From February 2015 to March 2016, we utilized Twitter’s streaming application programming interface (API) to continuously collect a random 1% sample of publicly available tweets with latitude and longitude coordinates. Given that neighborhood researchers differ in their use and interest in data at the census tract and zip code level, we constructed neighborhood indicators at both levels thereby increasing the flexibility of our dataset to address the potential data needs of other researchers. In total, we collected 79,848,992 million geotagged tweets from 603,363 unique Twitter users in the contiguous United States (including District of Columbia). The median number of tweets per user was 4. Job postings (identified through hashtags #hiring, #jobs, and #job) were removed from the final analytic sample of tweets because these were pervasive and not central to the neighborhood variables we constructed.

Each geotagged tweet was assigned a corresponding census tract and zip code it falls within, based on the latitude and longitude coordinates of where the tweet was sent. This spatial join procedure was implemented in Python (version 2.7.12; Python Software Foundation), a popular programming language for spatial data processing [50]. Specifically, Python libraries were used to read shapefile format vector data (PyShp 1.1.4), build an R-tree index on the polygon data (Rtree 0.8.2), and perform a spatial join operation (Shapely 1.5.12 and Fiona 1.6.1). The R-Tree was used to build a spatial index [51] on the national census tract and zip code polygon data to speed computation. Tweets that were not assigned a census tract or zip code location included those with destinations bordering the United States (ie, Mexico and Canada). We linked 99.8% of tweets with geocordinates to their respective 2010 census tract and zip code locations. The term neighborhood used in this paper refers to both zip codes and census tracts. We mapped tweets to these two geographic boundaries because they are among the most popular neighborhood definitions utilized by public health researchers [52-54].

Duplicate tweets (ie, tweets with the same tweet ID, <1%) were removed computationally. Although Twitter’s API collects a random subset of 1% of publicly available tweets, users (especially spam accounts) who tweet often have potentially greater influence on variable values we construct. We examined outliers in our datasets (defined as the users whose tweets accounted for more than 1% of tweets in our dataset) and eliminated automated accounts and accounts for which the majority of tweets were advertisements. Processing and statistical analysis tasks were performed with Stata MP13 (StataCorp LP).

From geotagged tweets, we derived variables that characterize happiness, food, and physical activity. Each tweet was divided into tokens using the Stanford tokenizer [55]. For processing of English text, tokens roughly correspond to words. We then built various algorithms utilizing tokens to create variables that characterize happiness and make references to food and physical activity. Below we describe in more detail our algorithms.

To conduct sentiment analysis, we utilized the Machine Learning for Language Toolkit (MALLET; AK McCallum, 2002), a Java-based package for statistical natural language processing, document classification, clustering, topic modeling, information extraction, and other machine learning applications to text. We leveraged the Maximum Entropy text classifier in MALLET to classify tweets as happy and not happy [56]. In order to train our classifier, we obtained training sets from the following resources: Sentiment140 [57], Sanders Analytics [58], and Kaggle [59]. We trained our classifier to differentiate between happy and not happy sentiments. We then ran our classifier on our national Twitter data to compute a happy score (range 0-1) for each tweet, where higher happiness scores indicate more positive sentiment. MALLET estimates predicted probabilities that a tweet is happy based upon word-level features. The classifier uses search-based optimization to assign weights that maximize the likelihood of the training data. However, unlike Naïve Bayes, the Maximum Entropy classifier does not assume conditional independence among features.

To calibrate the generated happiness scores with human generated labels, two raters manually read a random subset of 1200 tweets and assigned a value of 1 to happy tweets and 0 to not happy tweets. The initial interrater reliability was 92%, and discordant values were reviewed until a 100% agreement between raters was reached. To decide on a cut point for MALLET scores at which we would classify tweets as happy, we computed accuracy levels at different cut points of MALLET scores (Multimedia Appendix 1). Increasing the MALLET score improves the accuracy against human annotations but also reduces the calculated prevalence of tweets deemed as happy. A MALLET score of 0.80 achieves the highest level of accuracy while still maintaining a prevalence of happy tweets of 19% (which approximates the prevalence obtained by human annotations). Area under the receiver operating characteristic curve is approximately 0.7 for all MALLET cut points between 60 and 85.

