The NEW ENGLAND JOURNAL of MEDICINE
`
`SPECIAL REPORT
`
`Statistics in Medicine — Reporting of Su bgroup
`Analyses in Clinical Trials
`Rui Wang, M.S., Stephen W. Lagakos, Ph.D.,James H. Ware, Ph.D., David]. Hunter, M.S., B.S.,
`andJeffrey M. Drazen, MD.
`
`Medical research relies on clinical trials to as-
`
`sess therapeutic benefits. Because of the effort
`and cost involved in these studies, investigators
`frequently use analyses of subgroups of study
`participants to extract as much information as
`possible. Such analyses, which assess the heter-
`ogeneity of treatment effects in subgroups of pa-
`tients, may provide usefiil information for the care
`of patients and for future research. However, sub-
`group analyses also introduce analytic challeng-
`es and can lead to overstated and misleading
`results.“7 This report outlines the challenges as-
`sociated with conducting and reporting subgroup
`analyses, and it sets forth guidelines for their use
`in the journal. Although this report focuses on the
`reporting of clinical trials, many of the issues dis-
`cussed also apply to observational studies.
`
`SUBGROUP ANALYSES
`AND RELATED CONCEPTS
`
`
`SUBGROUP ANALYSIS
`
`By “subgroup analysis,” we mean any evaluation
`of treatment effects for a specific end point in sub-
`groups of patients defined by baseline character-
`istics. The end point may be a measure of treat-
`ment efficacy or safety. For a given end point, the
`treatment effect — a comparison between the
`treatment groups — is typically measured by a
`relative risk, odds ratio, or arithmetic difference.
`The research question usually posed is this: Do the
`treatment effects vary among the levels of a base-
`line factor?
`
`A subgroup analysis is sometimes undertaken
`to assess treatment effects for a specific patient
`characteristic; this assessment is often listed as
`a primary or secondary study objective. For exam-
`ple, Sacks et al.8 conducted a placebo-controlled
`trial in which the reduction in the incidence of
`
`coronary events with the use of pravastatin was
`examined in a diverse population of persons who
`had survived a myocardial infarction. In sub-
`group analyses, the invesn'gators further examined
`whether the efficacy of pravastatin relative to pla-
`cebo in preventing coronary events varied accord-
`ing to the patients baseline low-density lipopro-
`tein (LDL) levels.
`Subgroup analyses are also undertaken to in-
`vestigate the consistency of the trial conclusions
`among different subpopulations defined by each
`of multiple baseline characteristics of the patients.
`For example, Jackson et al.9 reported the outcomes
`of a study in which 36,282 postmenopausal
`women 50 to 79 years of age were randomly as-
`signed to receive 1000 mg of elemental calcium
`with 400 IU of vitamin D3 daily or placebo. Frac-
`tures, the primary outcome, were ascertained over
`an average follow-up period of 7.0 years; bone den-
`sity was a secondary outcome. Overall, no treat-
`ment effect was found for the primary outcome;
`that is, the active treatment was not shown to pre-
`vent fractures. The effect of calcium plus vitamin
`D supplementation relative to placebo on the risk
`ofeach of four fracture outcomes was further ana-
`
`lyzed for consistency in subgroups defined by 15
`characteristics of the participants.
`
`HETEROGENEITY AND STATISTICAL INTERACTIONS
`
`The heterogeneity of treatment effects across the
`levels of a baseline variable refers to the circum-
`
`stance in which the treatment effects vary across
`the levels of the baseline characteristic. Heteroge-
`neity is sometimes further classified as being ei-
`ther quantitative or qualitative. In the first case,
`one treatment is always better than the other, but
`by various degrees, whereas in the second case,
`one treatment is better than the other for one sub-
`
`group of patients and worse than the other for
`
`Biogen Exhibit 2071
`Mylan v. Biogen
`” ENG” “”35””
`IPR 2018-01403
`The New England Journal ofMedicine
`oggvnloaded from nejmorg on April 16, 2019. For personal use only. No other uses without permission.
`Copyright © 2007 Massachusetts Medical Society. All rights reserved.
`
`NOVEMBER 22, 2007
`
`WWW.NEJM.ORC.
