Average vitamin A per food item (per 100 g) | 548.8 (734.5) |
Average vitamin C per food item (per 100 g) | 7.1 (15.8) |
Average number of calories per healthy food item (per 100 g) | 67.4 (61.5) |
Average number of calories per unhealthy food item (per 100 g) | 189.8 (125.9) |
Descriptive statistics of census tract–level demographics and socioeconomic status features extracted from the 2019 American Community Survey.
Characteristic | Values, mean (SD) |
Percentage White and non-Hispanic | 62.7 (23.4) |
Percentage Black or African American | 15.6 (21.0) |
Percentage other race | 8.9 (12.3) |
Percentage Asian | 7.4 (9.2) |
Percentage American Indian or Alaska Native | 1.0 (1.9) |
Percentage owner-occupied housing units | 49.3 (24.8) |
Percentage of population living below the federal poverty line | 16.2 (12.1) |
Number of housing units | 1788.4 (863.5) |
Number of households | 1628.0 (799.1) |
Median family income (US $, 2019) | 82,371.4 (42,680.1) |
Median age (years) | 37.0 (6.8) |
Population | 4283.1 (2243.6) |
The adjusted linear regression models confirmed this hypothesis, revealing significant associations between food desert status and 5 of the Twitter-derived food characteristics (
Although we did not expect to see an association between living in a food desert and an increase in mentions of healthy foods with positive sentiment, we hypothesize that such an association might reflect aspirational tweets of individuals who long for healthy food that is not present in their neighborhood (for example, the positive sentiment does not reflect food consumption but rather a wish to increase accessibility).
Adjusted linear regression model results examining the associations between living in a food desert and food ingestion language of Twitter users.
Twitter-derived food features | β coefficient | SE | R-squared | |
Percentage of tweets that mention healthy foods, positive sentiment | .077 | .03 | 0.036 | 0.003 |
Percentage of tweets that mention healthy foods, negative sentiment | .023 | .44 | 0.031 | 3.45×10–5 |
Percentage of tweets that mention unhealthy foods, positive sentiment | –0.051 | .06 | 0.027 | 0.001 |
Percentage of tweets that mention unhealthy foods, negative sentiment | .022 | .32 | 0.022 | 3.98×10–4 |
Percentage of tweets that mention fast-food restaurants, positive sentiment | .096 | .01 | 0.039 | 0.005 |
Percentage of tweets that mention fast-food restaurants, negative sentiment | .010 | .74 | 0.032 | 8.88×10–5 |
Average number of healthy food mentions | –0.002 | .54 | 0.003 | 9.57×10–5 |
Average number of unhealthy food mentions | .014 | .03 | 0.006 | 0.001 |
Average number of fast-food mentions | –0.003 | .76 | 0.010 | 2.45×10–5 |
Average number of calories per food item (per 100 g) | .005 | .58 | 0.009 | 7.93×10–5 |
Average calcium per food item (per 100 g) | –0.001 | .60 | 0.002 | 7.36×10–5 |
Average carbohydrates per food item (per 100 g) | –0.009 | .19 | 0.007 | 4.46×10–4 |
Average cholesterol per food item (per 100 g) | .005 | .02 | 0.002 | 0.001 |
Average energy per food item (per 100 g) | .004 | .60 | 0.007 | 7.37×10–5 |
Average fat per food item (per 100 g) | –0.005 | .69 | 0.012 | 4.27×10–5 |
Average fiber per food item (per 100 g) | –0.014 | .10 | 0.008 | 7.26×10–4 |
Average iron per food item (per 100 g) | –6.44×10–4 | .56 | 0.001 | 9.04×10–5 |
Average potassium per food item (per 100 g) | –0.008 | .01 | 0.003 | 0.002 |
Average protein per food item (per 100 g) | –0.002 | .88 | 0.010 | 6.11×10–6 |
Average saturated fatty acids per food item (per 100 g) | .007 | .31 | 0.007 | 2.70×10–4 |
Average sodium per food item (per 100 g) | –0.005 | .06 | 0.002 | 9.13×10–4 |
Average sugar per food item (per 100 g) | –0.005 | .35 | 0.005 | 2.29×10–4 |
Average trans fatty acids per food item (per 100 g) | –0.002 | .79 | 0.007 | 1.78×10–5 |
Average unsaturated fatty acids per food item (per 100 g) | .002 | .72 | 0.006 | 3.39×10–5 |
Average vitamin A per food item (per 100 g) | .004 | .58 | 0.007 | 8.19×10–5 |
Average vitamin C per food item (per 100 g) | –5.53×10–4 | .71 | 0.002 | 3.52×10–5 |
Average number of calories per healthy food item (per 100 g) | 9.58×10–4 | .95 | 0.017 | 1.92×10–6 |
Average number of calories per unhealthy food item (per 100 g) | .007 | .64 | 0.015 | 8.42×10–5 |
To test the hypothesis that food ingestion language found in tweets can be used to infer census tract–level food desert status, we used various machine learning methods to compare the performance of 5 classification models (
Model performance.
