Evaluating personal narratives of interpersonal relationships via text mining to predict nonsuicidal self-injury

Abstract

Asking open-ended questions can assess constructs similar to traditional risk measures while capturing a richer portrait of current, prior, or future cognitive and affective processes. This allows researchers to identify important aspects of risk for psychological distress often overlooked in closed-ended questions. This may be particularly important for serious behavioral outcomes such as nonsuicidal self-injury (NSSI). Indeed, prior research has demonstrated that disruptions in interpersonal relationships often precede the occurrence of NSSI (e.g., Nock, Prinstein & Sterba, 2010), highlighting these events as prominent risk factors for the behavior (Kim et al., 2015). Despite this, limited research has focused on identifying which aspects (e.g., cognitive, affective, and/or social) of such interpersonal interactions may be most salient in conferring NSSI risk. We utilized text-based data from 74 individuals, 41 who reported a history of NSS. Via in-person interview, participants described an interpersonal interaction and completed a measure of emotion dysregulation, an oft-cited risk factor for NSSI (Difficulties in Emotion Regulation Scale [DERS]). Participant narrative descriptions of their interpersonal interaction were transcribed and analyzed to see if the topical content of their interpersonal interactions provided incremental (predictive) validity of NSSI status. By developing an extension of the supervised topic model that simultaneously incorporates both topics and other predictors (i.e., DERS scores), we predicted NSSI history presence. Selecting four topics as a best fitting model, both the DERS (b = 0.05, 95% BCI: [0.02, 0.08]) and a topic representing high arousal negative affect during conflict (b = 0.96, 95% BCI: [0.08, 2.01]) were positively associated with NSSI history presence with credible intervals that excluded 0 (the most representative words of this topic were “friend,” “mom,” “upset”, “mad,” “live”, “kitchen”, “annoy,” “house”, “angry”, and “start.”). The correlation between prevalence of this topic and DERS scores was 0.4, providing evidence of concurrent and incremental validity for the topic, as suggested by the gain in predictive performance. Overall, findings demonstrate the advantages of combining traditional psychological scales with text-based responses. Moreover, the use of interpersonal-focused text highlights the potential of high arousal negative affective responses during interpersonal disruptions as a driver of NSSI behavior.

Kenneth Tyler Wilcox

Statistical Consultant

My research interests include integrative data analysis, meta-analysis, topic modeling, Bayesian statistics, multilevel modeling, statistical programming, and psychology.