We examine the relationship between the content of a corpus of geo-tagged tweets (not retweets) and the obesity rate of 189 urban areas in the contiguous United States during the calendar years 2012 and 2013. Our data collection procedure adheres to Twitter’s terms of use/service. It uses Twitter’s streaming API which provides low latency access to Twitter’s global stream of Tweet data. The data we collected reflects a ∼ 10% random sample of all tweets in 2012–2013. From that random sample, 1.5% of the tweets were geo-tagged resulting in a corpus of over 25 million geo-tagged tweets. The geographic boundaries of the urban areas we explore reflect the MSAs defined by U.S. Office of Management and Budget. It is important to note that these urban area boundaries often agglomerate small towns together, particularly when there are small towns geographically close to larger towns or cities.

The obesity rates of the MSAs are provided by the 2012–2013 Gallup Healthways Wellbeing Survey. While other sources of geographic obesity rates exist (i.e. BRFSS and NHANES)[13, 14] we use the GHWS because its data was collected during the same time frame (2012–2013) as our Twitter corpus and (2) it measures other MSA-level variables related to happiness, diet and physical activity which allow us to evaluate additional aspects of our work (i.e. Question 4).

The relationship between these datasets is examined using six measures derived from our Twitter corpus: (a) one related to happiness, (b) three related to diet and (c) two related to physical activity. We define each of these measures next.

To quantify the happiness of a tweet we employ Mitchell et al.’s measure h_avg which reflects the happiness of a tweet. In previous work Mitchell et al. showed that the happiness of tweets are correlated with several population-scale measures including household income, education levels and the 2011 obesity rate in MSAs [12].

The happiness of a tweet is measured using the Language Assessment by Mechanical Turk (LabMT) word list, assembled by combining the 5,000 most frequent words occurring in each of four text sources: Google Books (English), music lyrics, the New York Times and Twitter. Ten thousand of these individual words have been scored by users of Amazon’s Mechanical Turk service on a scale of 1 (sad) to 9 (happy), resulting in a measure of happiness, h, for each given word [9]. For example, ‘rainbow’ is one of the happiest words in the list with a score of 8.10, while ‘earthquake’ is one of the saddest, with a score of 1.90. Neutral words like ‘the’ or ‘thereof’ tend to score in the middle of the scale, with h(the) = 4.98 and h(thereof) = 5.00 respectively.

For a given tweet T containing N unique words the average happiness, h_avg. is calculated by: havg(T)=∑i=1Nh(wi)fi∑i=1Nfi=∑i=1Nh(wi)pi (1)

In Eq 1, f_i is the frequency of the ith word w_i in T for which we have a happiness value h(w_i) and pi=fi/∑i=1Nfi is the normalized frequency of the word w_i.

To quantify the dietary content of the foods one tweets about we explore three different measures at varying degrees of granularity. Each of these three measures require that we partition our corpus of tweets using the following binary criteria: if a tweet contains a word(s) describing at least one food in the USDA National Nutrient Database (USDANDB) [15] it is placed in the Food Tweets set FT; otherwise it is placed in the Non-Food Tweets set NFT.

Given this partitioning, the Food Tweet % (FT%) of a MSA, is the ratio of Food Tweets in the MSA compared to the total number of tweets within the MSA. This reflects our first measure of diet and is shown in Eq 2. F T % = | F T | ( | F T | + | N F T | ) (2)

While, the FT% of a MSA quantifies how frequently people tweet about food, it does not offer any insight into the actual food about which people tweet. To measure how nutritious each food included in each tweet is we measure the average nutrient density, nd_avg, of the tweet by using the Nutrient-Rich Foods Index (NRF) formula [16].

While other formulae to determine the nutrient density of foods exist, we use the NRF because its’ scores have been shown to be highly correlated with the recommendations of the USDA’s Healthy Eating Index [17] and diets featuring high nutrient dense foods on the NRF have been been shown to reduce obesity, while diets consisting of low nutrient dense foods increase the prevalence of obesity [18, 19]. Furthermore the NRF is not restricted to any subset of foods. It is generalizable to any food in the USDANDB [20].

