SPECIAL COMMUNICATION

Composite Outcomes in Randomized Trials

Greater Precision But With Greater Uncertainty?
Nick Freemantle, PhD
Composite outcomes, in which multiple end points are combined, are fre-
Melanie Calvert, PhD
quently used as primary outcome measures in randomized trials and are of-
John Wood, MSc
ten associated with increased statistical efficiency. However, such mea-
Joanne Eastaugh, PhD sures may prove challenging for the interpretation of results. In this article,
Carl Griffin, MSc we examine the use of composite outcomes in major clinical trials, assess

R
ANDOMIZED CONTROLLED the arguments for and against them, and provide guidance on their appli-
trials are central to the cation and reporting. To assess incidence and quality of reporting, we sys-
evaluation of pharmaceuti- tematically reviewed the use of composite end points in clinical trials in An-
cals. They are used to pro- nals of Internal Medicine, BMJ, Circulation, Clinical Infectious Diseases,
vide evidence for the efficacy of phar- Journal of the American College of Cardiology, JAMA, Lancet, New En-
maceuticals in the licensing process gland Journal of Medicine, and Stroke from 1997 through 2001 using a
that aims to ensure that drugs only
sensitive search strategy. We selected for review 167 original reports of ran-
become available for widespread pre-
scription if they have a positive domized trials (with a total of 300276 patients) that included a composite
impact on symptoms, prognosis, or primary outcome that incorporated all-cause mortality. Sixty-three trials (38%)
both. Furthermore, they play a major were neutral both for the primary end point and the mortality component.
role in the evaluation of the effective- Sixty trials (36%) reported significant results for the primary outcome mea-
ness of treatments, providing the evi- sure but not for the mortality component. Only 6 trials (4%) were signifi-
dence base on which, for example, cant for the mortality component but not for the primary composite out-
decisions on the inclusion of treat-
come, whereas 19 trials (11%) were significant for both. Twenty-two trials
ments in clinical guidelines are
made.1 (13%) were inadequately reported. Our review suggests that reporting of
The choice of outcomes measured composite outcomes is generally inadequate, implying that the results ap-
(the outcome variables or end points) ply to the individual components of the composite outcome rather than only
in clinical trials is an important design to the overall composite. Current guidelines for the undertaking and report-
consideration. The primary outcome ing of clinical trials could be revised to reflect the common use of compos-
in particular has much invested in it, ite outcomes in clinical trials.
because it is normally the outcome JAMA. 2003;289:2554-2559 www.jama.com
alone that indicates whether or not
the trial provides evidence at an
acceptable level that the treatment is ing on those that include mortality. We COMPOSITE OUTCOMES
efficacious.2 assess the arguments for and against IN CLINICAL TRIALS
Trials that examine treatments that them and provide guidance on their ap- To assess the extent and nature of com-
are expected to have an effect on mor- plication and reporting. posite outcomes, we searched 9 jour-
tality and major morbidity often adopt
Author Affiliations: Department of Primary Care and and Medical Research Council. Dr Calvert has re-
a primary composite outcome mea- General Practice, University of Birmingham, Birming- ceived funding for research from Medtronic Inc and
sure that includes mortality along with ham, England. Servier Laboratories. Mr Wood undertook this work in
Financial Disclosures: Dr Freemantle has received fund- his role as visiting senior research fellow, University of
other nonfatal end points. This article ing for research from many of the companies that spon- Birmingham, and is employed as a statistician by No-
examines the use of composite out- sored trials included in this review; he has also re- vatis Pharma. This study received no external funding.
comes in major clinical trials, focus- ceived funding from the Department of Health in Corresponding Author and Reprints: Nick Freemantle,
England, the UK Medical Research Council, and other PhD, Department of Primary Care and General Prac-
medical charities. Dr Eastaugh has received funding for tice, Primary Care Clinical Sciences Bldg, University of
See also pp 2545 and 2575. research from Servier Laboratories, Medtronic Inc, Orion Birmingham, Edgbaston, Birmingham B15 2TT, En-
Pharmaceuticals, and the English Department of Health gland (e-mail: [email protected]).

