The National Institutes of Health describes chronic pain as pain persisting for >3 to 6 months [1,2], which is substantially longer than a “typical” acute response to injury [3]. Chronic pain can be attributed to a wide variety of underlying medical conditions [3], and it is one of the most common health complaints in the United States [4,5]. A recent report from the National Center for Health Statistics found that the prevalence of chronic pain among adults in the United States was approximately 20%, while an additional 7.4% of survey respondents reported pain severe enough to impact daily functioning [4,5]. Moreover, prevalence rates are similarly high in high-income [6-9] and low- and middle-income countries [9,10] throughout the world, contributing to a global public health problem [11,12].

Due to the complexities of pain processing, successful management often necessitates a multifaceted approach personalized to meet the needs of the individual [13]. Chronic pain is considered to be a complex process, whereby psychosocial factors such as emotions, expectations, and social relationships interact with an individual’s physical health to contribute to the maintenance of symptoms [3,14,15]. Thus, it is encouraged to engage in health self-management strategies that contribute to mental and physical well-being [16-18]. One particularly impactful self-management strategy is participation in peer support, which occurs when individuals facing similar health challenges exchange advice, emotional validation, and educational resources [19-22]. Those experiencing chronic pain frequently report emotional distress and feelings of social isolation, and peer support can help alleviate these stressors [23].

Peer support, and pain self-management [23] more broadly, is increasingly taking place on social media platforms such as Reddit (Reddit, Inc), Facebook (Meta Platforms, Inc), and Instagram (Meta Platforms, Inc) [24]. Social media users with chronic health conditions frequently self-organize into large peer-to-peer communities that function as social and informational hubs [25]. For example, there are many peer-driven support communities on Reddit in which tens of thousands of members share private health information and respond to medical inquiries. Support interactions on social media platforms are unique in that they consist of naturally occurring discussions driven by users themselves, contrasting formally organized support groups [26] and structured web-based interventions [27]. Although users tend to have positive experiences with online support [28], interactions on the internet are not universally positive. As demonstrated by the COVID-19 pandemic, social media platforms have the potential to spread misinformation and hostility toward medical experts [29,30]. Thus, although writing about pain [31] and having supportive relationships [32] can benefit mental health, it is unclear how naturally occurring web-based health interactions influence chronic pain [24,33,34].

It is reasonable to assume that the impact of online peer support not only depends on the amount of support an individual receives but also on the affective qualities of their social interactions [14,35]. Negative affective thoughts and pain catastrophizing are theorized to contribute to the maintenance of chronic pain (eg, [3,14,36-38]), and social interactions can drive users to adopt or reject those beliefs [14]. Furthermore, the manner in which people communicate about their experiences plays an important role in shaping their emotions and expectations for the future [31,39]. It follows that health communities perpetuating overly pessimistic attitudes or hostility toward health care providers can be harmful to well-being, just as those providing emotional validation can be beneficial [22]. Thus, examining the affective qualities of web-based interactions using sentiment analysis may serve as an entry point to understanding how health is influenced by participation in online health communities (OHCs). Moreover, the analysis of online health discussions can provide broader insight into how social media users experiencing chronic pain communicate about their pain.

This study investigated peer support interactions in OHCs by focusing on the sentiment of individual support interactions. Applying sentiment analysis to the comment histories of users in a popular chronic pain forum on Reddit, we measured the degree to which users synchronized the sentiment of their comments to match their peers, and we modeled the trajectories of their sentiment over the course of their community participation. This work addresses 3 questions regarding the use of online peer support for chronic pain: Do users who engage in online peer support synchronize their pain expressions to match the language of other community members? How does a user’s sentiment progress over time in response to online support interactions? And finally, how do specific interaction dynamics, such as affective synchrony with other users, relate to changes in sentiment?