Method and modela | AUCb | |
|
||
|
1 (baseline) | 0.759 |
|
2 | 0.749 |
|
3 | 0.738 |
|
4 | 0.650 |
|
5 | 0.723 |
|
||
|
1 (baseline) | 0.766 |
|
2 | 0.797 |
|
3 | 0.823 |
|
4 | 0.777 |
|
5 | 0.699 |
|
||
|
1 (baseline) | 0.682 |
|
2 | 0.720 |
|
3 | 0.777 |
|
4 | 0.809 |
|
5 | 0.663 |
|
||
|
1 (baseline) | 0.769 |
|
2 | 0.771 |
|
3 | 0.760 |
|
4 | 0.641 |
|
5 | 0.740 |
aModel descriptions (refer to
bAUC: area under the receiver operating characteristic curve.
In this study, we sought to address two key hypotheses: (1) living in a food desert is associated with positive mentions of unhealthy foods, such as tweets that mention foods that are high in caloric content or low in vital nutrients such as fiber and calcium, and (2) food ingestion language among Twitter users in a census tract can be used to infer census tract–level food desert status. The study found significant associations between living in a food desert and tweeting about unhealthy foods, including foods high in cholesterol content or low in key nutrients such as potassium. We also found that supplementing classification models with features derived from food ingestion language found in tweets, such as positive sentiment toward mentions of healthy foods and fast-food restaurants, improves baseline models that only include demographic and SES features by up to 19%, with AUC scores >0.8.
Assessing and understanding the food environment in neighborhoods is key to addressing the issue of food insecurity in the United States. The USDA conducts the official identification of food deserts in the United States but this assessment is infrequent and the latest assessment from 2015 is outdated. Other methods such as GIS technology, surveys, and food store assessments, although effective, can be costly and time consuming. Although conducting assessments of food stores provides important insights into the food environment, this study suggests that perhaps residents of census tracts unknowingly provide important information regarding the food environment on Twitter through the food ingestion language found in tweets. Using social media data for food insecurity research allows researchers to examine food consumption in various regions, allowing a comparison of how food ingestion differs between areas where residents have sufficient access to healthy foods and areas where residents do not have sufficient access to healthy foods.
The findings of this study contribute to the literature on food insecurity in the United States by examining the potential effects of living in a food desert on food consumption using Twitter-derived food ingestion features as a proxy to examine food consumption. In this study, we found that food desert status is associated with not only the sentiment toward the types of foods mentioned in tweets but also the nutritional content of foods mentioned in tweets. More specifically, a census tract being classified as a food desert was associated with an increase in the average cholesterol concentration and a decrease in the average potassium concentration (per 100 g) per food item mentioned in tweets, as well as an increase in the proportion of tweets that mention unhealthy foods. A census tract classified as a food desert was also associated with an increase in the proportion of tweets that mentioned healthy foods and fast-food restaurants with positive sentiment. These findings support prior studies that also found associations between neighborhood characteristics, such as food desert status or fast-food density, and the
This study makes further contributions by examining the predictive ability of food ingestion language derived from tweets on census tract food desert status. This builds upon a similar study that used Instagram posts to understand dietary choices and nutritional challenges in food deserts [
Other similar studies that sought to examine food consumption using tweets across various geographic regions suggest that many of the food-related tweets in an area may be an artifact of visitors to the area, not residents. For example, a study conducted by Mitchell et al [
Developing an algorithm that predicts food deserts by extracting information from tweets allows researchers to monitor food insecurity more frequently than current methods allow. The use of tweets for research related to food insecurity provides researchers with more frequently updated information, thereby addressing the “lag between capturing information about newly opened and recently closed food retail businesses” [
Although prior research has proved social media to be a rich data source, it does have some limitations. The ability to pull millions of tweets from a single data source is an attractive characteristic of Twitter data, but a study conducted by Pew Research Center showed that Twitter users are more likely to be younger than the general population (29% of Twitter users are aged 18 to 29 years compared with 21% of the general population in the United States), more highly educated (42% of Twitter users are college graduates compared with 31% of the general population in the United States), have higher incomes (41% of Twitter users earn at least US $75,000 per year compared with 32% of the general population in the United States), and are more likely to consider themselves Democrats (36% of Twitter users consider themselves Democrats compared with 30% of the general population in the United States) [
Adding to the lack of representation among Twitter users is the disparity in Twitter activity among Twitter users. The median number of tweets for Twitter users is only 2 tweets per month. Just 10% of Twitter users account for 80% of the tweets across users in the United States [
Tweets were collected using the Twitter streaming API, which is limited to a random sample of 1% of all tweets sent by Twitter users at any given time. Of this limited sample of tweets, studies have shown that only approximately 1% to 2% of the tweets from the Twitter streaming API include geolocation information [
Despite these limitations, the results of this study confirm both our hypotheses, demonstrating that food ingestion language found in tweets provides a signal that differentiates food deserts from non–food deserts.
The issue of food insecurity is an important public health issue because of the adverse health outcomes and underlying racial and economic disparities that are associated with insufficient access to healthy foods [
application programming interface
area under the receiver operating characteristic curve
geographic information systems
socioeconomic status
United States Department of Agriculture
The authors thank Nhat Pham, Daniela Nganjo, and Pauline Comising for their assistance with quality control activities associated with the Twitter data.
None declared.