Journal of Affective Disorders 118 (2009) 18

Contents lists available at ScienceDirect

Journal of Affective Disorders

j o u r n a l h o m e p a g e : w w w. e l s ev i e r. c o m / l o c a t e / j a d

Review

Meta-analysis of the placebo response in antidepressant trials

Winfried Rief a,, Yvonne Nestoriuc a, Sarah Weiss a, Eva Welzel a, Arthur J. Barsky b, Stefan G. Hofmann c
a
Division of Clinical Psychology and Psychotherapy, University of Marburg, Germany
b
Brigham and Women's Hospital/Harvard Medical School, Boston, MA, USA
c
Boston University, Boston, MA, USA

a r t i c l e i n f o a b s t r a c t

Article history: Improvements in placebo groups of antidepressant trials account for a major part of the
Received 8 December 2008 expected drug effects. We aimed to determine overall effect sizes of placebo and drug effects in
Received in revised form 19 January 2009 antidepressant trials, and to analyze whether the placebo effect in antidepressant trials also
Accepted 19 January 2009
occurs for patient self-perception, general psychopathology, and quality of life.
Available online 26 February 2009
Methods: Search terms covered different variants of pharmacotherapy for patients with
depressive disorders from January 1980 to December 2005 in the databases Medline/Pubmed,
Keywords:
Placebo PsychInfo and CENTRAL, a.o. We included RCTs with a placebo group and an antidepressant
Antidepressant group in people with depression.
Depression Results: We computed within group effect sizes for several outcome variables and integrated
Meta-analysis them using random-effect models. A total of 96 studies were included. Mean effect size in the
placebo group for primary outcome variables was d = 1.69 (95% CI= 1.541.84) compared to 2.50
in the drug group (95% CI= 2.302.69). There was a major difference between placebo effect sizes
assessed with observer ratings (d = 1.85, 95% CI = 1.692.01) versus patient self-perception
(d = 0.67; 95% CI = 0.490.85). The effect sizes in placebo groups in 2005 were more than twice
as great as those in 1980, but only for observer ratings, not for patient self-ratings. The result was
partly due to increased homogeneity of samples of recently published trials.
Conclusions: The placebo effect accounted for 68% of the effect in the drug groups. Whereas
clinical trials need to control the placebo effect, clinical practice should attempt to use its full
power.
2009 Elsevier B.V. All rights reserved.

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2. Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.1. Searching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.2. Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
2.3. Data abstraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.4. Validity assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.5. Data synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.1. Study characteristics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.1.1. Validity ratings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.1.2. Overall effect size of the placebo effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.1.3. Self-rating versus observer rating . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
3.1.4. Publication year effect . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

Corresponding author. Philipps University of Marburg, Clinical Psychology and Psychotherapy, Gutenbergstrasse18 D35032 Marburg. Fax: +49 6421 282 8904.
E-mail address: riefw@staff.uni-marburg.de (W. Rief).

0165-0327/$ see front matter 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.jad.2009.01.029

Downloaded for Anonymous User (n/a) at Chinese University of Hong Kong from ClinicalKey.com by Elsevier on September 22, 2017.
For personal use only. No other uses without permission. Copyright 2017. Elsevier Inc. All rights reserved.
2 W. Rief et al. / Journal of Affective Disorders 118 (2009) 18

3.1.5. Further moderators of the placebo response. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3.1.6. Sensitivity analysis. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.1.7. Sample homogeneity, inclusion criteria, and publication year . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.1. Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
4.2. Implications for research and clinical practice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Role of funding source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Conict of interest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1. Introduction studies show a decrease in effectiveness of the drugs with

