PMA11: COMPARISON OF STATISTICAL TESTS FOR THE COST/EFFECTIVENESS RATIOS

Abstract

BACKGROUND: Results of a cost-effectiveness analysis are summarized in a series of cost-effectiveness (C/E) ratios. As with other statistics, the C/E ratio is subject to sampling variation. However, constructing a confidence interval for the cost-effectiveness ratio is complicated because both numerator and denominator of the ratio are stochastic in nature. Several statistical methods have been proposed lately, yet, the systematic method of handling uncertainty is generally underdeveloped in economic evaluation. OBJECTIVE: This study is to compare the statistical methods proposed in constructing confidence intervals for the various ratio distributions, for the various correlation coefficients between numerator and denominator, for the various coefficient of variations, and the various sample sizes. METHODS: The ratio distributions are formed from the combinations of normal, log normal, and gamma distributions, which frequently arise in cost-effective studies. In evaluation the performance of statistical methods, a simulation study was conducted for the various ratio distributions. For each sample, the confidence intervals were constructed by five statistical procedures; the Box, Taylor's, Fieller's, bootstrap, and jackknife methods. RESULTS: First, since the ratio distributions are largely dependent on the distribution of its denominator, none of the statistical tests work if the mean of denominator is close to zero. Second, if the sample size is small, none of the statistical tests perform well enough regardless of correlation and coefficient of variation. Third, for the large sample size, all methods, except the box method, constructed the confidence interval well. Among them, Fieller's method is the first choice of selection for the estimation of the confidence interval. CONCLUSION: None of the statistical tests work if the mean of denominator is close to zero. This is problematic for Incremental Cost-effectiveness Ratio (ICER). If the “net effect” of the new procedure is insignificant, then the statistical test for the ICER is not stable.