Physical Activity Tweets on Audience
Engagement
Ying Lin
Student Number:10171800
Thesis Supervisor
Dr.Christin Scholz
Graduate School of Communication
University van Amsterdam
Abstract
Some researchers have believed that social media engagement in physical activities can contribute to offline engagement in physical activities. It is well established that there are robust relationships between emotionality (specifically valence and arousal of con-tent) and audience engagement on social media, but the directionality of these effects is not clear. This study reconsiders the relationship between arousal and valence and engagement with physical activity posts, and considers the moderating effect from in-tensity of exercise. 32354 tweets were sampled from 20 Twitter accounts related to high intensity exercise and 20 Twitter accounts related to low intensity exercise. The results of multilevel regression indicated that there was an interaction effect of intensity and arousal. More specifically, lower arousal predicted more retweeting in low intensity exercise-promoting tweets whereas higher arousal predicted more retweeting in high intensity exercise- promoting tweets. No effects of valence were found. In sum, only arousal predicted information-sharing behaviors in physical activities promotion on so-cial media, and the effect is moderated by intensity of exercise.
Keywords: emotion, physical activities, social media, intensity of exercise
Insufficient Physical activity
Physical activity contributes to personal health by consuming energy (Foster, Shilton, Westerman, Varney, & Bull, 2018). During the last few decades, insufficient physical activity among people, resulted from the prevalence of sedentary lifestyle, has been a major global problem given its strong association with the incidence of cancers, diabetes and cardiovascular diseases. Insufficient physical activity thus causes a high burden on individual well-being and the health-care systems. According to Foster et al. (2018), 25 percent of the total population engages in enough physical activity. In particular, more than 80 percent of the global adolescent population does not meet the recommended minimum levels of physical activity. A meta-analysis from Maher et al. (2014) has shown that social media-based interventions have significant effects on health-related behavior
change including physical activity promotion, and social media engagement behaviors can be considered as an important indicator of the effectiveness of social media-based interventions (Dolan, Conduit, Fahy, & Goodman, 2016). Moreover, previous studies have identified that the emotions expressed by posts can lead to the increase of social media engagement (e.g Poecze, Ebster, & Strauss, 2018), but the directionality of those effects are mixed and need more investigations. It can be assumed that one source of this inconsistency is the moderating effect of content type (Dolan et al., 2016). Therefore, applying the affective circumplex model, this study attempts to test the emotional effects in the context of physical activity by comparing effects of arousal and valence between Twitter accounts with high-and low intensity exercise.
Social Media-Based Interventions and Twitter
As information technology developed and shaped society rapidly, social media such as Facebook, Twitter and Instagram has been an integral part of many people’s daily life. From uses and gratification theory perspective, Whiting and Williams (2013) pointed out that social media users utilized social media for many purposes, including information seeking, information sharing, social interaction, passing time, entertainment, relaxation, expression of opinions. Recent studies have identified the effectiveness of Internet-based or social media-based exercise promotion (Davies, Spence, Vandelanotte, Caperchione, & Mummery, 2012; G. Williams, Hamm, Shulhan, Vandermeer, & Hartling, 2014). It implies a large potential for social media to be a platform for promoting physical activities in a large number of people.
In particular, Twitter has been considered as an effective tool for health promotion considering its effectiveness to spread health-related information (Myers & Leskovec, 2014). It allows people to interact with others by generating or sharing short texts. In particular, the retweet mechanism can spread the information to not only friends and followers, but also strangers (Kwak, Lee, Park, & Moon, 2010), and the retweeting be-havior is considered as a type of social media engagement bebe-havior (Dolan et al., 2016).
Therefore, in this case, the number of retweeting can be considered as the indicator of audience’s social media engagement in physical activities.
Considering its effectiveness to reach a broader audience, some organizations or in-dividuals use twitter to provide up-to-date information to communicate health-related issues with their audiences (e.g Park, Rodgers, & Stemmle, 2013). For example, a con-tent analysis investigating how local health departments use Twitter for health commu-nication showed that, apart from posting information related to their events and services, health departments also provide useful suggestions for improving health, encourage peo-ple to take actions for health benefits (Neiger, Thackeray, Burton, Thackeray, & Reese, 2013). Apart from getting information from specific accounts, people can search for in-formation they need by using queries. Moreover, Fox and Jones (2009) pointed out that such a searching process can change how people think about physical activity. How can we best make use of the potential of the Twitter platform to increase physical activity?
Social marketing has been concerned as an effective tool in promoting physical ac-tivities by some researchers (Gordon, McDermott, Stead, & Angus, 2006). It considers physical activity as a product and applies the commercial marketing techniques to pro-moting behavior changes in target audience and thus improve their well-being. Since the last decade, it can be observed that many social media communities that focus on health and fitness, offer a series of fitness activities and diet as an efficient method that can help people build ‘fit’ body, which always refer to well-trained, fat-free, slimmed body. More-over, when generating information, they always associate ‘fit’ body and health. More specifically, they present an idea that the ‘health’ can be achieved through governing body through regular physical activities and diet (Andreasson & Johansson, 2014). And those efforts are identified to be effective in promoting physical activities among adults (Xia, Deshpande, & Bonates, 2016). Suarez-Almazor (2011) explained that compared with other interventions that focus on facilitating rational and cognitive responses, the social marketing intervention may pay more attention to emotional and social response. In other words, the promotion of physical activity in social marketing setting may be more likely to influence people through evoking emotion that may motivate people to
participate in physical activity, and refining physical activities as desired behaviors for improving health.
What makes physical activity-related posts on Twitter successful? A previous study has identified some content characteristics of message such as argument quality can in-crease people’s social media engagement behavior (Chang, Yu, & Lu, 2015), which can refer to different types of responses such as like, share and comment to messages (Dolan et al., 2016). Although the empirical evidence of positive effects from social media en-gagement behavior on physical activity is ambiguous (Althoff, Jindal, & Leskovec, 2017), many researchers still believe that social media engagement behavior has potential to influence health behavior change, just needs more investigations (e.g. Heldman, Schin-delar, & Weaver, 2013; Keller, Labrique, Jain, Pekosz, & Levine, 2014). In addition, many previous studies have identified the impacts from different types of text messages that promote participation in physical activity (e.g. Buchholz, Wilbur, Ingram, & Fogg, 2013; Bull, Kreuter, & Scharff, 1999; Latimer, Brawley, & Bassett, 2010; Latimer, Rench, et al., 2008; Latimer, Rivers, et al., 2008; Priebe & Spink, 2012; Smeets, Brug, & de Vries, 2006). Accounting for those considerations, the following sections attempts to identify some features that may contribute to the virality of physical activities promoting tweets.