Nutrient density in the NRF is determined by computing the daily recommended intake value of protein, dietary fiber, vitamin A, vitamin C, vitamin E, calcium, magnesium, iron and potassium provided per 100 kCals of a given food and then subtracting the daily recommended intake values for saturated fat, sodium and added sugars in 100 kCals of the food. Using this formula, fruits and vegetables are some of the most nutrient dense foods (nrf(spinach) = 694.8; nrf(strawberries) = 375.9) while soda is one of the least (nrf(soda) = −55.8). For a given tweet T containing N unique foods we calculate the average nutrient density nd_avg using Eq 3. ndavg(T)=∑i=1Nnrf(foodi)fi∑i=1Nfi=∑i=1Nnrf(foodi)pi (3)

The calculation of nd_avg in Eq 3 is similar to the calculation of h_avg. In Eq 3 f_i is the frequency of the ith food food_i in T with NRF value nrf(food_i) and pi=fi∑i=1Nfi is the normalized frequency of the food food_i. The result is a measure of the average nutrient density of the foods mentioned in a single tweet.

There is a significant difference between the level of granularity in our first measure (FT%) and our second (nd_avg). To bridge this gap we formulate one more measure of the diet of an MSA: Produce % (Prod%). Prod% marries together the nutritional aspects of nd_avg with the coarse granularity of FT%.

Recall, fruits and vegetables are among the most nutritionally dense items on the NRF Index. Any tweet that mentions at least one food listed in either Fruits and Fruit Juices or Vegetable and Vegetable Products sections of the USDANDB is in set Prod. Given this partitioning, Prod% is the ratio of tweets in set Prod the compared to the total number of tweets in the MSA. This measure is shown in Eq 4. P r o d % = | P r o d | ( | F T | + | N F T | ) (4)

Along with happiness and diet, research has shown that the physical activity level of individuals affects obesity [21–23]. With this foundation we explore two different measures to quantify discussions of physical activity within our Twitter data set. Each of these measures require that we partition our corpus of tweets into those that discuss physical activities and those that do not. To do this partition we use a binary criteria similar to our food tweet criteria. If a tweet contains a word(s) discussing at least one physical activity in the guidelines for exercise testing published by the American College of Sports Medicine (ACSM) and the Center for Disease Control and Prevention (CDC) [24] it is placed in the Physical Activity Tweets set PA; otherwise it is placed in the Non-Physical Activity Tweets set NPA. While the guidelines for exercise published by the ACSM and CDC are not exhaustive and do not contain every possible physical activity descriptor we employ them in our work because they list over 400 activities and are well established. They been used by the American Heart Association [25], national cross-sectional studies [26] and public health recommendations [27].

Our first physical activity metric, Physical Activity % (PA%) is shown in Eq 5. It measures the ratio of Physical Activity Tweets compared to the total number of tweets. P A % = | P A | ( | P A | + | N P A | ) (5)

The guidelines of physical activities from the ACSM and CDC divides activities into two categories which serve as the basis for our second measure. The two categories of activities are: (1) moderately intense activities that burn 3.5 kCals a minute and (2) strenuously intense activities that burn 7.0 kCals a minute. Moderately intense physical activities include yoga, walking and stretching while strenuously intense physical activities include jogging, mountain climbing and aerobics. For a given tweet T discussing M moderately intense physical activities and S strenuously intense physical activities we calculate, pa_weighted in Eq 6. pa_weighted is the weighted number of calories burned by participating in all the physical activities discussed in the tweet for one minute. p a w e i g h t e d ( T ) = ( 3 . 5 × M ) + ( 7 . 0 × S ) (6)

All of the measures in Eqs 2–6 make no attempt to take the context of words or the meaning of a tweet into account. While this may limit the ability of our measures to appropriately score tweets containing only a few words, previous researchers have employed this approach and obtained reliable results. Furthermore, by ignoring the context of words we gain a degree of impartiality. We are not the one’s deciding a priori whether a given word, food or activity is associated with obesity. This strategy reduces experimental bias and maintains objectivity.

