OBJECTIVES: Health economic models commonly include costs calculated via regression equations based on clinical outcomes derived from real world data. When the inputs of these regression equations are functions of treatment effects from NMA, there may be situations where failure to capture correlation in treatment effects across outcomes can lead to important bias.

METHODS: We use a simple simulation of two correlated binary outcomes in a three treatment network where the outcomes are subequently input into a poisson regression specifying the log of the mean population cost being a function of an intercept term and the two binary variables. We show that in this situation (and any situation where the regression equation is non-linear) failure to capture correlation in outcomes within the baseline model results in inaccurate estimates of mean population cost, and failure to capture correlation in treatment effects leads to bias in the estimated cost differences. This could lead to important implications for decision making, particularly in the extreme case where treatment modifies the correlation between variables.

RESULTS: (Same as above)

CONCLUSIONS: In models where costs or benefits are calculated as some non-linear function of comparative efficacy outputs, multivariate NMA methods may be required to accurately capture differences in costs and benefits across therapies.