Making sense of complex interventions: application of hierarchical meta-regression in a meta-analysis of diabetes quality improvement (QI) interventions

Article type
Authors
Danko K1, Dahabreh I2, Ivers NM3, Trikalinos TA2, Tricco AC4, Edwards A5, Hillmer M6, Lavis J7, Manns B5, Moher D1, Paprica A8, Ramsay T1, Sargious P5, Shojania K9, Straus SE1, Tonelli M5, Yu CH4, Grimshaw J1
1Ottawa Hospital Research Institute
2Brown University
3Women's College Hospital
4St. Michael's Hospital
5University of Calgary
6Ontario Ministry of Health and Long-Term Care
7McMaster University
8University of Toronto
9Sunnybrook Research Institute
Abstract
Background: Systematic reviews often address complex interventions that have multiple components. Standard meta-analysis methods often do not adequately reflect the complexity of these interventions because compromises must be made to facilitate synthesis (e.g. multiarm studies are reduced to single-pairwise comparisons and only components that differ between arms are modelled in the observed difference of effect). As a result, the meta-analysis fails to include all available data and cannot isolate the effects of components that may be of interest to decision makers.

Objectives: To explore the utility of hierarchical meta-regression models in a meta-analysis of complex QI interventions for diabetes.

Methods: Systematic review of QI programmes for diabetes that included at least one of 12 QI strategies. We implemented a series of hierarchical models to assess the effects of QI strategies. We explored extensions of the models to evaluate interactions among QI components and with contextual and programme-level covariates. Finally, we used the models to predict the combined effects of QI strategies previously not evaluated in the same QI programme while accounting for other features of the available data (e.g. large number of cluster randomised trials with missing data on the intra-class correlation coefficient).

Results: We included 278 RCTs. Hierarchical meta-regression models estimated effects of individual QI components, producing different rankings compared to standard methods. For example, while the 3 QI strategies Promotion of Self Management (PSM), Team Changes (TC), and Case Management remained the most effective strategies for reducing glycated haemoglobin, the effects of each strategy were smaller (presumably due to the better isolation of their individual contributions) and TC emerged above PSM as most effective. The models allowed the assessment of interactions and effect modification; model selection is ongoing and additional results will be presented at the Summit.

Conclusions: Background knowledge combined with flexible synthesis models can allow fuller use of available data in reviews of complex interventions such as QI programmes.