The first measure we explore is the happiness conveyed in individual words from tweets. Mitchell et al. showed that the happiness of tweets are correlated with the 2011 obesity rate in MSAs [12]. To validate this result we explore the correlation between the happiness of a tweet and the obesity rate of MSAs in our random sample of Twitter data. Recall, our Twitter data contains ∼ 25 million tweets collected during 2012 and 2013 while Mitchell et al.’s data contains ∼ 10 million tweets collected during 2011. Also Mitchell et al. used GHWS obesity rates collected during 2011 while we use obesity rates collected during 2012 and 2013.

Fig 1 shows the correlation of h_avg and the obesity rate in all the MSAs for: (a) 2011 (Mitchell et al.) and (b) 2012–2013 (our work). The data shows that the happiness people express in tweets generally decreases as the obesity rate increases. This result holds true in 2011 as well as in 2012–2013. Furthermore, the strength of the relationship and the subtleties of the data points are similar. For example, Boulder, CO is the city with the lowest obesity rate and is among the three most happy cities each year. Furthermore Beaumont, TX is in the top 10 MSAs in terms of obesity rate in both data sets and bottom five happiest cities. The Spearman correlation coefficients are similar (r = -0.339 in 2011, r = -0.318 in 2012–2013) and each have p-values far below.001 indicating that the negative correlations are statistically significant. Next, we explore the relationship of five measures of other factors affecting obesity (diet and physical activity) that can be gleamed from Twitter data in a manner similar to the happiness metric, h_avg.

Research has shown that diet influences obesity [28, 29]. However, the happiness metric, h_avg, does not account for diet. Many foods that are widely considered unhealthy have high happiness values (h). For example, the term cake has a h value = 7.58 Also, healthy foods can have relatively low happiness values. The term vegan has a h value of 4.82 despite reflecting a diet featuring fruits and vegetables. Furthermore, many healthy and unhealthy foods are not included in the list of terms scored for happiness. As a result, they are completely ignored in the previous analysis.

To gather insight into the relationship between the foods one tweets about and obesity we explore the correlation between three different measures of the dietary content of a tweet and the obesity rate of MSAs. The first measure we explore is nd_avg shown in Eq 3. Recall, nd_avg reflects the average nutrient density of a tweet. The twitter data we use for this analysis includes more than two million tweets from 2012–2013 mentioning more than six hundred of the 8,000 different foods listed in the USDANDB. The Spearman correlation between nd_avg and obesity rate in all MSAs over 2012–2013 is shown in Fig 2.

Fig 2 shows that there is not a statistically significant relationship between the nutrient density of the foods people discuss in their tweets and obesity rate. This result is unexpected. Given our previous result related to the happiness of tweets and the established relationship between diet and obesity, we anticipated a statistically significant negative correlation. We pursue an explanation by identifying the ten foods that are most strongly negatively and positively correlated with obesity. These results are shown in Table 1.

Table 1 elucidates several insights into the set of tweets that discuss food. The first is that areas with lower obesity rates do not exclusively discuss foods that are nutritionally dense. Similarly areas with high obesity rates discuss a mix of nutritionally dense and non-nutritionally dense foods. Specifically, both lists contain multiple foods with positive and negative NRF Index values and the food with the highest NRF Index value (collards) is correlated with high obesity rates.

It is important to note that our nutrient density metric ignores the quantity and preparation of the food consumed in the tweet. These limitations could explain the lack of a significant relationship between the nutrient density of foods people discuss in tweets and their obesity rate. However, the correlation coefficients and p-values in Table 1 reveal that tweets that discuss food, regardless of their nutritional density, are more likely to be negatively correlated with obesity rate than positively correlated. The absolute value of the correlation coefficient of the food tenth most negatively correlated with obesity is ∼ 25% larger than the absolute value of the correlation coefficient for the food most positively correlated with obesity rate. The p-values in Table 1 also reflect this trend. The relationship between all the foods negatively correlated with obesity rate are statistically significant (p <.05) while only the top four foods positively correlated with obesity rate are statistically significant.