We analyzed comments posted to a popular chronic pain support community on Reddit. Reddit is a platform where users can create personalized forums dedicated to discussing specific topics [86] and within each user-created forum (ie, “subreddit”), users define their own rules regarding content and membership. The subreddit analyzed in this study describes itself as a forum for users to discuss their conditions with their peers and share advice; however, we have opted to withhold the exact name of the community out of concern for user privacy. The community guidelines provided by the moderators discourage direct medical advice and suggest consultation with professionals before participation. However, it is unclear how strictly these guidelines are enforced. In terms of content, discussion threads contain a mixture of advice-seeking, informational resources, and personal anecdotes. A recent paper using latent Dirichlet allocation to analyze chronic pain subreddits on Reddit similar to the one used in this study found that users most frequently mentioned phrases related to lower back pain in their posts in addition to words such as “doctors,” “help,” and “work” [87], providing evidence that discussions within these communities are highly focused on the topic of chronic pain.

We collected the post histories of 200 randomly selected Reddit users active on the subreddit and calculated the sentiment of their comments using a dictionary-based approach [88]. Sentiment analysis is a common natural language processing method used to classify and describe the emotional expressiveness of text by analyzing the valence, intensity, and structural features of language [89,90]. Next, we used bivariate wavelet analysis to estimate the degree of sentiment synchrony between users during interactions, which is a technique popularized for its applications on complex systems [91]. Unlike a standard linear approach (eg, cross-correlation), wavelet analysis is robust against nonstationarity and describes multiple types of synchronies, including matched intensity, comovement, and leader-follower dynamics [91,92]. Moreover, wavelet analysis can effectively separate changes in sentiment that are occurring reactively (eg, comment to comment), from those occurring globally over hundreds of interactions. Finally, we examined whether discourse on the subreddit was associated with negative or positive changes in the health attitudes of users by using growth curve analysis to model trajectories in comment sentiment.

This study was reviewed by the institutional review board at the University of Central Florida (00001430). Observational public data from Reddit were collected, and there were no direct interactions with human participants. Reddit is a social media platform where users typically provide pseudonyms in place of their real names, which are not reported, and we did not collect personally identifiable information. We randomly sampled users who fit our eligibility criteria, meaning a particular profile’s inclusion is not discernable. In addition, we have opted to withhold the name of the subreddit that we collected data from to further obfuscate the identity of its members. Given the large number of comments we collected, it was possible that some of the text contained identifiable information indirectly in the form of self-disclosures. To protect user anonymity, each user in our sample received a randomized identifier in place of their Reddit username, and the text content of their comments was aggregated using the procedures detailed in the sections that follow such that it would not be possible to link the stored data to any individual Reddit user or real-life individual.

We collected publicly available Reddit data using the Pushshift database [93] and Reddit’s application programming interface [94]. On Reddit, the term submission refers to user-created discussion threads posted within each community, while comments refer to the text replies within each discussion thread. To be included in our final sample, users had to be moderately active members of the support subreddit, with at least 100 total comments and submissions after excluding self-deleted posts and those deleted through moderation. This minimum activity threshold was applied to ensure there was an adequate number of data points for wavelet analysis at lower frequencies [92]. In addition, we excluded users serving as moderators and Reddit profiles posting automatic replies (eg, apparent bot accounts). To identify Reddit profiles meeting this criterion, we downloaded every publicly available comment and submission made to the subreddit in the last 10 years. We then calculated each user’s total number of posts in the community and randomly selected 200 users with more than the minimum required activity. The sample size was determined using an a priori power analysis for a growth curve model with a small effect size, f²=0.10, and α=.05, estimated based on similar previous research examining sentiment in online discussion forums [95-97]. Finally, we collected data from the public interactions these users had on Reddit. We define an interaction as the pairing between a user’s personal comment and the post they were replying to when making that comment. For example, an interaction could involve a user posting a top-level comment in response to a discussion thread. An interaction could also involve posting a reply to someone else’s comment inside of a discussion thread. Thus, for each user, we collected the following: (1) the text of all their comments, (2) the corresponding text of posts they interacted with, (3) comment scores (upvotes minus downvotes), and (4) time of posting.

