Predicting Mental Health From Followed Accounts on Twitter

The past decade has seen rapid growth in research linking stable psychological characteristics (i.e., traits) to digital records of online behavior in Online Social Networks (OSNs) like Facebook and Twitter, which has implications for basic and applied behavioral sciences. Findings indicate that a broad range of psychological characteristics can be predicted from various behavioral residue online, including language used in posts on Facebook (Park et al., 2015) and Twitter (Reece et al., 2017), and which pages a person ‘likes’ on Facebook (e.g., Kosinski, Stillwell, & Graepel, 2013). The present proposal seeks to examine the extent to which the accounts a user follows on Twitter – their Twitter friends – can predict individual differences in self-reported anxiety, depression, post-traumatic stress, and anger. Studying Twitter friends offers distinct theoretical and practical advantages for researchers, including the potential for less overt impression management and better capturing passive users. By incorporating best practices in open science and machine learning, we aim to provide unbiased estimates of predictive accuracy for predicting Mental Health from Twitter friends. Our findings will have implications for theories linking psychological traits to behavior online, applications seeking to infer psychological characteristics from records of online behavior, and for informing discussions of how such applications could affect users’ privacy. Predicting Mental Health from Followed Accounts 3 Predicting Mental Health from Followed Accounts on Twitter Stable psychological characteristics are expressed behaviorally in many domains, including online, where they often leave more or less permanent digital records in their wake. The extent to which stable individual differences in mental health are expressed online, imprinted in corresponding digital records, and ultimately recoverable from these records has wide-ranging implications for basic and applied behavioral sciences. Inferring individuals’ mental health status from online records with an appreciable degree of accuracy could accelerate advancements in clinical science, easing the burdens for researchers and participants imposed by traditional survey-based research. In time, such approaches could be developed into tools useful for clinical practice and public health. At the same time, the promise of inferring mental health from digital records of behavior is accompanied by potential threats to individuals’ privacy, as such tools could be used to infer a person’s mental health without their explicit consent. Given both the promise and risks, we need to better understand how mental health is reflected in, and recoverable from, digital records of online behavior. Our focus here is on inferring depression, anxiety, anger, and post-traumatic stress from the accounts users choose to follow on the popular online social network (OSN), Twitter. Psychological Traits can be Inferred from Digital Records The theory of behavioral residue holds that one by-product of the expression of traits is the accumulation of lasting residual traces of past behavior in the physical or digital spaces a person occupies. Early work demonstrated that human judges could infer psychological traits from behavioral residue in physical living and working spaces with considerable accuracy (Gosling, Ko, Mannarelli, & Morris, 2002). More recently, researchers have trained machine learning algorithms to do so with behavioral residue found in OSNs such as Facebook (Kosinski, Predicting Mental Health from Followed Accounts 4 Stillwell, & Graepel, 2013; Park et al., 2015; Schwartz et al., 2013). Behavioral residue in OSNs has included linguistic content (e.g., Facebook status updates; Park et al., 2015; Schwartz et al., 2013) and which pages a person has ‘liked’ on Facebook (Facebook-like ties; e.g., Kosinski, Stillwell, & Graepel, 2013), both having demonstrated considerable predictive accuracy. Indeed, the accuracy of inferences based on Facebook-like ties can even exceed that of knowledgeable human judges (Youyou, Kosinski, & Stillwell, 2015). We focus here on using behavioral residue from Twitter, an OSN that differs from Facebook in ways relevant to basic psychological theory, public health applications, and privacy concerns. Twitter is an OSN service and microblogging platform used by approximately 24% of US adults (as of January 2018; Pew Research Center, 2018). Users post short messages of no more than 280 characters called “tweets” that other users can see, respond to, share (called “retweeting”), or react to (via a “like” button). Unlike Facebook, accounts are public by default, and most users choose to keep their accounts public; Twitter does not release the percentage of public accounts, but a 2009 report found that 90% of accounts were public, with a trend towards even fewer private accounts (Moore, 2009). The public nature of Twitter makes it an especially