How an Interest in Mindfulness Influences Linguistic Markers in Online Microblogging Discourse

This study used a dataset of public Twitter messages posted by two groups: a mindfulness interest group (n = 10,347) and a control group consisting of a random selection of Twitter users from the Sentiment140 dataset (n = 9385). The aim of this selection process was to create participant categories of those who follow a public mindfulness expert, vs. a random sample of Twitter users. Twitter users were included in the mindfulness interest group if they followed at least one of five public mindfulness and meditation experts. The number of followers was collected at the time of selecting the mindfulness public figures for the study, so it may vary and only reflects a moment in time. The inclusion criteria to be considered publicly recognised mindfulness experts were (1) they had published books on mindfulness, (2) had a PhD, (3) were actively teaching on the subject of mindfulness or meditation at the time of the study, (4) were active on Twitter, and (5) had at least 30,000 followers at the time of selection (June 2019). The public mindfulness figures were identified by first creating a list of teachers from internationally renowned teaching and retreat centres including the Insight Meditation Society and Spirit Rock in North America or Gaia House in Europe, as well as from looking up the queries “top mindfulness teachers globally”, “global renowned mindfulness teachers”, “global mindfulness experts”, “most respected mindfulness teachers globally” and “best mindfulness teachers globally” on Google Search and extracting names from the resulting pages. This list was curated by looking up whether the teachers were published authors through a search on Amazon’s book section and Google Search using queries like the name and “ + book”. To identify if they held a PhD, Google Search was queried manually with search criteria like the name and “ + PhD” as well as the Wikipedia pages of the public figures. For those that met the “teaching”, “published author”, and “holds a PhD” criteria — and were alive — it was manually confirmed on Twitter whether they had an active account and how many followers they had. This number was finally used to stack rank them and identify the five that met these criteria and had the greatest number of followers. The mindfulness experts that met these criteria were Jack Kornfield (@JackKornfield, 131 k followers), Danny Penman (@DrDannyPenman, 110 k followers), Tara Brach (@tarabrach, 63.9 k followers), Jon Kabat-Zinn (@jonkabatzinn, 54.5 k followers), and Joan Halifax (@jhalifax, 36.2 k followers). The control group dataset was comparable in size to the mindfulness interest group. It included a random selection of users from the Sentiment140, a publicly available dataset that was accessed through the online data science platform Kaggle. Users with fewer than 20 public tweets were removed from the datasets for both groups.

Ethical approval was obtained from the Ethics Board at Bangor University prior to data collection. Public Twitter updates can be seen by anyone who has chosen to follow the publisher if they are not from password-protected conversations in moderated or closed groups. This means that it is reasonable to expect that Twitter users whose profile is public are likely to have no expectation of privacy. For this reason, neither the public figures nor the members of the mindfulness interest or control group were contacted directly for the study. This approach is in accordance with the British Psychological Society Ethics Guidelines for Internet-mediated Research (2017). In accordance with the Twitter Terms of Service and Privacy Policy (Twitter, 2021a, b, c), which users accept when creating an account, no demographic information was available about participants. These policies regulate the relationship between Twitter as a service provider and its users and clearly state that such information will not be made public. In addition, the data collection for this study complied with the Developer Terms (Twitter, 2021a, b, c). These terms define how Twitter content may be used, and specifically allow the analysis of aggregate data that does not store personal information such as user IDs.

All data were collected using the standard Twitter application programming interfaces or APIs (services to access and request Twitter data systematically). Specifically, the “GET followers/ids” endpoint was used to retrieve followers of the public mindfulness figures; the “GET users/lookup” endpoint was used to gather Twitter user information such as total number of followers or account creation date; and the “GET statuses/user_timeline” endpoint was used to obtain the content of the 200 most recent tweets (Twitter, 2021a, b, 1c). The Twitter APIs were accessed using the open-source programming language Python v3, and the wrapper Twython for Python (McGrath, 2014), which is a set of functions used for programming convenience.

The follower list for each of these five public figures was obtained in batches of 75,000 every 15 min. Once the complete list had been collected, the list of followers was randomised. A maximum of 200 of the most recent, public tweets were downloaded for each user from a random sample of 15,000 followers, at a rate of 900 API calls every 15 min. The dataset for the control group was accessed through Kaggle using the search query “Twitter” and activating the filter “dataset”. The resulting dataset, Sentiment140 (Go et al., 2009), was built by sending recurring search queries to the Twitter API for emoticons:) and:( and it is formed of the individual tweet results that the service returned. The downloadable csv file contains a total of 1.6 million tweets along with their labels and metadata that includes the user who published the tweet. A sample of 15,000 of these Twitter users was randomised and a maximum of the 200 most recent, public tweets per user was downloaded, at a rate of 900 API calls every 15 min.