Placebo responses are general phenomena in medicine,
and they seem to be particularly powerful in depression.
However, the exact size of improvement in placebo groups
compared to drug groups remains unclear because of serious
shortcomings of existing reports (e.g., strong selection of
included studies, poor methodology based on percentage
scores, and inclusion of many studies with missing data).
Moreover, it is unclear whether the improvements in placebo
groups can only be found for primary outcome variables (e.g.,
depression), or whether they can also be found for other
psychopathological variables and quality of life.
For the purpose of this paper, we will use the term placebo
response to refer to improvements that occur in the placebo
groups of clinical trials. Despite some critical evaluations
(Hrbjartsson and Gtzsche, 2001), the placebo response is
still impressive in medicine and psychiatry. The most provok-
ing examples of placebo effects in medicine come from sham
surgery (Moseley et al., 2002; McRae et al., 2004), but the
placebo effect is also impressive in relieving pain (Petrovic
et al., 2002) and reducing hypertension (Preston et al., 2000).
The placebo response was also examined in antidepres-
sant trials. Half of the drug group and 30% of the placebo
group met the responder criterion of a 50%-reduction in
depression scores from baseline (Rush, 1993; Walsh et al.,
2002). The positive effects attained with selective serotonin
reuptake inhibitors (SSRIs) that are explained by placebo are
estimated to be as high as 82% depending on the study
inclusion strategy and analysis methods used (Joffe et al.,
1996; Kirsch and Sapirstein, 1998; Kahn et al., 2000; Kirsch
et al., 2002). A recently published study shows an inverse
relationship of placebo response and depression severity,
with less placebo response in very seriously depressed
patients (Kirsch et al., 2008). When antidepressants were
compared not only to inert placebos, but also to active placebo
conditions imitating the side effects of antidepressants, the
differences in the efcacy of antidepressants and active
placebos were much less pronounced (Moncrieff et al.,
2004). However, this analysis was based on only a few, older
studies.
When comparing different antidepressant trials, it became
evident that there was tremendous variation in the size of the
placebo response. An intriguing change of the placebo
response with time was reported by Walsh and colleagues
(Walsh et al., 2002) who analyzed 57 trials and found a
correlation of r = 0.45 between the effect size in the placebo
group and publication year. The reasons for this publication
year effect are still poorly understood. In contrast, other
publication year (e.g., statins, Gehr et al., 2006).
Treatment efcacy in psychiatry can be determined using 2
different methods, namely observer ratings and patient's self-
ratings. A special characteristic of antidepressant trials is the use
of observer ratings to assess depression. Most drug trials focus
on results of the Hamilton Depression Scale (HAMD) or the
Montgomery Asberg Depression Rating Scale (MADRS). Some
studies, however, use additional self-rating scales that allow
investigators to examine improvement from the patient's
perspective. More information is needed to determine whether
placebo effects in antidepressant trials are limited to observer
ratings, or whether they also occur in patient self-ratings.
Moreover, it is unclear whether the publication year effect can
also be found for patient self-ratings.
The aim of this meta-analysis was to analyze the placebo
effect when rated by patient self-report as compared to
observer ratings. Moreover, the placebo response was not
only computed for depression, but also for general psycho-
pathology and quality of life. Special emphasis was given to
possible moderating effects, such as publication year, diag-
nosis, drug groups, and others.
2. Method
The reporting of methods follows the QUORUM Checklist.
2.1. Searching
We searched Medline/Pubmed, PsychInfo, and the
Cochrane Central Register of Controlled Trials (CENTRAL),
using the following search procedures and key words:
pharmacotherapy or drug therapy or drug treatment or
antidepressant or antidepressive and depression or affective
disorder or depressive or depressed or dysthymia or dysthy-
mic. The publication language was limited to English or
German, publication type = journal article. Results were
restricted to hits including the search terms: randomized or
randomised, and placebo, and double blind or triple blind.
In the Cochrane database of systematic reviews, all reviews
dealing with pharmacotherapy of depression were checked for
further references.
2.2. Selection
Publication year was limited to January 1980 (introduction
of DSM-III) until December 2005. Only studies addressing the
treatment of adult patients with unipolar depressive disorder
were considered. Studies were excluded if depression was

Downloaded for Anonymous User (n/a) at Chinese University of Hong Kong from ClinicalKey.com by Elsevier on September 22, 2017.
For personal use only. No other uses without permission. Copyright 2017. Elsevier Inc. All rights reserved.
 W. Rief et al. / Journal of Affective Disorders 118 (2009) 18 3

exclusively a comorbid disorder of a medical condition (e.g., 2.3. Data abstraction

post-stroke depression; depression after heart attacks).
Concurrent medication was allowed if less than 20% of
patients received it. The placebo groups had to have a
minimum of 20 patients. In typical RCT designs with 2 groups
(drug vs placebo) and 2 assessment times (pre vs post), this
allows us to detect effect sizes of the interaction term group
time according to Cohen's d of N0.33 (p b .05; power = .80);
therefore, even moderate treatment effects were detectable in
the included studies (Cohen, 1988). Additionally, the studies
had to report sufcient data to compute prepost-effect sizes
(means, at least one standard deviation or prepost-test
statistics). The utilization of at least one instrument assessing
depression was required.
A structured coding plan was developed prior to the meta-
analysis. The coding plan included a systematic denition of
43 items addressing general study characteristics (e.g.,
authors, publication year), study setting, study design (e.g.,
treatment duration, medication, randomization, blinding,
assessment instruments), report of detailed results, and
sample characteristics (e.g., diagnosis, comorbidity, drop-out).
2.4. Validity assessment
Because the frequently used JADAD-score (Jadad et al.,
1996) does not cover all aspects of validity, we expanded it

Fig. 1. Trial search.

