Methodologies for network meta-analysis of randomised controlled trials in pain, anaesthesia, and perioperative medicine: a narrative review

Abstract

Network meta-analysis has emerged as a method for analysing clinical trials, with a large increase in the number of publications over the past decade. Network meta-analysis offers advantages over traditional pairwise meta-analysis, including increased power, the ability to compare treatments not compared in the original trials, and the ability to rank treatments. However, network meta-analyses are inherently more complex than pairwise meta-analyses, requiring additional statistical expertise and assumptions. Many factors can affect the certainty of evidence from pairwise meta-analysis and can often lead to unreliable results. Network meta-analysis is prone to all these issues, although it has the additional assumption of transitivity. Here we review network meta-analyses, problems with their conduct and reporting, and methodological strategies that can be used by those conducting reviews to help improve the reliability of their findings. We provide evidence that violation of the assumption of transitivity is relatively common and inadequately considered in published network meta-analyses. We explain key concepts with clinically relevant examples for those unfamiliar with network meta-analysis to facilitate their appraisal and application of their results to clinical practice.

Keywords: network meta-analysis; review methodology; statistics in anaesthesia; systematic review; transitivity.

Fig 1

(a) Network plot showing the hypothetical network meta-analysis. Each direct comparison of lidocaine (LID) and gabapentinoids (GAB) with placebo (PLA) is shown. The dashed line shows the indirect comparison between GAB and LID. Using the formulae in the text, an estimated reduction of 5 mg in morphine dose with LID can be achieved based on indirect comparisons derived from direct comparisons with a common comparator. (b) Network plot of interventions for reducing the incidence of chronic postsurgical pain (CPSP). Compared with (a), there are loops of evidence showing direct and indirect estimates for GAB, ketamine (KET), and PLA. This can be used to assess inconsistency as opposed to the evidence from the same plot for previous interventions as there is no trial comparing these agents directly. ALP, alpha-2 agonists; GAB, gabapentinoids; GABGLU, gabapentinoids and glucocorticoids; GABKET, gabapentinoids and ketamine; GLU, glucocorticoids; KET, ketamine; KETGLU, ketamine and glucocorticoids; LID, lidocaine; LIDKET, lidocaine and ketamine; MAG, magnesium; NEF, nefopam; NSA, NSAIDs and COX-2 inhibitors; PAR, paracetamol; PLA, placebo.

Fig 2

(a) Network meta-regression of α₂ adrenergic agonists for the outcome of 24-h morphine consumption. The x-axis is baseline risk (control group morphine consumption), and the y-axis is the mean reduction in 24-h morphine consumption. As baseline risk increases, efficacy also increases. Grey dots are the effects for each study, the solid blue line is the regression line estimate, and the dashed blue line is the 95% confidence interval (CI). Clinical significance changes with the value of baseline risk (x), so a single conclusion cannot be drawn on clinical significance. Baseline risk is a potential effect modifier, and unequal distribution between comparisons could violate transitivity, as shown. (b) Stacked bar chart showing the absolute morphine consumption after gabapentinoids (GAB) and lidocaine (LID) relative to placebo. The purple regions highlight the true difference in morphine consumption between the GAB and LID groups (i.e. LID more effective). The green region shows the effect of a difference in the distribution of effect modifiers (baseline risk or risk of bias) between comparisons of active interventions with placebo. Despite the ‘known’ underlying benefit of LID (purple region), a difference in effect modifiers (green region) makes them falsely appear to have the same efficacy.

Fig 3

Comparison adjusted funnel plot for nonopioid analgesics in reducing the incidence of chronic postsurgical pain. The y-axis represents the standard error on a reverse scale, and the x-axis represents the effect in each study, a direct summary effect for that comparison. The specific comparison here is active vs control interventions, with a different symbol for each comparison (blue points). The funnel plot appears asymmetric, which is corroborated by the test for funnel plot asymmetry (P<0.001).

Fig 4

(a) Surface under the cumulative ranking curve (SUCRA) plot for the reduction in chronic postsurgical pain for lidocaine (LID), ketamine (KET), and gabapentinoids (GAB). The area under the curve normalised from 0 to 1 gives the probability that the treatment is ranked as the best (or one of the best). Lidocaine has the largest area (close to 1) and therefore is likely to be the most effective in reducing chronic postsurgical pain (CPSP). Placebo (PLA) has the lowest area under the curve and is therefore likely to be the least effective. (b) Rankogram including four nonopioids for reducing CPSP (using frequentist SUCRA simulations) showing the probability of each treatment being at a certain rank. Lidocaine is ranked first with a probability of 0.94. Ketamine is ranked at least second, with a probability of 0.63 and with a corresponding probability of 0.46 for gabapentinoids. Placebo has a probability of 1 (i.e. certain) of being ranked last.