OBJECTIVES: Microsimulation using risk equations to convert biomarker values into event risks is the norm in T2DM. Though risk factor clustering (whereby individuals with one unfavorable risk factor are likely to have other unfavorable risk factors as well) is common in diabetic populations, accounting for it in empirical applications is rare despite the longstanding example of the Global Diabetes Model (GDM). This absence can potentially bias cost-effectiveness estimates. While the GDM approach is data-intensive, the problem can be addressed simply by allowing correlation of baseline patient characteristics. This study aims to leverage National Health and Nutrition Examination Survey (NHANES) data to estimate correlation coefficients and fill a gap in the literature and to explore bias potential using examples from the US 3^rd party payer perspective.

METHODS: Two cohorts of individuals with T2DM—biguanide only and sulfonylurea + biguanide—were identified in the five NHANES cross-sections between 2007 and 2016. Baseline characteristics and correlation coefficients were estimated using stratification weights. Baseline population characteristics were entered into ECHO-T2DM, a validated microsimulation model of T2DM, and 20-year cost-effectiveness simulations were run including and then excluding the correlation coefficients. While hypothetical, the intervention reflects common scenarios.

RESULTS: The correlation coefficients spanned from tightly correlated (-0.91 for HDL and triglycerides) to almost uncorrelated (0.01 for HDL and BMI), with some variation by cohort. While estimated cost-effectiveness ratios did not change qualitatively when correlations were added, there were important numerical differences. It should be acknowledged that the NHANES sample sizes are relatively small and that conditioning on HbA1c failure (typical in T2DM modeling) was not possible.

CONCLUSIONS: Risk factor clustering may be important for modeling the cost-effectiveness of T2DM interventions, correlation in sampling baseline characteristics is easy, and now two sets of correlation coefficients (albeit crude) are available for other researchers.