The meta-analysts studied the estimated effectiveness of these drugs when data were combined from the FDA records and when data were combined from the published literature. For all drugs, the published literature inflated the effect sizes. The inflation varied from 11% to 69% and it was 32% on average. The FDA data would suggest that these agents had small, modest benefits (standardized effect size [ES] = 0.31 on average). Conversely, for 4 of the 12 agents, if one were to perform unawares only a meta-analysis of the published data, the summary result would suggest clinically important effectiveness (ES>0.5). This was not true for any agent based on more complete FDA data.

The meta-analysts found 74 eligible FDA-registered trials with 12,564 patients. Among them, a third (n = 26 trials [31%] with 3449 patients) had remained unpublished. The FDA had determined that half of the registered trials (38/74) had found statistically significant benefits for the antidepressant ("positive" trials). All but one of these trials had been published in journals. Conversely, of the other half trials (36/74) that were deemed to be "negative" by the FDA, one in three were published as "negative" results; another 11 trials were published, but the results were presented in such a way so as to seem "positive" and 22 "negative" trials were silenced and never appeared in the literature.

The first meta-analysis used data that were submitted to the U.S. Food and Drug Administration (FDA) for 12 antidepressant drugs that were approved between 1987 and 2004 [ 7 ]. These were bupropion SR (Wellbutrin SR, GlaxoSmithKline), citalopram (Celexa, Forest), duloxetine (Cymbalta, Eli Lilly), escitalopram (Lexapro, Forest), fluoxetine (Prozac, Eli Lilly), mirtazapine (Remeron, Organon), nefazodone (Serzone, Bristol-Meyers Squibb), paroxetine (Paxil CR, GlaxoSmithKline), sertraline (Zoloft, Pfizer), venlafaxine (Effexor, Wyeth), and venlafaxine XR (Effexor XR, Wyeth). The major advantage of using data submitted to FDA is that this includes all the trials that each company had registered as evidence in support for marketing approval or change in labelling. This registration allows one to have knowledge of all these trials, regardless of whether they were eventually published or not. Moreover, the process of regulatory review is such that there is less room for manipulating analyses and distorting results in data entered in the FDA registry tables.

The PLoS Medicine meta-analysis asked the question: is there a relationship between the baseline severity of depression and the difference in effectiveness between drug and placebo? Meta-regression analyses identified such a relationship. Drug-placebo differences were generally small, but they increased with increasing baseline severity. The meta-analysts used a previous consensus [ 9 ] to propose that a clinically important difference needs to be at least 3 points in the Hamilton scale or ES>0.50. The difference between drug and placebo became large enough to be clinically important only in the small minority of patient populations with severe depression (baseline score exceeding 28 in the study population). Even in these severely depressed patients, the difference between drug and placebo was due to the fact that placebo became less effective; there was no evidence that the antidepressants became more effective. The authors concluded that most of the benefit from antidepressants is duplicated by the placebo effect. This is a conclusion that had been proposed also based on earlier meta-analysis [ 9 ]. Moreover, the current meta-analysis added the insight that these agents may be of clinical use only in severely depressed people, a small minority compared with the vast populations who take antidepressants currently. Even in the few extremely depressed patients, the eventual benefit was due to lack of responsiveness of placebo, not due to increased responsiveness to antidepressants.

A second meta-analysis in PLoS Medicine used also data that were submitted to FDA on 6 new generation antidepressants and eventually used the information on four of them (5 trials on fluoxetine, 4 on venlafaxine, 8 on nefazodone, and 16 on paroxetine) [ 8 ]. For 2 other drugs (sertraline and citalopram), some trials were simply reported even in the FDA databases as having non-significant results. In contrast with the other meta-analysis, the investigators of this meta-analysis did not wish to impute data when there was such missing information.

