Using machine learning to retrospectively predict self-reported gambling problems in Quebec.

Area of Science:

• Computational Social Science
• Machine Learning in Behavioral Science
• Online Gambling Research

Background:

• Online gambling participation correlates with increased gambling-related harms.
• Effective harm prevention requires early detection of at-risk individuals.
• Machine learning offers potential for personalized intervention models.

Purpose of the Study:

• To assess machine learning algorithms' ability to detect at-risk online gamblers.
• To utilize site-generated data for retrospective identification of problem gambling.
• To compare the efficacy of various supervised machine learning methods.

Main Methods:

• Exploratory comparison of six supervised machine learning algorithms.
• Prediction of problem gambling risk using the Problem Gambling Severity Index (PGSI).
• Analysis of 144 predictor variables from user transaction and behavior data on Lotoquebec.com.

Main Results:

• Random forest models achieved high accuracy (84.33% for PGSI 5+, 82.52% for PGSI 8+).
• Key predictors included betting frequency, variability, and repeat engagement.
• Model performance was evaluated using receiver operating characteristic curves.

Conclusions:

• Machine learning effectively classifies at-risk online gamblers using platform data.
• Personalized harm prevention initiatives are feasible.
• Model performance involves trade-offs between sensitivity and precision.