2554 JAMA, May 21, 2003—Vol 289, No. 19 (Reprinted) ©2003 American Medical Association. All rights reserved.

Downloaded from jama.ama-assn.org at University of California - Los Angeles DCS on February 6, 2011
 COMPOSITE OUTCOMES IN RANDOMIZED TRIALS

nals, Annals of Internal Medicine, BMJ, thors (M.C., J.E., N.F.). TABLE 3 de- and analysis of clinical trials,2 this may
Circulation, Clinical Infectious Diseases, scribes the assessment of each trial by not be the only motivation behind the
Journal of the American College of Cardi- journal of publication. popularity of composite outcome mea-
ology, JAMA, Lancet, New England Jour- The finding that such a high propor- sures. Instead, issues of statistical effi-
nal of Medicine, and Stroke, to identify tion of trials that measure composite ciency appear to be prominent, with
trials that use composite outcomes, in- outcomes, including mortality, pro- composite outcomes in time to event
cluding mortality and major morbidity, vide neutral results on the primary out- trials leading to higher event rates and
for the 5 years from 1997 through 2001. come may be unsurprising. However, thus enabling smaller sample sizes or
We identified randomized trials using the finding that a similar proportion are shorter follow-up (or both). Thus, to
hand searching and a sensitive elec- positive yet fail independently to iden- have power (1−!) of .9 of finding a haz-
tronic search strategy within journals, tify an effect on the mortality compo- ard ratio (relative risk) of 0.65 signifi-
searching on the words death, survival, nent is striking and requires further cant at the .05 level requires approxi-
and mortality as text words mentioned consideration. mately 1400 patients with a 20% control
in the title or abstract. Altogether we event rate but only 700 with a 40% con-
identified 167 randomized trials pub- RATIONALE FOR trol event rate.4 Of course, the increase
lished during this period, which to- COMPOSITE OUTCOMES in power identified here depends on a
gether randomized 300276 patients. The Randomized clinical trials can provide common hazard ratio; this may or may
most common interventions across the unbiased estimates of treatment effect. not be the case.
journals we included were in the area of To achieve this, it is essential that trials
cardiovascular disease, and the most follow a predetermined protocol de- PROBLEMS WITH
common single intervention during the scribing the population to be studied, an COMPOSITE OUTCOMES
period involved the use of stents. The appropriate method for random alloca- A substantive risk associated with the re-
therapeutic areas of the included trials tion, the procedure for follow-up, the porting of composite outcomes is that the
are described in TABLE 1. outcomes to be analyzed, and the sta- benefits described may be presumed to
The frequency of the use of compos- tistical methods that will be used. relate to all of the components. For ex-
ite primary outcomes that included If there is no obvious choice of pri- ample, Le May et al5 state in the ab-
mortality during the 5-year period was mary outcome in a trial, then it is pos-
fairly consistent. The New England Jour- sible to adopt the composite of several Table 1. Therapeutic Areas of Included
nal of Medicine published the most, with outcomes. This situation was described Trials
58 during the period, whereas Clinical by the International Conference on Har- No. (%)
Infectious Diseases published only 1. The monisation of Technical Requirements Clinical Area of Trials
number of trials per journal per year is for Registration of Pharmaceuticals for Cardiovascular trials
Nonstenting 63 (38)
described in TABLE 2. Human Use (ICH)3 as follows: Stenting 20 (12)
In 63 trials (38%) neither the com- If a single primary variable cannot be se- Stroke 24 (14)
posite outcome nor the mortality com- Oncology 15 (9)
lected from multiple measurements asso- Neonatal medicine 11 (7)
ponent was statistically significant at the ciated with the primary objective, another Human immunodeficiency virus 10 (6)
conventional P=.05 level. In 60 trials useful strategy is to integrate or combine the Chronic obstructive pulmonary 3 (2)
multiple measurements into a single or disease
(36%) the overall composite outcome Diabetes 2 (1)
‘composite’ variable, using a predefined al-
was significant, but the mortality com- Leukemia 2 (1)
gorithm. . . . This approach addresses the Neutropenia 2 (1)
ponent was not significant. In only 19 multiplicity problem without requiring ad- Pediatrics 2 (1)
trials (11%) were both the primary com- justment to the type 1 error. Peripheral vascular disease 2 (1)
posite outcome and its mortality com- Variceal bleeding 2 (1)
Although dealing with multiple test- Other 9 (5)
ponent significant. In 6 trials (4%) the ing is an important factor in the design Total 167 (100)
composite primary outcome was not sta-
tistically significant, whereas the mor-
tality component was. Nine trials (5%) Table 2. Number of Included Trials per Journal per Year
reported multiple composite primary Year Annals BMJ Circulation CID JACC JAMA Lancet NEJM Stroke Total
outcomes. Twenty-two trials (13%) did 2001 0 1 3 0 5 5 10 14 2 40
not report the statistical significance of 2000 1 1 8 0 3 6 9 5 3 36
the mortality component separately, al- 1999 1 0 3 0 1 0 9 12 2 28
though P values could be calculated for 1998 0 0 3 1 4 1 8 12 2 31
12 of these trials. The results of the prin- 1997 1 1 4 0 3 1 7 15 0 32
cipal analyses of the remaining 10 trials Total 3 3 21 1 16 13 43 58 9 167
could not be interpreted, despite being Abbreviations: Annals, Annals of Internal Medicine; CID, Clinical Infectious Diseases; JACC, Journal of the American
College of Cardiology; NEJM, New England Journal of Medicine.
assessed independently by 3 of the au-
©2003 American Medical Association. All rights reserved. (Reprinted) JAMA, May 21, 2003—Vol 289, No. 19 2555