The Role of Emotion
Emotion has been considered as a potential driver in facilitating social media engage-ment behaviors and thus motivate people to participate in physical activity in a social marketing setting. Bagozzi, Gopinath, and Nyer (1999) pointed out that consumer may conduct different behaviors in response to anticipated emotions. Furthermore, when gen-erators intend to evaluate certain topics, express their opinions or just show their mental states, they may integrate emotions into their posts (Stieglitz & Dang-Xuan, 2013). In addition, Harris and Paradice (2007) indicated that emotions can be transferred through computer-mediated communication (CMC) when message receivers associate message content with emotions or use emotional cues. Once they received the emotions involved
in social media content, they may react to them via sharing content on the social me-dia. In addition, previous studies found that different exercise can trigger different emo-tions and those emoemo-tions can affect following exercise’s participation or maintenance (e.g. Smith, Crabbe, et al., 2000; Vallerand & Blanchard, 2000; D. M. Williams, 2008). Therefore, it can be assumed that the relationship between emotionally charged physi-cal activities-promoting tweets and information sharing may be influenced by different types of exercise.
To set the theoretical foundation of this study, the researcher reviews literatures re-garding emotional valence and arousal effects, and the moderating effects from different types of exercise. More specifically, the researcher attempts to link these effects with peo-ple’s information sharing behaviors. Considering the effectiveness of computer-mediated communication (CMC) in transferring emotions (e.g. Harris & Paradice, 2007; Riordan & Kreuz, 2010), the researcher can argue that the emotions transferred by social media content related to promoting different physical activity can affect audience’s social media engagement behaviors differently and thus provide evidences for conducting sentiment analysis.
Affective Circumplex Model
Emotions involved in the text messages can be transferred to audiences and subsequently influenced their emotions (Hatfield, Cacioppo, & Rapson, 1993). Therefore, it can be as-sumed that social media content may activate audiences’ more information sharing be-haviors considering its emotional effects. Previous study indicated that the transferred emotions, in turn, influence people’s information sharing behaviors (Kramer, Guillory, & Hancock, 2014). As a result, many scholars have begun to investigate the effects from different types of emotions on information sharing behaviors. The affective circumplex model proposed by Russell (1980) may shed light on the interaction between different types of emotions and responses to messages. The application of this model allows all affective states to be characterized by two independent dimensions: valence (sentiment)
and arousal, which represent polarity and intensity of the content respectively. There-fore, any emotions can be cateogrized according to valence and arousal.
Existing work on the effects of valence and arousal on social media engagement is inconclusive. Rozin and Royzman (2001) suggested that there should be a negativity bias in which people tend to give more weight on negative events for its involved poten-tial risks. In other words, people’s attitudes or behaviors are more likely to influenced by negative input than positive input (Ito, Larsen, Smith, & Cacioppo, 1998). As a re-sult, negative information may be more viral than positive information (e.g. Tsugawa & Ohsaki, 2015, 2017).However, not all findings from previous studies are consistent with this proposition. For example, two prior studies from Barasch and Berger (2014); Berger and Milkman (2012) found that articles written with positive emotions were more viral than articles written with negative emotions. Another study exploring informa-tion diffusion on Twitter found that news that containing negative emoinforma-tions was more retweeted whereas non-news content that containing positive emotions were more pop-ular (Hansen, Arvidsson, Nielsen, Colleoni, & Etter, 2011).
In the context of physical activity, a study from Pasco et al. (2011) reported that there was a positive link between positive emotions and level of habitual physical activity but no association between negative emotions and physical activity. Therefore, it is possible to post the first hypothesis as follows:
H1: Physical activities promoting messages which express a more positive sentiment are more likely to be shared on Twitter.
Berger and Milkman (2012) also argued that psychological arousal is another driver of information diffusion. More specifically, Berger (2011); Dunlop, Wakefield, and Kashima (2008) found that more arousing messages can increase more information sharing be-haviors than less arousing messages, regardless of emotional valence. Therefore, if the physical activity-related messages induce high levels of arousal, they may activate more information sharing behaviors. In other words, people induced to feel emotions with high arousal such as happy or anxiety are more likely to share physical-related messages than people induced to feel emotions with low arousal such as calm or depressed. Therefore,
a positive relationship was expected between the emotional arousal and the number of retweeting:
H2: Physical activities promoting messages with high level of arousal are more likely to be shared on Twitter.
There have been different concerns about the interaction between valence and arousal. Some scholars considered their contributions to emotional effects independently (e.g. Russell & Barrett, 1999) whereas others argued that some combinations of valence and arousal can enhance emotional effects (e.g. Eder & Rothermund, 2010). For example, a study from Kuppens, Tuerlinckx, Russell, and Barrett (2013) investigating people’s affective experience with laboratory experiments found that when people experienced more positive or negative emotions, they were more likely to experience a high level of emotional arousal simultaneously. When it comes to emotional effects on informa-tion diffusion on social media, many previous studies focus more on valence and arousal separately, regardless of their interactions (e.g. Berger & Milkman, 2013; Stieglitz & Dang-Xuan, 2013). For more understandings about emotional effects, in this case, it is necessary to investigate the effects from the interaction between valence and arousal evoked by physical activities promoting content:
RQ1: What are the effects of valence and arousal of physical activities promoting content on Twitter on information sharing?
On the basis of previous literature review, there is considerable inconsistency in exist-ing research on the effect of emotional valence on social media engagement. What may explain these inconsistent findings?
One possible explanation is that different types of emotions are positively linked to sharing in different contexts. In the context of health communication, expressed emo-tions have been considered as a key predictor of information diffusion on the social media (e.g. Kim, 2015; Kim, Lee, Cappella, Vera, & Emery, 2013; McLaughlin et al., 2016). Kim et al. (2013) found that anti-smoking messages containing positive sentiment were more likely to be spread whereas McLaughlin et al. (2016) found that HIV prevention- related tweets containing negative sentiment were more likely to be spread. Whether positive or
negative emotions lead to higher level of information sharing may depend on different health-related behaviors.
Moderated by Intensity of Exercise
When it comes to physical activities, many previous studies have associated different types of emotions and intensity of exercise (e.g. Kerr & Kuk, 2001; Kilpatrick, Kraemer, Bartholomew, Acevedo, & Jarreau, 2007; Steptoe & Bolton, 1988). Despite the mixed findings, it has been widely believed that compared to low intensity exercise, high in-tensity exercise is more likely to elicit negative feelings with such as tension or anxiety during and after exercise. Therefore, it can be assumed that high intensity exercise such as high intensity interval training may trigger more negative emotions whereas low in-tensity exercise such as yoga and pilates may trigger more positive emotions.
Furthermore, each physical activities-promoting tweet, in some sense, can be consid-ered as an eliciting episode (Rimé, 2009). Once exposed to such an emotion-eliciting episode, people may share the episode with other people. It can be described as a process of social sharing of emotion (Christophe & Rimé, 1997; Rimé, Finkenauer, Lu-minet, Zech, & Philippot, 1998). Moreover, Rimé (2009) pointed out that such a process can be explained by the perception-action model of empathy (PAM) (Preston, 2007). According to the perception-action model of empathy (PAM), people will automatically activate representations, which can be understood as a particular response from the brain and body to a particular state, to understand and feel the emotions conveyed by the tweets. If their representations activate similar emotions to the emotions expressed by tweets, the empathy will be activated. Moreover, (Lee, Kim, & Kim, 2015) pointed out that empathy is a strong motivation of retweeting behaviors.
