Confounding and missing data in cost-effectiveness analysis : comparing different methods

Introduction: Common approaches in cost-effectiveness analyses do not adjust for confounders. In nonrandomized studies this can result in biased results. Parametric models such as regression models are commonly applied to adjust for confounding, but there are several issues which need to be accounted for. The distribution of costs is often skewed and there can be a considerable proportion of observations of zero costs, which cannot be well handled using simple linear models. Associations between costs and effectiveness cannot usually be explained using observed background information alone, which also requires special attention in parametric modeling. Furthermore, in longitudinal panel data, missing observations are a growing problem also with nonparametric methods when cumulative outcome measures are used. Methods: We compare two methods, which can handle the aforementioned issues, in addition to the standard unadjusted bootstrap techniques for assessing cost-effectiveness in the Helsinki Psychotherapy Study based on five repeated measurements of the Global Severity Index (SCL-90-GSI) and direct costs during one year of follow-up in two groups defined by the Defence Style Questionnaire (DSQ) at baseline. The first method models cumulative costs and effectiveness using generalized linear models, multiple imputation and bootstrap techniques. The second method deals with repeated measurement data directly using a hierarchical two-part logistic and gamma regression model for costs, a hierarchical linear model for effectiveness, and Bayesian inference. Results : The adjustment for confounders mitigated the differences of the DSQ groups. Our method, based on Bayesian inference, revealed the unexplained association of costs and effectiveness. Furthermore, the method also demonstrated strong heteroscedasticity in positive costs. Conclusions: Confounders should be accounted for in cost-effectiveness analyses, if the comparison groups are not randomized.