Estimating Determinants of Attrition in Eating Disorder Community on Social Media: An Instrumental Variables Approach Tao Wang 1,2* , Emmanouil Mentzakis 1 , Markus Brede 3 , and Antonella Ianni 1 1 Department of Economics, University of Southampton, UK 2 The Alan Turing Institute, UK 3 Department of Electronics and Computer Science, University of Southampton, UK * t.wang@soton.ac.uk ABSTRACT High attrition is a major problem in online health interventions. However, little is known about risk factors of people dropping out online. Challenges exist due to the lack of baseline knowledge on how attrition naturally happens in online communities, and reliable methods that can identify causality between past traits and future dropout. Here, we examine characteristics of naturally occurring attrition in online health communities and use longitudinal statistical models to assess whether emotion and online social networks influence dropout behaviors in the future. From three comparable subpopulations sampled from Twitter, we find that individuals who self-identified as eating disordered have shorter active durations than general populations, with a half of cohort dropping out in 6 months after creating a Twitter profile. Applying instrumental variables estimation and survival analysis to longitudinal data on users’ activities spanning 1.5 year, we identify that negative emotions increase forthcoming dropout in general populations, while positive emotions instead increase dropout in disordered populations. Individuals who are surrounded by many active peers and those who are central in the network tend to stay longer in the future. We interpret our findings with clinical evidence and discuss their implications for designing network interventions that can promote organizational well-being in online communities. 1 Introduction Eating disorders (ED), such as anorexia nervosa and bulimia, are a major public health concern due to a high mortality rate (highest of any mental illness) 1 , intractable co-morbidities 2 and worldwide prevalence 3, 4 . More than 2.7% of 13-17 year olds in the US 3 and 725,000 people in the UK 4 have been affected by ED, with a trend that is increasing over time. Although health interventions have been proposed, ED population are very hard to reach and as such study, plan and administer intervention to those in need 5 . Individuals often conceal their ED symptoms due to feelings of shame or fear of stigma 6, 7 and many never disclose their struggles with professionals 8 . Due to the secretive nature and the need to ensure anonymity, people suffering from ED often seek for social support and resources from peer-communities online 9 , particularly via social networking sites (SNS) such as Twitter and Facebook. Engagement in these online communities is common among individuals with ED 10–15 and has recently been suggested as a screening factor for ED 2 . Given these facts, vast and growing research has focused on whether public health and policy can harness the power of such social networks for the benefit of ED sufferers and how to leverage online interventions over SNS to promote healthy behaviors and improve community-level well-being 16–18 . Compared with traditional approaches, online interventions appear to be more accessible for broad audiences and more cost-effective in achieving short- or long-term goals in public health 19 . Existing interventions delivered via online health websites have shown a positive effect on healthy behavioral outcomes 20, 21 . However, using online communities to develop and deliver successful interventions requires stability and frequency of interactions within these communities themselves 17, 22 . For communities with a very high dropout rate, it is unlikely that members will have adequate opportunity to promote a target behavior change. Attrition (i.e. participants stopping usage or are lost in follow-ups) has been identified as a crucial issue in the efficacy of online interventions 21, 23, 24 , since cost-effectiveness is largely reduced for population-level interventions as the number of people reaping their benefits goes down 25 . A recent meta-analysis of 22 studies found that all studies suffered from decreased participation throughout the intervention period, with 12 studies reporting rates of more than 20% 24 . Despite such high attrition rates, characteristics that differentiate dropouts from completers at various time points in an online intervention are still unknown in the literature 26, 27 , even under-explored in the research on traditional face-to-face interventions on various behavior-related conditions, such as obesity, smoking and alcohol misuse 25, 28 . Although understanding these characteristics is important to implement alternative strategies that can improve retentions