More specifically, Firdaus, Ding, and Sadeghian (2018) argued that Twitter users retweet tweets not only for sharing information but also for sharing emotions. More specifically, they may retweet tweets that expressed their emotions towards specific top-ics. In this case, it may happen that Twitter users retweet physical activities tweets
ex-pressing emotions they experienced when taking similar fitness activities. For example, if a person participated in high intensity interval training before, he or she may be more likely to retweet high intensity interval training-related tweets that involved negative emotions. Therefore, it can be assumed that the relationships between emotional effects and retweeting behaviors are moderated by intensity of exercise. In detail, this study listed two hypotheses as follows:
H3a: Higher valence leads to more retweeting, but this effect is stronger in accounts related to low intensity exercise compared to accounts related to high intensity exercise. H3b: Higher arousal leads to more retweeting, but this effect is stronger in accounts related to high intensity exercise compared to accounts related to low intensity exercise. To test the moderating effects of intensity of exercise, the following section will com-pare effects of arousal and valence between Twitter accounts with high-and low inten-sity exercise. Findings from such a comparison may provide a deeper understanding for the relationship between emotionally charged physical activities-promoting tweets and information sharing behaviors on Twitter. And findings may also imply that physi-cal activities promoters should consider the associations between different emotions and different types of exercise when designing physical activities-promoting messages.
Methodology
Sampling
Twitter accounts with high-and low intensity exercise often involve various emotions into messages that advertise for product or services related to exercise, provide benefits from recommended exercise and diet, plan for exercising, diet, general attitudes towards physical activities, inspiration, representations of ‘fit’ bodies. For example, a tweet ‘ You don’t have to be strong to FEEL strong. Cultivate your strength. ’ reflects high positivity whereas a tweet ‘ Feeling shocked that the Japan disaster keeps getting worse... Video footage of blast at Japan nuclear power plant ‘reflects high negativity. And a tweet‘ Stay focused. Your power lies in this moment where you can take action to create your future’
reflects a high level of arousal whereas a tweet ‘ Surround yourself with those who lift you higher ’reflects a low level of arousal.
For testing moderated effects from intensity of exercise, this paper decided to sample Twitter accounts related to high intensity exercise and Twitter accounts related to low intensity exercise respectively. Twitter provide an official search service ‘who to follow’ service for users to search for topic-related accounts based on queries. This service rec-ommended ‘experts’ based on several factors including users’ profile information, users’ social interactions and users’ engagement in Twitter. To build a sampling list, the study used the service with the queries as Table 1 showed. As Zafar, Bhattacharya, Ganguly, Gummadi, and Ghosh (2015) mentioned, Twitter provided several metrics for ranking users including the number of a user’s followers (follower-rank), the popularity of a user in the social network and the number of times the user was listed by other users (List-rank). In particular, the list function launched by Twitter provides a way for users to gather Twitter accounts that focus on specific topics. And the names and descriptions of the lists may help us infer whether specific account focus on specific topic or not (Ghosh et al., 2013).
High Intensity Exercise Low Intensity Exercise high intensity interval training yoga
strength training pilates body weight training mindfulness functional fitness meditation weight loss mobility core training
circuit training zumba
crossfit running
Table 1: Queries for ’who to follow’
In implementation, the researcher applied those queries and got 15 lists of Twitter accounts. Then the researcher limited the sample by selecting the accounts if they pass the threshold of follower’s number or listed times. More specifically, the researcher only
chose Twitter accounts which were listed more than ten times on physical activities-related topics by other users or have more than 10K followers. After all, 20 Twitter accounts related to high intensity exercise and 20 Twitter accounts related to low inten-sity exercise were chosen. Because of the Twitter API’s limit, this study can only collect the last 3200 tweets from each public account. In addition, Some Twitter accounts may provide much less than 3200 tweets. Some of them did not generate or share more than 3200 tweets. And some others protected their tweets and only provided a few tweets for collection. In this case, 7 Twitter accounts provided much less than 3200 tweets with a range from 126 tweets to 2336 tweets. Finally, 116929 tweets from 40 accounts were collected. In detail, 57782 tweets were collected from Twitter accounts with high in-tensity exercise whereas 59147 tweets were collected from Twitter accounts with low intensity exercise.
It should be mentioned that not all tweets from those accounts are relevant to physi-cal activities. Therefore, the researcher consolidated the sample by removing irrelevant posts. Several studies have shown that there may be a number of words frequently oc-curred in the physical activities-promoting posts (e.g Jong & Drummond, 2016; Markula, 1997). Therefore, firstly, the researcher built a list with queries from Table 1 and some frequent words such as health and fitness. Then the researcher selected tweets that con-tain any words in the list by writing a python script. In total, 36199 tweets were selected for the following sentiment analysis.
In addition, Thelwall (2016) mentioned that the number of retweeting is measured by the number of times a tweet has been retweeted. Therefore, it may be difficult to identify the accurate number of retweeting of a tweet retweeted by specific Twitter ac-counts. Consequently, the emotional effects on retweeting from tweets retweeted by sampled Twitter accounts have a high probability to be distorted. To accurately estimate the relationship between emotion and retweeting, only tweet generated from sampled Twitter accounts were chosen. Finally, 3749 retweeted tweets were removed from the data, and 32450 tweets were kept for the following sentiment analysis.
Sentiment Analysis
Sentiment analysis can be considered as a task that automatically determine people’s emotions, opinions and attitudes towards specific topics, products, individuals, organiza-tions or events by analyzing the texts (Liu, 2012). In implementation, it always refers to automatic language processing techniques developed on different assumptions. It should be mentioned that some sentiment analysis techniques may be developed on the basis of emotion models (Alm, Roth, & Sproat, 2005). For instance, Bradley and Lang (1999) developed an affective word list with 1034 English words, called ‘Affective Norms for En-glish Words (ANEW)’, on the assumption that a particular word can be characterized by three affective measures (valence, arousal and dominance). More specifically, each word was rated by a scale from 1 to 9, where 1 represents the most negative/least arousal/least controlling and 9 represents the most. In other words, items rated around 1 or 9 can be considered as most negative or positive words whereas items rated around 5 can be considered as neutral words. Warriner, Kuperman, and Brysbaert (2013) extended the word list to 13915 words, and thus provide a much richer resource for researchers to estimate the sentiments expressed by texts. Considering the inconsistencies in annota-tions originated from rating scales, Mohammad (2018) applied a new scale to rating more than 20,000 English words with an interval from 0, which represents the lowest valence, arousal and dominance) to 1, which represents the highest valence, arousal and dominance, by repeatedly comparing four words at a time. For each word in the list, its score is the linear transformed result of the proportion of times the word was chosen as the item with the highest valence, arousal or dominance minus the proportion of times the word was chosen as the item with the lowest valence, arousal or dominance. Nor-mally, it works by finding all emotion rating for words in the document. In other words, it is a word-by-word approach for sentiment analysis.