We compiled a list of over 1430 popular food words from the US Department of Agriculture’s National Nutrient Database [60]. Each food item was associated with a measure of caloric density, operationalized as calories per 100 grams. Fruits, vegetables, nuts, and lean proteins (ie, fish, chicken, and turkey) were labeled as healthy foods (340 food terms in total). Fried foods were not considered healthy foods. Our food list also contained popular national fast food restaurants such as McDonald’s and Kentucky Fried Chicken (captured via 154 food terms including popular variations of restaurant names) to enable quantification of fast food references. From April 2015 to March 2016, we collected and processed 4,041,521 geotagged food tweets. In the food dataset, the median number of tweets per user was 12 tweets.

To analyze food culture, each tweet was examined for words or phrases matching those on our list. Each food item on our list was described by one or two words. Our text-matching algorithm first searched over a tweet for matches to two-word foods (eg, orange chicken). It then searched over the remaining words for matches to one-word food terms (eg, taco). We computed caloric density by summing up all the foods mentioned in the tweet. We also created a count of healthy food references and fast food restaurant references for each tweet. Moreover, we leveraged our sentiment analysis to assess sentiment toward food. Specifically, we tracked sentiment around healthy foods and fast food. These variables (any food references, healthy food references, fast food references, caloric density, and sentiment toward healthy foods and fast food) were then aggregated and summarized at the census tract and zip code level to create neighborhood indicators of food culture.

We created a list of physical activities using published lists of physical activity terms gathered from physical activity questionnaires, compendia of physical activities, and popularly available fitness programs [61,62]. Our physical activity list had 376 different activities that incorporate gym-related exercise (eg, treadmill, weight lifting), sports (eg, baseball), recreation (eg, hiking, scuba diving) and household chores (eg, gardening). We excluded popular phrases that generally do not relate to physical activity such as “walk away” and “running late.” Using metabolic equivalents associated with physical activities, we quantified the exercise intensity of each physical activity mention, scaled for a duration of 30 minutes and for a 155-pound individual [63], which approximates the weight of an average American adult [64,65].

Upon piloting our algorithm, we identified commonly used phrases or pop culture references that do not involve physical activity (eg, walking dead) which were manually coded and excluded. Moreover, in order to help reduce the possibility that the tweet was about watching rather than actually participating in the physical activity, we excluded the tweet if it contained any of the following terms: “watch,” “watching,” “watches,” “watched,” “attend,” “attending,” “attends,” and “attended.” In reviewing preliminary labeled physical activity data, we found that most tweets (over 90%) pertaining to team sports (eg, baseball, basketball, football, soccer) were about watching games rather than participating in them. Thus, for team sports, we required that the tweet include the words “play,” “playing,” or “played.”

Our algorithm created the following physical activity variables for each tweet: any physical activity mention, exercise intensity, and sentiment around physical activity. From April 2015 to March 2016, we collected 1,473,976 geotagged physical activity tweets. In the physical activity dataset, the median number of tweets per user was 5 tweets.

A total of 5000 tweets have been manually labeled by two of the authors for quality control activities on food and physical activity. The authors manually labeled whether each tweet was food-related (2000), non–food-related (500), physical activity-related (2000), or non–physical activity-related (500). Excellent interrater reliability was achieved with greater than 90% agreement in all categories, and differences were discussed and resolved.

Among tweets our algorithm had labeled as food-related, 83% were labeled accurately when compared to labels generated by manual categorization. Among tweets our algorithm had labeled as non–food-related, 81% were labeled accurately (ie, both algorithm and human categorizers labeled the tweet as non–food-related). Overall, accuracy for food tweets was 83% and the F score was 0.86. It should be noted our algorithm could label a food-related tweet as non–food-related if the food reference was not in our food dictionary. Food items that are often associated with non-related food meaning, such as “perch,” have been excluded from our food dictionary. For tweets that had been mislabeled as food-related, commons reasons included food term used as a metaphor, in a pun, or for food advertisement.

Among tweets our algorithm had labeled as physical activity-related, 82% of them were labeled accurately when compared to labels generated by human categorizers. An accuracy of 97% was found among tweets labeled as non-physical activity-related by our algorithm. The F score was 0.90 and the overall accuracy was 85% for physical activity tweets. Typical errors in classification of physical activity tweets included the use of an idiom (eg, running late) or the tweet was about watching sports games rather than playing sports.