Downloaded from jama.ama-assn.org at University of California - Los Angeles DCS on February 6, 2011
COMPOSITE OUTCOMES IN RANDOMIZED TRIALS

Table 3. Significance of Primary Composite Outcome and Mortality Component in Included Trials by Journal
Composite or Mortality Annals BMJ Circulation CID JACC JAMA Lancet NEJM Stroke Total No. (%)
Nonsignificant primary, nonsignificant mortality 1 1 11 0 8 5 15 17 5 63 (38)
Nonsignificant primary, significant mortality 0 0 1 0 0 1 2 2 0 6 (4)
Significant primary, nonsignificant mortality 1 1 7 0 7 5 14 23 2 60 (36)
Significant primary, significant mortality 0 0 1 0 1 0 8 9 0 19 (11)
Multiple primary outcomes 0 1 1 0 0 0 1 6 0 9 (5)
Not clear 1 0 0 1 0 2 3 1 2 10 (6)
Total 3 3 21 1 16 13 43 58 9 167 (100)
Abbreviations: Annals, Annals of Internal Medicine; CID, Clinical Infectious Diseases; JACC, Journal of the American College of Cardiology; NEJM, New England Journal of
Medicine.

stract to their article, “Compared with ac- Post-infaRct survIval COntRol in LV new and more stringent level deter-
celerated t-PA [tissue-type plasminogen dysfunctioN] trial investigated the ef- mined by the revised policy for allocat-
activator], primary stenting reduces fects of carvedilol, a !-blocker, in 1959 ing statistical power in the trial. Al-
death, reinfarction, stroke or repeat TVR patients with left ventricular dysfunc- though 12% of patients died in the
[target vessel revascularization] for is- tion following myocardial infarction.7 carvedilol group, compared with 15% in
chemia.” In fact, only 5 deaths contrib- Originally, the primary outcome iden- the placebo group, 23% of patients in the
uted to the primary end point, 3 in the tified in the trial protocol was all-cause carvedilol group and 22% of patients in
stenting group and 2 in the t-PA group. mortality. However, while the study was the placebo group qualified for the com-
Elsewhere in the abstract, the authors ongoing, the data and safety monitor- posite outcome on the basis of hospi-
make it clear that there is no observed ing board (which has the key role of pro- talizations alone, a result that under-
benefit on mortality, and the results for tecting the interests of patients in the mined the relatively modest numerical
the components of the primary out- trial)8 noted that the overall rate of ac- reduction in mortality in the carvedilol
come are given together in a table. How- crual of deaths was lower than that pre- group. Thus, strictly, CAPRICORN7 pro-
ever, the statement describing the pri- dicted. The board communicated its vides a neutral result, although ironi-
mary outcome may be considered concerns to the trial steering commit- cally CAPRICORN would have been
misleading, especially if quoted out of tee that the trial would not be powered modestly statistically significant had the
context. A more appropriate descrip- to identify the primary end point as original primary outcome of all-cause
tion of this form of primary outcome significant. The trial steering commit- mortality been maintained.
could be “major event-free survival,” al- tee then took the unusual step of chang- This example demonstrates the pos-
beit that descriptor, albeit technically cor- ing the primary outcome, dividing sibility that, by using composite out-
rect, might also appear to suggest a re- the available statistical power between comes, the measure of treatment effect