However, apart from normative words, the emotional valence involved in tweets are influenced by grammatical and syntactical conventions. Moreover, it can’t be ignored that emojis, slangs and emoticons played a significant role in expressing or emphasizing emotions, especially in the social media texts. Considering those limitation, Hutto and
Gilbert (2014) combined those features and subsequently developed the Valence Aware Dictionary for sEntiment Reasoning (VADER) for sentiment analysis of social media texts. In implementation, it returns several sentiment scores including positive score, negative score, neutral score and compound score for each document, with the interval from -1 to 1. In particular, the compound score represents a weight average of the positive, negative and neutral scores. More specifically, -1 represents the most negative sentiment whereas 1 represent the most positive sentiment. And the neutral sentiment is assigned around the midpoint of the interval. In implementation, as Hutto and Gilbert (2014) mentioned, if the compound score was between -0.05 and 0.05, the sentiment involved in the tweet can be categorized as neutral sentiment.
However, the VADER package only computes the valence value for each tweet. In ad-dition, most sentiment analysis techniques, especially lexicon-based techniques are de-veloped for identifying polarity or computing emotional valence in text (Hutto & Gilbert, 2014). Maybe the machine learning approach can be applied to computing arousal value for each tweet on the basis of training dataset with validated emotional lexicons (Neethu & Rajasree, 2013). However, it is difficult to acquire training dataset consisting of physi-cal activities promoting social media posts. Comparatively speaking, ANEW words is an effective way to compute arousal value for each tweet.
Based on those consideration, this study decided to use the VADER package and the ANEW list developed by Mohammad (2018) for rating valence and arousal for each fitness-relevant tweet respectively. Finally, 32354 tweets were rated with valence and arousal scores.
Regression Analysis
To test those hypotheses and answer the questions which postulated linear relationships between emotional effects and retweeting behaviors, the researcher determined to ap-ply regression analysis to all the data with the number of times a tweet was retweeted (retweet_count) as dependent variables, and computed valence value and arousal value
as independent variables.
In addition, Suh, Hong, Pirolli, and Chi (2010) showed that some features of tweets such as inclusion of URLs, inclusion of hashtags and number of followers had signifi-cant influences on people’s retweeting behaviors. And Jenders, Kasneci, and Naumann (2013) indicated that the inclusion of mentions can also increase retweeting. Those variables influenced retweeting in different ways. For instance, using hashtags can cate-gorize the tweets so people can find tweets they are interested in and some of them may share tweets afterwards. A tweet from a user with more followers may have a higher probability to be received by more people and thus increase its probability to be shared (Jenders et al., 2013). Therefore, this study also included hashtags, user_mentions, URLs and the number of followers as independent variables. In manipulation, hashtags, user_mentioned, and URLs were coded as binary variables depending on whether hash-tags, mentioned users or URLs were included in a tweet or not whereas the number of followers was coded as a numerical variable. If a tweet involved hashtags, mentioned users or URLs, it would be code ‘1’, otherwise it would be coded ‘0’.
Before applying regression analysis, it should be mentioned that many tweets may be nested with specific Twitter accounts. Consequently, the number of retweeting may be more correlated with specific accounts rather than the effects from valence value and arousal value (Peugh, 2010). It violated the assumption of independence for observations between groups and within groups and may lead to inaccurate estimations of regression coefficients when running traditional regression models such as ordinary least-squares (OLS) multiple regression. Austin, Goel, and van Walraven (2001) suggested that ap-plying multilevel regression can result in a more valid inference drawn from the data because it considers variations among specific Twitter accounts through estimation of random effects. In implementation, those random effects can be manipulated by adding random intercepts, which reflect the differences of outcome variables across Twitter ac-counts, or random slopes, which reflects the differences of effects from predictors across Twitter accounts, or both (Bauer, Preacher, & Gil, 2006). In this case, it may be arbitrary to postulate that emotional valence and emotional arousal influenced retweeting
differ-ently across Twitter accounts without any theoretical and empirical support. Therefore, the multilevel regression model only added a random intercept for each Twitter account. In addition, to distinguish effects from specific Twitter accounts from emotional ef-fects, a mean centering approach was applied to produce predictor variables in this model (Paccagnella, 2006). In implementation, the mean centered valance values were trans-formed by subtracting the mean across all valence values from the valence value of each tweet. And the mean centered arousal values were manipulated in the same way. Finally, the valence value and arousal value were replaced by mean centered valence value and mean centered arousal value respectively.
In conclusion, this study used multilevel regression approach to predict the number of retweeting by using mean centered valance value and mean centered arousal value, adding a random intercept for each Twitter account, and controlling the number of fol-lowers, hashtags, user mentioning and urls.
Results
Before reporting regression results, the result section starts with descriptive analysis for variables in the model.
Table 2 showed the distributions of tweet’s features including frequencies of using hashtags, frequencies of using mentions, frequencies of including URLs for Twitter ac-counts related to high intensity exercise and Twitter acac-counts related to low intensity exercise sample. It seems that there are some differences of features between those types of Twitter accounts. More specifically, Twitter accounts related to high intensity exercise involved more urls (n=11395) than Twitter accounts related to low intensity exercise (n=7176) in their tweets. Twitter accounts related to low intensity exercise (n=11977) added more hashtags than Twitter accounts related to high intensity exercise (n=5412) in their tweets.
As Table 3 indicated, there seemed some significantly statistic differences in relevant variables for Twitter accounts related to high intensity exercise and Twitter accounts
exercise with high intensity exercise with low or moderate intensity features number of tweets number of tweets
hashtags 5412 11977 mentions 3030 4295 urls 11395 7176 total 15392 17012
Table 2: Distributions of tweet’s features
related to low intensity exercise sample. For independent variables, there are also sig-nificant differences in valence value and arousal value for both samples. On average, a tweet from Twitter accounts related to high intensity exercise (M=0.28, SD=0.40) had a higher valence value than a tweet from Twitter accounts related to low intensity exer-cise (M=0.23, SD=0.41) (t (36077) = 11.89, p=.00). A tweet from Twitter accounts related to high intensity exercise (M=3.63, SD=2.00) also had a higher arousal value than a tweet from Twitter accounts related to low intensity exercise (M=3.45, SD=1.73) (t (36077) =8.64, p=.00). Similarly, Twitter accounts related to high intensity exercise (M= 354036.13, SD= 566214.86) had more followers than Twitter accounts related to low intensity exercise (M= 183618.83, SD= 337605.94).
