the newest treatment could be favored and disentangling the sources of bias operating on direct and indirect evidence would be difficult. Second, the choice of the FDAregistered trials as a reference standard could be debated but seems reasonable. Pair-wise effect sizes derived from FDA data should not be considered unbiased estimates of antidepressants efficacy per se but may be considered unbiased estimates of treatment effects via NMAs of placebo-controlled trials. In fact, during the application review process for new drugs, the FDA reanalyses the trial data using raw data from the sponsor in adherence to the pre-specified statistical methods in the trial protocols. This FDA dataset was previously described as ����an unbiased body of evidence����. Moreover, as usual, checking the required assumptions for the indirect treatment comparisons framework is difficult. However, there is no reason to suggest that these conditions are not met. Homogeneity was satisfied in our analysis. Trial similarity is likely because all trials were randomized, double-blind, placebocontrolled studies of drugs for the short-term treatment of depression, with close selection criteria. Other NMAs have been performed in this field and did not raise concerns about these assumptions. In addition, if one of these assumptions were not met, our analysis would not likely have been affected because it probably would have concerned both NMAs of published and FDA data that we compared. An additional required assumption for NMA is exchangeability, which implies that if all the RCTs had included all the treatments evaluated in the network, then each trial would have estimated the same pair wise effect sizes. The consistency assumption strictly follows from the exchangeability assumption. Star-networks do not allow for quantifying the amount of incoherence between indirect and direct evidence. Unequal availability of trials for different comparisons, because of reporting bias, may result in inconsistency. When reporting bias hypothetically affected only one drug, we basically assessed the consequences of violating the assumption of exchangeability and found that the ranking of all drugs could be modified. Differential reporting bias could occur across and within competing interventions. For instance, reporting bias may differ between trials conducted before and after the FDA Amendments Act of 2007, which expanded the legal requirements for trial reporting. NMA is a promising statistical tool, especially for comparative effectiveness research, but authors should be aware of the potential impact of reporting bias on the results of such analysis.