duction in mortality. a new composite outcome (all-cause can be diluted by an outcome that ex-
Some authors manage to communi- mortality or cardiovascular hospital hibits no effect being combined with a
cate clearly the findings of their stud- admissions), which the trial steering more critical measure that individually
ies using composite primary outcome committee awarded a critical P value of shows some evidence of benefit. It also
measures. For example, Leon et al,6 de- .045, and the original primary out- demonstrates some of the arbitrariness
scribing the results of a trial of local- come (all-cause mortality), which was of clinical trials; for example, the same
ized intracoronary "-radiation therapy, awarded a prespecified critical P value trial may be considered positive or neu-
state that the composite primary out- of .005 to achieve statistical signifi- tral on the basis of decisions concern-
come was statistically significant. Yet, cance. Thus, if the P value for either pri- ing the statistical design. Fortunately, for
they then state, “However, the reduc- mary outcome achieved statistical sig- those interpreting the results of trials, it
tion in the incidence of major adverse nificance at the critical level, the study is unusual for only 1 trial to be avail-
cardiac events was determined solely by would be deemed positive. able. Although on its own CAPRICORN7
a diminished need for revasculariza- The original primary end point (all- is neutral, it should be interpreted in the
tion of the target lesion, not by reduc- cause mortality) achieved a P value of context of data from all relevant clini-
tions in the incidence of death or myo- .03 (ie, substantially larger than the .005 cal trials, highlighting an important role
cardial infarction.” allocated to it), whereas the alternative for systematic review and quantitative
As our review establishes, the use of primary outcome (composite of all- meta-analysis.
a composite outcome does not always cause mortality and cardiovascular hos- Assessing the benefits of treatment
lead to an increase in the evidence for pital admissions) had a P value of .30. across relevant available trials can high-
benefit of an intervention. For ex- Thus, the original primary outcome did light further challenges for the inter-
ample, the CAPRICORN [CArvedilol not achieve statistical significance at the pretation of composite outcomes. Gly-
2556 JAMA, May 21, 2003—Vol 289, No. 19 (Reprinted) ©2003 American Medical Association. All rights reserved.

Downloaded from jama.ama-assn.org at University of California - Los Angeles DCS on February 6, 2011
 COMPOSITE OUTCOMES IN RANDOMIZED TRIALS

Figure. Glycoprotein IIb/IIIa Inhibitors in the Management of Acute Coronary Syndromes

A Primary Composite Outcome B All-Cause Mortality Component of Primary

Composite Outcome
Favors Favors Favors Favors
Treatment Control Treatment Control
No. of No. of
Source Composite Outcome Events Source Events

PRISM, 1998 Death, MI, or Refractory Ischemia Within 48 h 151 PRISM, 1998 9

PRISM PLUS, 1998 Death, MI, or Refractory Ischemia Within 7 d 98 PRISM PLUS, 1998 9

PURSUIT, 1998 Death or Nonfatal MI at 30 d 1414 PURSUIT, 1998 340

PARAGON, 1998 Death or Nonfatal MI at 30 d 261 PARAGON, 1998 73

Pooled 1924 Pooled 431

0.2 0.5 1 2 0.2 0.5 1 2 5 10 100

Odds Ratio Odds Ratio
(95% Confidence Interval) (95% Confidence Interval)

