Drug Information Journal - March 2009 - (Page 173) Trends in Meta-analysis BIOSTATISTICS 173 outcome definitions, specification of any subgroups of interest, and the meta-analysis techniques to be used. All trials would be prospectively registered so they could be chased down later if results were not forthcoming and, ideally, plans could be made to collect and analyze IPD, not just aggregate results, from each relevant trial. There would be a plan to publish the main results, and the analyses could permit additional questions (or trials) only if data were still blinded. This approach is not impossible or ideal only, but has been successfully adopted by several cooperative groups already. Examples range from cholesterol reduction (Prospective Pravastatin Pooling Project, 12; Cholesterol Treatment Trialists, 1 to interventions to reduce falls in the 3) elderly (Frailty and Injuries: Cooperative Studies of Intervention Techniques, 14), to antithrombotic prophylaxis in atrial fibrillation (SPORTIF III and V, 15). It is high time industry embraced this approach, as the benefits are obvious and large, with little downside. The main reasons to hesitate would be that (a) acceptance by regulators and policymakers is still uncertain, and (b) CMA methods are still evolving, in fact morphing into Bayesian methods. The former can only be influenced by top-down and bottom-up education of regulators, and the latter is a development that should be cheered. To understand why Bayesian meta-analysis is a positive development, one must first understand that there are two main models used in metaanalysis: fixed effects models (FEM) and random effects models (REM). Simply put, the main differences between the two are that FEM incorporates within-study variability, and REM incorporates both within- and between-study variability into estimates of effect. This sets the stage for a sense of how Bayesian methods differ, namely in that Bayesian meta-analyses incorporate yet another source of variability in a study set—the degree of uncertainty in the prior information. In Bayesian meta-analyses, where new information is factored into existing, or prior, information, the mean and variance of the underlying distribution of the outcomes of the prior studies are given their own underlying distributions (16–20). Practically speaking, Bayesian methods offer improved modeling frameworks for meta-analyses where there are studies with arms with zero events, or where you wish to analyze time-toevent data for rare events. Bayesian methods allow the analyst to include so-called informative priors. For instance, results of meta-analysis of observational studies can be used as prior information for a new meta-analysis of randomized controlled trials (RCTs). Bayesian methods also facilitate multiple comparisons simultaneously, as well as analyses when outcomes have multiple expressions (ie, correlated estimates acquire borrowed strength from one another). The main disadvantages of Bayesian meta-analysis are that it is not simple to do, although this has been rectified by the recent availability of off-theshelf software programs (eg, WinBugs). Bayesian methods are also not simple to explain or understand, which is not to be underestimated as a disadvantage when the results of meta-analyses are intended for use by various nonstatistician decision makers in health care. For most meta-analysis situations, however, readers should remember a simple bottom-line fact regarding Bayesian meta-analyses, and that is that results with Bayes approximate REM, which approximate FEM. A meta-analysis of the efficacy of lansoprazole in healing gastric ulcers illustrates this point nicely (21), wherein the treatment effect as calculated using Bayesian, REM, and FEM meta-analyses varied by only hundredths of a decimal point across the three methods, regardless of whether the primary outcome (gastric ulcer healing rates) was expressed as a risk ratio, a risk difference, or an odds ratio (OR). Internal consistency such as this indicates a very robust result, enhancing the credibility of the findings. N E T W O R K M E T A - A N A LY S I S ( M I X E D T R E AT M E N T C O M PA R I S O N S O R INDIRECT COMPARISONS) Q. Can we compare treatments when no headto-head trials exist? A. Yes, under certain assumptions. Increasingly we are seeing such analyses in the literature (22–24). Originally, these were relatively simple comparisons using indirect Drug Information Journal
For optimal viewing of this digital publication, please enable JavaScript and then refresh the page. If you would like to try to load the digital publication without using Flash Player detection, please click here.