Downloaded for Anonymous User (n/a) at Chinese University of Hong Kong from ClinicalKey.com by Elsevier on September 22, 2017.
For personal use only. No other uses without permission. Copyright 2017. Elsevier Inc. All rights reserved.
4 W. Rief et al. / Journal of Affective Disorders 118 (2009) 18
and coded additional aspects of internal validity (randomiza-
tion, intent to treat analysis, report of means and standard
deviations), construct validity (blinding, diagnosis, treatment
adherence, assessment instruments), and external validity
(report of selection of participants, participant characteristics,
exclusion criteria, treatment duration). A random selection of
30 original studies was presented to two raters (SW, EW) and
inter-rater reliability for study characteristics and validity
characteristics were computed.
2.5. Data synthesis
We computed within group effect sizes for the drug and
placebo groups. These effect sizes were dened as the mean
difference from pre-to post-treatment divided by the pooled
standard deviation (e.g., Hartmann et al., 1992). In cases of
missing post-scores, we used Becker's formula (Becker, 1988)
which is based on the difference of the post-and pre-scores
related to the standard deviation of the baseline. We used a
study weighting procedure according to sample size (Hedges
and Olkin, 1985). Homogeneity statistics using Q-statistics
(DerSimonian and Laird, 1986) revealed signicant deviation
of the mean effect. Therefore, we applied random effect
models for the integration of all effect sizes. Publication bias
was controlled for by computing safe-n rates, stem-leaf-
diagrams and funnel plots (Becker, 1988; Egger et al., 1997).
Predened subgroup analyses were performed to test
whether the ndings were sensitive to the type of anti-
depressants administered, the subtype of depression diag-
nosis, and the type of outcome variables. Furthermore, the
potential inuence of continuous moderator variables (such
as publication year, study validity, treatment duration, mean
age of patients) was analyzed in regression models.
3. Results
3.1. Study characteristics
As shown in Fig. 1, 3322 different studies fullled the rst
level of inclusion criteria. After excluding studies based on
information presented in study abstracts, 406 complete study
reports were considered in more detail. The nal sample
consisted of 96 trials1 that reported sufcient data to compute
effect sizes. The placebo groups of these studies comprised
9566 people. Approximately half of the studies were pub-
lished after 1996, 68% were conducted in the United States,
and the mean sample size was 86 participants. The mean
discontinuation rate was 30.4%. Seventy-nine of the 96 studies
addressed patients with major depression, 7 addressed
patients with dysthymia, and 10 studies addressed patients
with less specied or minor depressive disorders. In the active
drug groups, the following medications were administered:
SSRIs (36), tri-or tetra-cyclic antidepressants (18), MAO
inhibitors (5), herbal remedies (10), and others (27).
Most studies (86 of 97; 89%) utilized the Hamilton
Depression Scale (HAMD). Because various versions of the
HAMD were used, we focused on the HAMD-17 version and
estimated equivalent scores if a different version had been
1 Quality of Life
A complete list of included publications can be requested from the rst
author.
Downloaded for Anonymous User (n/a) at Chinese University of Hong Kong from ClinicalKey.com by Elsevier on September 22, 2017.
For personal use only. No other uses without permission. Copyright 2017. Elsevier Inc. All rights reserved.
used. Equivalent scores were computed according to Walsh et
al. (Walsh et al., 2002). Typical self-rating scales that were
used in the studies included the Depression subscale of the
Symptom Checklist SCL-90R (Derogatis, 1994) (10 studies),
the Beck Depression Inventory BDI (Beck and Steer, 1987) (9
studies), and the Zung Depression Scale (Zung, 1965) (4
studies).
3.1.1. Validity ratings
In contrast to previous reviews on the placebo effect in
antidepressant trials that reported substantially lower rates of
complete data sets, we successfully detected relevant means
and standard deviation scores for nearly three quarters of the
trials (72 of 96 trials). The inter-rater reliability for the study
quality characteristics was r = 0.95 which can be considered
excellent. The inter-rater reliability for other variables of the
data selection process ranged from 0.85 (strategy of analysis) to
1.00 (study identication characteristics, publication year, etc.).
3.1.2. Overall effect size of the placebo effect
The overall effect size of the placebo effect was 1.69 (95%
CI=1.541.85), as compared to d=2.50 (95% CI =2.302.69) in
the drug group. The ratio of the effect sizes suggests that 67.6%
of the improvements in the drug group were attributable to
the placebo effect. There was a high correlation between the
effect sizes in the placebo and drug groups (r = 0.69; p b .001),
indicating that reported improvements in placebo and drug
groups were strongly interdependent. Stem-leaf diagrams
revealed a normal distribution with most effect sizes lying in
the area between 1.00 and 2.8, which indicates that publica-
tion bias does not explain the results. The effect sizes of the
placebo response were not inuenced by the category of drug
that was investigated, with mean effect sizes ranging from
1.65 (SSRI-placebo) to 1.73 (other antidepressants). However,
the placebo response was signicantly greater in major
depression (d = 1.83; 95% CI = 1.671.99) than in dysthymia
(d = 1.11; 95% CI = 0.491.58). Interestingly, this diagnostic
specicity remained even after controlling for depression
severity (see below). The overall placebo effect was highest
for the primary outcome variable (depression). However,
substantial effect sizes also emerged for anxiety, general
psychopathology and quality of life (see Table 1).
3.1.3. Self-rating versus observer rating
In the placebo groups, there was a substantial difference
between effect sizes for improvements rated by observers
(d = 1.85; 95% CI = 1.692.01; 93 studies) compared to those
rated by patients (d = 0.67; 95% CI = 0.490.85; 28 studies).
Pearson's correlation between self-and observer ratings was
r = 0.57 for the 24 studies that reported both types of ratings.
The difference between self-ratings and observer ratings was
Table 1
Average effect sizes for improvements in the placebo groups.
Variable K/n Mean d 95% CI
Depression 97/9,469 1.69 1.541.84
Anxiety 17/2,633 0.98 0.711.25
General Psychopathology 4/258 0.68 0.081.29
12/1,214 1.00 0.571.43
K = number of studies; n = total sample size; mean d = average effect size.
 W. Rief et al. / Journal of Affective Disorders 118 (2009) 18 5