Results of pivotal trials for primary composite outcomes and for all-cause mortality alone as a component of the primary composite outcome. Error bars indicate 95%
confidence intervals. MI indicates myocardial infarction; PRISM, the Platelet Receptor Inhibition in Ischemic Syndrome Management trial; PRISM-PLUS, the Platelet
Receptor Inhibition in Ischemic Syndrome Management in Patients Limited by Unstable Signs and Symptoms study; PURSUIT, Platelet Glycoprotein IIb/IIIa in Unstable
Angina: Receptor Suppression Using Integrilin Therapy trial; and PARAGON, Platelet IIb/IIIa Antagonist for the Reduction of Acute coronary syndrome events in a
Global Organization Network.

coprotein IIb/IIIa inhibitors are used Although all-cause mortality is in- this basis to claim efficacy of a treat-
increasingly as medical therapy in the cluded in the composite outcome mea- ment on the individual components of
management of acute coronary syn- sure for every trial, the evidence that a composite primary outcome is mis-
dromes. The pivotal trials used for li- glycoprotein IIb/IIIa inhibitors are ef- leading. For example, the abstract de-
censing these agents used composite fective in reducing mortality is lack- scribing the results of the TARGET [Do
primary outcomes, including death, ing. We pooled the all-cause mortality Tirofiban and ReoPro Give Similar Ef-
nonfatal myocardial infarction, and, oc- component in each of the trials, and the ficacy Outcomes] trial14 states, “The
casionally, refractory ischemia. The US results are shown in the Figure. For the magnitude and the direction of the effect
product insert for tirofiban hydrochlo- 4 trials, the exact OR for the effect of were similar for each component of the
ride, one of the glycoprotein IIb/IIIa in- glycoprotein IIb/IIIa inhibitors is 1.00 composite end point (hazard ratio for
hibitors, states that: (95% CI, 0.82-1.22). Thus, although the death, 1.21; hazard ratio for myocar-
AGGRASTAT [tirofiban], in combination results do not exclude a benefit for gly- dial infarction, 1.27; and hazard ratio for
with heparin, is indicated for the treat- coprotein IIb/IIIa inhibitors on mor- urgent target-vessel revascularization,
ment of acute coronary syndrome, includ- tality, our best estimate of the effect is 1.26.” However, the estimate for mor-
ing patients who are to be managed zero, and the plausible range runs from tality is based on approximately 20
medically and those undergoing PTCA [per-
cutaneous transluminal coronary angio-
an 18% reduction in the odds of death deaths and the 95% CI stretches from
plasty] or atherectomy. In this setting, to a 22% increase in the odds of death. 0.57 to 2.8 (P =.66), indicating that the
AGGRASTAT has been shown to decrease Investigation into the use of stents play of chance is a highly plausible ex-
the rate of a combined endpoint of death, was the most common intervention in planation for the result.
new myocardial infarction or refractory is- the trials we identified, with 20 (12%) We hypothesized that the type of out-
chemia/repeat cardiac procedure.9
overall. Twenty of the 24 predefined come included might influence the sta-
Reviewing the evidence for the medi- composite primary outcome measures tistical significance of the composite pri-
cal use of glycoprotein IIb/IIIa inhibi- (83%), including a mortality compo- mary outcome, specifically where
tors for acute coronary syndromes, Bhatt nent, described in these trials were sta- components of the outcome were de-
and Topol10 identified 4 trials that they tistically significant. However, only 2 termined at least in part by the actions
state provide evidence for their use. The (8%) of 24 were significant for the mor- of clinicians rather than describing di-
results of these trials, on the primary out- tality component of the analysis, and rectly the disease. Thus, we consid-
come identified by the protocols for each in both cases the comparison was with ered that outcomes such as hospital-
trial, are shown in the FIGURE. The angioplasty, which was associated with ization or revascularization might be
(theoretically) exact pooled odds ratio increased hazard of death.12,13 more amenable to change than a dis-
(OR)11 for the effects of treatment at pri- Simply identifying that the compo- ease outcome such as nonfatal myocar-
mary outcome is 0.86 (95% confidence nents of a primary outcome are consis- dial infarction.
interval [CI], 0.78-0.95), a benefit in fa- tent without considering statistical pre- Overall, 78 of 179 comparisons (in-
vor of the treatment. cision is insufficient, and indeed using cluding 20 primary outcomes from
©2003 American Medical Association. All rights reserved. (Reprinted) JAMA, May 21, 2003—Vol 289, No. 19 2557