Additionally, prior to analysis, software was used to clean the text in both groups and discard retweets (publication of content coming from other Twitter users), responses, tweets in languages other than English, and media files (images or videos). The content of the tweets was analysed to replace URLs with the text “[URL]”, and mentions of Twitter users with the text “[USER]”, as a means of preserving privacy. In order to avoid bias, no attempt was made manually or by computer programme to post-process the control group on the basis of having tweeted mindfulness-related content or following any or all of the five public figures in question.

Linguistic Inquiry and Word Count (LIWC) built-in dictionaries were used to measure and compare the frequency of linguistic markers associated with affective processes, social orientation, and the “being” mode of mind. These dictionaries consist of words, word stems, emojis, and verbal constructions that have been iteratively identified and tested by experts to reflect psychological categories (Pennebaker et al., 2015). To measure linguistic markers of affective processes, this study examined the frequency of words in the LIWC “affective processes” category including the subcategories of “positive emotions” and “negative emotions”. It operationalised a metric for “happiness” by calculating the difference in the frequency of use of words in the positive and negative emotion subcategories (Happiness = positive emotions − negative emotions) (see example words in Table 1). Social orientation was measured computing the frequency of words in the LIWC “social processes” category. Within this category, the subcategories of “friend orientation” and “family orientation” were explored to provide granularity on social orientation (see Table 1 for example words). To explore linguistic markers associated with the “being” mode of mind, the study examined the frequency of words in the LIWC “Mental time travelling versus present moment orientation” category. Specifically, the subcategories of “time”, “present focus”, “past focus”, and “future focus” were explored.

In addition, a third-party dictionary developed by a panel of experts (Collins et al., 2009) which includes words describing the mindfulness experience (e.g., “observe”, “accept”, “calm”) as well as challenges associated with the practice (e.g., “react”, “autopilot”, “judge”) was used to measure and compare the frequency of linguistic markers associated with dispositional mindfulness among Twitter users in the dataset. Table 1 shows examples of words in this category.

Linguistic analyses were performed with the 2015 version of the software LIWC, which is available in open access (Pennebaker et al., 2015). Psycholinguistics researchers developed this programme to analyse human, social, and psychological states through language. LIWC calculates the use frequency of certain specific words, word stems, or emoticons that are associated with psychological and cognitive processes, and it compares them against psychometrically validated categories, transforming text into numerical values. These values are expressed as the percentage of the total number of words in a text that belong to a specific category or subcategory, whilst controlling for absolute text length. For example, LIWC includes the subcategory “sadness” nested under “negative emotions” and, in turn, under the category “affective processes”. An output value of 56.15 for the category “sadness” means that 56.15% of the words in the analysed text are included in the “sadness” category. The language markers in the “sadness” category are also included, without limitation, in the parent category “negative emotions”, and in the overarching category “affective processes”. In addition to the built-in categories, LIWC 2015 can also be used to process text using custom-made dictionaries built by users, to explore other psychometric categories or for data in languages other than English.

Prior to conducting statistical analyses, all variables were reviewed to confirm a normal distribution. Then, descriptive statistics were calculated and independent t-tests were conducted to analyse the differences between the mindfulness interest group and the control group in relation to the variables of interest. For rigour, twelve separate linear regression analyses (weighted least squares) were performed to determine whether belonging to the mindfulness interest group had an effect on the variables explored when controlling, for each participant, the following eight factors as dependent variables: the date of their first and last tweets in the dataset, the number of replies, retweets, followers, and accounts followed, the word count of tweets, and when the Twitter account was first created. Lastly, when t-test results were statistically significant, the standardised mean difference (Cohen’s d) was estimated to assess effect size. Given the large sample, a significance criterion of p < 0.001 was adopted for all analyses.

Perhaps the most salient limitation in this study is that it is unknown whether those who follow mindfulness public figures on Twitter are also likely to practise mindfulness, or to what extent. Additionally, it is unknown whether followers of public mindfulness figures do so because they are more broadly interested in positive mental health and well-being, rather than mindfulness per se. The results of this present study are therefore associated with an assumed interest in mindfulness based on participants following mindfulness public figures. In future research, it would be useful to include a measure of dispositional mindfulness or gather data on frequency of mindfulness practice. Additionally, a replication of this study but with different groups could help unpack whether it is solely followers of mindfulness public figures that have these linguistic markers. It is not clear from this study if the observed results are because of a common interest in mindfulness, or whether there could be an alternative explanation, such as an interest in wellness more generally. Future research could investigate different groups, for example, Twitter users that follow key “healthy living” public figures which are not associated explicitly with mindfulness to see if the effects observed in this study are specific to an interest in mindfulness.

