Article Text

Download PDFPDF

Is placebo response in antidepressant trials rising or not? A reanalysis of datasets to conclude this long-lasting controversy
  1. Toshi A Furukawa1,
  2. Andrea Cipriani2,3,
  3. Stefan Leucht4,
  4. Lauren Z Atkinson2,3,
  5. Yusuke Ogawa1,
  6. Nozomi Takeshima1,
  7. Yu Hayasaka1,
  8. Anna Chaimani5,6,7,
  9. Georgia Salanti8
  1. 1 Departments of Health Promotion and Human Behavior and Clinical Epidemiology, Kyoto University Graduate School of Medicine/ School of Public Health, Kyoto, Japan
  2. 2 Department of Psychiatry, University of Oxford, Oxford, UK
  3. 3 Oxford Health NHS Foundation Trust, Oxford, UK
  4. 4 Department of Psychiatry and Psychotherapy, Technische Universität München, Klinikum rechts der Isar, Munich, Germany
  5. 5 School of Medicine, Paris Descartes University, Paris, France
  6. 6 INSERM, UMR1153 Epidemiology and Statistics, Sorbonne Paris Cité Research Center (CRESS), METHODS Team, Paris, France
  7. 7 Cochrane France, Paris, France
  8. 8 Institute of Social and Preventive Medicine, University of Bern, Bern, Switzerland
  1. Correspondence to Dr Toshi A Furukawa, Departments of Health Promotion and Human Behavior and of Clinical Epidemiology, Kyoto University Graduate School of Medicine and School of Public Health, Yoshida Konoe-cho, Sakyo-ku, Kyoto 606-8501, Japan; furukawa{at}


It had long been believed that placebo response rates in antidepressant trials have been increasing and that they were responsible for rising numbers of so-called failed antidepressant trials. Two recent systematic reviews examined this issue and reached completely opposite findings. Furukawa and colleagues in a paper published in 2016 found that the placebo response rates are stable since 1991 and the apparent increase up to 2000 was confounded by changes in trial design features. By contrast, Khan and colleagues more recently concluded that placebo response rates had grown steadily in the past 30 years. The two reviews differed in the datasets they used, definitions of placebo response and statistical analyses. In this perspective article, we examined if such differences were responsible for the two reviews’ contrasting conclusions. Our reanalyses confirmed our previous results. We found that in any dataset and for any placebo response definition, there was no increase in placebo response over the years when the analysis was adjusted for the confounders related to study design features or when it was limited to studies published after 1990s. We conclude that placebo response in antidepressant trials has remained stable for the past 25 years, during which time the large majority of the studies have come to share similar design features.

  • depression & mood disorders

Statistics from

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.

There is ongoing debate whether placebo response in antidepressant trials has been increasing over the years.1 2 As the collaborative efforts in evidence synthesis have progressed and less biased datasets of antidepressant trials become available,3 two new analyses including both published and unpublished studies revisited this issue but reached opposing conclusions.

Furukawa et al 4 examined this issue based on 252 published and unpublished double-blind randomised controlled trials of 20 antidepressants conducted between 1978 and 2015 and concluded that (1) response rates on placebo have remained steady since 1991 and (ii) there was apparent increase up to 2000 but this trend became non-significant when adjusted for length of trial, number of study centres and use of fixed dosing.

Khan et al 5 re-examined this issue by studying 85 trials contained in FDA reviews for 16 antidepressants (1987–2013) and found that the magnitude of placebo response has been steadily growing in the past 30 years. To these authors’ merit, they also examined the time trend in the efficacy estimates of antidepressants, namely the difference in symptom reduction between antidepressants and placebo. In this EBMH Perspective, we will focus on exploring the reasons for contradicting conclusions between the two research articles regarding the time trends in placebo response.

There are several important differences between these two studies which may explain their contradictory findings.

  1. Furukawa et al 4 included all published and unpublished studies, while Khan et al 5 included studies reported in FDA reviews only. However comprehensive it may be, the former dataset cannot be exempt from risk of publication bias, while the latter apparently suffers from some selective reporting bias, as the FDA reviews sometimes did not provide details especially for ‘failed’ studies, which then cannot be included in the analyses.