Downloaded from jama.ama-assn.org at University of California - Los Angeles DCS on February 6, 2011
COMPOSITE OUTCOMES IN RANDOMIZED TRIALS

studies with multiple comparisons) in- vestigators who are having difficulty in ite primary outcome measure of “time
cluded the following clinician-driven deciding which outcome to elect as the to treatment failure” was used.17 The
outcomes: revascularization, percuta primary outcome measure in a trial and composite was defined as “death; liver
neous mitral valvuloplasty, mechani- deal with the issue of multiplicity in an transplantation; histological progres-
cal ventilation, hospitalization, trans- efficient manner, avoiding the need for sion by two stages (of four) or progres-
plantation (cardiac and liver), shunt, arbitrary choices. sion to cirrhosis; the development of
use of rescue therapy, institutionaliza- The disadvantages of composite out- varices, ascites, or encephalopathy; sus-
tion, initiation of new antibiotics, use comes may arise when the constituents tained quadrupling of the serum biliru-
of shock therapy, amputation, extra cor- do not move in line with each other. This bin concentration; marked worsening of
poreal membrane oxygenation, and di- appears most starkly when there is a fatigue or pruritus; inability to tolerate
alysis. We used a nonlinear mixed principal end point (often all-cause mor- the drug; or voluntary withdrawal from
model15 to estimate the effects of clini- tality) supported by an additional and the study.”
cian-driven outcomes on the likeli- more common end point or end points. The more clearly a composite refers
hood of finding a statistically signifi- This disadvantage can happen in 3 ways. to an overall disease process, the less
cant result, accounting for journal as a First, as in CAPRICORN,7 if an impor- there is any problem of interpretation.
random effect. The inclusion of a cli- tant component of the composite out- The more likely it is too that the com-
nician-driven outcome was predictive come is not substantially modified by the ponents of the composite will move in
of a statistically significant result for the effects of treatment, then less rather than the same direction given an effective
primary composite outcome (OR, 2.24; more statistical power to detect effects treatment. However, this is not guar-
95% CI, 1.15-4.34; P= .02). on the principal end point may be the anteed, as seen in CAPRICORN,7 the
result. Second, although the composite glycoprotein IIb/IIIa inhibitor trials,10
COMMENT as a whole may appear to be affected by and coronary stenting trials, all of which
There are 2 main advantages in using treatment, the evidence for benefit re- used composite end points that were,
composite outcomes. First, the correct lating to its most important constitu- presumably, a priori plausible.
(a priori) identification of a composite ent may not exist or lack persuasive- It is particularly in the dissection of
primary end point can increase the sta- ness within the trial. This is the case with a composite that the differing inter-
tistical precision and thus the efficiency the trials of glycoprotein IIb/IIIa inhibi- ests of sponsors, licensing authorities,
of a trial. When the aim of a trial is to tors described herein. Third, when cli- and interpreters become manifest.
provide proof of the efficacy of a treat- nician-driven outcomes, such as revas- Sponsors are attracted by the hoped-
ment, a composite outcome can prove cularization or hospitalization, are used, for gains in statistical precision to im-
helpful, enabling a positive result from these appear generally to be more ame- prove their chances of a positive trial,
a smaller number of patients random- nable to change, presenting further chal- and the increase in the use of compos-
ized. Thus, more trials may be con- lenges for interpretation. ite outcome measures is largely due to
ducted for a given investment in re- We have demonstrated that compos- their acceptance by licensing authori-
search, and results that provide guidance ite outcomes are commonly used in trials ties. The latter might argue that their
on the efficacy, or otherwise, of inter- of mortality and major morbidity and role is to ensure that licensed pharma-
ventions may be available more quickly. identified several potential challenges in ceuticals have an effect on an underly-
Trials are costly to society and ideally the interpretation of their results. A fun- ing disease process or symptoms rather
they should provide reliable estimates of damental issue for composite out- than specifically on components of a
treatment effects in a timely and cost- comes, correctly identified by the ICH composite primary outcome measure.
effective manner. Although some trial- harmonised tripartite guideline (ICH Licensing authorities, however, have a
ists have advocated the use of large trials, E9),3 is that the outcomes that contrib- role in the manner in which sponsors
these may also prove problematic. An ex- ute to a composite outcome must be “as- are permitted to market their prod-
ample is the Heart Outcomes Preven- sociated with the primary objective.” ucts (the indication being agreed as a
tion Evaluation trial, which estimated the When this relationship is in question, the result of the evidence submitted by the
effects of ramipril in patients who were interpretation of composite outcomes is sponsor). It is unsatisfactory for mar-
at high risk but who “did not have left also in question. This requirement can keting claims to be made purely on the
ventricular dysfunction or heart fail- be translated as being that the compo- grounds of composite end points. In-
ure.”16 Because of the scale of this trial nents should be measurable events that deed, the product insert for tirofiban9
(9297 patients), measurement of left ven- can sensibly be added together as being in the United States contains some in-
tricular function was only available for aspects of the same underlying disease formation on the constituents of the pri-
a minority of patients, undermining to process, and it helps if the composite can mary outcome measure used in the
some extent the interpretation of the be given a single name. For example, in trials that support licensing, although
findings. Second, as the ICH identi- a trial of ursodiol for primary scleros- it is possible that this may be lost (lit-
fied,3 a composite outcome may help in- ing cholangitis, a predefined compos- erally) in the small print.
2558 JAMA, May 21, 2003—Vol 289, No. 19 (Reprinted) ©2003 American Medical Association. All rights reserved.

