Augmenting Predictive Models in Forensic Psychiatry with Cultural Consensus Theory

Forensic psychiatric hospitals regularly monitor the mental health and forensic risk factors of their patients. As part of this monitoring, staff score patients on various items. Common practice is to aggregate these scores across staff members. However, this is suboptimal because it assumes that assessors are interchangeable and that patients are independent. An improvement over averaging scores is the use of Cultural Consensus Theory (CCT), which imposes a hierarchical model across patients, staff members, and items. While accounting for differences between patients and staff members, CCT estimates a ‘true’ score for each patient on each item based on the consensus among staff members. Here, we apply a CCT model to data from a Dutch maximum-security forensic psychiatric centre and use the inferences to predict violent behaviour in patients. The CCT model outpredicts several alternatives, such as random forest and boosted regression trees, albeit by a small margin. We discuss practical limitations and directions for how future monitoring of patients could be adapted to maximize the added value of a CCT-based approach.