As there is developing research interest in whether mindfulness practice increases pro-social behaviours (e.g., Golubickis et al., 2022), further exploration around the use of linguistic markers may be a useful additional methodological approach. First steps here might be for future research to establish whether there are correlations between the language markers explored in this study and established self-reporting measures of dispositional mindfulness or pro-social inclinations such as compassion for others (Brown & Ryan, 2003; López et al., 2018). The use of experimental study designs to explore whether there are changes in use of linguistic markers after attending a mindfulness-based programme or engaging in regular mindfulness practice may also further elucidate the links between language use and trait mindfulness.

The question of the general applicability of the results is a further limitation since the Twitter population is not demographically representative of the population as a whole. For example, Twitter users differ from the general population in that they tend to be younger, more educated, and politically liberal (Mellon & Prosser, 2017; Wojcik & Hughes, 2019). The demographic information that is available on Twitter is also limited, and precludes any analysis of the data by gender, age, geography, and so on. Lastly on demographics, a potential limitation may be the fact that public figures are only from European and North American countries and not globally representative, which may also include a bias in their follower base. It may be useful in future research to also compare the linguistic markers of followers of mindfulness public figures from different geographical locations, for example followers of mindfulness public figures based in Asia with those who follow those who are based in Europe and North America. Lastly, one of the criteria to select mindfulness experts is to hold a PhD which excludes other popular spiritually oriented mindfulness and meditation experts, or those teaching out of a particular religious or cultural context. This means the results may be specific to individuals interested in scientific or evidence-based framings of mindfulness and more research to verify the general applicability of results would be required.

A broader point to consider is how ideas about mindfulness are showing up in public discourse and understanding. Although this study did not look at this issue directly, from the perspective of linguistic analysis, whilst there is clarity about the sequential development of the LIWC built-in dictionaries, little is known about how Collins et al. (2009) developed the mindfulness dictionary used in this study and few studies have tested it since it was developed. It is therefore unknown how closely the mindfulness dictionary may relate to lay understanding about mindfulness, there is some evidence which suggests that there is little convergence on key facets of mindfulness (for example, acceptance) between lay understanding and the constructs used in common measures of mindfulness (Choi et al., 2021). For example, University students, when asked what mindfulness is, often associated mindfulness with controlling emotions, and rarely with psychological acceptance (Hitchcock et al., 2016). It must be noted as a further limitation that recent studies have shown inconsistencies in the relationship between self-reported emotion and LIWC metrics (Sun et al., 2020). More research focusing on the analysis of the lexical markers included in the different categories that are built into the LIWC software, and their suitability for mindfulness research, would be beneficial. In addition, the mindfulness dictionary would need to be rigorously evaluated from a psychometric perspective to confirm its reliability and validity. Furthermore, the development of additional, specific, mindfulness-related metrics such as the seven attitudinal foundations, such as acceptance and non-striving (Kabat-Zinn, 1990), all dimensions of the being mode of mind (Williams & Penman, 2011; Williams, 2008), and nonduality, would be a welcome addition to the research field.

A final limitation of the present study is the use of computerised linguistic word count analysis. Whilst this methodology reveals significant lexical patterns, it does not account for semantics, for complex linguistic features like sarcasm and situational context, or for more qualitative analyses. Future research could employ more sophisticated natural language processing techniques to get a deeper understanding not only of lexical features but also of semantics, context, etc. For example, programmatic content clustering could be used to understand the relationship between an interest in mindfulness and topics that Twitter users tweet about, or more advanced sentiment analysis tools could provide insights that account for tone, sarcasm, irony, and the like. This would contribute to the emergent discussions about popular conceptions of mindfulness (as contrasted to scholarly definitions), and to how ideas about mindfulness are showing up in public discourse. Applied research might explore the development of an absentminded writing detection system to alert individuals that they may want to take a pause before publishing or sending the text. Finally, the very limited existing research in this area is largely based on written language. By contrast, most mindfulness research is based on programmes that rely primarily on spoken language. Therefore, future research should be extended to spoken language to deepen understanding of how mindfully the general population speaks, or to detect and analyse any changes in speech before and after interventions. Such research could also extend our understanding of the characteristics of inquiry as a key element of mindfulness training (Crane et al., 2015).