As H1 and H2 mentioned, this study expected, the emotional valence and emotional arousal are expected to predict the increase of the retweeting for the whole data. In other words, a tweet with higher valence (arousal) value should produce a greater num-ber of retweeting. The results indicated that the main effect from emotional valence ap-pears to be insignificant (β=.65, SE=.42,p=.013,CI=[-0.18, 1.48]) but the main effect from emotional arousal appears to be significant (β=-.70, SE=.10, p=.000, CI=[-0.89,-0.50]). It implies that a physical activity promoting tweet with lower level of arousal may be more likely to be shared on Twitter. This study also found that interaction between emotional valence and emotional arousal appears to be significant and positive (β=.67, SE=.21, p=.002, CI= [0.25, 1.08]). Moreover, there appears to be a significant in-teraction between arousal value and types of account (β=.71, SE=.14, p=.000, CI= [0.44, 0.98]). It may be necessary to discuss whether emotional valence and emotional arousal influenced the number of retweeting differently between the tweets from Twitter accounts related high intensity exercise and tweets from Twitter accounts related low
exercise with high intensit y exercise with lo w or m oderate intensit y tot a l Mean St andard Devi ati o n Mean St andard Devi ati o n Mean St andard Dev ati o n Dependent V ari a b le n umber o f ret w eeting 5.66 16.42 19.60 33.89 12.99 27.93 Independent V ari a b les v a len ce 0.28 0.40 0.23 0.41 0.25 0.41 aro u sa l 3.63 2.00 3.45 1.73 3.54 1.86 ha sht ags(d ummy) 0.35 0.48 0.70 0.46 0.54 0.50 u ser_menti o ns(d ummy) 0.20 0.40 0.25 0.43 0.23 0.42 urls(d ummy) 0.74 0.44 0.46 0.50 0.59 0.49 n umber o f foll o w ers 354036.13 566214.86 183618.83 337605.94 247058.58 451767.74 T a b le 3: Descripti ve st atisti cs for rel ev ant v ari a b les in regressi o n m odel
intensity exercise.
Following the third hypothesis, the researcher expected the lower valence value can predict more retweeting in tweets from Twitter accounts related to high intensity ex-ercise and higher valence value can predict more retweeting in tweets from Twitter ac-counts related to low intensity exercise. The results showed that the emotional arousal has a negative effect on the number of retweeting for tweets from Twitter accounts re-lated to low intensity exercise (β=-.27, SE=.12, p=.029, CI=[-0.51,-0.03]), but has a positive effect for tweets from Twitter accounts related to high intensity exercise(β=.27, SE=.06, p=.000, CI=[0.14, 0.39]), and emotional valence does not have any impacts on retweeting for both samples(β=.64, SE=.51, p=.211, CI=[-0.36, 1.64] and β=.47, SE=.31, p=.125, CI=[-0.13, 1.07]). Therefore, only the effects from emotional arousal were moderated by the intensity of exercise and the results would be discussed in the next section.
When it comes to other independent variables, in general, both user mentioning and inclusion of urls had negative effects on predicting increase of retweeting, which are opposite with previous findings.
witho ut interacti o n β SE p F df p R ˆ 2 types o f acco unt(d ummy(l o w intensit y)) -1.61 2.73 .554 210.68 10, 32344 .000 .43 aro u sa l -0.71 0.10 .000 aro u sa l:t ypes o f acco unt(d ummy(l o w intensit y)) 0.76 0.14 .000 v a len ce 0.99 0.41 .015 v a len ce:t ypes o f acco unt(d ummy(l o w intensit y)) -0.28 0.32 .645 ha sht ags 0.27 0.32 .389 u ser_menti o ns -6.03 0.31 .000 urls -11.80 0.31 .000 n umber o f foll o w ers 0.00 0.00 .000 with interacti o n types o f acco unt(d ummy(l o w intensit y)) -1.65 2.73 .546 176.72 12, 32342 .000 0.43 aro u sa l -0.69 0.10 .000 aro u sa l:t ypes o f acco unt(d ummy(l o w intensit y)) 0.71 0.14 .000 v a len ce 0.65 0.42 .125 v a len ce:t ypes o f acco unt(d ummy(l o w intensit y)) 0.14 0.62 .819 v a len ce:aro u sa l 0.67 0.21 .002 v a len ce:t ypes o f acco unt(d ummy(l o w intensit y):aro u sa l -0.30 0.30 .300 ha sht ags 0.24 0.32 .442 u ser_menti o ns -5.98 0.31 .000 urls -11.81 0.31 .000 n umber o f foll o w ers 0.00 0.00 .000 T a b le 4: R esults o f the Multil ev el R egressi o n An a ly sis (f or a ll the dat a)
witho ut interacti o n β SE p F df p R ˆ 2 v a len ce 0.97 0.49 .048 294.20 7, 17005 .000 .44 aro u sa l -0.29 0.12 .020 ha sht ags 0.93 0.56 .093 u ser_menti o ns -8.38 0.49 .000 urls -19.16 0.50 .000 n umber o f foll o w ers 0.00 0.00 .000 with interacti o n v a len ce 0.63 0.51 .211 258.30 8, 17004 .000 .44 aro u sa -0.27 0.12 .029 v a len ce:aro u sa l 0.65 0.26 .011 ha sht ags 0.90 0.56 .166 u ser_menti o ns -8.29 0.49 .000 urls -19.20 0.50 .000 n umber o f foll o w ers 0.00 0.00 .000 T a b le 5: R esults o f the Multil ev el R egressi o n An a ly sis (f or lo w intensit y exercise)
witho ut interacti o n β SE p F df p R ˆ 2 v a len ce 0.75 0.30 .012 14.62 7, 15335 .000 .27 aro u sa l 0.31 0.06 .000 ha sht ags 0.60 0.29 .040 u ser_menti o ns -0.66 0.33 .047 urls -1.55 0.32 .000 n umber o f foll o w ers 0.00 0.00 .000 with interacti o n v a len ce 0.47 0.31 .125 14.51 8, 15334 .000 .27 aro u sa l 0.27 0.06 .000 v a len ce:aro u sa l 0.50 0.13 .000 ha sht ags 0.57 0.29 .051 u ser_menti o ns -0.66 0.33 .048 urls -1.52 0.32 .000 n umber o f foll o w ers 0.00 0.00 .000 T a b le 6: R esults o f the Multil ev el R egressi o n An a ly sis (f or high intensit y exercise)
Figure 1: Interaction Plot for Valence and Types of Account
Discussion
This study attempted to explore the role of emotion in driving diffusion of physical activities-promoting information on Twitter. Applying the affective circumplex model, it developed several hypotheses and a research question about effects from two dimen-sions of emotion: valence and arousal, and tested those effects with multilevel regression model. Moreover, to test the moderating effect from intensity of exercise in social me-dia setting, this study employed the quasi-systematic sampling to select tweets from 40 Twitter accounts related to high intensity exercise or low intensity exercise (Wilhelm, Tillé, & Qualité, 2017). Compared with previous studies, the comparison of emotional effects between two samples may provide new insights into the spreading of physical-activities-promoting information in the context of social media.
