What is a Network Meta-Analysis

A Network Meta-Analysis, sometimes called a Mixed Treatment Comparison, synthesises the results of disparate intervention studies. Typically it is used to overcome the fact that no head-to-head evaluations have been made for all the available options e.g. drugs, so indirect comparisons are needed to link the studies in the evidential 'network'. 
It is important to distinguish two types of NMA, in the same way as is necessary for Health Technology Assessments.
The inputs into, and outputs from, a science-oriented NMA are determined broadly by the standards of science and in any particular case by the views of the analysts as to what falls below the threshold of what constitutes 'scientific evidence'. The resulting strict inclusion criteria inevitably mean there are no performance ratings for some options on attributes relevant to the decision and none at all for some criteria, including many person/patient important ones, because 'robust' studies have not been carried out.
In contrast the inclusion criteria for the inputs into a decision- or practice-oriented NMA are determined solely by the need of the decision maker to fill each and every cell of the decision matrix with the Best Estimate Available Now (BEAN), no external threshold being relevant in pursuing this task. Inputs are taken from the highest point in the evidential hierarchy as possible, but as low as is necessary. In many cases the resulting 'NMA' will be the result of tapping into the relevant network of expertise and eliciting expert judgements. It may be necessary to decide whether a weak decision-oriented NMA of the conventional sort is better or worse than the results of an expert elicitation process. No hard and fast rule can be applied to this choice, value judgements always being necessary in assessing the multi-criterial concept of quality.
The comparator for the inputs into a decision oriented NMA is always those supplied in normal or standard practice - typically by the clinician or clinical team and usually in the form of implicit verbal magnitudes. In such comparative evaluation it is important to reject the argument  that option/criterion cells with no 'acceptable' evidence by scientific standards, can and should be ignored. There must be agreement on the relevant criteria - including all person/patient important outcomes - prior to consideration of the state of the evidence base. 
Here is a good example of statement in a science-oriented NMA undertaken for a practice guideline. Note that the 'other outcomes' are clearly decision-relevant and ratings will need to be supplied somehow in clinical practice.

"The current analysis only reports results from NMA for six outcomes: all-cause stroke or SE, major bleeding, all-cause mortality, ICH (including intracerebral hemorrhage), extracranial bleeding, and MI. Network meta-analysis for other outcomes were not conducted because data was either sparsely reported (e.g., PE, life-threatening bleeds, transient ischemic stroke), outcome definitions varied considerably (e.g., minor bleeds), or outcomes were redundant (all-cause stroke or SE versus ischemic stroke or SE). Further, there was limited subgroup data reported for some outcomes considered in NMA." Canadian Agency for Drugs and Technology in Health Antithrombotic Agents for the Prevention of Stroke and Systemic Embolism in Patients With Atrial Fibrillation p49 . Available at