Performance of statistical models to predict mental health and substance abuse cost
2006

Predicting Mental Health and Substance Abuse Costs

Sample size: 525620 publication Evidence: high

Author Information

Author(s): Maria Montez-Rath, Cindy L. Christiansen, Susan L. Ettner, Susan Loveland, Amy K. Rosen

Primary Institution: Boston University School of Public Health

Hypothesis

Which statistical model works best for predicting costs in a large sample of patients with mental health and substance abuse disorders?

Conclusion

Models with square-root transformation or link fit the data best for predicting costs in patients with mental health and substance abuse disorders.

Supporting Evidence

  • The Square-root Normal model had the lowest RMSE and MAPE values.
  • The Gamma with square-root link model performed well across all deciles of predicted costs.
  • Models tailored to deal with skewed data performed better than the OLS model.

Takeaway

This study looked at different ways to predict how much money patients with mental health and substance abuse issues will cost. They found that using a special math trick called square-root helps make better predictions.

Methodology

The study analyzed cost data from 525,620 Veterans Health Administration patients using various statistical models, including OLS, Log Normal, and Gamma models.

Potential Biases

The OLS model produced a high percentage of negative predictions, which could mislead individual cost predictions.

Limitations

The Gamma with square-root link model had convergence problems with smaller samples, and the study only used one risk-adjustment system.

Participant Demographics

95% of participants were male, with a mean age of 57 years; about 30% were age 65 and older.

Digital Object Identifier (DOI)

10.1186/1471-2288-6-53

Want to read the original?

Access the complete publication on the publisher's website

View Original Publication