To sum up, some interesting findings emerged from the results. Indeed, emotion is found to be associated with physical-activities-promoting information diffusion on Twit-ter. Unexpectedly, the emotional valence did not predict the retweeting of physical activ-ities promoting tweets across all samples. More specifically, the polarity of emotion did not have impacts on the virality of physical-activities-promoting information on Twitter. In contrast, the arousal predicted the retweeting of physical activities promoting tweets. In other words, the tweets with higher arousal make them more viral on Twitter. It is partly consistent with the findings from Berger and Milkman (2012). It found that there is a large effect from emotional arousal on information sharing, but no effect from emotional valence. For its underlying mechanism, Delaney-Busch, Wilkie, and Kuper-berg (2016) provided a possible explanation that the presence or absence of effect from emotional valence or emotional arousal, or their interaction may depend on situation de-mands. That is, because information sharing behaviors is not directly related to emotion, the emotional arousal, is more sufficient to make people pay attention to messages and motivate them to retweet the tweets.
More interestingly, in this case, the arousal had different impacts on retweeting de-pending on intensity of exercise. In tweets generated from Twitter accounts related to low intensity exercise, tweets with lower arousal predicted more retweeting rather
than tweets with higher arousal. For such an opposite finding, Ekkekakis, Parfitt, and Petruzzello (2011) suggested that the emotions induced by specific exercise may con-tribute to the formation of people’s memories for specific exercise. In turn, those memo-ries, may influence their retweeting behaviors when exposed to tweets related to exercise they experienced before.
More specifically, for low intensity exercise, Koopmann-Holm, Sze, Ochs, and Tsai (2013) and Batchelor (2011) emphasized the importance of calm in meditation prac-tices. In other words, people may be more likely to experience calm or other similar emotions such as relaxation and contentment when participating in low intensity exer-cise such as yoga and pilates. According to the affective circumplex model, calm and relaxation are categorized as emotions with positive valence and low arousal. In turn, low intense exercise-related tweets expressing lower emotional arousal may be more likely to be shared on Twitter. For high intensity exercise, although previous studies argued that high intense exercise may trigger more positive or negative emotions than low intense exercise during or after exercise, the emotions they found or examined, such as tension, anxiety and enjoyment are usually conceived as emotions with high arousal (e.g. Bartlett et al., 2011; Steptoe & Bolton, 1988). Similarly, people may be more likely to share high intense exercise-related tweets expressing emotion with higher arousal on Twitter.
Differently, previous studies that investigated the role of emotion in the spread of political messages (e.g Stieglitz & Dang-Xuan, 2012, 2013) have found significant re-lationships between emotional valence and information sharing. The discrepancies be-tween those findings may imply that the effects from emotional valence and arousal on information sharing in the social media context vary depending on specific domain, and contribute to existing knowledge.
The findings about the associations between emotional arousal and intensity of ex-ercise may contribute more to practical domains. As Ekkekakis et al. (2011) mentioned, previous physical activities promotion in many settings focused more on providing health benefits of taking exercise to change people’s beliefs and thus improve their intentions to
participate in recommended exercise. They also argued that the ineffectiveness of those interventions can be attributed to the ignorance of emotional motives. Therefore, involv-ing emotion in physical activity promotinvolv-ing messages may improve the effectiveness of interventions. Moreover, the identification for the moderating effects from intensity of exercise may help them to optimize the messages with connecting specific emotions with specific exercise.
Limitation
There were also some limitations that raised concerns. Firstly, the quasi-experimental design led to various differences among Twitter accounts. Most specifically, the sample consists of 40 Twitter accounts with different types of exercise. In other words, the differences among exercise may decrease the validity of this study.
Moreover, it was found that only exercisers paid attention to exercise-related mes-sages (Berry, 2006). Therefore, it can happen that only exercisers are more likely to fol-low physical activities-related Twitter accounts to search for useful information and thus have a higher probability to share physical-activities-promoting tweets. In other words, people living a sedentary lifestyle, who are the main target audiences for physical activity promotion, have a lower probability to be exposed to such emotionally charged physi-cal activities promoting messages. Therefore, the effectiveness of social media-based exercise promotion will be limited. To improve the effectiveness of social media-based exercise promotion, the next step of this study is to investigate the emotional effects from physical-activities-promoting information on sedentary people’s intentions to participate in physical activities.
Conclusion
Despite the limit of sampled Twitter accounts, the study found that the emotion can play an important role in spreading physical-activities-promoting information on Twitter. In particular, the emotional arousal influenced the information sharing behaviors in terms
of retweeting quantity, but emotional valence did not. And the effect from emotional arousal is moderated by the intensity of exercise. Implied by those findings, the physical activities promoters can make use of emotional arousal when promoting specific exercise on social media. More generally, researchers may gain some insights into the effects from emotion on information diffusion in different domains on social media.
References
Alm, C. O., Roth, D., & Sproat, R. (2005). Emotions from text: machine learning for text-based emotion prediction. In Proceedings of the conference on human language
technology and empirical methods in natural language processing (pp. 579–586).
Althoff, T., Jindal, P., & Leskovec, J. (2017). Online actions with offline impact: How online social networks influence online and offline user behavior. In Proceedings
of the tenth acm international conference on web search and data mining (pp. 537–
546).
Andreasson, J., & Johansson, T. (2014). The fitness revolution. historical transformations in the global gym and fitness culture. Sport science review, 23(3-4), 91–111. Austin, P. C., Goel, V., & van Walraven, C. (2001). An introduction to multilevel
regres-sion models. Canadian journal of public health, 92(2), 150–154.
Bagozzi, R. P., Gopinath, M., & Nyer, P. U. (1999). The role of emotions in marketing.
Journal of the academy of marketing science, 27(2), 184.
Barasch, A., & Berger, J. (2014). Broadcasting and narrowcasting: How audience size affects what people share. Journal of Marketing Research, 51(3), 286–299.
Bartlett, J. D., Close, G. L., MacLaren, D. P., Gregson, W., Drust, B., & Morton, J. P. (2011). High-intensity interval running is perceived to be more enjoyable than moderate-intensity continuous exercise: implications for exercise adherence.
Jour-nal of sports sciences, 29(6), 547–553.
Batchelor, M. (2011). Meditation and mindfulness. Contemporary Buddhism, 12(01), 157–164.
Bauer, D. J., Preacher, K. J., & Gil, K. M. (2006). Conceptualizing and testing random indirect effects and moderated mediation in multilevel models: new procedures and recommendations. Psychological methods, 11(2), 142.
Berger, J. (2011). Arousal increases social transmission of information. Psychological
science, 22(7), 891–893.
Berger, J., & Milkman, K. L. (2012). What makes online content viral? Journal of marketing research, 49(2), 192–205.
Berger, J., & Milkman, K. L. (2013). Emotion and virality: what makes online content go viral? Gf K Marketing Intelligence Review, 5(1), 18–23.
Berry, T. R. (2006). Who’s even interested in the exercise message? attentional bias for exercise and sedentary-lifestyle related words. Journal of Sport and Exercise
Psychology, 28(1), 4–17.
Bradley, M. M., & Lang, P. J. (1999). Affective norms for english words (anew): Instruction
manual and affective ratings (Tech. Rep.). Citeseer.