also found in the drug groups (self-rating d = 1.12 versus
observer rating d = 2.89). Interestingly, the self-ratings and
observer ratings of proportion of positive effects in the
placebo group relative to the drug group were comparable
(observer: 66.5%; self-rating: 66.6%).
analyzed (see Fig. 2b). Patients' perception of self-improve-
ment in the placebo groups was not inuenced by publication
year (r =.08; NS). In the drug groups, the same tendency was
found (see Fig. 2a, b; bottom), although due to more variance
this effect was less clear and failed to reach signicance.

3.1.4. Publication year effect
With our large sample size, we were able to replicate the
strong linear association between publication year and effect
sizes of the placebo group (see Fig. 2a). In fact, the mean effect
sizes in the placebo groups of the studies published around
2005 were more than twice as great as those published before
1985. The correlation between effect size and publication year
was 0.41 (p b .001; compared to 0.45 in the original study
reporting this effect (Walsh et al., 2002)). However, this effect
was not found when self-rating scales for depression were
3.1.5. Further moderators of the placebo response
Using bivariate regression models, we analyzed further
correlates of the effect size of the placebo response. We could
not conrm any signicant inuence for the average age of
the sample ( = .18), percentage of female participants
( = 0.17), validity of the study ( = 0.06), or treatment
duration in weeks ( = 0.06). However, there seemed to be a
signicant inuence of diagnosis (major depression versus
dysthymia; = 0.30; p b 0.003), and severity of depression
( = 0.26; p b 0.02). Therefore, we entered publication year,

Fig. 2. a) Association of effect size with publication year (observer ratings; placebo groups at top, drug groups at the bottom). b) association of effect size with
publication year (self-report; placebo groups at top, drug groups at the bottom).