`
`P
`
`age
`
`1
`
`2189
`
`

`

`T h e ne w e ngl a nd jou r na l o f m e dicine
`
`another subgroup of patients. Such variation, also
`called “effect modification,” is typically expressed
`in a statistical model as an interaction term or
`terms between the treatment group and the base-
`line variable. The presence or absence of interac-
`tion is specific to the measure of the treatment
`effect.
`The appropriate statistical method for assess-
`ing the heterogeneity of treatment effects among
`the levels of a baseline variable begins with a sta-
`tistical test for interaction.10‑13 For example, Sacks
`et al.8 showed the heterogeneity in pravastatin
`efficacy by reporting a statistically significant
`(P = 0.03) result of testing for the interaction be-
`tween the treatment and baseline LDL level when
`the measure of the treatment effect was the rel-
`ative risk. Many trials lack the power to detect het-
`erogeneity in treatment effect; thus, the inability
`to find significant interactions does not show that
`the treatment effect seen overall necessarily ap-
`plies to all subjects. A common mistake is to
`claim heterogeneity on the basis of separate tests
`of treatment effects within each of the levels of
`the baseline variable.6,7,14 For example, testing the
`hypothesis that there is no treatment effect in
`women and then testing it separately in men does
`not address the question of whether treatment dif-
`ferences vary according to sex. Another common
`error is to claim heterogeneity on the basis of the
`observed treatment-effect sizes within each sub-
`group, ignoring the uncertainty of these esti-
`mates.
`
`Multiplicity
`It is common practice to conduct a subgroup analy-
`sis for each of several — and often many — base-
`line characteristics, for each of several end points,
`or for both. For example, the analysis by Jackson
`and colleagues9 of the effect of calcium plus vi-
`tamin D supplementation relative to placebo on
`the risk of each of four fracture outcomes for 15
`participant characteristics resulted in a total of
`60 subgroup analyses.
`When multiple subgroup analyses are per-
`formed, the probability of a false positive finding
`can be substantial.7 For example, if the null hy-
`pothesis is true for each of 10 independent tests
`for interaction at the 0.05 significance level, the
`chance of at least one false positive result exceeds
`40%. Thus, one must be cautious in the interpre-
`
`tation of such results. There are several methods
`for addressing multiplicity that are based on the
`use of more stringent criteria for statistical sig-
`nificance than the customary P<0.05.7,15 A less
`formal approach for addressing multiplicity is to
`note the number of nominally significant inter-
`action tests that would be expected to occur by
`chance alone. For example, after noting that 60
`subgroup analyses were planned, Jackson et al.9
`pointed out that “Up to three statistically signifi-
`cant interaction tests (P<0.05) would be expected
`on the basis of chance alone,” and then they in-
`corporated this consideration in their interpre-
`tation of the results.
`
`Prespecified Analysis versus Post hoc
`Analysis
`A prespecified subgroup analysis is one that is
`planned and documented before any examination
`of the data, preferably in the study protocol. This
`analysis includes specification of the end point,
`the baseline characteristic, and the statistical
`method used to test for an interaction. For exam-
`ple, the Heart Outcomes Prevention Evaluation 2
`investigators16 conducted a study involving 5522
`patients with vascular disease or diabetes to as-
`sess the effect of homocysteine lowering with fo-
`lic acid and B vitamins on the risk of a major car-
`diovascular event. The primary outcome was a
`composite of death from cardiovascular causes,
`myocardial infarction, and stroke. In the Methods
`section of their article, the authors noted that “Pre-
`specified subgroup analyses involving Cox mod-
`els were used to evaluate outcomes in patients
`from regions with folate fortification of food and
`regions without folate fortification, according to
`the baseline plasma homocysteine level and the
`baseline serum creatinine level.” Post hoc analy-
`ses refer to those in which the hypotheses being
`tested are not specified before any examination
`of the data. Such analyses are of particular con-
`cern because it is often unclear how many were
`undertaken and whether some were motivated by
`inspection of the data. However, both prespeci-
`fied and post hoc subgroup analyses are subject
`to inflated false positive rates arising from mul-
`tiple testing. Investigators should avoid the ten-
`dency to prespecify many subgroup analyses in the
`mistaken belief that these analyses are free of
`the multiplicity problem.