Downloaded from jama.ama-assn.org at University of California - Los Angeles DCS on February 6, 2011
 COMPOSITE OUTCOMES IN RANDOMIZED TRIALS

The licensing of pharmaceuticals is

necessarily a game of fixed rules, and Box. Recommendations
the judicial use of a composite out- 1. Trialists should follow the CONSORT guidelines18 and identify precisely the
come may serve this aim. However, the prespecified primary and secondary outcome measures, reporting the results
interpretation of the results of trials for clearly in publications that describe the trial.
other purposes should not allow rules 2. When trials report composite variables as primary outcomes, these should be
to dominate well-informed good sense. interpreted together rather than as demonstrating efficacy of individual com-
A substantial number of major trials use ponents of the composite.
composite primary outcomes that in- 3. Components of composite outcomes should always be defined as secondary out-
clude all-cause mortality. The ICH E93 comes and reported alongside the results of the primary analysis, preferably in
states, “When a composite variable is a table.
used as a primary variable, the compo- 4. Authors and journal editors should ensure that the reporting of composite out-
nents of this variable may sometimes comes is clear and avoids the suggestion that individual components of the com-
be analysed separately.” To avoid the posite have been demonstrated to be effective.
burying of important components of 5. Systematic overviews and quantitative meta-analysis should be used to iden-
composite primary outcomes for which tify the effects of treatments on rare but important end points that may be in-
on their own no effect is observed, such cluded as part of composite outcomes in individual trials.
as arguably may have been the case with
the glycoprotein IIb/IIIa inhibitors de-
scribed herein, the components of a form the treatment of patients with strategy of systematic overview and
composite primary outcome should al- similar disease in the future. How- quantitative summary of the results of
ways be described and analyzed indi- ever, in more than 1 in 10 trials in our clinical trials to provide the best pos-
vidually. Thus, the components of a review the reporting of data was inad- sible estimates on the effects on impor-
composite outcome should always be equate. Recommendations on the ap- tant outcomes.
declared as secondary outcomes, and propriate use of composite outcomes
their results described alongside the re- are described in the BOX. Author Contributions: Study concept and design:
Freemantle.
sult for the composite outcome. In ad- When describing the results of clini- Acquisition of data: Freemantle, Eastaugh, Calvert,
dition, ICH E9 should be revised to ad- cal trials, authors and journal editors Griffin.
dress this point. The current version of could ensure that attempts are made to Analysis and interpretation of data: Freemantle, Wood,
Eastaugh, Calvert.
the CONSORT recommendations 18 place the findings in the context of other Drafting of the manuscript: Freemantle, Calvert.
makes no comment on composite out- available relevant research, as re- Critical revision of the manuscript for important in-
tellectual content: Freemantle, Wood, Eastaugh,
comes and could usefully be revised ac- quired by the CONSORT recommen- Calvert, Griffin.
cordingly. Patients often agree to par- dations.18 The use of composite pri- Statistical expertise: Freemantle, Wood.
Administrative, technical, or material support: Calvert,
ticipate in randomized trials because mary end points in clinical trials Griffin.
they believe their experience will in- provides considerable support for the Study supervision: Freemantle.