Each user’s comment history was formatted as a time series sorted in chronological order, and each point in those time series represented 1 interaction with the community. Activity occurring outside of the subreddit of interest was excluded from the analysis. In addition, the Reddit application programming interface returns an error message instead of the original text when comments are removed by moderators or users themselves, and these error messages were excluded. A small portion of comments contained single-word phrases, such as a web address or emoji. Sentiment calculations for these items would have been unreliable, and we opted to remove comments containing <3 words. A total of 73,876 of comment interactions were collected from the pain support subreddit. Of those comments, 5670 (7.68%) were excluded based on low word count or missing data. Shortly before analysis, we also discovered 1 user acting as a moderator, and this individual’s data were removed due to their unique role in the community. Thus, our final sample size included 68,206 (92.32%) of the 73,876 peer-to-peer interactions from 199 users and across 29,360 unique discussion threads.

After cleaning the data, we calculated sentiment scores using a lexicon-based approach [88], such that there were separate scores for both sides of each social interaction. To improve the reliability of these estimates, we compared 2 popular techniques in the psychology and social media literature. The Linguistic Inquiry and Word Count (LIWC) [98] dictionary is a natural language processing tool used to extract psychological language attributes and has been applied in a variety of digital health and social media contexts (eg, [99-102]). The LIWC dictionary calculates the percentage of text pertaining to specific topics (eg, the percentage of health-related words) and provides summary variables measuring higher-level cognitive processes. For our purposes, we used the LIWC dictionary to calculate sentiment, number of words, and percentage of health-related words in each comment. Sentiment scores from the LIWC dictionary ranged from –100 to 100, with higher scores indicative of more positive language, lower scores indicative of more negative language, and scores around 0 indicative of neutral language [98,103]. For comparison, we also calculated sentiment using the Valence Aware Dictionary for Sentiment Reasoning (VADER) [104]. Sentiment scores from VADER have been validated against manual human reviewers and were designed for the analysis of brief web-based internet interactions [104]. Like the LIWC dictionary, compound sentiment scores from VADER ranged from –1 to +1, with higher scores indicative of more positive language, lower scores indicative of more negative language, and scores around 0 indicative of neutral language [105]. We found both methods were highly consistent (α=.93) and strongly correlated (r=0.70; P<.001); consequently, we created a compound score by standardizing and averaging the results of both. Generating compound scores through aggregation has been shown to improve reliability [106], and, in our case, this approach minimized the influence of text passages where the 2 methods diverged. As a result, each user had one-time series representing their personal comment sentiment and one-time series for the sentiment of comments they interacted with when making their personal comments.

Bivariate wavelet analysis was used to quantify the sentiment synchrony between users and their peers. We used the cross-wavelet transform (XWT) and wavelet coherence (WTC) to gain information about matching sentiment intensity and phase-locked (ie, correlated) changes in sentiment. Compared to cross-correlation, wavelet analysis describes changes in sentiment at multiple frequencies (eg, short- vs long-term changes in sentiment). This allowed us to isolate fluctuations in sentiment occurring comment to comment (ie, high-frequency changes) from those occurring over dozens or hundreds of interactions (ie, lower-frequency changes). In other words, this analysis provided insight into short-term reactionary changes in sentiment versus long-term global changes. We briefly discuss XWT and WTC subsequently; however, readers should refer to the work of Torrence and Compo [107] for a comprehensive guide to wavelet analysis or Issartel et al [92] for applications on behavioral synchrony. A discussion of the method and theoretical implications in the context of social media is provided by Necaise et al [95].

Wavelet calculations were completed using the biwavelet package in R (R Foundation for Statistical Computing) [108] with a Morlet wavelet function (ω₀=6). The Morlet wavelet was selected because it provides good resolution compared to other functions [107]. Furthermore, cross-wavelet calculations can be sensitive to large spikes in the data, so the data were transformed using a cumulative distribution function, so the values represented percentiles. This transformation has been suggested in previous literature [91]. The XWT was applied using the method outlined by Grinsted et al [91] and identified points in time (x-axis of Figure 1) and frequency (“period” on the y-axis of Figure 1) where users had fluctuations in sentiment that were similar in intensity as their conversation partners. Significance is determined in comparison to AR(1) background processes [91,92], and the significant points (circled in black in panel A of Figure 1) are referred to as regions of high common power [91]. The XWT also provided information about the relative phase (RP) relationship between users and their conversation partners within regions of high common power (depicted by the orientation of arrows in Figure 1). The RP angle was extracted from the XWT and described how user sentiment changed relative to their peers [91], providing information about leader-follower dynamics.

