Pooling systematic reviews of systematic reviews: a Bayesian panoramic meta-analysis

Systematic reviews and meta-analyses usually synthesise evidence from studies reporting outcomes from particular interventions in specific diseases. For example, a meta-analysis of prophylactic antibiotics (intervention) in elective arterial reconstruction (disease) for rates of wound infection (outcome). However, because systematic reviews and meta-analyses are so widespread, a body of evidence often exists around specific intervention effects on particular outcomes over a range of diseases. So for example, a multitude of independent meta-analyses have evaluated rates of wound infection with and without the use of prophylactic antibiotics over multiple surgery types. A systematic review of systematic reviews is a means of synthesising evidence for the same intervention over multiple disease types. We propose a panoramic meta-analysis as a means of pooling effect estimates over systematic reviews of systematic reviews. We explore several methods ranging from a simple two-step approach, to a meta-regression or mixed effects approach, where variation between diseases are modelled as fixed covariate effects and between-study variation by random effects, and to a three-level hierarchical model in which exchangeability is assumed, which allows both a between-disease component of variance and a between-study (within disease) component of variance. In the surgery example, we pool 18 meta-analyses (each including between 4 and 26 studies) of prophylactic antibiotics reporting rates of wound infection from 18 different surgery sites to obtain a single pooled estimate of effect and estimates of between-disease, within-disease and within-study variability.