Thiazolidinediones and cancer: results of a meta-analysis of randomized clinical trials

ORIGINAL ARTICLE
Thiazolidinediones and cancer: results of a meta-analysis
of randomized clinical trials
Matteo Monami •Ilaria Dicembrini •
Edoardo Mannucci
Received: 1 May 2012 / Accepted: 6 July 2013
ÓSpringer-Verlag Italia 2013
Abstract Recent epidemiological data have contributed
to the formulation of the hypothesis about the long-term
safety of pioglitazone, a thiazolidinedione (TZD), with
respect to malignancies, in particular bladder cancer. The
primary aim of this meta-analysis of randomized clinical
trials, not designed a priori to test this hypothesis, was to
explore whether TZDs affect the risk of cancer. A meta-
analysis was performed including published and unpub-
lished randomized trials with a duration of at least
52 weeks, enrolling patients with or without diabetes,
comparing TZDs with either placebo or other drug thera-
pies on various different outcomes. We found 22 trials
reporting at least one cancer and enrolling 13,197 patients
to TZD (pioglitazone: n=3,710 and rosiglitazone:
n=9,487) and 12,359 to placebo or active comparator
groups. The mean follow-up was 26.1 months. Overall,
those assigned at random to TZDs had a significant
reduction (MH-OR 0.85 [0.73–0.98]; p=0.027) in the
incidence of malignancies, with no significant difference in
effect between pioglitazone and rosiglitazone. Specifically,
subgroup analyses showed a significant reduction for ros-
iglitazone (MH-OR 0.82 [0.69–0.98]; p=0.029), but not
for pioglitazone (MH-OR 0.66 [0.34–1.28]; p=0.22). In
further subgroup analyses of site-specific malignancies
based on the data from four trials, the risk of bladder cancer
with pioglitazone (MH-OR) was 2.05 [0.84–5.02];
p=0.12. Further, rosiglitazone, but not pioglitazone, was
associated with a significantly reduced risk of bowel can-
cer. In contrast, pioglitazone, but not rosiglitazone, was
associated with a significant reduction in breast cancer. The
present meta-analysis of trials, not designed a priori to test
the hypothesis, provides reassuring evidence that TZDs are
not associated with risk of overall malignancies. In fact,
they are compatible with the possibility of a decreased risk
of cancer. In site-specific subgroup analyses, for rosiglit-
azone, there was a significant decreased risk of bowel
cancer. Subgroup analyses for pioglitazone did not allow to
exclude an increased risk of bladder cancer, while the risk
of breast cancer was significantly decreased. While these
data are also useful to formulate not test hypotheses, they
provide somewhat more cogent evidence than the previ-
ously published epidemiological data.
Keywords Thiazolidinediones Cancer 
Type 2 diabetes
Introduction
Recent epidemiological data have raised concerns on the
long-term safety of pioglitazone with respect to malig-
nancies, showing an association with bladder cancer [1,2].
These results, which were in line with reports from some
large-scale randomized trials [3–6], prompted regulatory
Communicated by Antonio Secchi.
Electronic supplementary material The online version of this
article (doi:10.1007/s00592-013-0504-8) contains supplementary
material, which is available to authorized users.
M. Monami (&)
Geriatric Cardiology, Careggi Teaching Hospital, Florence, Italy
e-mail: [email protected]
I. Dicembrini
Obesity Agency, Careggi Teaching Hospital, Florence, Italy
E. Mannucci
Diabetes Agency, Careggi Teaching Hospital, Azienda
Ospedaliero-Universitaria Careggi, Via delle Oblate 4,
50141 Florence, Italy
e-mail: edoardo.mannucci@unifi.it
123
Acta Diabetol
DOI 10.1007/s00592-013-0504-8

authorities to modify the labels of pioglitazone-based
products; in some countries (i.e., France), the drug was
withdrawn from the market.
Little systematic information is available on the effect of
thiazolidinediones on other types of malignancies, different
from bladder cancer. Many experimental studies had
shown that thiazolidinediones, as a class, could inhibit
proliferation of some types of cancer cells in vitro and
in vivo [7,8], although pilot studies with these drugs as
anti-cancer therapy provided unsatisfactory results [9,10].
