What Counts as Mental Illness in Suicide Records, and For Whom?

• In Event: Computational Methods and Models for Social Analysis

Abstract

Although mental illness is considered a core risk factor for suicide, prior work observes surprisingly low rates of mental distress among suicides in our national surveillance system for suicide (the National Violent Death Reporting System, or NVDRS). Some argue that this reflects an undercount of the true prevalence of mental distress and caution that this undercount might disproportionately affect men and minoritized racial/ethnic groups. Meanwhile, other work hints at a meaningful portion of suicides that reflect mechanisms beyond mental illness. Drawing on data from the NVDRS, we first estimate the potential undercount in mental distress among suicides and examine how this undercount varies by sex and race/ethnicity. To do so, we use a language model trained to predict structured measures of mental health in the NVDRS based on summaries of coroner/medical examiner reports. Second, we analyze the textual signals of mental distress embedded in these summaries. Our preliminary results hint that among the suicide decedents who were not coded as having mental distress, roughly 28% had evidence of mental distress in their coroner/medical examiner narrative that should have warranted positive coding. Quantitative and qualitative examination of model classifications indicates that undercounts may arise not only from missing information but also from inconsistencies in how mental distress is captured in the structured NVDRS variables. Preliminary results also hint that classification processes are uneven across racial and ethnic groups but not sex. Overall, this study provides new insight into how mental distress is recorded and sometimes misclassified in national suicide statistics. Methodologically, this study contributes to an emerging computational social science framework that treats model misclassification as analytically meaningful rather than purely technical error, a strategy that is becoming increasingly powerful as predictive models improve.

Authors

• Alina Arseniev-Koehler, Purdue University

• Wesley Wang, Purdue University

• Vickie Mays, Department of Psychology, UCLA and Department of Health Policy and Management, UCLA Fielding School of Public Health

• Christina Chance, Department of Computer Science, Samueli School of Engineering, UCLA

• Susan Cochran, Department of Epidemiology, UCLA Fielding School of Public Health and Department of Statistics and Data Science, UCLA