Downloaded for Anonymous User (n/a) at Chinese University of Hong Kong from ClinicalKey.com by Elsevier on September 22, 2017.
For personal use only. No other uses without permission. Copyright 2017. Elsevier Inc. All rights reserved.
6 W. Rief et al. / Journal of Affective Disorders 118 (2009) 18
diagnosis, and severity of depression in one regression
analysis model. This analysis conrmed the major inuence
of publication year ( = 0.48; p b 0.001) and diagnosis
( = 0.29; p b 0.003), but failed to show an additional
inuence of depression severity ( = 0.14; p = 0 .07).
3.1.6. Sensitivity analysis
A funnel plot conrmed that the greater the sample size,
the closer the effect size was to the average effect size. This
argues against a publication bias; if publication bias played a
major role, large sample size trials would supposedly reveal
lower effect sizes than small sample size trials (Egger et al.,
1997). Failsafe-rates were computed to determine how many
unpublished studies would be needed to reduce the effect size
to d = 0.01 thereby eliminating the positive effect in the
placebo group. It was found that 16,300 unpublished studies
reporting no positive placebo effects would be needed. To
reduce the mean effect size of 1.69 to a low effect size of 0.21, a
failsafe-rate of 683 studies would be needed. Both rates
conrm the stability of improvements in the placebo groups.
3.1.7. Sample homogeneity, inclusion criteria, and publication year
Effect sizes and signicant ndings for improvements can
be increased if more homogeneous samples are included.
Homogeneity of samples is reected in the standard deviation
of scores. The lower the standard deviation (e.g., at baseline),
the higher the effect size if the average improvement is
comparable. Therefore, we analyzed whether there was a
linear association between publication year and standard
deviation of depression scores at baseline. Fig. 3 shows that, in
fact, a linear association between publication year and
homogeneity of samples existed (r = 0.30; p b .01). However,
the variation was less than 25% of the values from 1980;
therefore, this reduction of heterogeneity can only partly
account for the substantial changes of the effect sizes for
Fig. 3. Decreasing heterogeneity of depression with publication year
(r = 0.30; p b .01). Abbreviation: SD = standard deviation.
Downloaded for Anonymous User (n/a) at Chinese University of Hong Kong from ClinicalKey.com by Elsevier on September 22, 2017.
For personal use only. No other uses without permission. Copyright 2017. Elsevier Inc. All rights reserved.
improvements in the placebo groups. In the studies included
in our analysis, we could not conrm signicant associations
between inclusion cut-off scores for depression and publica-
tion year (r = 0.01; ns) or mean baseline depression severity
and publication year (r = .195; ns). Therefore the associa-
tion between depression standard deviation and publication
year can be only attributed to increasing homogeneity.
4. Discussion
We conducted a meta-analysis of improvements in
placebo groups in 96 trials containing a total of 9,566
depressed patients. We found that the improvements in the
placebo groups corresponded to 67% of the improvements in
the active drug groups. We further conrmed that the
reported placebo effect sizes increased tremendously from
1980 to 2005. However, this strong publication year effect was
only found when the primary outcome variable was an
observer rating of depression. Self-ratings of depression were
less prone to publication year effects. Indeed, the placebo
response in self-ratings was only of medium magnitude, but
corresponded to other outcome measures (psychopathology,
quality of life).
Our analysis conrms that the placebo effect can differ
substantially depending on study design and outcome
measures (Antonaci et al., 2007). The difference between
self-ratings and observer ratings in evaluating clinical out-
comes of depression trials is intriguing, and raises questions
about which is the more valid assessment approach. Some
scientists question whether people with depression can
accurately detect small changes in mood, and therefore
argue in favor of observer ratings. Others, however, including
Fava and colleagues (Fava et al., 2003), believe that clinicians
tend to overestimate changes in patients, and therefore
endorse the use of additional assessment strategies.
The strong publication year effect found for observer
ratings questions their validity. Following the view of Fava
and others (Fava et al., 2003), this effect might indicate
increasing overestimation of positive effects by clinicians. One
possibility is that doctors became enthusiastic about the
efcacy of the antidepressant drugs, and developed higher
expectations for improvement. It is unlikely that the drug-and
placebo effects actually increased more than 100% over the
last 2 decades. Walsh et al. (Walsh et al., 2002) mention that
the Hamilton cut-off score for study inclusion has also
increased over the years, and that investigators might have
tended to overestimate depression severity at baseline in
order to include patients in studies. This might be associated
with a substantial decrease in depression scores between the
rst and second assessments (Khan et al., 2007). While this
effect might contribute in part to increasing effect sizes over
time, it is unlikely that it explains the entire publication year
effect. In most studies, the placebo response not only occurs
between the rst and second assessments, but rather,
improvements increase continuously. Furthermore, we were
able to show that increasing effect sizes for improvements in
antidepressant trials were partly due to the decreasing
heterogeneity of the samples. While this process might
improve internal validity, samples of clinical trials differ
more and more from clinical practice. However, the change of
heterogeneity over the years is smaller than the increase of
 W. Rief et al. / Journal of Affective Disorders 118 (2009) 18
improvement effect sizes, therefore indicating that variations
of heterogeneity can only partly explain the publication year
effect.
Interestingly, we found differences in placebo responses
depending on diagnosis. Patients with dysthymia reported
lower placebo responses than patients with major depression.
This is not only due to depression severity (which was
controlled for in regression analyses), but depended on
depression type. Dysthymia is chronic by denition, while
major depression tends to follow a more cyclic course,