  2. The outcome was the proportion of responders, defined as showing 50% or greater reduction from baseline, in the former review,4 while it was the percentage of symptom reduction from baseline in the latter study.5 Both outcomes included imputations according to the last-observation-carried-forward method for study participants that provided at least one postbaseline measurement before dropping out of the study. In the former approach, participants without postbaseline measurements were assumed to be non-responders. Such missing data rate was small (on average 3.6% across all studies) and previous studies have shown that this technique provides robust results.6 The latter approach ignores such early dropouts completely but has the advantage of potentially being more sensitive because it is a continuous outcome.

Khan et al suggested that these two factors might explain the discrepancies between the two studies. However, there are some other differences as well.

  1. In the statistical analyses of the relationship between placebo response and year, Furukawa et al 4 used meta-regression to take account of the different sample sizes of the included studies, while Khan et al 5 used linear regression without weighting for precision of each study. Moreover, because there is no guarantee that the relationship between placebo response and year was linear, Furukawa et al 4 studied the structural break point.

  2. Furukawa et al 4 defined the year as the year of study completion or, when this was not reported, the year of publication. The year variable was considered missing when neither was available. Khan et al 5 defined year as the FDA approval year of the drug. This apparently risks some inaccuracies because studies for each drug must have been conducted several and variable years ahead of the drug approval year.

  3. Khan et al 5 examined three additional covariates: duration of the trial, number of trial arms and dosing schedule. Furukawa et al 4 examined 14 covariates including the above three, as prespecified in the study protocol for the systematic review7 and found that length of trial, number of study centres and use of fixed dosing impacted on the placebo response.

In this reappraisal, we will examine whether the above differences can explain the discrepancies in the findings of the two studies and thereby elucidate the true underlying relationship between placebo response and year. First, we will examine the differences in the datasets. Second, after determining the dataset with possibly the least publication bias, we will examine the two primary outcomes, namely the proportion of responders and the percentage reduction in symptoms. Last, we will perform some sensitivity analyses to see whether our final findings are robust.


In order to prepare the common dataset so that the ensuing analyses will be interpretable, we decided to concentrate on the studies contained in the FDA reviews of the 16 antidepressants that Khan et al had included. In comparing the eligible studies in Furukawa et al 4 and Khan et al,5 we made the following observations:

  1. There were six studies that Khan et al 5 included in their study but Furukawa et al 4 did not. One study may not have been eligible because 27% of the participants did not suffer from major depression (Study 86 141 of citalopram). Five studies were excluded in Furukawa et al 4 because the active drugs were administered outside the range of the FDA-licensed doses (Studies 223, 304, 308, 320 of desvenlafaxine and 85 027 of mirtazapine). The latter studies might not provide valid estimates of placebo response as it would be harder to maintain double-blindness in studies using extreme high doses; by contrast, it may be argued that including such studies would provide a better estimate of placebo response across antidepressant trials including ‘failed’ trials.

    On the other hand, Khan et al 5 seem to have missed 14 studies that are mentioned in the FDA reviews (Studies 84  023, 84  62, 003–023 of mirtazapine; CN104-002, 030A2-0004, CN104-045, CN104-054 of nefazodone; PAR-274 of paroxetine; SER 103, SER 310 of sertraline; 304, 12541A of vortioxetine).

  1. Furukawa et al 4 had more outcome data, whether proportion of responders or percentage of symptom reduction, than Khan et al 5 for the same studies. The latter restricted the data to be extracted from FDA reviews which often did not report details of trials (eg, baseline or endpoint scores), especially when the trials were ‘negative’. In contrast, the former review searched for all published and unpublished data for the trials by searching Cochrane Central Register of Controlled Trials, CINAHL, EMBASE, LiLACS, MEDLINE, PSYCINFO, regulatory agencies' websites including FDA, EMA and PMDA, several drug companies’ websites and international registers of clinical trials. When there were discrepancies in the reported outcome between published and unpublished reports, Furukawa et al 4prioritised the unpublished data according to the prespecified review protocol to minimise the risk of selective reporting.7

We judged that the primary dataset that is least susceptible to study publication bias and selective outcome reporting bias should therefore contain: all the studies included in Khan et al,5 except for 86 141 of citalopram, plus the 14 studies that they may have missed, with the outcome information supplemented from all available published and unpublished sources in addition to the FDA documents. The year of the study was defined as the year of study completion, the year of study publication or the year of drug approval from FDA, where available in this order. All in all, there were 98 placebo arms (9440 participants) representing 16 antidepressants (study year range: 1985–2012). We will however run a sensitivity analysis to see if the findings would be different if we restricted the included studies and data to those by Khan et al.5


Figure 1 shows the secular changes in (1) proportion of responders and (2) percentage of symptom reduction, meta-analytically synthesised for each year between 1985 and 2012.