Epidemiological studies failed to detect any association of
thiazolidinedione use with the overall incidence of cancer
[11]. We have previously shown that rosiglitazone was not
associated with any increase in the overall incidence of
cancer in randomized trials available as of December 1,
2007 [12].
Interpretation of epidemiological data on the relation-
ship between treatments and adverse events is always
problematic, because patients who receive any specific
drug are different from those who are prescribed alternative
therapies. Although many confounders can be controlled
and adjusted for in analysis, a complete elimination of this
prescription bias is impossible. On the other hand, ran-
domized trials, which do not suffer from this limitation,
usually lack a sufficient sample size to verify the effects of
diabetes treatments on the incidence of cancer. Thiazolid-
inediones are the class of anti-hyperglycemic drugs, which
has been most widely studied, due to a relatively huge
number of longer-term available trials (either with meta-
bolic or with cardiovascular outcomes) [3,13–15]. Incident
malignancies reported as serious adverse events in those
studies can be a valuable source of information on the
effects of thiazolidinediones on cancer.
Materials and methods
Data sources and searches
An extensive MEDLINE and Embase search for ‘‘pioglit-
azone,’’ ‘‘rosiglitazone,’’ ‘‘troglitazone,’’ ‘‘rivoglitazone,’’
and ‘‘balaglitazone’’ was performed, collecting all ran-
domized clinical trials on humans up to August 1, 2011.
The identification of relevant abstracts, the selection of
studies based on the criteria described above, and the
subsequent data extraction were performed independently
by two of the authors (I.D. and M.M.), and conflicts
resolved by the third investigator (E.M.). Completed but
still unpublished trials were identified through a search of
www.clinicaltrials.gov Web site. FDA (www.fda.gov) and
European Medicines Agency (EMA, www.ema.europa.eu)
reviews of approved drugs, as well as published informa-
tion provided to FDA in response to queries during the
approval process, were also searched for retrieval of
unpublished trials.
Study selection
A meta-analysis was performed including all randomized
clinical trials with a duration of at least 52 weeks, enrolling
patients with or without diabetes, comparing thiazolidinedi-
ones with placebo or active drugs (oral hypoglycemic agents
and/or insulin) different from other thiazolidinediones. Trials
with a shorter duration were excluded, due to the fact that
they could not yield relevant information on cancer inci-
dence, which had been chosen as the principal outcome
variable. No review protocol was published elsewhere.
Data extraction and quality assessment
Results of unpublished trials were retrieved, if available, on
www.clinicaltrials.gov,www.clinicalstudyresults.org, FDA
(http://www.accessdata.fda.gov/scripts/cder/drugsatfda),
EMA (www.clinicaltrialsregister.eu), and GSK (http://
www.gsk-clinicalstudyregister.com) Web sites. All those
sources were also used to complete information on results of
published trials, when not reported in publications. A request
for further information was addressed to Takeda for trials on
pioglitazone with missing data, but the company declined.
For all published trials, results reported in papers were used
as the primary source of information, when available.
The quality of trials was assessed using some of the
parameters proposed by Jadad et al. [16]. In particular,
adequate reporting of randomization, allocation, blinding,
and patient flow, together with the description of intention-
to-treat safety analysis, were independently assessed by
two investigators (M.M. and I.D.); conflicts were resolved
by a third investigator (E.M.). The score was not used as a
criterion for the selection of trials, whereas some items
were used only for descriptive purposes.
Data synthesis and analysis
The principal outcome was the effect of thiazolidinediones,
compared either with placebo or with active drugs, on the
incidence of cancer. Secondary outcomes included all-
cause and non-cardiovascular mortality, and cancer-related
mortality.
Pre-defined separate analyses were performed for trials
with different comparators and for site-specific malignancies.