thereby allowing more variability in placebo responses.
Different placebo responses depending on diagnostic group
have also been shown to exist in other areas of mental
disorders (Huppert et al., 2004).
The question arises as to which mechanisms might be
responsible for the placebo response. Rapaport and others
(Rapaport et al., 2000) assume that placebo responders differ
from drug responders, therefore indicating that the mechan-
isms of improvement under placebo might differ from the
mechanisms of improvement under drug inuence. However,
other studies have found that many predictors and correlates of
change are similar between placebo and antidepressant drug
responders (Leuchter et al., 2002; Mayberg et al., 2002); this
notion is conrmed by the high correlation of placebo group
effect sizes and drug group effect sizes in our analysis (r = 0.69).
A positive placebo response depends on the activation of
specic brain areas (Benedetti, 1996; Bingel et al., 2006).
Classical conditioning might also explain the placebo response.
Classical conditioning depends on prior experience. While prior
experience might facilitate the placebo response, it has also
been shown that the placebo response can be triggered by
expectation alone (Benedetti et al., 2003; Levine et al., 2006).
A crucial question for the placebo response in antidepres-
sant trials is whether the effect is simply a regression to the
mean. It has been shown that the higher the depression score,
the greater the placebo response (r = 0.27, Walsh et al.,
2002). However, our results demonstrate that depression
severity does not explain the major effects, and the results of
Kirsch et al. (Kirsch et al., 2008) also contradict the
assumption that a mere regression-to-the mean effect
explains placebo responses in depression. Moreover, regres-
sion to the mean cannot account for the observed publication
year effect, the differences depending on diagnosis, or the
differences depending on ascertainment strategy.
Another methodological shortcoming is the lack of control
groups that show the natural course of the disorder. In fact, in a
critical analysis of the placebo response, Hrobjartsson and
Gtzsche (Hrbjartsson and Gtzsche, 2001) compared results
of placebo groups to natural course groups and found
substantially lower placebo effects than expected. Three years
after assessment, remission rates for major depression can be
up to 70% (Rhebergen et al., 2009); in contrast, other studies
have described stable depression severity for samples with
depression scores comparable to those in our included studies
(Stoolmiller et al., 2005). Therefore only natural course control
groups can elucidate the long-term inuence of placebo
reactions, but these groups are rarely included in depression
trials due to ethical concerns.
Differences between the mechanisms acting in placebo
groups compared to those acting in the drug groups have
further implications for the rationale of placebo-controlled
Downloaded for Anonymous User (n/a) at Chinese University of Hong Kong from ClinicalKey.com by Elsevier on September 22, 2017.
For personal use only. No other uses without permission. Copyright 2017. Elsevier Inc. All rights reserved.
7
trials. It is highly questionable whether the nonspecic effects
acting in placebo groups are the same as the nonspecic
effects acting in active drug groups. If placebos are not
completely inert, but instead mimic the side effects of the
drug (active placebos), they are more effective. Therefore, it
is not surprising that the difference between the improve-
ments in drug groups compared to placebo groups diminish to
insignicant levels if the effect of active placebos is compared
to that of antidepressants (Moncrieff et al., 2004). In other
words, if people experience side effects, they might experi-
ence more non-specic, placebo effects. However, in typical
placebo controlled trials, this increase of nonspecic effects
only occurs in the drug group, but not in the placebo group.
4.1. Limitations
A general limitation of meta-analyses is the quality of the
included studies. However, we did not nd any association
between study quality and the reported results. The total
sample size and the number of studies included in our meta-
analysis were sufciently high for the primary outcome
variable. Publication bias is also a point of consideration in
meta-analyses, and has been shown to inuence system-
atically the report of positive effects for antidepressants
(Turner et al., 2008). We controlled for publication bias
computing fail-to-safe rates, and we were able to show that
publication bias was unlikely to account for the substantial
improvements in placebo groups. Although we had to exclude
many studies because of insufcient data (Fig. 1), we still
included more studies than other meta-analyses on anti-
depressants, and the reported fail-to-safe rates conrm that
the results are robust. However, the results of self-rating
scales in our analysis were based on only one-quarter of the
studies included, which increases the risk of type-II-errors.
Furthermore, different self-rating scales were investigated,
while observer ratings were mainly based on the Hamilton
scale. This might have contributed to a blunting of systematic
inuences such as the publication year effect in self-ratings.
Further studies should directly compare self-and expert
ratings using the same items (e.g., the MADRS versus
MADRS self-report). Therefore, our meta-analysis leaves
some questions unanswered. Although we can demonstrate
a tremendous difference between the results of patient self-
reports and observer ratings, we cannot completely explain
this difference. The same is true for the publication year effect,
which is only partly explained by ascertainment strategy and
reduction of sample heterogeneity. Therefore, these effects
should be analyzed further in future studies.
4.2. Implications for research and clinical practice
These results have several implications for research and
clinical practice. As long as it remains unclear whether self-
ratings or observer ratings are the more valid instrument,
both types of instruments should be used in clinical trials.
However, only one fourth of the studies included in this
overview used self-rating scales for depression. For clinical
practice, it should be kept in mind that the placebo response
in the treatment of depression is substantial. Therefore,
clinicians should attempt to retain and optimize the non-
specic effects. Indeed, Benedetti (Benedetti et al., 2006)
8 W. Rief et al. / Journal of Affective Disorders 118 (2009) 18