Additionally, we evaluated our algorithm on its ability to identify relevant food and physical activity terms within tweets. To do this, we examined a random subset of tweets that the algorithm had identified as positive for food (n=200) and physical activity (n=200). Here we focused on the accuracy of our algorithm to conduct string detection. We manually read the tweets to verify that manual annotations agreed with the terms detected. For food tweets, 87% of manual annotations matched all detected terms from the algorithm. Errors for nondetection of terms occurred when the tweet included a hashtag that had multiple food terms without spacing (eg, #chocolatebrownie) or when there were misspellings (eg, sandwhich) or when the food was not on the food list. String detection for physical activity-related terms was more accurate with 98% of manual annotations matching detected terms from the algorithm. Errors included the omission of certain terms from the dictionary (eg, cycling) and use of hashtags without spacing of terms (#runrunrun).

We further evaluated our sentiment analysis activities through Amazon Mechanical Turk (Mturk; Amazon.com Inc, Seattle, WA, USA), an online crowdsourcing marketplace [66]. We randomly selected 500 tweets with 50% labeled as happy and 50% as not happy by our algorithm. Then, we created 20 online surveys through random sorting, with each survey consisting of 25 tweets. We asked participants to rate the sentiment of each tweet. All 20 surveys were live on April 1, 2015. Each online survey closed itself when 15 responses had been reached; the last survey closed on April 5, 2015. For each completed survey, 25 cents ($0.25) was deposited into the participant’s Mturk account. A total of 32 participants completed 300 surveys (ie, 15 responses per survey, 20 surveys). Some participants completed multiple surveys rather than just one. Each tweet was then assigned a label of either happy or not happy based on the modal response from Mturkers (participants from Amazon Mturk). We found an accuracy of 69% for happy tweets and 80% for nonhappy tweets when compared to responses from Mturkers. The overall accuracy for sentiment was 78%, with an F score of 0.54.

We additionally compared performance of MALLET with two other sentiment analysis techniques: a popular bag-of-words technique involving the use of a 10,000 word list [67] and Sentiment140, a machine-learning classifier [68]. Among the 500 control tweets from our LabMT experiment, the bag-of-words algorithm had an accuracy of 73% (F score 0.55) and Sentiment140 was had an accuracy of 77% (F score 0.47).

To examine how Twitter-derived neighborhood variables relate to more traditional neighborhood variables, we merged our social media dataset with the 2010 Census and 2014 American Community Survey data which comprised the following demographic, household, and economic characteristics: household size, median family income and percent of the following: 65 years and older age group, 10-24 years, male, African American, white, Hispanic, households with relatives (other than spouse and children), households with unmarried partner, single female-headed households, householder living alone, owner-occupied housing, college graduates, unemployed, less than a high school degree and families living in poverty. A census tract was urban if the geographic centroid of the tract was in an area with more than 2500 people; all other tracts are rural. A zip code was defined as urban if the majority (75% or more) of its land area was characterized as urban (ie, containing at least 2500 people).

Data on business types at the zip code level were obtained from the 2013 US Census Bureau zip code business patterns accessed via American FactFinder [69]. The following North American Industry Classification System (NAICS) codes were utilized to categorize businesses: 722410 (drinking places [alcoholic beverages]; these places are also known as bars, taverns, night clubs and primarily serve alcohol and may have limited food services) and 722511 (full-service restaurants; these include, for instance, diners and steakhouses). Fast food was defined by the following NAICS codes: 722513 (limited-service restaurants; these include carryout restaurants, drive-in restaurants, and other fast food restaurants) and 722515 (snack and nonalcoholic beverage bars). We also tracked supermarkets and grocery stores (NAICS code 445110) and convenience stores (NAICS code 445120). To examine associations between Twitter physical activity mentions and presence of recreational facilities, we retrieved business data for the following types of establishments: fitness and recreational sports centers (NAICS code 713940), nature parks (NAICS code 712190), zoos and botanical gardens (NAICS code 712130), golf courses and country clubs (NAICS code 713910), skiing facilities (NAICS code 713920), and bowling centers (NAICS code 713950).

