dc.description.abstract | Abstract Personality disorders (PDs) are broadly understood to exhibit high service utilisation, however individual patterns of engagement within complex needs are poorly understood. The clinical relevance of engagement behaviour was delineated, and the causal-predictive relevance of engagement archetypes on clinical outcomes was modelled. Methods A sample of 7,897 episodes comprising 3,941 patients with a diagnosis of personality disorder (PD) receiving care in South London & Maudsley secondary services was assembled. Data were recorded at first and final episodes of care (Time 1 versus Time 2), with an exposure window of 11 years (2007-2018). PLS SEM modelled the indirect effect of engagement behaviour on HoNOS clinical score change. Engagement comprised face-to-face contacts (F2F) and number of did-not-attends (DNA), per episode of care. LPA was conducted on Time 1 data and causal mediation analysis tested the direct effect of derived latent profiles on clinical outcomes (clinical score and length of episode). Results Linear mixed effect model showed that greater engagement improved outcomes only for those with higher Total clinical scores (> 15). PLS-SEM considered three candidate models: mediation, explanation, and prediction. Mediation was not upheld in model testing, and in the explanatory model, Direct effects between exogenous constructs and the outcome generally failed to reach statistical significance. For Demographic adversity, the path was β = .009, 95% CI[-.05, .04]. For Diagnosis, β = .001, 95% CI[-.03, .03]. Predictive appraisal fared better, with Mean Absolute Error and Root Mean Square Error values demonstrating superior model performance, compared to naïve benchmarks. Demographic and diagnostic factors poorly explained clinical outcome, although 4 predictive out-of-sample modelling showed clinical score change as having good generalisability, outperforming naïve benchmarks. LPA identified two subgroups, characterised by low and high service receipt, denoted by Profile 1 (n = 2,879, 73.05%) and Profile 2 (n = 1,062, 26.95%), respectively. A two-profile solution (P < .01) was preferred over a three-profile solution, which was non-significant. In unconditional (t = 19.53, P < .001, B = 7.27, CI 6.54 – 8) and conditional models (t = -3.31, P <.001, B = -74.94, CI -119.34 – -30.56), cluster membership was significantly related to receipt of nursing contacts, over and above other team contacts. Causal mediation analysis demonstrated that counterfactual classification of low engagers may adversely impact the treatment effect, increasing clinical scores. For high engagers, such classification extended episodes by 400 days. Conclusions Results suggest that routinely collected data may be used effectively to classify likely engagement archetypes. This and future algorithms could enable services to tailor treatment pathways to individual need. | en_US |