Figure 1

Secular changes in proportion of responders (A) and percentage of symptom reduction (B). The size of the bubble is proportional to the sample size.

Furukawa et al 4 had found that there was a structural break in the trends in placebo response rates in 1991 and that the placebo response has stabilised since 1990s. The present dataset was too small to examine the structural break point, so we simply examined if the meta-regression revealed a significant influence of the year if we limited the included studies to those after 1991. For both outcomes, the year was not statistically significant (OR=1.00, 95% CI 0.98 to 1.01, P=0.73 and β=0.003, 95% CI −0.0003 to 0.005, P=0.08, respectively). The lack of influence of the year on placebo response rates after 1990s is clearly visible in figure 1.

Conventional meta-regression of the year on these two outcomes over the four decades since 1980s through 2010s revealed that the year was a statistically significant predictor of placebo response (OR=1.01, 95% CI 1.00 to 1.02, P=0.025 for proportion of responders, β=0.004, 95% CI 0.002 to 0.006, P<0.001 for percentage of symptom reduction).

Furukawa et al 4 suggested, however, that the apparent increase may be confounded by the increase in trial length, in the number of multicentred trials and in the use of fixed dosing schedule over the years. We therefore reran the above meta-regressions by adjusting for these three covariates (table 1). The influence of the year was no longer significant (OR=1.01, 95% CI 0.99 to 1.02, P=0.38 and β=0.001, 95% CI −0.002 to 0.004, P=0.59, respectively): in both instances, the increase in multicentred trials was the strongest confounder.

Table 1

Influence of covariates on the secular trends in placebo response

Sensitivity analyses

We ran several sensitivity analyses to confirm Khan et al 5 findings by (1) restricting the dataset to what they used and (2) using simple linear regression without weighting for study precision.

The meta-regression of the year on the percentage of symptom reduction in Khan’s dataset revealed a statistically significant influence of the study year (P<0.001). However, when adjusted for the length of trial, number of study centres and dosing schedule, it was no longer significant (P=0.28). The year was also non-significant when we limited the studies to those after 1991 (P=0.17).

Using linear regression as in Khan et al, the results were similar: significant in the crude analysis (P<0.001) but no longer so when adjusted for the confounders (P=0.28) or when limited to studies after 1991 (P=0.16).


The opposing conclusions regarding the secular trend in placebo response in antidepressant trials reached by Furukawa et al 4 and Khan et al 5 were not due to the differences in the datasets they used (all available published and unpublished studies vs FDA reviews), the definition of response they examined (proportion of responders vs percentage reduction in symptoms) or the analytical methods they used (weighted meta-regression vs simple linear regression). Regardless of these differences, there was no increase in the placebo response since 1990s. There was apparent increase especially between 1980s and 1990s but it became non-detectable when adjusted for several confounders notably the number of study centres. We conclude that the placebo response in antidepressant trials has remained stable for the past 25 years, during which time the large majority of the studies have come to share similar design features.


  1. 1.
  2. 2.
  3. 3.
  4. 4.
  5. 5.
  6. 6.
  7. 7.


  • Competing interests TAF has received lecture fees from Eli Lilly, Janssen, Meiji, MSD, Pfizer and Tanabe-Mitsubishi and consultancy fees from Takeda Science Foundation. He has received royalties from Igaku-Shoin and Nihon Bunka Kagaku-sha publishers. He has received grant or research support from Mochida and Tanabe-Mitsubishi. AC was expert witness for Accord Healthcare for a patent issue about quetiapine extended release. SL has received honoraria for lectures from Eli Lilly, Lundbeck (Institute), Pfizer, Janssen, BMS, Johnson and Johnson, Otsuka, Roche, SanofiAventis, ICON, Abbvie, AOP Orphan, Servier; for consulting/advisory boards from Roche, Janssen, Lundbeck, Eli Lilly, Otsuka, TEVA; for the preparation of educational material and publications from Lundbeck Institute and Roche. Eli Lilly has provided medication for a clinical trial led by SL as principal investigator. All the other authors declare no competing interests.

  • Provenance and peer review Commissioned; externally peer reviewed.