In addition to measuring common power with the XWT, we also calculated user coherence with their peers using WTC. The formula for WTC is similar to Pearson correlation except localized in frequency and time [91]. Therefore, it is helpful to think about coherence as an indicator of correlation or “comovement” between 2 signals. We used a standard smoothing factor of 0.6 for Morlet wavelets in the WTC calculation and tested for significant coherence against simulated data via Monte Carlo methods with 2000 random initializations [91]. Regions of significant coherence are circled in black in panel B of Figure 1.

Table 2 contains descriptive statistics related to the Reddit activity of users in our sample, the sentiment of their comments, feedback received from their peers, and wavelet synchrony metrics. Frequency distributions for comment sentiment, common power, and coherence can be found in Figure 2. Although we only analyzed interactions within the support community, it is worth noting users made a substantial number of posts to other subreddits. On average, activity inside of the support group constituted 22.03% (SD 24.01%) of users’ total Reddit activity.

We found users had significantly more common power during their real support interactions (mean 0.06, SD 0.02) compared to a shuffled baseline (mean 0.04, SD 0.01; t₁₉₈=4.02; P<.001; Cohen d=0.40). Likewise, there was significantly more sentiment coherence during real interactions (mean 0.08, SD 0.04) as compared to shuffled data (mean 0.05, SD 0.03; t₁₉₈=6.57; P<.001; Cohen d=0.64). These results suggest users synchronized in terms of matching each other’s intensity (common power), and changes in user sentiment corresponded to changes in peer sentiment (coherence).

Comparing these results to a traditional Pearson correlation approach (Table 3), we found significant positive correlations between Pearson correlation coefficients, real-data coherence (r₁₉₇=0.46; P<.001), and real-data common power (r₁₉₇=0.18; P<.001). However, there were no relationships between Pearson coefficients and wavelet calculations in the shuffled data (P=.77 and .50). Taken together, we found evidence that users synchronized their sentiment during support interactions beyond what could be explained by random variance in their data. The results of our correlation analysis also supported the validity of our synchrony analysis.

Across all users, we found an average circular RP angle of 2.81°, suggesting users were in-phase with the sentiment of their peers during interactions (Figure 3). Thus, the sentiment intensity of users fluctuated in the same direction without a clear leader or follower, and periods of positive posting by a user were mirrored by periods of positive posting by their peers.

We found that the average common power varied significantly depending on frequency (F_2,198=17.70; P<.001). The common power in low-frequency bands (mean 0.08, SD 0.10) was significantly higher than medium- (mean 0.05, SD 0.05; t₁₉₉=3.66; P<.001) or high-frequency bands (mean 0.05, SD 0.01; t₁₉₉=5.05; P<.001). The same was true for coherence (F_2,198=3.79; P=.01), as there was significantly more coherence in low (mean 0.10, SD 0.15) compared to high frequencies (mean 0.07, SD 0.03; t₁₉₉=2.36; P=.02). There were no differences in coherence between medium and high frequencies (P=.13). These findings suggest that synchrony was most evident in long-term sentiment fluctuations occurring over hundreds of interactions as opposed to the more reactionary changes in sentiment occurring from comment to comment. There were no differences in RP angle (P=.11), indicating the sentiment of users and their peers were in-phase regardless of scale.

Across all users, sentiment decreased significantly when interacting with the subreddit (β=–.02; P=.003), and this negative trend was observed while controlling for differences in average word count (β=–.10; P<.001) and average health-related language (β=–.20; P<.001; Table 4). In addition, there was a significant 3-way interaction between interaction number, synchrony, and the number of days active (β=.02; P=.02), indicating the trajectory of comment sentiment depended on the degree of emotional synchrony and the duration of subreddit participation. Users who were active on the subreddit for a greater number of days exhibited a steeper decline in sentiment compared to those who were active for shorter durations. Thus, sentiment not only decreased as users interacted with their peers but also decreased longitudinally with the total number of days spent seeking online support. However, this compounding effect depended on affective synchrony. Greater affective synchrony was associated with negative changes in sentiment for those who spent less time on the subreddit, but it functioned as a protective factor for those who were active for longer (Figure 4).

