Confidence intervals for cost-effectiveness ratios

The problem of variability in computed cost-effectiveness ratios (CERs) is usually addressed by performing sensitivity analyses to determine the effects on these ratios of plausible ranges of values of input parameters. However, the sampling variation that exists in these estimated parameters can be utilized to obtain confidence intervals for cost-effectiveness ratios. As cost-effectiveness analysis becomes more widely used, new techniques need to be developed for establishing when a difference in strategies evaluated is meaningful. A first step is to establish the precision of the CER itself. The authors estimate the precision of a CER in the context of a statistical model in which the primary outcome is survival, with cost and effectiveness defined in terms of the underlying survival distribution (S). Effectiveness (alpha) is measured by life expectancy, restricted to a finite time horizon and discounted at a fixed rate r, alpha = integral of e-rtS(t)dt. Cumulative cost (beta) per patient is regarded as resource utilization and incurred randomly over time depending on the survival experience of the patient, beta = integral of e-rtS(t)dC(t), where C(t) is the total potential resources utilized up to time t. Average cost-effectiveness (ACE) of a single strategy is beta/alpha, and when comparing two strategies, the CER is delta beta/delta alpha, the ratio of the incremental cost to the difference in mean survival. Utilizing the sampling distribution of the Kaplan-Meier estimate of S yields standard errors and confidence intervals for ACE and CER. The technique is applied to survival data from 218 previously studied patients to assess 95% confidence intervals for the CER and ACE of the implantable cardioverter defibrillator as compared with electrophysiology-guided therapy.