Purpose: We provide a method for decomposing a Markov cohort analysis into multiple simpler cohort analyses that are not only simpler but computationally more efficient.
Methods: Markov models for medical cost-effectiveness often have multiple factors or components that model different features of a decision problem. For example, Col and colleagues' recent model of the survival benefit of tamoxifen had components describing breast cancer, endometrial cancer, venous thrombosis, pulmonary embolism, and other-cause mortality. In such models the number of health states can quickly proliferate. We show here that when independence holds across components, then a standard cohort analysis may be decomposed into several simpler cohort analyses, one for each factor, and the results may be combined to produce desired expected costs and QALYs.
Results: As an illustration, we replicate a study by Schousboe and colleagues, who conducted a cost-effectiveness analysis of alendronate therapy for fracture risk in osteopenic postmenopausal women. A full cohort analysis of Schousboe's model would require 33 health states. Schousboe's analysis identified what was deemed to be the seven most important states and restricted transition to those, but the result was the loss of some model features such as the ability to record the history of multiple fracture types. Our re-analysis instead decomposes the original model into 6 simple factors, most containing only two states. This approach retains desirable features of the original 33-state model with over a 50% reduction in model complexity.
Conclusion: Cohort decomposition can not only reduce computation but simplify model formulation and presentation, thereby opening models to peer inspection and critique. 