We used the search strategy from an earlier paper 10 on venlafaxine and fluoxetine to conduct this meta-analysis on aggregated data.

Eligible studies were identified from PubMed/Medline, the Cochrane library and Embase, including congress abstracts. A three-step search was used for each component of this review. In a first step, an initial search on Medline was carried out in order to refresh optimal search terms and include possible changes in the databases. The search terms used were double-checked before starting the main search. In a second step, all keywords identified were used to search all the above-mentioned databases. A third search was undertaken by searching the reference lists of articles identified. Initial keywords used were: “Depressive Disorder NOT Depression, Postpartum NOT Seasonal Affective Disorder”; “Antidepressive Agents”; “Fluoxetine”; “Venlafaxine”. In addition, manual searches of articles were performed in previous meta-analyses.

Unpublished studies were searched for by communication with key organizations such as the Food and Drug Administration (FDA) and the European Medicines Agency (EMEA), and key researchers in the area. A search on clinicaltrials.gov and Current Controlled Trial was performed.

Authors of abstracts or meta-analysis were contacted for further information and were asked for references of the studies when needed. If no response was obtained to a first solicitation, they were then contacted a second time.

Eligibility assessment was performed independently in a blinded standardized manner by two reviewers (NF and MAS). Disagreements were resolved by consensus or in consultation with a third reviewer (FB).

A comparison across the studies, checking author names, treatment comparisons, sample sizes and outcomes was performed to avoid duplicates and compilations of data from several reports of the same study.

Each paper was then assessed for methodological quality prior to inclusion in the review using an appropriate standardized critical appraisal instrument from the Joanna Briggs Institute 11 .

A data extraction sheet based on the Cochrane Handbook for Systematic Reviews of Interventions’ guidelines (Version 5.0.2, updated September 2009) 12 was used. One review author (NF) extracted these data from the studies included.

The dichotomous outcomes used here are considered as robust outcome measures of treatment efficacy 7 . When these outcomes were not reported, the studies were excluded, and no imputation method was used. Responder and remitter data were extracted as the original study investigators analyzed the data, mainly using the LOCF (Last Observation Carried Forward) method.

The search of Medline, Cochrane and Embase databases provided a total of 11,051 citations. An additional 66 studies were identified by manual search. After adjusting for duplicates, 4,615 remained. Of these, 4,063 studies were discarded because, after review of the abstracts, it appeared that these papers did not meet the criteria. A total of 114 studies were excluded because of the language (among which 97 were in Chinese). Of 33 unpublished relevant studies identified, only 3 were made available by pharmaceutical firms. Thirty-one studies were included in the quantitative review. A flow-chart detailing the study selection process is given in Figure 1.

A summary of the methodologies of these 31 studies is given in Table 1. The quality of the different studies is presented as a table in a web appendix (Additional file 1: Table S1). The “response” analysis involved 29 RCTs including 7,459 participants and the “remission” analysis involved 20 RCTs including 5,096 participants.

n, number of studies. Results are presented in (minimum, first quartile, median, third quartile, maximum) for quantitative data except for V vs F vs P where all the data are given. Qualitative data are presented as a number (percentage). NA represents the number of studies where data were unavailable.

V vs F vs P: studies comparing venlafaxine to fluoxetine to placebo.

V vs F: studies comparing venlafaxine to fluoxetine.

V vs P: studies comparing venlafaxine to placebo.

F vs P: studies comparing fluoxetine to placebo.

HDRS, Hamilton Depression Rating Scale; MADRS, Montgomery-Åsberg Depression Rating Scale.

Six funnel plots were drawn (three for response and three for remission for each head-to-head direct meta-analysis, Figure 4). The funnel plots of the meta-analysis comparing venlafaxine to fluoxetine showed some asymmetry (P <0.03 for response and P <0.01 for remission). The asymmetry was less striking for funnel plots of meta-analyses for both active treatments versus placebo.

The three placebos considered were not significantly different in terms of response or remission. Antidepressant agents were significantly more efficacious than placebos, and venlafaxine was more efficacious than fluoxetine. This was coherent with previous meta-analyses 7 24 . Thus, venlafaxine placebo appears as a “me-too” placebo without greater benefit in relation to fluoxetine placebo and/or fluoxetine and venlafaxine placebo.

Since, obviously, no direct evidence of well-powered, randomized, direct-comparisons exist, indirect comparisons were the only option for comparing the three different placebos. Such indirect evidence is not totally the same as direct evidence and in some cases it has been shown that indirect comparisons led to conflicting results as compared with direct evidence. Such a paradox has been recently shown concerning antidepressants in a recent paper comparing citalopram with its “me-too” escitalopram which found an inconsistency between direct evidence (showing a superiority of escitalopram) and indirect evidence (which did not find any significant difference) 25 . Additionally, Song et al. have suggested that in some cases indirect evidence is less biased than direct evidence 26 . Moreover, the validity of indirect comparison is dependent on the distribution of relative treatment effect modifiers across different comparisons 27 . In our study, initial severity (HDRS score ≥25) could be an important effect modifier 4 5 but despite slight variations, its distribution seemed well-balanced across the different direct comparisons (that is, there was no systematic difference in its distribution between the different direct comparisons).

Nevertheless, results of our indirect comparisons were consistent with the implicit conception that if the three placebos (all made of the same sucrose) were compared in a double blind randomized trial, no difference would be observed. These results are good news indeed: 1) for the supporters of placebo because they conclude that the choice of the placebo is not really important, 2) for the supporters of antidepressants because antidepressants prove superior to placebos, and 3) for supporters of rationality because a validated method has not led to invalid conclusions, that is, we have managed to conclude that “sucrose equals sucrose” in the treatment of major depressive disorder, which is wholly reassuring. Yet significant limitations question these findings.