interesting setting for the present investigation for two reasons. First, its public nature eases the burden of collecting users’ data: one of several off-the-shelf Python (e.g., Tweepy; Roesslein, 2009) or R libraries (e.g., twitter; Gentry, 2015) can be used to download any of these many public accounts’ data, including their recent tweets, whom they follow, and who follows them. Thus, there is at least one fewer barrier to people outside of the Twitter company for implementing beneficial (e.g., public-health) or harmful (e.g., discriminatory) applications on Twitter than less public-facing OSNs like Facebook. Second, its public nature could affect the 1 Prior to November 2017, tweets were limited to 140 characters. Predicting Mental Health from Followed Accounts 5 relative candor of behavior on Twitter, since efforts to manage others’ impressions can be stronger in more public settings (Leary & Kowalski, 1990; Paulhus & Trapnell, 2008). Previous work has attempted to infer or predict psychological traits from behavioral residue on Twitter, focusing primarily on linguistic analyses of tweets. This growing body of work demonstrates that tweets can be used to predict a wide range of psychological characteristics, including self-reported personality traits, affective states, depression, posttraumatic stress, and the onset of suicidal ideation (Coppersmith, Harman, & Dredze, 2014; De Choudhury, Counts, & Horvitz, 2013; De Choudhury, Gamon, Counts, & Horvitz, 2013; De Choudhury, Kiciman, Dredze, Coppersmith, & Kumar, 2016; Dodds, Harris, Kloumann, Bliss, & Danforth, 2011; Nadeem, Horn, & Coppersmith, 2016; Park, Cha, & Cha, 2012; Qiu, Lin, Ramsay, & Yang, 2012; Reece et al., 2017). Although the heterogeneity in how mental health is measured can make interpretation challenging, in broad terms these previous studies suggest that behavior on Twitter relates meaningfully to psychological traits. In contrast to the emphasis on linguistic analyses, there has been relatively little work using network ties on Twitter to predict psychological traits. The few attempts have looked at abstract structural characteristics of ties (e.g., tie counts or social network density; Golbeck, Robles, Edmonson, & Turner, 2011; Quercia, Kosinski, Stillwell, & Crowcroft, 2011) rather than treating the specific accounts to which a user is tied as meaningful. We focus here on the specific accounts that users follow. Ties or connections on Twitter are directed, meaning that users can initiate outgoing ties (called “following” on Twitter) and receive incoming ties (called “being followed” on Twitter) which are not necessarily reciprocal. In keeping with the terminology of Twitter’s Application Programming Interface (API), we refer to the group of users that a person follows as their friends Predicting Mental Health from Followed Accounts 6 and the group of users that follow a person as their followers. While both ties are likely rich in psychological meaning, we focus on friends in the present investigation for several reasons. First, a user has nearly complete control over the accounts they follow, making friends a more direct product of the user’s own behavior. While most of a users’ followers likely reflect their own behavior and relationships, some might be unrelated (e.g., spam accounts, bots, users looking for reciprocated following, etc.), increasing the relative noise among followers (vs. friends). Second, following accounts is the primary way users’ curate their feed – what they see when they log into the app – and so their choice of friends likely reflects the information they are seeking out on Twitter. In this way, following accounts on Twitter is similar to liking pages on Facebook, a behavior which has been previously demonstrated to robustly predict psychological characteristics (Kosinski et al., 2013; Youyou et al., 2015). Third, an important practical consideration for a predictive modeling approach is that friends on Twitter often include famous brands, celebrities, politicians, or other high in-degree accounts, which appeal to similar interests or motivations in many users. Users are thus likely to share some friends even in moderately small samples, whereas they may have no followers in common because there aren’t parallel high out-degree accounts that appear in many users’ follower networks. Consequently, friends are far less likely to be zero-variance predictors than followers in moderately-sized, random samples. Predicting Mental Health from Twitter-Friend Ties Twitter-friend ties are an important next step in studying behavioral residue online for both theoretical and practical reasons. In contrast to tweets, Twitter-friend ties are not explicit signs or displays intended to be consumed by an audience of other people, and so they may be less subject to impression management goals. For this reason, Twitter-friend ties may be Predicting Mental Health from Followed Accounts 7 especially apt for predicting more evaluative psychological traits lik