`
`2190
`
`n engl j med 357;21 www.nejm.org november 22, 2007
`
`The New England Journal of Medicine
`
`Downloaded from nejm.org on April 16, 2019. For personal use only. No other uses without permission.
`
` Copyright © 2007 Massachusetts Medical Society. All rights reserved.
`
`Page 2 of 6
`
`

`

`Special Report
`
`subgroup analyses in the
`journal — a ssessment of
`rep or ting pr ac tices
`
`As part of internal quality-control activities at the
`Journal, we assessed the completeness and qual-
`ity of subgroup analyses reported in the Journal
`during the period from July 1, 2005, through June
`30, 2006. A detailed description of the study meth-
`ods can be found in the Supplementary Appen-
`dix, available with the full text of this article at
`www.nejm.org. In this report, we describe the
`clarity and completeness of subgroup-analysis re-
`porting, evaluate the authors’ interpretation and
`justification of the results of subgroup analyses,
`and recommend guidelines for reporting subgroup
`analyses.
`Among the original articles published in the
`Journal during the period from July 1, 2005,
`through June 30, 2006, a total of 95 articles re-
`ported primary outcome results from randomized
`clinical trials. Among these 95 articles, 93 report-
`ed results from one clinical trial; the remaining
`2 articles reported results from two trials. Thus,
`results from 97 trials were reported, from which
`subgroup analyses were reported for 59 trials
`(61%). Table 1 summarizes the characteristics of
`the trials. We found that larger trials and multi-
`center trials were significantly more likely to re-
`port subgroup analyses than smaller trials and
`single-center trials, respectively. With the use of
`multivariate logistic-regression models, when
`ranked according to the number of participants
`enrolled in a trial and compared with trials with
`the fewest participants, the odds ratio for report-
`ing subgroup analyses for the second quartile was
`1.38 (95% confidence interval [CI], 0.45 to 4.20),
`for the third quartile was 1.98 (95% CI, 0.62 to
`6.24), and for the fourth quartile was 8.90 (95%
`CI, 2.10 to 37.78) (P = 0.02, trend test). The odds
`ratio for reporting subgroup analyses in multi-
`center trials as compared with single-center trials
`was 4.33 (95% CI, 1.56 to 12.16).
`Among the 59 trials that reported subgroup
`analyses, these analyses were mentioned in the
`Methods section for 21 trials (36%), in the Results
`section for 57 trials (97%), and in the Discussion
`section for 37 trials (63%); subgroup analyses were
`reported in both the text and a figure or table for
`39 trials (66%). Other characteristics of the reports
`
`are shown in Figure 1. In general, we are unable
`to determine the number of subgroup analyses
`conducted; we attempted to count the number of
`subgroup analyses reported in the article and
`found that this number was unclear in nine ar-
`ticles (15%). For example, Lees et al.17 reported
`that “We explored analyses of numerous other
`subgroups to assess the effect of baseline prog-
`nostic factors or coexisting conditions on the
`
`Table 1. Characteristics and Predictors of Reporting Subgroup Analyses
` in 97 Clinical Trials.*
`
`Variable
`
`No. of subjects
`
`≤218
`
`219–429
`
`430–1012
`
`>1012
`
`Superiority trial
`
`Yes
`
`No
`
`Trial sites
`
`Single-center
`
`Multicenter
`
`Trials Reporting
`Subgroup
`Analyses
`
`P Value†
`
`No. of Trials/
`Total No. (%)
`
`Univariate
`Odds Ratio
`
`Multivariate
`Odds Ratio
`
`0.002†
`
`0.02†
`
`11/25 (44)
`
`13/25 (52)
`
`14/23 (61)
`
`21/24 (88)
`
`53/84 (63)
`
`6/13 (46)
`
`7/21 (33)
`
`52/76 (68)
`
`0.25
`
`0.89
`
`0.005
`
`0.05
`
`Type of disease studied
`
`0.18
`
`0.37
`
`Cardiovascular
`
`16/20 (80)
`
`Infectious
`
`Oncologic
`
`Respiratory
`
`Pediatric
`
`Psychiatric or neurologic
`
`Metabolic, endocrine,
`or gastrointestinal
`
`Gynecologic
`
`Other
`
`Statistically significant primary
`end point
`
`Yes
`
`No
`
`2/7 (29)
`
`9/11 (82)
`
`7/10 (70)
`
`5/10 (50)
`
`6/10 (60)
`
`5/10 (50)
`
`3/6 (50)
`
`6/13 (46)
`
`35/62 (56)
`
`24/35 (69)
`
`* A total of 59 trials reported subgroup analyses.
