Bootstrap analyses of cost effectiveness in antidepressant pharmacotherapy

In this study, we describe 'bootstrap' methodology for placing statistical confidence limits around an incremental cost effectiveness ratio (ICER). This approach was applied to a retrospective study of annual charges for patients undergoing pharmacotherapy for depression. We used MarketScanSM (service mark) data from 1990 to 1992, which includes medical and pharmacy claims for a privately insured group of employed individuals and their families in the US. Our primary effectiveness measure was the proportion of patients who remained stable on their initial antidepressant medication for at least 6 consecutive months. Our primary cost measure was the total annual charge incurred by patients taking the selective serotonin reuptake inhibitor fluoxetine, a tricyclic antidepressant or a heterocyclic antidepressant. On average, fluoxetine pharmacotherapy tended to decrease annual charges by $US16.48 per patient for each percentage increase in depressed patients remaining stable on initial pharmacotherapy for 6 months, resulting in a negative ICER point-estimate. However, the upper ICER confidence limit is positive, which means that fluoxetine treatment may possibly increase annual per patient charges. With 95% confidence, any such increase was no more than $US130 per patient for each percentage increase in patients remaining stable on initial pharmacotherapy for at least 6 months. One advantage of using a bootstrap approach to ICER analysis is that it does not require restrictive distributional assumptions about cost and outcome measures. Bootstrapping also yields a dramatic graphical display of the variability in cost and effectiveness outcomes that result when a study is literally 'redone' hundreds of times. This graphic also displays the ICER confidence interval as a 'wedge-shaped' region on the cost-effectiveness plane. In fact, bootstrapping is easier to explain and appreciate than the elaborate calculations and approximations otherwise involved in ICER estimation. Our discussion addresses key technical questions, such as the role of logarithmic transformation in symmetrising highly skewed cost distributions. We hope that our discussion contributes to a dialogue, leading ultimately to a consensus on analysis of ICERs.