We obtained state-level health outcome data including age-adjusted death rates due to all-causes and homicides from 2013 National Vital Statistics Reports. Data in this report was based on information from all resident death certificates filed in the 50 states and the District of Columbia. Death certificates are generally completed by funeral directors, attending physicians, medical examiners, and coroners. Age-adjusted death rates expressed per 100,000 population were based on the 2000 US standard population. Causes of death statistics were classified by the International Classification of Diseases, Tenth Revision, and based on the underlying cause of death.

We obtained age-adjusted prevalences of health risk behaviors and chronic conditions of US adult residents for the 50 states from the 2013 Behavioral Risk Factor Surveillance System (BRFSS), the nation's premier system of health-related telephone surveys. The questionnaires were created by BRFSS state coordinators and Centers for Disease Control and Prevention staff. BRFSS data includes self-reported physical activity, self-rated health, body mass index (BMI, kg/m²), and medical diagnoses of chronic conditions aggregated to the state level. Data from a national health survey suggests that BMI estimates derived from self-reported height and weight were lower than those are derived from measured height and weight, although BMI differences were generally less than 1.0 kg/m² across sex and age groups [70]. State-level BRFSS data is publicly available. Smaller area aggregations can require data use agreements. In addition to state-level BRFSS data, we also utilized restricted-access zip-code–level data from the 2009-2014 Utah BRFSS survey to examine zip-code–level health outcomes [71,72].

We implemented adjusted linear regression models to examine associations between area-level Twitter characteristics and other area-level characteristics (demographics, business characteristics, and health outcomes). To facilitate interpretation of findings for different variables, we standardized all variables to have a mean of zero and standard deviation of one. We investigated spatial autocorrelation and found that Moran’s I was highest for census tract Twitter happiness (0.12) and less than 0.04 for other Twitter tract and zip code summaries. To account for spatial autocorrelation of area-level values in linear regression analyses, we adjusted standard errors for clustering of census tract and zip code values within a county. Statistical analyses were implemented with Stata MP13 (StataCorp LP) and ArcGIS Desktop version 10.1-10.3 (Esri).

In this paper, we detail the building of a new national neighborhood data repository constructed from Twitter data which addresses a pressing need for neighborhood data that are available across large geographies and can be updated efficiently and cost-effectively. We demonstrate that simple machine learning algorithms for the construction of indicators for happiness, food, and physical activity can agree extremely well with manually generated labels. About one-fifth of tweets were identified as happy. There was substantial spatial variation in happiness across the United States. For instance, the proportion of tweets that were happy in Montana (the most happy state) was 10% greater than in Louisiana (the least happy state). Only a few terms are needed to capture the majority of tweets on food and physical activity. Economic disadvantage, urbanicity, and presence of fast food restaurants predicted lower area level happiness and lower frequency of healthy behavior mentions on Twitter. Moreover, we find that Twitter area-level characteristics are correlated with area-level health outcomes relating to health behaviors, chronic diseases, mortality, and self-rated health.

Social media represents an important new data resource that is increasingly being harnessed for public health efforts such as surveillance of smoking behavior and sentiment toward tobacco products [74]. However, few studies are leveraging social media data for the investigation of local area characteristics. More commonly, studies utilizing social media data examine patterns at the city, county, or state level [67,75] rather than at finer levels of aggregation, which is necessary for understanding the potential impacts of neighborhood conditions.

Neighborhoods can impact health through a myriad of pathways. Disadvantaged neighborhoods may have fewer resources that support physical activity and healthy diets. Poor and minority neighborhoods have fewer large supermarkets (where healthy foods are more abundant and affordable) compared to wealthy and majority white neighborhoods. Studies have documented increased fruit and vegetable consumption with more supermarket availability [17]. Poor neighborhoods, which have been labeled food deserts, also tend to have more fast food restaurants, which can contribute to weight gain [6]. In this study, we found that higher numbers of fast food restaurants were associated with higher frequency of fast food mentions, lower frequency of healthy food mentions, and less positive sentiment about healthy foods on Twitter. Our results align with a recent study conducted analyzing Instagram posts, which found that posts originating from census tracts deemed as food deserts contained fewer mentions of fruits and vegetables compared to Instagram posts outside food deserts [76]. Additionally, neighborhoods may promote poor health through psychosocial pathways. Living in neighborhoods that are unclean, noisy, and violent can be psychologically harmful through over-activation of the stress response [77,78].

