Exploring treatment by covariate interactions in Cochrane Reviews: recent practice

Article type
Authors
Donegan S1, Williams L1, Dias S2, Welton N2, Tudur-Smith C1
1University of Liverpool, United Kingdom
2University of Bristol, United Kingdom
Abstract
Background:
Treatment by covariate interactions can be explored in reviews using interaction analyses (i.e. subgroup analysis or meta-regression techniques). Such analyses provide information regarding how the covariate modifies the treatment effect. Guidance exists regarding how to apply such analyses but there is limited research detailing how review authors explore interactions.

Objectives:
To review Cochrane Reviews (CRs) to establish how well interaction analyses are designed, applied, interpreted and reported.

Methods:
For each Cochrane Review Group, we included the most recently published review for which the protocol was accessible. We excluded review updates, diagnostic test accuracy reviews, withdrawn reviews and overviews of reviews. The Cochrane Database of Systematic Reviews was searched on 19 August 2013. We devised criteria using literature to assess how well interaction analyses were designed, applied, interpreted and reported. Data were extracted and summarized regarding review and covariate characteristics and analyses.

Results:
No review reported how or why each covariate was chosen. 42% of reviews reported each covariate in the protocol. No review labelled post hoc covariates as such. 21% of reviews mentioned five covariates or less. 2% of reviews sought missing covariate data from study authors. 15% of reviews that did aggregate data analyses included at least 10 trials. 2% of reviews planned to use a method to detect interactions (e.g. interaction test) but 3% of reviews reported whether an interaction was detected. No review discussed the importance, or plausibility, of the results, or the possibility of confounding. 39% of reviews reported the number of trials with each covariate value. 58% of reviews reported results from the interaction analysis. 8% of reviews reported results from methods to detect interactions.

Conclusions:
This review highlights flaws of interaction analyses in recent Cochrane Reviews and suggested areas where such analyses can be improved. We recommend that authors use our devised criteria to more carefully consider the design, application, interpretation, and reporting of analyses.