MTM and Bias

rankbupropionIt is possible to integrate the concept of bias within MTM models using Bayesian ‘multiple-treatments meta-regression’ approaches. Several bias sources can be studied.

Study-quality characteristics (e.g. allocation concealment, blinding) can be considered as possible sources of bias. Several empirical studies in head-to-head meta-analyses suggest that poor quality characteristics (such as lack of blinding) relate to overestimation of the treatment effects but their role in MTM is uncertain. Less studied sources of bias, such as novelty bias (the bias created by the belief that novel agents perform better) and sponsorship bias can be evaluated only when information for the relative effectiveness of several different interventions is available. MTM provides an exciting opportunity to tackle these two emerging sources of bias, as well as to clarify the role of the established sources that relate to quality characteristics.

Novel agent bias: Consider two treatments equally effective A and B; A being an established intervention and B being an improved version of A, or a novel agent. As there is usually an underlying assumption among clinicians and consumers that novel agents might work better than old ones (‘wish’ or ‘novel agent’ bias), studies A versus B may spuriously favour B. In a meta-analysis of A versus B studies it is not possible to distinguish between the effect of bias and the potentially genuine advantage of intervention B. Consider that later on, a further intervention C appears and the same sort of bias operates; both A versus C and B versus C studies favour C.

Although bias will exaggerate the effect of C, there is no particular reason for the amount of bias to be larger depending on the comparator A or B. In this case indirect comparison of A versus B through C will decrease or even wash out bias. Comparison of direct (possibly biased) and indirect (unbiased) A vs B estimates can indicate the presence of bias. This idea applies to a more common scenario: interventions would be compared with older reference treatments or suboptimal dosages of a reference treatment but not between themselves in an attempt to show more impressive effect sizes. 

Sponsorship bias: Sponsoring can play a very important role in biasing the results. Consider three treatments equally effective A, B and C and n studies comparing pairs of them, so that each head-to-head comparison is evaluated in n/3 studies. Suppose now that A versus B studies are sponsored by the manufacturing company of B; A versus C studies are sponsored by the manufacturing company of C and so are the B versus C studies. If bias has the same magnitude for all three comparisons, indirect evidence for A versus B would be free of bias compared to the direct (and therefore preferable!) as it will allow the bias components to ‘cancel each other’. On the other side, evidence for A versus C through comparator B has the double amount of bias. For more details on MTM bias models see (Dias et al., 2010a, Dias et al., 2010b, Salanti et al., 2010b).


Dias, S., Welton, N., & Ades, A. (2010a). Study designs to detect sponsorship and other biases in systematic reviews. Journal of Clinical Epidemiology 63, 587-588.

Dias, S., Welton, N., Marinho, V., Salanti, G., & Ades, A. (2010b). Estimation and adjustment of Bias in randomised evidence using Mixed Treatment Comparison Meta-analysis. Journal of the Royal Statistical Society (A) 173

Salanti, G., Dias, S., Welton, N. J., Ades, A. E., Golfinopoulos, V., Kyrgiou, M., Mauri, D., & Ioannidis, J. P. (2010b). Evaluating novel agent effects in multiple-treatments meta-regression. Stat. Med. 29, 2369-2383

Additional information