`† P values were determined with the use of trend tests.
`
`0.24
`
`0.38
`
`n engl j med 357;21 www.nejm.org november 22, 2007
`
`2191
`
`The New England Journal of Medicine
`
`Downloaded from nejm.org on April 16, 2019. For personal use only. No other uses without permission.
`
` Copyright © 2007 Massachusetts Medical Society. All rights reserved.
`
`Page 3 of 6
`
`

`

`The NEW ENGLAND JOURNAL of MEDICINE
`
`No. ofSubgroup Analyses
`
`Clear about Prespecified or Post Hoe
`40
`
`3
`
`Sometimes
`
`Always
`
`Interaction Test Reported
`
`32
`
`Information Repon within Subgroups
`21
`
`
`
`Tr'IIIs(no.)
`
`
`
`Trials(no.)
`
`
`
`Triak(no.)
`
`
`
`Trials(no.)
`
`‘1-
`OCv
`_"a
`'E
`..
`
`
`
`Trials(no.)
`
`
`
`Sometimes
`
`Always
`
`Heterogeneity and Multiplicity
`44
`
`Subgroup Analyses Reported in Abstract
`
`37
`
`No
`
`Yes
`
`Heterogeneity
`Not Claimed
`
`Heterogeneity
`
`Not Claimed
`
`2
`
`Yes
`
`No
`
`Multiplicity Issues
`Addressed and
`Haerogeneity Claimed
`
`Figure 1. Reporting of Subgroup Analyses from 59 Clinical Trials.
`The specific reporting characteristics examined in this quality-improvement exercise are indicated in each panel.
`CI denotes confidence interval.
`
`treatment effect but found no evidence of nomi-
`
`nal significance for any biologically likely factor.”
`For four of these nine articles, we were able to de-
`termine that at least eight subgroup analyses were
`reported. In 40 trials (68%), it was unclear wheth-
`er any of the subgroup analyses were prespecified
`or post hoc, and in 3 others (5%) it was unclear
`whether some were prespecified or post hoc. In-
`teraction tests were reported to have been used to
`assess the heterogeneity of treatment effects for
`all subgroup analyses in only 16 trials (27%), and
`
`they were reported to be used for some, but not
`all, subgroup analyses in 11 trials (19%).
`We assessed whether information was provided
`about treatment effects within the levels of each
`
`subgroup variable (Fig. 1). In 25 trials (42%), in-
`formation about treatment effects was reported
`consistently for all of the reported subgroup analy-
`ses, and in 13 trials (22%), nothing was reported.
`Investigators in 15 trials (25%), all using supe-
`riority designs,1° claimed heterogeneity of treat-
`ment effects between at least one subject sub-
`
`2192
`
`N ENGLJ MED 357:21 WWW.NE]M.ORG NOVEMBER 22, 2007
`
`The New England Journal of Medicine
`Page 4 osgvuloaded from nejmorg on
`April 16, 2019. For personal use only. No other uses without permission.
`Copyright O 2007 Massachusetts Medical Society. All rights reserved
`
`

`

`SPECIAL REPORT
`
`group and the overall study population (see Table 1
`of the Supplementary Appendix). For 4 of these
`15 trials, this claim was based on a nominally sig-
`nificant interaction test, and for 4 others it was
`based on within-subgroup comparisons only. In
`the remaining seven trials, significant results of
`interaction tests were reported for some but not
`all subgroup analyses. When heterogeneity in the
`treatment effect was reported, for two trials (13%),
`investigators offered caution about multiplicity,
`and for four trials (27%), investigators noted the
`heterogeneity in the Abstract section.