The quality of a meta-analysis is linked to the quality of the individual studies included 28 . In this respect, the National Institute for Health and Clinical Excellence (NICE) guideline on the treatment and management of depression in adults 29 advises caution in the application of results from RCTs and meta-analyses in routine practice. In particular, our model is based on the postulate that all placebo responders should be antidepressant responders (additive model). This key assumption has never been proved to be true. Indeed, antidepressant response and placebo response could be independent or at least substantially overlapping phenomena (non-additive model), with four different types of patients: 1) placebo-only responders, 2) treatment-only responders, 3) placebo and treatment responders, and 4) never responders 30 . Moreover, the classic logic of the randomized controlled trial casts the placebo effect as a negative foil for measuring therapeutic efficacy and a large amount of important information concerning placebo is not reported in these studies, such as the appearance of the medication (size, shape and color of the pills) 31 .

As well, the indirect comparison may have had low statistical power. Even if no trend toward statistical significance was observed within the indirect comparisons of placebos, insignificant P-values never tell much about equivalence 32 .

Above all, a publication bias and a selective outcome reporting bias might account for some of the effects we observed. The funnel plot for comparison between venlafaxine and fluoxetine shows some asymmetry in favor of venlafaxine. As “true heterogeneity” could not explain this result, and since studies on antidepressants generate substantial conflicts of interest (these drugs generate vast sales revenues), the result is very much open to suspicion of reporting bias 17 . Concerning comparisons between active antidepressants and placebo, no evidence was found for a publication bias, but as statistical tests for asymmetry typically have low power 17 , this bias cannot be excluded. The bias is well known 33 and, for example, it has led to considering reboxetine as a serious antidepressant agent, whereas it is probably ineffective and potentially harmful 34 . It has been recently demonstrated that the selective reporting of studies in network meta-analysis of antidepressants biases estimates of relative treatment efficacy 35 .

Various barriers were encountered in our meta-analytic quest for exhaustiveness: many trials were carried out in China and published in Chinese journals. Nevertheless, the quality of many of these studies could be expected to be poor 36 , and excluding these trials, although it means a loss of randomized evidence, thus avoids other major biases.

Moreover - and this is probably the main problem - antidepressant research is completely controlled by the pharmaceutical industry 37 : 1) the firms that promoted some of the trials we identified refused to communicate results from these studies; 2) in this meta-analysis all the studies were sponsored by the pharmaceutical industry. Such studies have been shown to be more likely to demonstrate positive effects for the sponsor’s drug than independent studies 38 .

Thus, in view of these limitations, a reasonable measure of skepticism should discourage hasty conclusions. They illustrate the fact that every scientific result is uncertain and that it is difficult to be sure of an individual study conclusion, even if it explored something as patently obvious as “sucrose = sucrose”, however rigorous the method. Nevertheless, although published research findings can be erroneous 39 , Science often generates representations that leave no room for skepticism. This is probably the most insidious pitfall in Evidence-Based Medicine; it does not concern the findings of Science, it concerns academics’ understanding of Science (the knowledge-producing activity). The present-day context of medicalization of modern society 40 implicitly dictates that scientific results should have the status of Truth. Concerning major depressive disorder, although it is likely to be untrue, clinicians and a great number of patients 41 strongly believe that antidepressant drugs target a specific biological state that produces depression 42 and the pleasing serotonin hypothesis is often taken as gospel 43 .

As for the study by Cipriani et al., it was disputed with similar arguments to those set out in the present study 37 44 45 46 47 and its results were not replicated by Gartlehner et al. 48 . Moreover, whereas in our study insignificant P-values do not tell much about equivalence, in over-powered studies, like Cipriani’s study, statistically significant differences never tell much about clinically significant differences 32 . It is nevertheless mentioned in the NICE guideline 29 , with some kind of double bind: qualitatively no recommendations for ranking antidepressants are made, but quantitatively special emphasis is placed on the study (Cipriani’s name is cited 23 times versus 3 times for Gartlehner, tables present the results, and so on). What is more, in day-to-day practice, clinicians generally consider that Cipriani’s study is solid evidence for choosing antidepressants when treating a patient with newly diagnosed depression 49 . Uncertainty is sometimes acknowledged theoretically, but not in clinical practice.

This epistemological position translates into the well-known anthropological observation that the hopes and expectations of the physician are just as crucial as those of the patient in the healing process 50 . This suggests that while the opposite may not be true, the best placebos to treat Major Depressive disorder could be antidepressants because they are believed to be effective, which is probably an important determinant in the placebo effect 51 .

Nonetheless, we cannot simply assume that, because patients appear to improve on placebos in the short term, as we observe in randomized controlled trials, placebos have demonstrated the required cost-benefit balance. When one uses a treatment that relies on expectation, one must also be careful as to its possible harmful consequences, which could be linked with the corollary, disappointment. But here again this is an unsolved central question in Evidence Based Medicine.

As in pre-scientific medicine, modern physicians need to believe in the effectiveness of their techniques, and current medical literature, with its strengths and also its limitations, appears as the sophisticated way to generate such beliefs. It also raises the ethical issue of the dissemination of scientific evidence, for example, when editors permit reprints for the pharmaceutical industry intended to present results to doctors via a more commercial than epistemological approach, as was the case for the Cipriani study 52 .

We did not find any superiority of one placebo over the other. However, a critical approach to our results prevents any firm conclusion on this apparently obvious result. This result should remind clinicians to step back to take a more objective view when interpreting a scientific result, keeping in mind that Science can never be actually sure that “sucrose = sucrose” in the treatment of major depressive disorder. It is of crucial importance for their practice, far more so than ranking antidepressant efficacy.