Heterogeneity was assessed by using I
2
statistics. If a
low heterogeneity was detected, both a random-effects
model and a fixed-effects model were applied. We report
the results of the random-effects models because the
validity of tests of heterogeneity can be limited with a
small number of component studies. To estimate possible
Acta Diabetol
123

publication/disclosure bias, we used funnel plots (Fig. 1)
and the Begg-adjusted rank correlation test [17,18],
including published and unpublished, but disclosed, trials.
However, because these tests have low statistical power
when the number of trials is small [19], undetected bias
may still be present. The Duval and Tweedie’s trim and fill
method was used for the estimation of results after cor-
rection for publication bias. Mantel–Haenszel odds ratio
with 95 % confidence interval (MH-OR) was calculated for
all the adverse events defined above, on an intention-to-
treat basis, excluding trials with zero events. A sensitivity
analysis was performed, for the main analysis and for the
predefined separate analyses, including also trials with zero
events, using continuity correction.
The meta-analysis was reported following the PRISMA
checklist [20]. All analyses were performed using Com-
prehensive Meta-analysis version 2, Biostat, (Englewood,
NJ, USA) and SPSS 18.1.
This research was performed independently of any fund-
ing, as part of the institutional activity of the investigators.
Results
The trial flow summary is reported in Fig. 2. Of the 72
identified trials that fulfilled the inclusion criteria, infor-
mation on cancer incidence was available in 46 studies. In
particular, the number of cases of malignancies was dis-
closed in 28 out of 33 trials with rosiglitazone, 18 out of 36
trials with pioglitazone; none of the three trials with trog-
litazone was disclosed. The characteristics of the studies
included in the analysis are summarized in Table 1. The
majority of the retrieved trials showed an adequate quality
for the parameters considered (Table 1). Among the studies
actually retrieved, 24 reported zero events, and therefore,
they were not included in the main analysis, but only in the
confirmatory analysis with continuity correction.
In the 46 trials reporting information on cancer, which
enrolled 14,666 and 13,970 on thiazolidinediones and
comparators, respectively, with a mean follow-up of
83.7 weeks, the number of incident cases was 342 and 457,
respectively; therefore, the observed yearly incidence of
cancer was 1.5 % with thiazolidinediones and 5.3 % with
comparators. The 22 trials reporting at least one event
enrolled 13,197 and 12,359 patients in thiazolidinedione
(pioglitazone: n=3,710 and rosiglitazone: n=9,487) and
control groups, respectively, with a mean follow-up of
26.1 months. Recorded cases of incident cancer were 342
and 457 in the thiazolidinedione and control groups,
respectively, with an yearly incidence of 0.95 and 1.12
cases/patient*year for drug and comparators, respectively.
I
2
for all thiazolidinediones, rosiglitazone, and pioglitazone
was 0.1, 0.1, and 28.4, respectively (all p[0.10). Overall,
thiazolidinediones were associated with a significant
reduction (MH-OR 0.85 [0.73–0.98]; p=0.027) in the
incidence of malignancies (Fig. 3); a separate analysis for
rosiglitazone (17 trials) yielded similar results (MH-OR
0.82 [0.69–0.98]; p=0.029), whereas results for pioglit-
azone were not statistically significant (MH-OR 0.66
[0.34–1.28]; p=0.22), due to the smaller number of
available disclosed trials (n=5). Similar figures were
obtained in the confirmatory analysis with continuity cor-
rection performed on 46 trials, including those with zero
events (MH-OR 0.86 [0.74–0.99], p=0.034; 0.83
[0.70–0.98], p=0.034; 0.92 [0.71–1.19], p=0.52, for all
thiazolidinediones, rosiglitazone, and pioglitazone,
respectively). Kendall’s tau was -0.34 (p=0.026), sug-
gesting the possibility of publication bias; this was con-
firmed by the visual analysis of funnel plot (Fig. 1). Using
the Duval and Tweedie’s trim and fill method, which
suggested the addition of 8 undisclosed trials, the estimated
MH-OR was 0.88 [0.76–1.02].
