OBJECTIVES Network meta-analyses (NMA) provide estimates of comparative efficacy for multiple treatments based on an analysis of connected networks of trial comparisons. A key concern is the comparability of treatment effect estimates from different trials. Where there is both indirect and direct evidence for one or more comparisons (‘loops’ in the network) it is possible to evaluate empirically the ‘consistency’ of the network. METHODS A variety of methods have been proposed to examine inconsistency including: (i) node-splitting where the direct and indirect estimates are compared across the network (ii) comparison to an ‘inconsistency’ model where estimates for each treatment comparison are allowed to be independent, (iii) inclusion of treatment by design interaction terms, (iv) investigation of residual deviance estimates for individual trial arms, and (v) investigation of mixed predictive p-values.   We compare the implementation and, most importantly, the interpretation of these methods using a previously published NMA in type 2 diabetes.  In this analysis HbA1c was compared across six treatments in a network of 22 studies with multiple ‘loops’. RESULTS The methods agreed in showing the presence of inconsistency with the network. For example, the inconsistency model showed an improved fit (DIC -62.35) compared to the consistency model (DIC -60.25).  The node splitting method identified statistically significant inconsistency in two treatment arcs (liraglutide 1.8mg vs placebo and liraglutide 1.8mg vs exenatide QW). CONCLUSIONS The alternative methods vary in their ability to provide an omnibus ‘test’ of inconsistency across the network and their ability to identify which parts of the network contain inconsistencies.  We highlight that none of the methods alone can identify individual studies as being the cause of inconsistencies and argue that we need to consider the whole structure of the network and the characteristics of the studies (in terms of treatments, subjects and design) within the network.