Buchholz, S. W., Wilbur, J., Ingram, D., & Fogg, L. (2013). Physical activity text mes-saging interventions in adults: a systematic review. Worldviews on Evidence-Based
Nursing, 10(3), 163–173.
Bull, F. C., Kreuter, M. W., & Scharff, D. P. (1999). Effects of tailored, personalized and general health messages on physical activity. Patient education and counseling,
36(2), 181–192.
Chang, Y.-T., Yu, H., & Lu, H.-P. (2015). Persuasive messages, popularity cohesion, and message diffusion in social media marketing. Journal of Business Research, 68(4), 777–782.
Christophe, V., & Rimé, B. (1997). Exposure to the social sharing of emotion: Emotional impact, listener responses and secondary social sharing. European Journal of Social
Psychology, 27(1), 37–54.
Davies, C. A., Spence, J. C., Vandelanotte, C., Caperchione, C. M., & Mummery, W. K. (2012). Meta-analysis of internet-delivered interventions to increase physical activ-ity levels. International Journal of Behavioral Nutrition and Physical Activactiv-ity, 9(1),
1.
Delaney-Busch, N., Wilkie, G., & Kuperberg, G. (2016). Vivid: how valence and arousal influence word processing under different task demands. Cognitive, Affective, &
Behavioral Neuroscience, 16(3), 415–432.
Dolan, R., Conduit, J., Fahy, J., & Goodman, S. (2016). Social media engagement be-haviour: a uses and gratifications perspective. Journal of Strategic Marketing, 24(3-4), 261–277.
Dunlop, S., Wakefield, M., & Kashima, Y. (2008). Can you feel it? negative emotion, risk, and narrative in health communication. Media Psychology, 11(1), 52–75. Eder, A. B., & Rothermund, K. (2010). Automatic influence of arousal information on
evaluative processing: Valence–arousal interactions in an affective simon task.
Cog-nition and Emotion, 24(6), 1053–1061.
Ekkekakis, P., Parfitt, G., & Petruzzello, S. J. (2011). The pleasure and displeasure people feel when they exercise at different intensities. Sports medicine, 41(8), 641–671. Firdaus, S. N., Ding, C., & Sadeghian, A. (2018). Topic specific emotion detection for
retweet prediction. International Journal of Machine Learning and Cybernetics, 1– 13.
Foster, C., Shilton, T., Westerman, L., Varney, J., & Bull, F. (2018). World health
organi-sation to develop global action plan to promote physical activity: time for action. BMJ
Publishing Group Ltd and British Association of Sport and Exercise Medicine. Fox, S., & Jones, S. (2009). The social life of health information. pew research center. Ghosh, S., Zafar, M. B., Bhattacharya, P., Sharma, N., Ganguly, N., & Gummadi, K.
(2013). On sampling the wisdom of crowds: random vs. expert sampling of the twitter stream. In Proceedings of the 22nd acm international conference on
informa-tion & knowledge management (pp. 1739–1744).
Gordon, R., McDermott, L., Stead, M., & Angus, K. (2006). The effectiveness of so-cial marketing interventions for health improvement: what’s the evidence? Public
health, 120(12), 1133–1139.
friends, bad news-affect and virality in twitter. In Future information technology (pp. 34–43). Springer.
Harris, R. B., & Paradice, D. (2007). An investigation of the computer-mediated com-munication of emotions. Journal of Applied Sciences Research, 3(12), 2081–2090. Hatfield, E., Cacioppo, J. T., & Rapson, R. L. (1993). Emotional contagion. Current
directions in psychological science, 2(3), 96–100.
Heldman, A. B., Schindelar, J., & Weaver, J. B. (2013). Social media engagement and public health communication: implications for public health organizations being truly social. Public Health Reviews, 35(1), 13.
Hutto, C. J., & Gilbert, E. (2014). Vader: A parsimonious rule-based model for sentiment analysis of social media text. In Eighth international aaai conference on weblogs and
social media.
Ito, T. A., Larsen, J. T., Smith, N. K., & Cacioppo, J. T. (1998). Negative information weighs more heavily on the brain: the negativity bias in evaluative categorizations.
Journal of personality and social psychology, 75(4), 887.
Jenders, M., Kasneci, G., & Naumann, F. (2013). Analyzing and predicting viral tweets. In Proceedings of the 22nd international conference on world wide web (pp. 657– 664).
Jong, S. T., & Drummond, M. J. (2016). Exploring online fitness culture and young females. Leisure Studies, 35(6), 758–770.
Keller, B., Labrique, A., Jain, K. M., Pekosz, A., & Levine, O. (2014). Mind the gap: social media engagement by public health researchers. Journal of medical Internet
research, 16(1).
Kerr, J. H., & Kuk, G. (2001). The effects of low and high intensity exercise on emotions, stress and effort. Psychology of sport and exercise, 2(3), 173–186.
Kilpatrick, M., Kraemer, R., Bartholomew, J., Acevedo, E., & Jarreau, D. (2007). Affective responses to exercise are dependent on intensity rather than total work. Medicine
and science in sports and exercise, 39(8), 1417.
news-sharing channels affect health news diffusion. Journal of Communication, 65(3), 512–534.
Kim, H. S., Lee, S., Cappella, J. N., Vera, L., & Emery, S. (2013). Content characteristics driving the diffusion of antismoking messages: Implications for cancer prevention in the emerging public communication environment. Journal of the National Cancer
Institute Monographs, 2013(47), 182–187.
Koopmann-Holm, B., Sze, J., Ochs, C., & Tsai, J. L. (2013). Buddhist-inspired meditation increases the value of calm. Emotion, 13(3), 497.
Kramer, A. D., Guillory, J. E., & Hancock, J. T. (2014). Experimental evidence of massive-scale emotional contagion through social networks. Proceedings of the National
Academy of Sciences, 111(24), 8788–8790.
Kuppens, P., Tuerlinckx, F., Russell, J. A., & Barrett, L. F. (2013). The relation between valence and arousal in subjective experience. Psychological Bulletin, 139(4), 917. Kwak, H., Lee, C., Park, H., & Moon, S. (2010). What is twitter, a social network or a
news media? In Proceedings of the 19th international conference on world wide web (pp. 591–600).
Latimer, A. E., Brawley, L. R., & Bassett, R. L. (2010). A systematic review of three approaches for constructing physical activity messages: what messages work and what improvements are needed? International Journal of Behavioral Nutrition and
Physical Activity, 7(1), 36.
Latimer, A. E., Rench, T. A., Rivers, S. E., Katulak, N. A., Materese, S. A., Cadmus, L., . . . Salovey, P. (2008). Promoting participation in physical activity using framed messages: An application of prospect theory. British Journal of Health Psychology,
13(4), 659–681.
Latimer, A. E., Rivers, S. E., Rench, T. A., Katulak, N. A., Hicks, A., Hodorowski, J. K., . . . Salovey, P. (2008). A field experiment testing the utility of regulatory fit messages for promoting physical activity. Journal of experimental social psychology, 44(3), 826–832.
and reciprocity motivations for retweeting. Psychologia, 58(4), 189–201.
