Standard methodology for meta-analysis (MA) only focuses on deriving a summary estimate but remains implicit about the patient population for which it describes the treatment effect. However, the evidence about treatment effect in one population cannot always be applied to another population that differs possibly in many aspects from that of the original trial, in which case we say there is lack of transportability. In practice, standard MAs do not always ensure homogeneity in the baseline characteristics of the different trial populations; thereby do not rule out a potential lack of transportability. Besides, standard MAs often overlook the under-representation of low- and middle-income countries in the available evidence for several medical interventions. As a consequence, simply meta-analyzing results of trials conducted in developed countries will be not informative for policymakers in low-income countries to implement the most efficient interventions for their populations.

In this paper, we remedy these issues by proposing an alternative MA approach for randomized clinical trials, which uses individual patient data (IPD) from all trials to infer the treatment effect for the patient population in a given trial, based on case-mix standardization. Denote Y(1_k) and Y(0_k) as the counterfactual outcome that would be observed for a given individual if (version of) treatment and control used in trial S=k were administered, respectively. Our focus is on estimating the probabilities P{Y(x_k)=1|S=j} (x=0,1), which are the chance of successfully treated if the patients in population S=j was given the version of treatment (x=1) or control (x=0) used in study S=k. Based on these, the effect of the treatment version in population can be measured–for instance, on the relative risk scale RR(j,k)=P{Y(1_k)=1|S=j}/P{Y(0_k)=1|S=j}. Under certain conditions, RR(j, k) can be estimated by two different approaches, namely:

– outcome regression;

– inverse probability weighting.

The estimates log{RR(j, k)} obtained from the same population j are then summarized by conducting a conventional random-effect MA. The resulting summary estimate thereby represents the effect of treatment for the target population j, while the heterogeneity variance estimate expresses how much results from different trials vary even when considered for the same patient population.

We illustrate the new approach by considering a MA of numerically simulated RCTs that evaluate a binary treatment versus control with respect to a binary outcome. Results of this simulation point out that the proposed framework does enable better understood heterogeneity assessments. The overall heterogeneity across trial results can now be decomposed into 2 different parts, namely case-mix (CM) and beyond case-mix (BCM) heterogeneity. The new framework is then applied to reanalyze a published IPD MA evaluating the effect of vitamin D on the risk of respiratory infection. Finally, we extend the framework to settings where the IPD is only available for some but not for all trials in the MA.

In conclusion, case-mix standardization prior to evidence synthesis is a good strategy to overcome the lack of transportability and other related issues that complicate most MAs. Not only resulting in more interpretable summary results, this strategy also sheds insightful light onto how and why results of different trials are heterogeneous in practice.