Given these two observations we explore the data to see if the frequency with which individuals tweet about food, regardless of its nutritional density, is correlated with obesity. We use the same twitter data as our previous analysis. However, in this version we measure the ratio of Food Tweets compared to the total number of tweets. This metric, FT% is shown in Eq 2. The Spearman correlation between FT% and obesity over all MSAs for each is shown in Fig 3.

Fig 3 shows that the frequency with which people discuss foods in tweets generally decreases as obesity rate increases. For example, San Francisco, CA is the MSA with one of the highest FT% and is among the ten MSAs with the lowest obesity rate. Similarly, several of the MSAs with top twenty obesity levels (Flint, MI; Mobile, AL; Rockford, IL) are amongst the bottom twenty MSAs in terms of FT%. However, the negative correlation between FT% and obesity rate is not as strong as the negative correlation between h_avg and obesity rate. Furthermore, the negative correlation between FT% and obesity is not immediately obvious. There is not a quorum of established evidence that shows that the more people discuss food the less obese they are.

In order to examine our data further we explore our final measure of the diet of a MSA: Produce % (Prod%). Recall, Prod% marries together the nutritional aspects of nd_avg with the coarse granularity of FT%. It reflects the percentage of total tweets that discuss at least one of the foods listed in either the Fruits and Fruit Juices or Vegetable and Vegetable Products sections of the USDANDB. The twitter data we use for this analysis includes more than one million tweets from 2012–2013 mentioning more than 150 different fruits, vegetables or fruit/vegetable related products. The Spearman correlation between Prod% and obesity rate over all MSAs is shown in Fig 4.

Fig 4 shows that the Prod% metric reconciles the trends we saw in our previous explorations with the measures nd_avg and FT%. The frequency with which people tweet about fruits, vegetables or related products increases as obesity decreases.

Intuitively this makes sense. Fruits and vegetables are some of the highest scoring items on the NRF Index, so eating them regularly should decrease the obesity rate. The previous measure, nd_avg, attempted to account for this but over penalized tweeters for mentioning average and below average foods on the NRF Index. The FT% metric offered a much coarser level of granularity but did not consider the nutritional density of the foods being discussed in a tweet at all. By including nutritional density at a coarse level of granularity we are able to reveal a correlation with obesity rate (r = -0.344) that is similar in magnitude to the correlation between h_avg and obesity rate. Next, we investigate the discussion of physical activity levels on Twitter and their relationship to the obesity rate in MSAs.

Along with happiness and diet, research has shown that the physical activity level of individuals affects obesity [21–23]. However, none of our previously explored measures (h_avg, nd_avg, FT% and Prod%) account for discussions of physical activities within tweets. As a result, we explore two different measures of discussions of physical activity within our Twitter data set.

Our first physical activity measure, Physical Activity % (PA%) measures the ratio of Physical Activity related tweets compared to the total number of tweets. Our second measure weights physical activities according to the intensity levels published by guidelines of the ACSM and CDC. These two measures are shown in Eqs 5 and 6. The Spearman correlation between PA% and obesity rate and pa_weighted and obesity rate in all MSAs over 2012 and 2013 is shown in Fig 5(a) and 5(b).

The twitter data we use for this analysis includes more than three million tweets from 2012 and 2013 mentioning more than eighty of the physical activities listed by the ACSM and CDC. Almost two million tweets discuss forty-eight different activities of moderate intensity and more than one million tweets discuss thirty-six different activities of strenuous intensity.

The pa_weighted values of the tweets in our data set vary. The minimum is zero, which reflects a tweet that does not discuss any physical activities from the list published by the ACSM and CDC. The maximum pa_weighted observed in our data set is 24.5. However, over 99% of the tweets in our data set have pa_weighted values of either: 0, 3.5 or 7.

Fig 5 shows that there is a statistically significant negative correlation between both PA% and pa_weighted and the obesity rate in MSAs. However, the relationship between PA% and obesity rate is stronger (r = -0.330) than the relationship between pa_weighted (r = -0.190) and obesity rate. This result may seem unexpected. The pa_weighted metric offers the capability to combine the calories burned from multiple activities based on their intensity level. Given these additional capabilities one might expect it to correlate better with obesity rate than the basic PA% metric. To gather additional insight we calculate the activities most positively and negatively correlated with obesity rate in Table 2. Table 2 only includes five activities in each column because there are so few physical activities that have a positive statistically significant correlation with obesity rate.

