Selecting the best scale for measuring treatment effect in a network meta-analysis: a case study in childhood nocturnal enuresis.

Dichotomous outcomes in pairwise meta-analysis are typically summarised using the odds ratio (OR), relative risk (RR) or risk difference (RD). The hazard ratio (HR) may also be used where events occur over time. Choice of scale is often determined by ease of interpretation or mathematical properties of a measure, as there is frequently insufficient power to compare goodness-of-fit across different scales. Network meta-analysis (NMA) combines evidence across a network of treatment comparisons. NMA allows the combination of a greater numbers of trials, so there is greater potential to use goodness-of-fit statistics to determine an appropriate scale on which the effects of treatments are additive. In this paper, we explore choice of scale in an NMA of childhood nocturnal enuresis for the outcome 'failure to achieve 14 days consecutive dry nights'. We compare OR, RR of both the harmful (RRH) and the beneficial (RRB) outcomes, RD and HR. Using a Bayesian framework, the posterior mean residual deviance and deviance information criterion are used to evaluate model fit and selection between the different summary effect measures. We use I(2) to examine within-comparison heterogeneity for the pairwise analyses. The results suggest that the HR and RRB provide the best fit. We conclude that choice of scale should be based on physiological and epidemiological understanding of the disease process, together with an empirical assessment of model adequacy. HR should be given greater consideration where there is an underlying time-to-event process. We demonstrate how results can be transformed to an alternative scale to aid interpretability.