We found that economic disadvantage was related to lower frequency of happy tweets. Previous research by Mitchell and colleagues found that higher socioeconomic status was associated with higher Twitter happiness scores at the city level. Moreover, they identified mild correlations (r=−0.34) between happiness and obesity rates for 190 metropolitan statistical areas [67] and that Twitter happiness scores were moderately correlated with other state-level indicators of well-being including shootings, the Peace index, America’s Health Ranking, and the Gallup-Healthways Well-Being Index (correlations ranged between 0.51 and 0.64) [67].

In this paper, we describe the creation of a new neighborhood data repository constructed from Twitter data and merged with publicly available administrative datasets. However, this study is subject to several limitations. For instance, users of social media tend to be younger; in 2014, 37% of individuals aged 18 to 29 years old used Twitter compared to 12% of individuals aged 50 to 64 years and 10% among those 65 years and older. Nonetheless, adoption rates of social media have been steadily increasing [79]. Tweets also include information rarely found in other neighborhood sources. Twitter users are composed of individuals as well as groups of individuals, organizations, companies, and news outlets. Thus, compiling such information may allow for a more comprehensive examination of the social environment.

Moreover, we are only collecting a subset of publicly available tweets, and thus conclusions from our analytic sample may not generalize to the full population of tweets [80]. Our construction of neighborhood indicators from Twitter data necessitated that we restricted our data collection to geolocated tweets. We utilized Twitter’s API which allows the retrieval of a maximum resulting volume of 1% of the total tweets at any given time point. Previous studies suggest that about 1% to 2% of tweets may contain global positioning system location information [81,82] and that use of Twitter’s streaming API may obtain 40% to 90% of all geotagged tweets [81,82]. Tweets with location information may be different from those without. For example, tweets in which users share their locations may be more likely to contain public and social activities such as friends tweeting from a restaurant or an event. However, in sensitivity analyses with a subset of control tweets (n=138,152 tweets) collected from July 9 to July 14, 2015, we did not detect any statistically significant differences in happiness scores between tweets with and without geographic coordinates (not shown).

In creating our neighborhood indicators from Twitter data, we prioritized transparency and ease of implementation so that other researchers can replicate our algorithms. Our sentiment algorithm was trained to differentiate between happy and not happy sentiments (which encompasses neutral and sad sentiments). Thus, we were not able to specifically examine the prevalence of sad tweets, which may provide additionally useful information about the well-being of communities. In future work, we plan to target the identification of sadness. Our algorithms for food and physical activity implemented corpus-based classification with steps that are easily understandable. However, this technique does not take into account the entire context of sarcasm or humor in a tweet, challenges which still evade most natural language processing algorithms though some studies show promising results [83,84]. Our analysis of caloric density of food assumed calories per 100 grams. Most tweets do not specify the exact amount of food consumed, and thus our estimate is just an approximation.

Additionally, the content of tweets reflects the type of information that people feel comfortable reporting and may not represent the true spectrum of their feelings or their experiences. For instance, people may feel most comfortable presenting a neutral stance rather than voicing polarizing viewpoints. Certain foods (cupcakes) may get tweeted more often than others (celery). Additionally, we cannot be certain that the food that was tweeted was indeed consumed. Similarly, physical activity tweets may reflect a mixture of intentions, plans, and actual engagement in those physical activities. Also, exercise intensity for physical activities was assessed for 30 minutes of physical activity for an individual weighing 155 pounds, which can be an under- or overestimation depending on the type of activity and persons engaged in that activity.

The epidemic rise in obesity and related chronic diseases in recent decades signal the importance of structural forces and social processes, but the dearth of data on contextual factors limits the investigation of multilevel effects on health. Social media data can be uniquely harnessed to capture social and cultural processes with potential impacts on health [71,72,85-89]. For instance, public posts can be utilized to measure prevalent happiness which can impact health through emotional contagion and the interconnectedness between mental health and physical health. Additionally, public posts about health behaviors may help us understand the prevalence of those behaviors as well as local area social norms. We demonstrate that tweets can provide a means to assess prevalent sentiment and food behaviors and physical activity, which can inform health interventions and policies to meet the needs of different neighborhoods. In particular, as this study suggests, neighborhoods with social and economic disadvantage, high urbanicity, and those with more fast food restaurants may exhibit lower happiness and fewer healthy behaviors.