Table 2 shows that areas with low obesity and areas with high obesity engage in twitter discussions of a mixture of moderately and strenuously intense activities. Both lists include three moderately intense activities and two strenuously intense activities. However, Table 2 also shows that areas with lower obesity rates simply tweet more about physical activities than areas with high obesity rates. The absolute value of the correlation coefficient for the fifth most negatively correlated activity is higher than the absolute value of the correlation coefficient for the activity most positively correlated with obesity.

It is important to note that our physical activity measures ignore if an individual’s discussion of an activity reflects them physically engaging in it or merely witnessing it in some manner. The inability to make this distinction could explain the lack of a more significant relationship between the intensity levels of physical activities and obesity rate.

However, these insights do reveal similarities between the measures: (1) nd_avg and Prod% and (2) pa_weighted and PA%. In both cases adding too much detail to the measure derived from tweets diluted the relationship between the quantities of interest. This is a valuable lesson learned. Given the complexity of Mitchell et. al.’s happiness metric, h_avg, we assumed we would need measures of discussions of food and physical activities with a similar structure. However, this is not the case. The more coarse measures Prod% and PA% had a stronger relationship to obesity rate than the nuanced measures nd_avg and pa_weighted. Next, we explore the extent to which these measures provide different insight about the obesity rate of a MSA and evaluate the extent to which each correlates with a a MSA-level survey measure of a similar variable.

The results we have presented thus far demonstrate that three measures (h_avg, Prod% and PA%) which can be obtained from geo-tagged tweets have a statistically significant negative correlation with the obesity rate of a MSA and that correlation is on the order of -0.30. However, we have not presented any results which show that: (1) the three measures (h_avg, Prod% and PA%) have unique relationships with the obesity rate of a MSA and (2) the measures actually quantify the happiness, diet and physical activity level of a MSA.

We address both of these questions by computing the correlation among seven variables. Three of the seven variables are the measures of happiness, diet, and physical activity that can be gleamed from Twitter discussions within a MSA and are most correlated with obesity rate: h_avg, Prod% and PA%. The other four variables reflect MSA-level data collected by the GHWS survey data. These variables are the: (1) obesity rate of a MSA, (2) percentage of individuals in a MSA who report that they eat a healthy diet, (3) percentage of individuals in a MSA who report that they exercise frequently and (4) Well-Being Index of a MSA. The Well-Being Index is computed by aggregating the responses from participants to five statements. Each participant rates their agreement with each statement on a 0 (very strong disagreement) -10 (very strong agreement) scale. The statements are [7]:

I am satisfied with my present life situation and anticipated life situation.

Fig 6 visualizes the lower triangle of a matrix of the Spearman correlations among the seven variables. The blue boxes in Fig 6 reflect a positive correlation, red boxes reflect a negative correlation. This data shows that each of the measures we computed from Twitter discussions within a MSA (h_avg, Prod% and PA%) are more correlated with the obesity rate of a MSA than they are correlated with any of the other measures computed from Twitter data. This provides evidence that each of the three measures reflect different factors which are correlated with the obesity rate of a MSA. In other words, these three measures are not simply different methods of quantifying the same variable.

Furthermore, each of the three measures gleamed from our Twitter corpus is more correlated with the MSA-level measure of a similar variable from GHWS than any other variable. To help elucidate this trend we have underlined the correlation coefficient of the variables with the strongest correlation to happiness, Prod% and PA%. While, this trend does not completely rule out the existence of confounders within our Twitter-level measures, it provides evidence that h_avg, Prod% and PA% are actually reflecting the level of happiness/well-being, diet/healthy-eating and physical activity/exercise within a MSA as opposed to three completely unrelated variables. Next, we review related work, discuss the validity and limitations of our results and provide directions for future work.