Statistical approaches to adjusting weights for dependent arms in network meta-analysis.

Abstract

Network meta-analysis compares multiple treatments in terms of their efficacy and harm by including evidence from randomized controlled trials. Most clinical trials use parallel design, where patients are randomly allocated to different treatments and receive only 1 treatment. However, some trials use within person designs such as split-body, split-mouth, and crossover designs, where each patient may receive more than one treatment. Data from treatment arms within these trials are no longer independent, so the correlations between dependent arms need to be accounted for within the statistical analyses. Ignoring these correlations may result in incorrect conclusions. The main objective of this study is to develop statistical approaches to adjusting weights for dependent arms within special design trials. In this study, we demonstrate the following 3 approaches: the data augmentation approach, the adjusting variance approach, and the reducing weight approach. These 3 methods could be perfectly applied in current statistical tools such as R and STATA. An example of periodontal regeneration was used to demonstrate how these approaches could be undertaken and implemented within statistical software packages and to compare results from different approaches. The adjusting variance approach can be implemented within the network package in STATA, while reducing weight approach requires computer software programming to set up the within-study variance-covariance matrix.