Exploring predictive disciplinary exposure in police recruitment

A machine learning approach

Bibliographic Data

ID21372343
AuthorsIan Gibson (0000-0001-9154-8247, Research Department, Xenara-AI Research, corresponding author), Gareth Stubbs (0000-0002-7631-2823, Rabdan Academy)
Year2026
Publication date2026-07-06
Open AccessYes
TypeARTICLE
VenueThe Police Journal Theory Practice and Principles (JOURNAL)
Journal identifiersISSN: 0032-258X • E-ISSN: 1740-5599
PublisherSAGE Publications (PUBLISHER • US)
DOI10.1177/0032258x261457279
OpenAlexW7167494796
LanguageEN
References cited56

This exploratory study examines the connection between initial motivations for joining the police, social connections and subsequent misconduct investigation using a random forest machine learning model. Drawing upon a survey of retired officers from the US and the UK ( n = 228), this research assesses how pre-entry motivations and social exposure may influence later exposure to misconduct investigation. For analyses requiring complete cases across all variables, the analytical sample comprised 214 officers. Feature importance findings show a notable relationship between entry motivations, social ties and exposure to misconduct investigations, which highlights the potential of predictive analytics to identify risk factors for future disciplinary issues in new police recruits. The analysis indicates that motivations related to authority, economic benefits, job excitement and social ties to current or former police officers appeared among the strongest predictors of exposure to misconduct investigation in this sample. Conversely, desire to serve (a public service motivation indicator) showed lower predictive value in this sample. This study employed a clearly specified machine learning protocol, including feature engineering and ensemble learning, with a stratified five-fold cross-validation on the training set, Synthetic Minority Oversampling Technique (SMOTE) applied within folds only to avoid leakage, and hyperparameter tuning via grid search evaluated by Receiver Operating Characteristic–Area Under the Curve (ROC–AUC) as the primary metric, along with Matthews Correlation Coefficient (MCC) as secondary. Despite strong apparent performance, the exploratory dataset and class imbalance require confirmation on independent cohorts and careful calibration assessment before operational use. The implications of these findings are significant for policymakers, police organisations and governance agencies, suggesting a data-driven approach for enhancing recruitment processes. This research contributes to a more nuanced understanding of the factors influencing police officer behaviour and underscores the potential of predictive modelling in addressing disciplinary challenges within law enforcement. Future research should expand sample size, harmonise outcome definitions across jurisdictions, and evaluate model calibration and group fairness

Discipline · Feature (linguistics) · Misconduct · Predictive analytics · Predictive validity · Random forest · Sample (material) · Social learning · Crime Patterns and Interventions · Deception detection and forensic psychology · Policing Practices and Perceptions

Tools

Open DOIOpen Access