Content area
Full Text
Summary statistics derived from binary measures of treatment outcome can be divided into two categories. The first category comprises arithmetic measures such as absolute risk difference (ARD) and its reciprocal, the number-needed-to-treat (NNT); the second category comprises relative (ratio) measures such as relative risk (RR) and odds ratio (OR). Economic analysis and assessment of net clinical benefit typically require estimates of ARD; however, ARD is seldom taken directly from trials as it usually varies with patients' baseline risk of the outcome.[1-3] Instead, the convention is to adopt a two-step procedure. First, the baseline risk for the particular circumstances of the economic analysis is estimated. Then, under the assumption of constant relative treatment effect, this risk is modified by an appropriate algebraic form of the average relative treatment effect (e.g. RR, OR) from a trial (or meta-analysis of trials) to estimate an arithmetic measure of treatment effect (e.g. ARD, NNT) for circumstances relevant to decision making.
There is continuing debate about whether it is better to use OR or RR as the measure of relative treatment effect. Several clinical epidemiologists have stated their preference for RR on the basis that clinicians may be inclined to think on the RR scale and may mistakenly interpret OR as RR.[4-6] If OR is mistaken as RR, then the treatment effect can be overestimated as OR tends to be further away from the null value (1.0) where baseline risk is not small, as is usually the case in clinical trials. However, such arguments about misinterpreting OR as RR in interpreting relative treatment effect lose relevance where the purpose is to consistently estimate absolute differences in effects and cost as the basis for cost-effectiveness analysis, as in this article.
Furthermore, there are problems associated with using RR in modelling binary data. The core issue is that, unlike OR, RR is not symmetric.[7] Specifically, if the event of interest is switched to its complement (e.g. if survival is used as the outcome instead of mortality), then RR is not symmetrical around 1, in contrast to OR and ARD, which are symmetrical around 1 and 0, respectively. Walter[8] described this property of the RR as "very troubling" and Fleiss[9] argued that this property effectively rules RR out as a useful metric to...