A connection without placebo
A trial with 40 participants in Iran compared fluoxetine with nortriptyline. It had no placebo arm. Yet it increased the fluoxetine–placebo difference in two network meta-analyses of treatments for childhood and adolescent depression. In Richard Lyus and colleagues’ reanalysis, excluding the trial changed the standardized mean difference from −0.51 to −0.26 and −0.29. This measure expresses symptom-score differences in a common unit; moving farther from zero means a larger effect. Judging the trial’s influence from its participant count alone would be misleading here.[1]
The network’s structure explains that influence. Other trials compared nortriptyline with placebo, indirectly connecting the fluoxetine trial to placebo. In the small trial, using changes from baseline produced an effect of approximately −4.24, while using endpoint scores gave approximately −0.93. That discrepancy makes the choice of input consequential. Centering a funnel-plot comparison on its own mean can also conceal an outlier when the comparison contains only one trial. Pooling does not automatically neutralize a problematic input.[1]
The boundary of the reanalysis
The convergence of the two estimates with the reconstructed Cochrane result after exclusion supports an important contribution from the outlying data. Computational choices could also contribute: the team lacked all original code and datasets and built its own analyses. Bayesian and frequentist sensitivity checks supported the same direction, but the researchers do not claim exact reproduction of every number. The reappraisal was exploratory and not preregistered. Those limitations determine how confidently its clear sensitivity finding should be read.[1]
Readers of a pooled drug assessment need to see which comparator carries the estimate through the network. Here a small trial without placebo changed the estimated effect against placebo. The exclusion analysis therefore reveals the dependency behind the estimate as well as a smaller number. It cannot independently determine treatment for an individual patient or establish research fraud. Its strongest contribution is the explicit account of how far the result moves when one input is removed.[1]