| To Register SMDM Homepage | |
|
|
|
||||
Purpose: Probabilistic sensitivity analyses have frequently been used to characterize uncertainty in cost-effectiveness model results. However, other than variable correlation inherently introduced by Markov processes, correlations between the input variables are rarely explicitly included in such simulations, particularly for models that incorporate estimates obtained from different sources. We hypothesized that varying assumptions about the underlying correlation among input variables might influence the results of probabilistic sensitivity analyses. Methods: We created a cost-effectiveness Markov model comparing two hypothetical treatment strategies (treatments A and B), each with an associated annual probability of death and an annual treatment cost. We designated lognormal distributions for costs and logistic distributions for mortality probabilities with fixed parameters, and we generated 1000 samples for each input variable to represent typical data available for a model. Bootstrap samples from these populations were used to estimate distributional parameters for cost and effectiveness. We then performed Monte Carlo simulations in which different degrees of correlation were assumed between the four patient-level cost and effectiveness model inputs. Two hundred groups of 1000 first-order simulations were generated for each set of correlation assumptions. We subsequently calculated the probability that each strategy was truly the optimal alternative at selected cost-effectiveness thresholds (λ). We also calculated the expected value of perfect information (EVPI) for each threshold. Results: Probabilistic sensitivity analyses produced cost-effectiveness acceptability and EVPI curves that varied substantially depending on the underlying correlation structure. If costs and effectiveness were assumed to be uncorrelated, the probability that treatment B was optimal at λ=$100,000 per quality-adjusted life-year (QALY) was 0.77, and the EVPI was $10,353 per patient. However, if costs and benefits for treatment A were highly correlated, the probability that treatment B was “correct” at λ=$100,000/QALY was only 0.04, and the EVPI was $2075 per patient. If benefits and costs of both treatments were highly correlated both within and between treatments, then the probability treatment B was optimal at λ=$100,000/QALY was 0.36, and the EVPI was $18,063 per patient. Conclusion: The underlying correlation structure among costs and outcomes in cost-effectiveness models can profoundly influence the results of probabilistic sensitivity analyses. Inaccurate assumptions about correlation structure could greatly bias assessments of cost-effectiveness model uncertainty. The explicit assumptions about input variable correlation should therefore accompany reported results of probabilistic sensitivity analyses.
See more of Oral Concurrent Session B - Methodological Advances
See more of The 26th Annual Meeting of the Society for Medical Decision Making (October 17-20, 2004)