The current study was approved by the Swinburne University Human Research Ethics Committee (SUHREC) Project ID: 20190402-1922. In line with SUHREC advice, it was not possible to directly quote individual twitter usernames or their posts; thus, data are presented in aggregated form only.

This study utilized publicly available Twitter data from Pro-ED profiles collected between September 15, 2015 and July 1, 2018, adhering to university ethics requirements. To identify Pro-ED profiles, we used an online scraping tool to gather posts (tweets) and reposts (retweets) using Twitter's public Application Program Interface (API). Twitter offers a systematic collection of sampled tweets as they are posted through a public API filtered by specific criteria. For this research, the hashtag #proana was the qualifying criteria, which resulted in 54,506 tweets and retweets (tweets that are recirculated by other users) across 11,620 Twitter profiles from various time zones and geographic locations. These profiles then underwent a double pseudonymization process to preserve anonymity before a 2-step screening process using inclusion and exclusion criteria was imposed to reach the final sample of 967 profiles (see Figure 1).

The 967 profiles were further classified into 1 of 2 categories (with-mention or without-mention) based upon the user’s preferred posting style. A without-mention message pertained to the user sending a tweet not to a particular individual, whereas a mention relates to the user including another Twitter account in the message. In the initial phase, descriptive statistics were ascertained to compare the proportion of tweets or retweets in relation to the user’s most frequently used hashtag. A series of Mann Whitney U tests were then conducted to compare preferred posting style across number of followed, followers, tweets, and favorites. Finally, a multiple linear regression model, using ordinary least squares [27], was then used to estimate the number of followers based upon the number of followed users, tweets, and favorites. The criteria for stepwise selection were based upon changes in the adjusted R² values at each new step. Skewed predictor variables were trimmed by excluding extreme cases, as identified with a Cook distance >3 SDs from the mean. Analyses were conducted in SPSS Version 26.0 (IBM Corp, Armonk, NY).

Our study explored the common lexicon of the Pro-ED Twitter community by identifying popular key words and phrases tagged in tweets. Results of this analysis indicate #thinspo as the most prominent hashtag within the Pro-ED Twitter community, other than #proana, suggesting considerable overlap between the topics and their intent. This indicates that wider conversation involving #thinspo across other social media platforms needs to be further scrutinized and treated as ED-related discussion. Previous research [17,19,28] has found that thinspiration tweeters, that is individuals using #thinspo or #thinspiration to accompany appearance- or weight-related posts on Twitter, form part of a closely connected genuine virtual community and differ to those propagating fitspiration content. Indeed, #fitspo or #fitspiration did not feature as one of our top 10 most used hashtags, suggesting that this hashtag is potentially identifying a different community focused more on the promotion of fitness and muscle building [28]. Taken together, these findings suggest #thinspo as a salient aspect of a Pro-ED lifestyle, with #proana and #thinspo used interchangeably in online spaces to communicate a supposedly motivating weight loss message to other community members.

Observing communication within online communities provides insight into their structure, member roles, and tribal behavior [29-31]. Typically, the communication patterns and network structures of online ED communities are differentiated by their intentional online behavior. Members of pro-recovery communities who view EDs as an illness and are actively working towards recovery generate more original content and actively seek out new profiles to follow when compared to Pro-ED communities [31]. Our findings suggest the communication patterns within Pro-ED Twitter communities to be more community driven. The type of tweet (with-mention vs without-mention) did not differ significantly across followed, followers, tweets, and favorites. This implies that whether a member is the source of the content or merely sharing it, they are equally likely to contribute to the growing Pro-ED community and its formation. As previously suggested by Wang et al [31], members within the Pro-ED Twitter community use the platform as a tool for community engagement and not typically as a means of communication per se as indicated by the number of retweets within our findings. This is a crucial finding as there is a greater role that social media platforms can play in addressing the communication. In essence, social media platforms could fill this void with tools that can facilitate communication and extend ED-related discussions with the ED community users. One approach would be to channel them to external sites, such as the National Butterfly Foundation (official organization for ED-related matters in Australia), mediated by a chatbot for cost efficiency.

The Pro-ED community chatter was dominated by retweets, by 75%, rather than genuine threads of communication. For example, the exchanges featuring “41DaysofStarvation” were a passing superficial topic in a particular subgroup of users that garnered quick interest and then discontinued. This could be due to the members’ transient [32-34] nature, which prohibits them from building longer-lasting discussion threads, with over 10,484 profiles only engaging in no more than 6 tweets. It is possible that this “41 days” of extreme weight control led to a deterioration in their physical health and subsequent inpatient admission. However, this pattern of communication could also be indicative of the network structure. Previous findings indicate that Pro-ED communities have a far-reaching online community [15] but low reciprocity rates of communication with other users through replies and mentions [31]. This alludes to the allegiance the members have towards the topic and care shown among members but rarely are there extended discussion threads [35]. Additionally, the possibility of users being barred for violating the rules of engagement, especially if their postings included suicidal and self-harm messages, may account for hashtag attrition rates. This was evident within this study where 14.9% (169/1136) of the Twitter users had their account suspended or deleted between September 15, 2015 and July 1, 2018 as shown in Figure 1. A further analysis on the remaining 967 profiles revealed that only 632 profiles are currently still active, showing a 34% attrition from July 1, 2018 to September 15, 2020. For example, one Twitter user was barred for 2 months by Twitter for posting adverse Pro-ED content. This resulted in the removal of all previous postings and interactions. This account holder has since resumed being online with the same Twitter handle continuing posting Pro-ED content, however less active. This incident was documented by the authors due to the longitudinal data collected from Twitter, a strength of our study.

