Predicting and explaining improvement in work and social adjustment in clients attending police service psychological therapies

Police officers face traumatic experiences such as violence, verbal abuse, exposure to accidents and crime scenes. This can lead to mental health conditions which extend into retirement as well as impacting the officers’ immediate family, resulting in the potential for impairment in work and social adjustment. Some police services therefore offer psychological mental health services to retired and retiring officers and their families. This study focuses on analysing data from the digital administrative system of a psychological therapies service offered to police officers and their families in Northern Ireland. The aim was to explore the mental health and demographic factors that predict changes in work and social adjustment through attending the service. Whilst past studies have focused on the factors related to mental health issues in police officers, fewer studies have focused on retired police officers and their families. Additionally, few studies have focused on impairment in work and social adjustment in retired and retiring police officers and their families.To address these knowledge gaps, machine learning approaches were applied alongside traditional statistical techniques to predict changes in the clients score on the work and social adjustment scale. Data were from the services administrative system, with a total of 636 observations included in the study, split into a training set (80%) and test set (20%). Ten fold cross validation, repeated five times, was used to tune the model parameters of common machine learning algorithms including decision trees, gradient boosted machines, and random forests. Interpretable machine learning techniques were applied to gain additional insight, including partial dependence plots and permutation variable importance.Descriptive statistics indicated that clients attending the service have an average age of 51 years, with 70% male, and 54% married. The most frequent condition categories include ‚`combination’ (59%), ‘other’ (19%), and ‘psychological trauma’ (17%), with frequent subcategories including Post Traumatic Stress Disorder (37%) and Anxiety Disorder (17%). Clients have an average presenting score on the work and social adjustment scale of 21 (scale range 0 ‚Äì 40). Results show a statistically significant (p<0.05) improvement in work and social adjustment during attendance at the service, with an average reduction in impairment of 10 points. More complex machine learning algorithms were most accurate in modelling the determinants of work and social adjustment, with gradient boosting resulting in the most accurate predictions on the test dataset (RMSE 7.63, R-Squared 0.42). Important predictors of improvement include baseline characteristics, completion of the full episode of care, episode length, and the client’s motivation. These relationships are further explored using techniques from interpretable machine learning, including partial dependence plots, highlighting more complex non-linear relationships. The findings have important scientific and practical implications. The results reveal important determinants of improvement in work and social adjustment. The results also highlight the potential benefit from the application of machine learning approaches alongside traditional statistical techniques in analysing psychological therapy data. This can provide useful insights into service delivery and evaluation.