demonstrated that the elimination of positive expectation Jadad, A.R., Moore, R.A., Carroll, D., Jenkinson, C.P., Reynolds, D.J., Gavaghan,
D.J., McQuai, H.J., 1996. Assessing the quality of reports of randomized
effects made analgesic therapies less effective and that higher clinical trials: is blinding necessary? Controlled Clinical Trials 17, 112.
drug dosages were needed in those patients as compared to Joffe, R.T., Sokolov, S., Streiner, D., 1996. Antidepressant treatment of
others with positive treatment expectations. Waber et al. depression: a meta-analysis. Canadian Journal of Psychiatry 41, 613616.
Kahn, A., Warner, H.A., Brown, W.A., 2000. Symptom reduction and suicide
were able to show that the analgesic placebo effect is higher if risk in patients treated with placebo in antidepressant clinical trials.
people expect the drug to be expensive and valuable (Waber Archives of General Psychiatry 57, 311317.
et al., 2008). Therefore the systematic application of positive Khan, A., Schwartz, K., Kolts, R.L., Ridgway, D.L.C., 2007. Relationship between
depression severity entry criteria and antidepressant clinical trial
expectations can help to optimize medication use. outcomes. Biological Psychiatry 62, 6571.
Kirsch, I., Sapirstein, G., 1998. Listening to prozac but hearing placebo: a
Role of funding source meta-analysis of antidepressant medication. Prevention and treatment 1
Nothing declared. (article 0002a).
Kirsch, I., Moore, T.J., Scoboria, A., Nicholls, S.S., 2002. The emperor's new
Conict of interest drugs: an analysis of antidepressant medication data submitted to the U.S.
food and drug administration. Prevention and Treatment 5(article 23).
No conict declared.
Kirsch, I., Deacon, B.J., Huedo-Medina, T.B., Scoboria, A., Moore, T.J., Johnson,
B.T., 2008. Initial severity and antidepressant benets: a meta-analysis of
References data submitted to the Food and Drug Administration. PLOS Medicine 5
(2), e45.
Leuchter, A.F., Cook, I.A., Witte, E.A., Morgan, M., Abrams, M., 2002. Changes
Antonaci, F., Chimento, P., Diener, H.C., Sances, G., Bono, G., 2007. Lessons
in brain function of depressed subjects during treatment with placebo.
from placebo effects in migraine treatment. Journal of Headache and Pain
American Journal of Psychiatry 159, 122129.
8, 6366.
Levine, M.E., Stern, R.M., Koch, K.L., 2006. The effects of manipulating
Beck, A.T., Steer, R.A., 1987. Beck Depression Inventory. Manual. The
expectations through placebo and nocebo administration on gastric
Psychological Cooperation, San Antonio.
tachyarrhythmia and motion-induced nausea. Psychosomatic Medicine
Becker, B.J., 1988. Synthesizing standardized mean-change measures. British
68, 478486.
Journal of Mathematical and Statistical Psychology 41, 257278.
Mayberg, H.S., Silva, J.A., Brannan, S.K., Tekell, J.L., Mahurin, R.K., McGinnis, S.,
Benedetti, F., 1996. The opposite effect of the opiate antagonist naloxon and
Jerabek, P.A., 2002. The functional neuroanatomy of the placebo effect.
the cholecystokinin antagonist proglumide on placebo analgesia. Pain 64,
American Journal of Psychiatry 159, 728737.
535543.
McRae, C., Cherin, E., Yamazaki, T.G., Diem, G., Vo, A.H., Russell, D., Ellgring, J.H.,
Benedetti, F., Pollo, A., Lopiano, L., Lanotte, M., Vighetti, S., Rainero, I., 2003.
Fahn, S., Greene, P., Dillon, S., Wineld, H., Bjugstad, K.B., Freed, C.R., 2004.
Conscious expectation and unconscious conditioning in analgesic, motor,
Effects of perceived treatment on quality of life and medical outcomes in a
and hormonal placebo/nocebo responses. Journal of Neuroscience 23,
double-blind placebo surgery trial. Archives of General Psychiatry 61,
43154323.
412420.
Benedetti, F., Arduino, C., Costa, S., Vighetti, S., Tarenzi, L., Rainero, I.,
Moncrieff, J., Wessely, S., Hardy, R., 2004. Active placebos versus antidepres-
Asteggiano, G., 2006. Loss of expectation-related mechanisms in
sants for depression. The Cochrane Database of Systematic Review (1).
Alzheimer's disease makes analgesic therapies less effective. Pain 121,