The data also showed retweets of the original postings being still “alive” on Twitter despite the corrective action by the platform. This leads us to question the amount of ED-related messages that could be retweeted and commented upon long after an account has been deleted. Impact of these unhealthy messages could be everlasting to the society. This warrants further investigation but also highlights the complexity in removing postings beyond the immediate network of an individual where postings are transversed fluidly and randomly. Social media platforms will need to take heed of the fact that user content lives beyond the immediate layer of where a posting has been initially lodged and could be shared across different platforms. This could impact policy development for content removal and moderation to avoid similar incidents to the live streaming of a mass shooting in Christchurch, New Zealand via Facebook [36]. To best address these and other contextual issues, social media platforms need to work closely with external support organizations to adopt a best practice approach. In the context of ED, Twitter will need to soon adopt national ED bodies as safety partners [37] to continuously engage and be advised on matters relating to ED.

From our analysis, there appears to be a reciprocal effect between followers and the followed. This implies that the Pro-ED community is resilient [38] and gains traction, as more and more people may be influenced to be part of it. This emulates cybersectarian society behaviors [39], whereby niche sentiments appeal to only a select community of people who propagate information and are virtually enduring. While opinion leaders and influencers have been found to exist within online ED communities [31], dominating members are not typically apparent. Dominating members can exert constant enforcement or exhibit power that could encourage members to change their allegiance behavior or even abandon the community [35]. As indicated through the type of tweets and number of retweets, the Pro-ED community engages with content and propagates it, and while externally, the community may appear just as an avenue for individuals seeking social support, the focus is potentially more about aligning with the collective identity of the community. Both issues with identity and social roles have been noted as risk and maintenance factors of EDs [31]. Adverse health outcomes of these groups have been observed over time on social media platforms in their desire to become “thin,” hence the crucial need for an understanding of the community structure and development of innovative intervention methods. When faced with mediation, cybersectarian groups typically react impulsively to go incognito and reappear after a length of time or remain hidden forever. For the health and safety of the members of these groups, a more participatory treatment intervention would potentially generate better outcomes compared with an outright ban, as noted by Casilli et al [40].

Understanding the vitality of Pro-ED communities is relatively complex and is reliant on the emergence of health fads and the traction of passing themes. Here, social media platforms such as Twitter would need to play a proactive role in addressing these issues. For example, Twitter should directly communicate to the 632 active profiles reported in this study to reduce further ED-related discussion and minimize sharing of related content that has a negative impact, as reported by Tiggemann and Zaccardo [41]. While Twitter has already undertaken some action within the suicide and self-harm space [42], more would be expected to follow, as Boyd [43] noted that adolescent users frequently turn to social media platforms including Twitter as a coping mechanism to diffuse external pressures threatening their mental health. It would also be beneficial for this approach of analysis to be replicated on other social media platforms to observe whether Pro-ED communities behave in the same manner across platforms. Importantly, future research should address how hashtags and other message content can be utilized to identify and reach individuals who are struggling with EDs and provide them with much needed therapeutic interventions. However, a challenge for interventions is the rapidly changing lexicon of the community [44]. As our findings indicate, hashtags accompanying Pro-ED events such as #41DaysofStarvation were short lived; however, #proana persists as a consistent theme in an otherwise transient community, potentially providing an ideal starting point for intervention.

Our analysis showed that the number of accounts followed and number of Pro-ED tweets posted were significant predictors for the number of followers of a user compared to likes. Hence the “like” counter is an obsolete predictor for ED engagement and activity. This important finding could potentially be useful to social media platforms to understand which features could help or otherwise curtail the spread of ED messages. A recent report about Instagram’s decision to turn off the “like” counter [45] might be futile to curtail ED, though the number of likes has been reported to give some indication of support [46,47].

There are several limitations to the current study. First, some data may have been omitted in the data collection process owing to the free data access from the Twitter streaming API, which normally constitutes about 1% of the whole Twitter data stream [48]. However, as mentioned by Cavazos-Regh et al [28], the percentage of private Twitter accounts is very small, and Twitter accounts default to a public setting. Results might be more reflective if we had subscribed to Twitter premium API services [49] and targeted tweets from personal accounts, and a larger sample size would have made this study more generalizable across the board. Further, a suite of other ED-related hashtags described in [16] would have contributed to a larger data set. These factors will be considered in future to improve research outcomes.

Notwithstanding these limitations, our study contributes to the emerging literature examining Pro-ED content on social media platforms by providing an understanding of the Pro-ED communities and also the engagement of these groups. Continued research is needed to understand how we might use these messages and group dynamics to provide intervention and support to people with EDs in need.