Five antidepressants have been studied: A vs. placebo (3 trials), B vs. placebo (2 trials), A vs. B (1 trial), C vs. A (1 trial), and D vs. B (1 trial). No trial directly compared C to D. Can NMA estimate the C vs. D effect, and what is the evidence pathway?
ANo — NMA requires at least one direct comparison between all treatment pairs
BYes — the indirect pathway runs through the network: C → A → placebo → B → D, using the transitive chain of comparisons
CYes — but only if a Bayesian framework is used
DNo — indirect comparisons are never valid for treatments not directly compared
The power of NMA is that it can estimate any pairwise comparison that is connected through the network, regardless of whether a direct trial exists. C was compared to A, A to placebo, placebo to B, and B to D — the network is connected, so C vs. D can be estimated indirectly. The estimate combines all available evidence and is less precise than a direct comparison would be (it inherits uncertainty from each link in the chain), but it provides information that would otherwise be unavailable. The transitivity assumption requires that the relative effects are consistent across the indirect pathway.
Question 2 True / False
A network meta-analysis finds that the direct estimate of A vs. B from head-to-head trials is OR = 1.5, but the indirect estimate (through a common comparator C) is OR = 0.8. This inconsistency threatens the validity of the pooled NMA estimate.
TTrue
FFalse
Answer: True
Inconsistency between direct and indirect evidence indicates that the treatment effects may depend on the study populations, co-interventions, or other factors that differ between the trial networks — violating the transitivity assumption. The pooled NMA estimate for A vs. B would blend the direct (1.5) and indirect (0.8) evidence, producing a result that may not represent either the direct or indirect truth accurately. Inconsistency should be investigated: are the trials comparing A-B directly conducted in different populations than those forming the indirect comparison? Are there effect modifiers that explain the discrepancy?
Question 3 Short Answer
NMA produces treatment rankings (e.g., 'Treatment B has a 73% probability of being the best'). Explain why these rankings should be interpreted cautiously.
Think about your answer, then reveal below.
Model answer: Rankings are highly sensitive to the precision and number of comparisons available for each treatment. A treatment with only one small trial may rank highly due to an imprecisely estimated large effect — wide confidence intervals create high ranking probabilities without strong evidence. Rankings also ignore clinically important differences in magnitude: being ranked #1 versus #2 may correspond to a trivially small difference. Furthermore, rankings can be unstable across modeling choices (fixed vs. random effects, prior distributions). Treatment effect estimates with confidence intervals are more informative than rankings for clinical decision-making.
The Surface Under the Cumulative Ranking curve (SUCRA) is often reported alongside rankings, summarizing each treatment's probability of being among the best. But SUCRA values are vulnerable to the same problems as raw rankings. A treatment with SUCRA = 0.85 could be slightly better than all alternatives with high confidence, or dramatically better with very low confidence. The effect estimate and its precision are always more informative than the ranking statistic.