Our findings make several contributions to the literature. First, our study is unique in quantitatively examining the dynamics of online chronic pain support communities, contrasting the current literature that is primarily qualitative and focused on pain narratives [28,73,80,87,123,137,138]. Second, the qualities of pain are typically investigated through measures of central tendency (ie, by comparing average pain catastrophizing across groups). However, the biopsychosocial model describes chronic pain as a dynamic (changing over time) and complex (evolving from the interactions of multiple factors) process [14,139,140]. Consistent with this framing of chronic pain, we used wavelet analysis to analyze sentiment synchrony, which is an analytic technique better suited to describe complex system interactions and nonstationary behavior [92,107]. Next, previous studies of online pain support groups have often relied on small sample sizes and do not analyze data longitudinally. In comparison, we analyzed the entire comment histories of Reddit users consisting of over a million comments and approximately 70,000 interactions within 1 OHC. Thus, we present a comprehensive view of users’ web-based interactions inside the chronic pain subreddit over nearly a decade. Finally, as opposed to examining the impact of a structured intervention (eg, [27,85]), we examined naturally occurring conversations that are more representative of everyday social interactions. Unlike a structured intervention or professionally managed support group, our analyses considered undirected web-based conversations where participants likely encountered a wide variety of positive and negative actors. Participants were free to express their emotions with little inhibition related to research participation, especially due to the relatively anonymous nature of Reddit [33]. This type of longitudinal and naturalistic look into an individual’s pain expressions would be exceedingly difficult to collect in traditional settings [33].

Despite the strengths of our quantitative approach, there are several limitations. Notably, our investigations lacked data about user health outcomes. We investigated text sentiment as a correlate to chronic pain attitudes and emotions, and our results provided novel insights regarding how sentiment progressed in 1 OHC. However, without data on health outcomes, we cannot know if these trends were indicative of declines in physical or psychological well-being. Users could have viewed the forum as an opportunity for emotional catharsis, venting about their pain increasingly over time. There are some circumstances where negative emotional expression can lead to positive outcomes [141-143], particularly if those expressions are positively reframed by the audience [144]. In contrast, emotional catharsis can be counterproductive and lead to agitation [145,146]. Thus, it is necessary to exercise caution when drawing conclusions about the merits of support seeking on social media platforms based on the findings of this study.

Another potential limitation relates to our selection of Reddit as a data source. Reddit blends features from popular social networking sites (eg, the ability to personalize a user profile, add friends, or form groups) with the formatting of a web-based discussion board. As a result, the interactions on Reddit tend to surround topical discussions as opposed to resharing or reposting user-generated content. This format facilitates verbose and focused conversations, and it allows us to collect rich longitudinal data about pain-related social interactions. Reddit also affords users increased anonymity compared to social networking sites such as Facebook, as a user’s Reddit profile is not tied to their identity, friends, or family. There is evidence to suggest that social media users seeking chronic pain support prefer the ability to remain anonymous [28], thus, we suspected users would more candidly share their feelings on Reddit compared to a public-facing social networking site. However, given the unique features of Reddit, it is not clear how our findings generalize to more popular social media platforms that promote different styles of communication. It is also not clear if our findings concerning chronic pain support can be applied to OHCs for other types of chronic health conditions.

Although social media provides easy access to massive peer support networks, it also has the potential to spread negative attitudes and beliefs about pain. We found evidence that social media users synchronized the emotional intensity of their language during virtual conversations about chronic pain, which can be viewed as an indicator of social bonding [113] and highlights the influence of OHC interactions on pain expressions. Furthermore, affective synchrony functioned as a protective factor against continued negative language use among those with the most negative expressions of pain in the online community. Despite these promising findings, we noted a steady increase in negative sentiment averaged across users as they continued their community participation. This negative trend is consistent with the “downward spiral” of affect described by the FA model of pain [132], suggesting that the impact of OHCs on pain expressions varies to a large degree dependent on specific user behaviors. Our results highlight the importance of considering specific user and community dynamics when assessing the impact of OHCs. Furthermore, our synchrony indices show there is a high degree of peer influence on pain attitudes and expressions, consistent with the SCMP.