Limitations in the meta-analyses

Both meta-analyses have some limitations. Many more trials are conducted after approval or outside of the FDA approval process. Moreover, registries of approved agents do not include antidepressants that were possibly tested in clinical trials in the USA, but did not make it (presumably because of more "negative" results), although they made it and were approved in other countries, e.g. fluoxamine, milnacipran, or mianserin. Among antidepressive drugs tested in the USA, only the "luckier" ones, the ones with larger ES, went to the FDA and received approval. The lack of a comprehensive global database is a major deficit in that we may be missing trials done in countries where the overall results for a particular agent were not very promising or overtly negative. Figure shows a simple simulation: suppose that a drug is tested in 40 countries and 5 small trials are preformed for licensing purposes in each country. Let us suppose that on average the drug has a true effect that is small (ES = 0.20). Each of the perfectly unbiased studies is expected to find on average ES = 0.20 and there can be some variability. We can examine situations with different levels of variability around this average, corresponding to standard deviations of 0.20, 0.40, and 0.60. The smaller the trials and the larger the diversity of the populations and drug response, the more variability is expected around the mean of ES = 0.20. Suppose the drug is approved only in countries where the 5 trials show average ES at least 0.20. This is expected to happen in about half the countries. Figure shows what the average ES estimates are in the trials registered in countries where the drug was approved: ES is markedly inflated. Similar considerations apply, if we consider not only many countries, but also many drugs tested in many countries.

Even focusing on FDA-registered trials, their data are not necessarily totally unbiased. Inherent biases in the study design and analysis cannot be corrected by simple registration. Data collection, arbitration of measurements and outcomes and multiple analysis options leave room for selectivity and for presentation of more optimal conclusions – even in FDA-registered results. Second, even these data are eventually incomplete in important details. This is amply demonstrated by the considerable number of studies that were simply registered as having "negative" results without further details on effect sizes, and by the additional missing information that the meta-analysts had to impute even for FDA-registered data. Third, these trials did not have available individual-level information and the data collection and arbitration of outcomes and measurements remained out of reach of the meta-analysts. For the considerable proportion of patients who did not complete the trials, typically last observation carried forward (LOCF) methods were applied, but these have limitations and may lead to overestimation of treatment effects in some circumstances [10].

All these limitations are more likely to have resulted in inflation of the treatment benefit, although there is considerable uncertainty about the exact bias. Of note, the PLoS Medicine meta-analysis [8] noticed funnel plot asymmetry, i.e. smaller trials had larger effects than larger trials. Funnel plot asymmetry is typically considered a sign of publication bias (small "negative" trials remaining unpublished), but this is clearly a misleading simplification [11]. Here publication bias in theory is impossible for FDA-registered data. The authors attributed the asymmetry to confounding due to higher severity scores in smaller trials [8]. However, an alternative explanation is that even for FDA-registered trials, results may still be biased. Exclusion or inclusion of specific patients and data due to questionable eligibility criteria or grey measurements, selection of imputation techniques, use or not or adjustments, and selective reporting of outcomes allow for manipulation in effect size estimation. In small trials, the same amount of manipulation will inflate the effect size more than in large trials, in other words the vibration of the effect size is larger [12,13]. The FDA review process will of course decrease analytical flexibility, but evaluation of depression involves messy outcomes and analyses are not cut in stone. In all, if anything, expectation of these biases further reinforces the message about antidepressants being less effective than thought.

A more serious limitation is inherent in the use of meta-regression techniques in the PLoS Medicine meta-analysis. The main analysis used a fixed effects meta-regression, and only a secondary analysis used a mixed effects approach. The latter, which may be more appropriate than fixed effects [14], had less conclusive results. Meta-regression modelling can be biased [15]. When the trials have only small differences in the average values of severity (as in this case), the slope of the regression terms can be affected by outliers and leverage problems. The most important limitation stems from the ecological fallacy [16,17]. The regression used as a moderator variable the average baseline severity of depression in each group of participants in each trial arm. However, this is a proxy that does not represent equally well all participants. For example, the average baseline score may be 28, but this may include patients with scores of 17, 27, 32, and 36. The relationship may not have been the same, if data could have been analyzed for individual patients. This is to say, while net effectiveness (difference of drug from placebo) seemed to increase with increasing average severity, within a specific trial it could be that the effectiveness decreased with increasing severity. Ecological fallacy is the main reason why meta-regression analyses with group average are viewed with scepticism [17].

Finally, the PLoS Medicine meta-analysis describes the end of the severity spectrum in the analyzed trials as containing patients who are "most extremely depressed" or alternatively "very severe" depression. In fact, "very severe" depression would correspond to patients with even worse depression status, primarily those hospitalized because of major depression. In fact, the analyzed regulatory trials have typically avoided including hospitalized patients with truly so extreme depression, because these newer agents had been shown early on to be ineffective – or at least less effective than older agents – in such patients.