Comparing single-level and multilevel regression analysis for risk adjustment of treatment outcomes in common mental health disorders

Aim: The aim of this paper is to compare single-level and multilevel regression analysis to obtain risk-adjusted outcomes from mental health care providers. Subject and methods: The study population consisted of adult patients receiving treatment for common mental health disorders. The outcome was self-reported symptom level at post-test. Risk adjustment models were developed using single- and multilevel regression analysis. In the multilevel approach, a random intercept for each provider was included. The intraclass correlation coefficient was used to estimate the proportion of variability in treatment outcome between providers. Spearman correlation coefficient of ranks was used to compare results between the two approaches. Results: The effects of most casemix variables on outcomes were similar for the two models. The ranking of providers in both methods was also quite similar (ρ =.99). The multilevel model estimated that 5.4% of total variability in adjusted post-test scores was explained by the provider factor. Conclusions: The findings of risk adjustment of mental health outcomes are quite robust for the use of single-level or multilevel regression analysis in the current study. However, given the small but significant amount of variation in outcomes that is attributable to providers, the multilevel approach is recommended for dealing with outcomes when patients are clustered within providers.