Identifying Consistency in Evidence Networks
Identifying inconsistency in evidence networks is one of the major challenges in network meta-analysis. Inconsistency is a conflict between direct and indirect evidence that makes it difficult to estimate treatment effects. To address this issue, researchers have developed a number of methods. They include:
One method for identifying inconsistency is to use a heuristic approach, which involves ordering designs sequentially. For example, if a study has two treatment arms, then it is likely to have inconsistency. However, the problem becomes more complicated when multi-arm trials are included. Approximately a quarter of randomized trials involve more than two arms, which can complicate the definition of loop inconsistency.
Another method for identifying inconsistency is the use of Bayesian models. These models are part of a generalized linear modeling (GLM) framework. In this approach, an inconsistency is identified by calculating the priority of two similar matrices under different values of k. Then, the difference between the two matrices is placed on treatment B or C.
Alternatively, researchers have used global optimization models to identify inconsistencies. In this method, the inconsistency is located in the network by fitting a model with fixed effects. Then, the model is adjusted so that the inconsistencies are removed. This approach leads to models that fit well with fewer d.f. The method is also useful for locating large inconsistencies in networks. It is also possible to remove portions of the evidence network to address the issue.
Another method for identifying inconsistency involves the use of the Bucher method. The Bucher method is used to detect statistically significant inconsistency. This method is useful for checking the acceptability of a response, and can also be used to identify inconsistency in network meta-analysis. It is also useful for checking the network inconsistency and to determine whether the network can be regarded as statistically consistent.
For a three-arm trial, a loop inconsistency can be detected when the induced bias matrix C contains the largest value deviating from a value of 1. The largest value of C is then the most inconsistent element of the matrix A. This element is also the largest absolute value in the model.
A two-arm trial has a loop inconsistency when the comparisons between two BC pairs are not statistically independent. This can be detected by comparing the BC pairs from the two trials. A model with the two inconsistencies can then be constructed by removing the pairwise comparisons from the three-arm trial. Then, a new closed loop is formed. The inconsistency can then be estimated.
Inconsistency can also be detected by using the Gower plot. This method is useful for detecting inconsistency in networks that include both direct and indirect comparisons. In this model, the inconsistency is found in the three loops. This model can also be used to detect inconsistency in two-arm trials. In this model, the inconsistency relates to the pairwise comparison between the two edges of the loop.
Inconsistency can also be identified in a three-treatment triangular network. In this model, the inconsistency term is the sum of the variances of comparisons. The inconsistency term can also be calculated by applying consistency equations.