A meta-epidemiological approach for evaluating bias and small study effects in networks of interventions

Tags: Poster
Chaimani A1, Schmid C2, Vasiliadis H3, Salanti G1
1Department of Hygiene and Epidemiology, School of Medicine, University of Ioannina, Greece, 2Sackler School of Graduate Biomedical Sciences, Tufts University School of Medicine, Boston, MA, US, 3Molecular Cell Biology and Regenerative Medicine, Sahlgrenska Academy, University of Gothenburg, Sweden

Background: Investigation of the impact of study-specific biases typically requires a large number of studies. Meta-epidemiological approaches consider collections of independent pairwise meta-analyses, where the impact of bias in each one of them might be different, depending on the clinical area. Network meta-analyses include pairwise meta-analyses assumed 'exchangeable’ in everything but the comparison being made. Hence biases can be studied simultaneously by borrowing strength across comparisons even when each pairwise meta-analysis comprises a small number of trials.

Objectives: To suggest a meta-epidemiological technique for exploring study-specific bias sources and small-study effects using a collection of published networks of interventions.Methods: We searched for and collected all star-shaped networks of interventions published until March 2011. We applied various multiple-treatment meta-regression models that used as covariate a)the probability that a trial is at risk of bias in terms of adequacy of sequence generation, allocation concealment and blinding b)a measure of precision or variance to explore small study effects. We assume exchangeable bias parameters across comparisons within each network and across networks for similar outcomes. The fit of the different models and the amount of heterogeneity explained were compared. An average effect per bias source was estimated.

Results: The adjustment for small study effects or bias items showed in some networks a slight improvement in the fit of the model and a respective reduction in heterogeneity. Borrowing strength across networks showed that a)small study effects exhibit an important amount of heterogeneity across networks b)bias sources overestimate the effectiveness of the interventions compared to common comparator, in agreement with previous findings.

Conclusions: This meta-epidemiological approach enables investigation of bias and heterogeneity while adjusting for differences in effectiveness between treatments. As networks of interventions become popular, large-scale investigation of bias should consider them as source of evidence.