Full Program »
176. Identifying and Helping Suicidal Users on a Social Media Platform: Results from an Observational Study to Inform a Social Support RCT
Suicide is a leading cause of death and a major public health concern (Nock et al., 2008; Heron, 2017). Technological innovations, such as machine learning-driven semantic classification on social media/digital platforms, have made it possible to identify high-risk users in real time; however, little empirical work has investigated how to best help such users (Jaroszewski et al., 2019). To best design an intervention aimed at increasing social support for new users posting about suicide on TalkLife (TL), a peer-to-peer digital platform popular among teens experiencing suicidal thoughts and behavior (STB), we conducted an observational study on a sample of existing TL users posting about suicide (as identified by machine learning classifiers) between May 2017 – May 2018 (N=24,356). Many users' (n=6,179) first post ever was a suicide-related post (e.g. “I want to kill myself”). Notably, nearly 20% (n=1,217) of these first-time-suicide-posters never posted again, precluding further engagement/receipt of social support on TL. We sought to determine if metadata and/or sociodemographic variables differentiated users who stayed and engaged after making an initial post about suicide versus users who dropped off. Results indicated that, after controlling for various factors (e.g., gender identification, semantic content of posts, time of day), users whose first post received more views (OR=1.03, p<.001) and comments (OR=1.41, p<.001) were significantly more likely to stay on the platform and engage than those who dropped off. A separate analysis indicated that each additional view a post received was associated with about a 3% increase in the odds the post would receive a comment (OR=1.025, p<.001), suggesting that one potential ‘nudge’-style intervention may involve (a) identifying TL users making initial posts about suicide and (b) increasing the number of views these post receive, possibly increasing the number of comments received and, in turn, the chances these high-risk users will engage/receive social support from other TL users. We will use causal inference techniques (cf., Pearl, 2009) on these cross-sectional data to better estimate the effect of post-views and comments, respectively, informing the appropriate dosage (i.e., number of views) of the intervention, which will be tested with an upcoming RCT.
Heron, M., Ph, D., & Statistics, V. (2017). National Vital Statistics Reports Deaths : Leading Causes for 2015, 66(5).
Jaroszewski, A. C., Morris, R. R., & Nock, M. K. (2019). Randomized controlled trial of an online machine learning-driven risk assessment and intervention platform for increasing the use of crisis services. Journal of consulting and clinical psychology, 87(4), 370.
Nock, M. K., Borges, G., Bromet, E. J., Alonso, J., Angermeyer, M., Beautrais, A., ... & De Graaf, R. (2008). Cross-national prevalence and risk factors for suicidal ideation, plans and attempts. The British Journal of Psychiatry, 192(2), 98-105.
Pearl, J. (2009). Causality. Cambridge university press.