`
`ANALYSIS OF OUR FINDINGS
`AND GUIDELINES FOR REPORTING
`SUBGROUPS
`
`In the 1-year period studied, the reporting of sub-
`group analyses was neither uniform nor complete.
`Because the design of future clinical trials can
`depend on the results of subgroup analyses, uni-
`formity in reporting would strengthen the foun-
`dation on which such research is built. Further-
`
`more, uniformity of reporting will be ofvalue in
`the interval between recognition of a potential
`subgroup effect and the availability of adequate
`data on which to base clinical decisions.
`
`Problems in the reporting of subgroup analy-
`ses are not new.1'°-18 Assmann et al.2 reported
`shortcomings of subgroup analyses in a review of
`the results of 50 trials published in 1997 in four
`leading medicaljoumals. More recently, Heman-
`dez et al.4 reviewed the results of 63 cardiovascu-
`
`lar trials published in 2002 and 2004 and noted
`the same problems. To improve the quality of
`reports of parallel-group randomized trials, the
`Consolidated Standards of Reporting Trials state-
`ment was proposed in the mid-19903 and revised
`in 2001.19 Although there has been considerable
`discussion of the potential problems associated
`with subgroup analysis and recommendations on
`when and how subgroup analyses should be con-
`ducted and reported,19’1° our analysis of recent
`articles shows that problems and ambiguities per-
`sist in articles published in the journal. For exam-
`ple, we found that in about two thirds of the pub-
`lished trials, it was unclear whether any of the
`reported subgroup analyses were prespecified or
`post hoc. In more than half of the trials, it was
`unclear whether interaction tests were used, and
`in about one third of the trials, within-level results
`were not presented in a consistent way.
`
`Guidelines for Reporting Subgroup Analysis.
`
`In the Abstract
`Present subgroup results in the Abstract only ifthe subgroup analyses were
`based on a primary study outcome, ifthey were prespecified, and ifthey
`were interpreted in light of the totality of prespecified subgroup analyses
`undertaken.
`
`
`
`In the Methods section:
`Indicate the number of prespecified subgroup analyses that were performed
`and the number of prespecified subgroup analyses that are reported.
`Distinguish a specific subgroup analysis of special interest. such as that
`in the article by Sacks et al.,8 from the multiple subgroup analyses typical-
`ly done to assess the consistency of a treatment effect among various pa-
`tient characteristics, such as those in the article byjackson et al.9 For
`each reported analysis. indicate the end point that was assessed and the
`statistical method that was used to assess the heterogeneity of treatment
`differences.
`Indicate the number ofpost hoc subgroup analyses that were performed and
`the number of post hoc subgroup analyses that are reported. For each re-
`ported analysis, indicate the end point that was assessed and the statisti-
`cal method used to assess the heterogeneity oftreatment differences.
`Detailed descriptions may require a supplementary appendix.
`Indicate the potential effect on type | errors (false positives) due to multiple
`subgroup analyses and how this effect is addressed. If formal adjust-
`ments for multiplicity were used, describe them; ifno formal adjustment
`was made, indicate the magnitude of the problem informally, as done by
`Jackson et al.9
`In the Rsults section:
`When possible. base analyses of the heterogeneity of treatment effects on
`tests for interaction, and present them along with effect estimates (in-
`cluding confidence intervals) within each level of each baseline covariate
`analyzed. A forest plot“22 is an effective method for presenting this in-
`formation.
`
`In the Discussion section:
`Avoid overinterpretation of subgroup differences. Be properly cautious in ap-
`praising their credibility, acknowledge the limitations, and provide sup-
`porting or contradictory data from other studies. if any.
`
`When properly planned, reported, and inter-
`preted, subgroup analyses can provide valuable
`information. With the availability of Web supple-
`ments, the opportunity exists to present more de-
`tailed information about the results of a trial. The
`
`purpose of the guidelines (see box) is to encour-
`age more clear and complete reporting of sub-
`group analyses. In some settings, a trial is con-
`ducted with a subgroup analysis as one of the
`primary objectives. These guidelines are directly
`applicable to the reporting of subgroup analyses
`in the primary publication of a clinical trial when
`the subgroup analyses are not among the primary
`objectives. In other settings, including observa-
`tional studies, we encourage complete and thor-
`ough reporting of the subgroup analyses in the
`spirit of the guidelines listed.