REFERENCES
1. Eccles M, Freemantle N, Mason JM. Methods of de- 7. The CAPRICORN Investigators. Effects of carve- year follow-up results. J Am Coll Cardiol. 2001;37:
veloping guidelines for efficient drug use in primary care: dilol on outcome after myocardial infarction in 51-58.
North of England Evidence Based Guidelines Develop- patients with left ventricular dysfunction: the CAPRI- 14. Topol EJ, Moliterno DJ, Herrmann HC, et al. Com-
ment Project. BMJ. 1998;316:1232-1235. CORN randomised trial. Lancet. 2001;357:1385- parison of two platelet glycoprotein IIb/IIIa inhibitors,
2. Freemantle N. Interpreting the results of second- 1390. tirofiban and abciximab, for the prevention of ischemic
ary endpoints and subgroup analyses in clinical trials: 8. Hampton JR. Clinical trial safety committees: the events with percutaneous coronary revascularization. N
should we lock the crazy aunt in the attic? BMJ. 2001; devil’s spoon. BMJ. 2000;320:244-245. Engl J Med. 2001;344:1888-1894.
322:989-991. 9. Aggrastat Injection, Aggrastat Injection Premixed 15. SAS [computer program]. Version 8.1. Cary, NC:
3. International Conference on Harmonisation of Tech- [package insert]. West Point, Pa: Merck; 1998. SAS Institute Inc; 2002 update.
nical Requirements for Registration of Pharmaceuti- 10. Bhatt DL, Topol EJ. Current role of platelet gly- 16. The Heart Outcomes Prevention Evaluation Study
cals for Human Use. ICH harmonised tripartite guide- coprotein IIb/IIIa inhibitors in acute coronary syn- Investigators. Effects of an angiotensin-converting-
line: statistical principles for clinical trials. Stat Med. dromes. JAMA. 2000;284:1549-1558. enzyme inhibitor, ramipril, on cardiovascular events
1999;18:1905-1942. 11. Martin DO, Austin H. An exact method for meta- in high risk patients. N Engl J Med. 2000;342:145-
4. Shih JH. Sample size calculation for complex clini- analysis of case-control and follow-up studies. Epi- 153.
cal trials with survival endpoints. Control Clin Trials. demiology. 2000;11:255-260. 17. Lindor KD, for the Mayo Primary Sclerosing Chol-
1995;16:395-407. 12. Lincoff AM, Califf RM, Moliterno DJ, et al. angitis-Ursodeoxycholic Acid Study Group. Ursodiol
5. Le May MR, Labinaz M, Davies RF, et al. Stenting Complementary clinical benefits of coronary-artery for primary sclerosing cholangitis. N Engl J Med. 1997;
versus thrombolysis in acute myocardial infarction trial stenting and blockade of platelet glycoprotein IIb/ 336:691-695.
(STAT). J Am Coll Cardiol. 2001;37:985-991. IIIa receptors. N Engl J Med. 1999;341:319-327. 18. Moher D, Schulz KF, Altman D, for the CON-
6. Leon MB, Teirstein PS, Moses JW, et al. Localized 13. Rodriguez A, Bernardi V, Navia J, et al. Argen- SORT Group. The CONSORT statement: revised rec-
intracoronary gamma-radiation therapy to inhibit the tine randomized study: coronary angioplasty with stent- ommendations for improving the quality of reports of
recurrence of restenosis after stenting. N Engl J Med. ing versus coronary bypass surgery in patients with parallel-group randomized trials. JAMA. 2001;285:
2001;344:250-256. multiple-vessel disease (ERACI II): 30-day and one- 1987-1991.

©2003 American Medical Association. All rights reserved. (Reprinted) JAMA, May 21, 2003—Vol 289, No. 19 2559

Downloaded from jama.ama-assn.org at University of California - Los Angeles DCS on February 6, 2011