When different comparators were analyzed separately,
rosiglitazone was associated with a significantly lower
Standard Error
MH-Log Odds Ratio
Fig. 1 Funnel plot of standard
error by Mantel–Haenszel odds
ratio with 95 % confidence
interval (MH-OR, 95 % CI)
Acta Diabetol
123

incidence of malignancies in comparison with placebo (9
trials; MH-OR 0.58 [0.35–0.96]; p=0.036), whereas a
nonsignificant trend toward reduction was observed in
studies versus sulfonylureas (6 trials; 0.75 [0.55–1.02];
p=0.068). A lower number of disclosed trials (2 versus
placebo and 2 versus sulfonylureas) was available for
pioglitazone, with no significant result (MH-OR 0.75
[0.31–1.81], p=0.52 and 0.58 [0.15–2.24], p=0.43,
respectively). Similar results were obtained when per-
forming the same analyses using continuity correction
(data not shown).
Results on most frequent malignancies are summarized
in Fig. 4. In trials with available information with this
respect [3; see Electronic supplementary material for 26A,
27A, 34A], pioglitazone was associated with an increased
risk of bladder cancer, which did not reach statistical sig-
nificance; conversely, no apparent effect was detected for
rosiglitazone. Rosiglitazone, but not pioglitazone, was
associated with a significantly reduced risk of bowel can-
cer. Conversely, pioglitazone, but not rosiglitazone, was
associated with a significant reduction in breast cancer. No
other significant effect, either harmful or beneficial, was
Fig. 2 Trial flow summary
Acta Diabetol
123

Table 1 Moderators and outcome variables in individual studies included in the meta-analysis
Study (Reference)
a
Trial duration
(weeks)
Number of
patients (ID/C)
Age
(years)
Cancer mortality
(n, ID/C)
Malignancies
(n, ID/C)
Pioglitazone
Versus placebo
Non-diabetic patients
Ramachandran (1A) 156 204/203 45 0/0 NR/NR
Gavrila (2A) 52 7/7 48 0/0 0/0
Aithal (3A) 52 30/31 53 0/0 0/0
Han (4A) 52 40/43 39 0/0 0/0
Geldmaker (5A) 78 14/15 71 0/0 0/0
Sanyal (6A) 96 80/83 46 0/0 NR/NR
DeFronzo (7A) 125 303/299 52 2/0 3/8
Diabetic patients
Dormandy (8A) 156 2,605/2,633 62 NR/NR 97/99
Versus none
Diabetic patients
Jin (9A) 52 30/30 52 0/0 NR/NR
Kaku (10A) 123 294/294 58 0/0 NR/NR
Versus acarbose
Diabetic patients
Derosa (11A) 52 175/175 56 0/0 0/0
Derosa (12A) 108 175/175 56 0/0 NR/NR
Versus voglibose
Diabetic patients
Nakamura (13A) 52 15/15 57 0/0 0/0
Versus glibenclamide
Diabetic patients
Nakamura (13A)
b
52 15/15 57 0/0 0/0
Heliovaara (14A) 52 29/30 57 0/0 NR/NR
Tan (15A) 52 91/109 59 0/0 NR/NR
Giles (16A) 52 151/149 64 0/0 NR/NR
Jain (17A) 56 251/251 52 0/0 0/2
Tolman (18A) 144 1,051/1,046 54 0/0 NR/NR
Versus gliclazide
Diabetic patients
Perriello (19A) 52 146/137 58 0/0 NR/NR
Matthews (20A) 52 317/313 58 0/0 0/0
Versus glimepiride
Diabetic patients
Yamanouchi (21A) 52 38/37 55 0/0 0/0
Derosa (22A) 52 69/66 57 0/0 0/0
Tan 2004 (23A) 52 121/123 55 0/0 NR/NR
Mazzone (24A) 72 230/228 59 1/0 NR/NR
Nissen (25A) 78 270/273 60 0/0 3/4
Versus metformin
Diabetic patients
Yamanouchi (26A) 52 38/39 55 0/0 0/0
Derosa (27A) 52 69/67 54 0/0 0/0
Versus others
Diabetic patients
Acta Diabetol
123