Addressing problems of non-transportability when combining treatment effects across patient populations in meta-analyses: a causality framework




Poster session 3


Tuesday 18 September 2018 - 12:30 to 14:00

All authors in correct order:

Vo T1, Porcher R2, Vansteelandt S1
1 Department of Applied Mathematics, Computer Science and Statistics, Ghent University, Belgium
2 INSERM, UMR1153 Epidemiology and Statistics Sorbonne Paris Cité Research Center (CRESS), METHODS Team, Paris Descartes University, France
Presenting author and contact person

Presenting author:

Tat-Thang Vo

Contact person:

Abstract text
Standard methodology for meta-analysis often focuses too much on deriving a summary treatment effect but remains silent about the patient population for which this summary statistics is described. Furthermore, some common treatment effect measures – i.e. odd ratios (OR) and hazard ratios (HR) – are vulnerable to the problem of non-collapsibility, which makes them difficult to pool.

To develop novel meta-analysis approaches for randomized clinical trials, which infer treatment effects for well-defined patient populations, thereby delivering more clinically interpretable results.

Our novel individual patient data (IPD) meta-analysis approaches are based on direct standardization (DS), using either outcome regression (OCR) or inverse probability weighting (IPW). We develop accompanying random-effects meta-analyses that enable the disentangling of heterogeneity in treatment effects that is due to differential covariate distribution from that due to a differential outcome generating mechanism. We conduct simulation experiments across five settings to evaluate the two proposed approaches with respect to 1) consistency, and 2) the performance in heterogeneity assessment. We then apply the novel approaches to reanalyze a published IPD meta-analysis assessing the effect of vitamin D on the risk of respiratory tract infection.

Both DS-based and IPW-based approaches are effective for transforming the evidence across different populations. However, the new meta-analysis approach based on OCR is more successful at correctly capturing treatment effect heterogeneity. Meanwhile, the IPW-based approach is less influenced by modeling assumptions.

The novel framework yields treatment effect estimates that express the effectiveness for well-defined patient populations. Thus, it promises to overcome the important drawbacks of standard meta-analysis methodology. Future frameworks should focus on 1) improving the performance of the IPW approach, and 2) generalizing the new approaches to aggregated data, which makes them become more applicable in practise.

Patient or healthcare consumer involvement:
This study uses IPD collected in randomized controlled trials. Otherwise, there was no direct involvement of patients or healthcare consumers.

Relevance to patients and consumers: 

The framework proposes a more proper way to summarize available evidences regarding a treatment's effectiveness for a given condition or disease. This will support the health care professionals in optimizing the therapeutic strategies for their patients.