Liu, B. (2012). Sentiment analysis and opinion mining. Synthesis lectures on human
language technologies, 5(1), 1–167.
Maher, C. A., Lewis, L. K., Ferrar, K., Marshall, S., De Bourdeaudhuij, I., & Vandelanotte, C. (2014). Are health behavior change interventions that use online social networks effective? a systematic review. Journal of medical Internet research, 16(2).
Markula, P. (1997). Are fit people healthy? health, exercise, active living and the body in fitness discourse. Waikato journal of Education, 3.
McLaughlin, M. L., Hou, J., Meng, J., Hu, C.-W., An, Z., Park, M., & Nam, Y. (2016). Propagation of information about preexposure prophylaxis (prep) for hiv preven-tion through twitter. Health communicapreven-tion, 31(8), 998–1007.
Mohammad, S. (2018). Obtaining reliable human ratings of valence, arousal, and dom-inance for 20,000 english words. In Proceedings of the 56th annual meeting of the
association for computational linguistics (volume 1: Long papers) (pp. 174–184).
Myers, S. A., & Leskovec, J. (2014). The bursty dynamics of the twitter information network. In Proceedings of the 23rd international conference on world wide web (pp. 913–924).
Neethu, M., & Rajasree, R. (2013). Sentiment analysis in twitter using machine learning techniques. In 2013 fourth international conference on computing, communications
and networking technologies (icccnt) (pp. 1–5).
Neiger, B. L., Thackeray, R., Burton, S. H., Thackeray, C. R., & Reese, J. H. (2013). Use of twitter among local health departments: an analysis of information sharing, engagement, and action. Journal of medical Internet research, 15(8), e177.
Paccagnella, O. (2006). Centering or not centering in multilevel models? the role of the group mean and the assessment of group effects. Evaluation review, 30(1), 66–85. Park, H., Rodgers, S., & Stemmle, J. (2013). Analyzing health organizations’ use of
twitter for promoting health literacy. Journal of health communication, 18(4), 410– 425.
Don’t worry, be active: Positive affect and habitual physical activity. Australian &
New Zealand Journal of Psychiatry, 45(12), 1047–1052.
Peugh, J. L. (2010). A practical guide to multilevel modeling. Journal of school
psychol-ogy, 48(1), 85–112.
Poecze, F., Ebster, C., & Strauss, C. (2018). Social media metrics and sentiment anal-ysis to evaluate the effectiveness of social media posts. Procedia computer science,
130(C), 660–666.
Preston, S. D. (2007). A perception-action model for empathy. Empathy in mental illness, 428–447.
Priebe, C. S., & Spink, K. S. (2012). Using messages promoting descriptive norms to increase physical activity. Health Communication, 27(3), 284–291.
Rimé, B. (2009). Emotion elicits the social sharing of emotion: Theory and empirical review. Emotion review, 1(1), 60–85.
Rimé, B., Finkenauer, C., Luminet, O., Zech, E., & Philippot, P. (1998). Social sharing of emotion: New evidence and new questions. European review of social psychology,
9(1), 145–189.
Riordan, M. A., & Kreuz, R. J. (2010). Emotion encoding and interpretation in computer-mediated communication: Reasons for use. Computers in human behavior, 26(6), 1667–1673.
Rozin, P., & Royzman, E. B. (2001). Negativity bias, negativity dominance, and conta-gion. Personality and social psychology review, 5(4), 296–320.
Russell, J. A. (1980). A circumplex model of affect. Journal of personality and social
psychology, 39(6), 1161.
Russell, J. A., & Barrett, L. F. (1999). Core affect, prototypical emotional episodes, and other things called emotion: dissecting the elephant. Journal of personality and
social psychology, 76(5), 805.
Smeets, T., Brug, J., & de Vries, H. (2006). Effects of tailoring health messages on physical activity. Health Education Research, 23(3), 402–413.
Psychology, 31(2), 156–174.
Steptoe, A., & Bolton, J. (1988). The short-term influence of high and low intensity physical exercise on mood. Psychology and Health, 2(2), 91–106.
Stieglitz, S., & Dang-Xuan, L. (2012). Political communication and influence through microblogging–an empirical analysis of sentiment in twitter messages and retweet behavior. In 2012 45th hawaii international conference on system sciences (pp. 3500– 3509).
Stieglitz, S., & Dang-Xuan, L. (2013). Emotions and information diffusion in social media—sentiment of microblogs and sharing behavior. Journal of management
information systems, 29(4), 217–248.
Suarez-Almazor, M. (2011). Changing health behaviors with social marketing.
Osteo-porosis international, 22(3), 461.
Suh, B., Hong, L., Pirolli, P., & Chi, E. H. (2010). Want to be retweeted? large scale analytics on factors impacting retweet in twitter network. In 2010 ieee second
international conference on social computing (pp. 177–184).
Thelwall, M. (2016). Web indicators for research evaluation: A practical guide. Synthesis
Lectures on Information Concepts, Retrieval, and Services, 8(4), i–155.
Tsugawa, S., & Ohsaki, H. (2015). Negative messages spread rapidly and widely on social media. In Proceedings of the 2015 acm on conference on online social networks (pp. 151–160).
Tsugawa, S., & Ohsaki, H. (2017). On the relation between message sentiment and its virality on social media. Social Network Analysis and Mining, 7(1), 19.
Vallerand, R. J., & Blanchard, C. M. (2000). The study of emotion in sport and exercise: Historical, definitional, and conceptual perspectives.
Warriner, A. B., Kuperman, V., & Brysbaert, M. (2013). Norms of valence, arousal, and dominance for 13,915 english lemmas. Behavior research methods, 45(4), 1191– 1207.
Whiting, A., & Williams, D. (2013). Why people use social media: a uses and grati-fications approach. Qualitative Market Research: An International Journal, 16(4),
362–369.
Wilhelm, M., Tillé, Y., & Qualité, L. (2017). Quasi-systematic sampling from a continuous population. Computational Statistics & Data Analysis, 105, 11–23.
Williams, D. M. (2008). Exercise, affect, and adherence: an integrated model and a case for self-paced exercise. Journal of Sport and Exercise Psychology, 30(5), 471–496. Williams, G., Hamm, M. P., Shulhan, J., Vandermeer, B., & Hartling, L. (2014).
So-cial media interventions for diet and exercise behaviours: a systematic review and meta-analysis of randomised controlled trials. BMJ open, 4(2), e003926.
Xia, Y., Deshpande, S., & Bonates, T. (2016). Effectiveness of social marketing interven-tions to promote physical activity among adults: a systematic review. Journal of
Physical Activity and Health, 13(11), 1263–1274.
Zafar, M. B., Bhattacharya, P., Ganguly, N., Gummadi, K. P., & Ghosh, S. (2015). Sam-pling content from online social networks: Comparing random vs. expert samSam-pling of the twitter stream. ACM Transactions on the Web (TWEB), 9(3), 12.