Moseley, J.B., O'Malley, K., Petersen, N.J., Menke, T.J., Brody, B.A., Kuykendall,
133144.
D.H., Hollingsworth, J.C., Ashton, C.M., Wray, N.P., 2002. A controlled trial
Bingel, U., Lorenz, J., Schoell, E., Weiller, C., Bchel, C., 2006. Mechanisms of
of arthroscopic surgery for osteoarthritis of the knee. New England
placebo analgesia: rACC recruitment of a subcortical antinociceptive
Journal of Medicine 347, 8188.
network. Pain 120, 815.
Petrovic, P., Kalso, E., Petersson, K.M., Ingvar, M., 2002. Placebo and opioid
Cohen, J., 1988. Statistical Power Analysis for the Behavioral Sciences.
analgesia-imaging a shared neuronal network. Science 295, 17371740.
Erlbaum Academic Press, Hillsdale, NJ.
Preston, R.A., Materson, B.J., Reda, D.J., Williams, D.W., 2000. Placebo-
Derogatis, L.R., 1994. SCL-90. Administration, Scoring and Procedures. Author,
associated blood pressure response and adverse effects in the treatment
Massachussetts.
of hypertension. Archives of Internal Medicine 160, 14491454.
DerSimonian, R., Laird, N., 1986. Meta-analysis in clinical trials. Controlled
Rapaport, M.H., Pollack, M., Wolkow, R., Mardekian, J., Clary, C., 2000. Is
Clinical Trials 7, 177188.
placebo response the same as drug response in panic disorder? American
Egger, M., Smith, G.D., Schneider, M., Minder, C., 1997. Bias in meta-analysis
Journal of Psychiatry 157, 10141016.
detected by a simple, graphical test. British Medical Journal 315,
Rhebergen, D., Beekman, A.T.F., de Graaf, R., Nolen, W.A., Spijker, J.,
629634.
Hoogendijk, W.J., Pennix, B.W.J.H., 2009. The three-year naturalistic
Fava, M., Eden Evins, A., Dorer, D.J., Schoenfeld, D.A., 2003. The problem of the
course of major depressive disorder, dysthymic disorder and double
placebo response in clinical trials for psychiatric disorders: culprits,
depression. Journal of Affective Disorders 115 (3), 450459.
possible remedies, and a novel study design approach. Psychotherapy
Rush, A.J., 1993. Clinical Practice Guideline 5. Depression in Primary Care Vol. 2.
and Psychosomatics 72, 115127.
Treatment of Major Depression. AHCPR Silver Spring.
Gehr, B.T., Weiss, C., Porzsolt, F., 2006. The fading of reported effectiveness. A
Stoolmiller, M., Kim, H.K., Capaldi, D.M., 2005. The course of depressive
meta-analysis of randomised controlled trials. BMC Medical Research
symptoms in men from early adolescence to young adulthood:
Methodology 6. doi:10.1186/1471-2288-11861125.
identifying latent trajectories and early predictors. Journal of Abnormal
Hartmann, A., Herzog, T., Drinkmann, A., 1992. Psychotherapy of bulimia
Psychology 114, 331345.
nervosa: what is effective? Journal of Psychosomatic Research 36,
Turner, E.H., Mathhews, A.M., Linardatos, E., Tell, R.A., Rosenthal, R., 2008.
159167.
Selective publication of antidepressant trials and its inuence on
Hedges, L.V., Olkin, I., 1985. Statistical Methods for Meta-analysis. Academic
apparent efcacy. New England Journal of Medicine 358, 252260.
Press, Orlando, Fla.
Waber, R.L., Shiv, B., Carmon, Z., Ariely, D., 2008. Commercial features of
Hrbjartsson, A., Gtzsche, P.C., 2001. Is the placebo powerless? An analysis of
placebo and therapeutic efcacy. JAMA 299, 10161017.
clinical trials comparing placebo with no treatment. New England
Walsh, B.T., Seidman, S.N., Sysko, R., Gould, M., 2002. Placebo response in
Journal of Medicine 344, 15941602.
studies of major depression: variable, substantial, and growing. Journal
Huppert, J.D., Schultz, L.T., Foa, E.B., Barlow, D.H., Davidson, J.R.T., Gorman, J.M.,
of the American Medical Association 287, 18401847.
Shear, M.K., Simpson, H.B., Woods, S.W., 2004. Differential response to
Zung, W.W.K., 1965. A self-rating depression scale. Archives of General
placebo among patients with social phobia, panic disorder, and
Psychiatry 12, 6370.
obsessivecompulsive disorder. American Journal of Psychiatry 161,
14851487.

Downloaded for Anonymous User (n/a) at Chinese University of Hong Kong from ClinicalKey.com by Elsevier on September 22, 2017.
For personal use only. No other uses without permission. Copyright 2017. Elsevier Inc. All rights reserved.