Multivariate modelling for meta-epidemiological assessment of the association between trial quality and apparent treatment effects in randomised clinical trials

Siersma V, Als-Nielsen B, Chen W, Hilden J, Lotte Gluud L, Gluud C

Background: Studies have found that several trial quality components may be associated with apparent treatment effects in randomised clinical trials. However, inadequate quality components frequently co-occur. The possible biases are therefore to be evaluated simultaneously.

Objectives: To find an appropriately stable multivariate method for assessing from meta-analyses the association between trial quality components and estimated treatment effects in randomised clinical trials.

Methods: In the context of our study of 523 randomised trials from 48 meta-analyses in a random sample of Cochrane reviews (Als-Nielsen et al., Cochrane Colloquium, Ottawa 2004), we divided trials into six categories according to the type of experimental and control intervention (ie, drug trials versus active control, drug trials versus placebo control, drug trials versus no treatment, procedural trials versus active control, behavioral trials versus active control, behavioral/procedural trials versus no treatment). To estimate the association between trial quality components and apparent treatment effects, we used three different multivariate methods: a fixed effect approach in which the effects of trial quality are assumed constant across meta-analyses; a random effect approach in which the effects of trial quality are assumed different across meta-analyses; and an intermediate approach in which the effects of trial quality are assumed constant within each trial type (ie, fixed effect), but different across trial types (ie, random effects).

Results: The fixed effect approach produced models that were too rigid to incorporate a possible between meta-analysis heterogeneity of the bias. Conversely, the random effect approach produced models that were unstable. The influence of trial quality components on estimated treatment effects varied considerably across the trial types. The intermediate method allows for heterogeneity between groups of trials (that are of similar type according to the six categories mentioned above) but avoids instability.

Conclusions: We found that the intermediate approach produced the most appropriate model for evaluating the association between trial quality components and apparent treatment effects of randomised clinical trials.