`The editors and statistical consultants of the
`
`journal consider these guidelines to be important
`in the reporting of subgroup analyses. The goal
`is to provide transparency in the statistical meth-
`
`N ENGLJ MED 357:21 WWW.NEJM.ORC.
`
`NOVEMBER 22, 2007
`
`2193
`
`P
`
`age
`
`5
`
`The New England Journal of Medicine
`oggvnloaded from nejmorg on April 16, 2019. For personal use only. No other uses without permission.
`Copyright © 2007 Massachusetts Medical Society. All rights reserved.
`
`

`

`Special Report
`
`ods used in order to increase the clarity and com-
`pleteness of the information reported. As always,
`these are guidelines and not rules; additions and
`exemptions can be made as long as there is a clear
`case for such action.
`No potential conflict of interest relevant to this article was re-
`ported.
`We thank Doug Altman, John Bailar, Colin Begg, Mohan
`Beltangady, Marc Buyse, David DeMets, Stephen Evans, Thomas
`Fleming, David Harrington, Joe Heyse, David Hoaglin, Michael
`Hughes, John Ioannidis, Curtis Meinert, James Neaton, Robert
`O’Neill, Ross Prentice, Stuart Pocock, Robert Temple, Janet
`Wittes, and Marvin Zelen for their helpful comments.
`
`Yusuf S, Wittes J, Probstfield J, Tyroler HA. Analysis and in-
`1.
`terpretation of treatment effects in subgroups of patients in ran-
`domized clinical trials. JAMA 1991;266:93-8.
`Assmann SF, Pocock SJ, Enos LE, Kasten LE. Subgroup
`2.
`analysis and other (mis)uses of baseline data in clinical trials.
`Lancet 2000;355:1064-9.
`Pocock SJ, Assmann SF, Enos LE, Kasten LE. Subgroup analy-
`3.
`sis, covariate adjustment and baseline comparisons in clinical
`trial reporting: current practice and problems. Stat Med 2002;
`21:2917-30.
`Hernández A, Boersma E, Murray G, Habbema J, Steyerberg
`4.
`E. Subgroup analyses in therapeutic cardiovascular clinical tri-
`als: are most of them misleading? Am Heart J 2006;151:257-64.
`Parker AB, Naylor CD. Subgroups, treatment effects, and
`5.
`baseline risks: some lessons from major cardiovascular trials. Am
`Heart J 2000;139:952-61.
`Rothwell PM. Subgroup analysis in randomised controlled
`6.
`trials: importance, indications, and interpretation. Lancet 2005;
`365:176-86.
`Lagakos SW. The challenge of subgroup analyses — report-
`7.
`ing without distorting. N Engl J Med 2006;354:1667-9. [Erratum,
`N Engl J Med 2006;355:533.]
`Sacks FM, Pfeffer MA, Moye LA, et al. The effect of prava-
`8.
`statin on coronary events after myocardial infarction in patients
`with average cholesterol levels. N Engl J Med 1996;335:1001-9.
`
`Jackson RD, LaCroix AZ, Gass M, et al. Calcium plus vitamin
`9.
`D supplementation and the risk of fractures. N Engl J Med 2006;
`354:669-83. [Erratum, N Engl J Med 2006;354:1102.]
`Pocock SJ. Clinical trials: a practical approach. Chichester,
`10.
`England: John Wiley, 1983.
`Halperin M, Ware JH, Byar DP, et al. Testing for interaction
`11.
`in an I×J×K contingency table. Biometrika 1977;64:271-5.
`Simon R. Patient subsets and variation in therapeutic effi-
`12.
`cacy. Br J Clin Pharmacol 1982;14:473-82.
`Gail M, Simon R. Testing for qualitative interactions between
`13.
`treatment effects and patient subsets. Biometrics 1985;41:361-72.
`Brookes ST, Whitely E, Egger M, Smith GD, Mulheran PA,
`14.
`Peters T. Subgroup analyses in randomized trials: risks of sub-
`group-specific analyses; power and sample size for the interac-
`tion test. J Clin Epidemiol 2004;57:229-36.
`Bailar JC III, Mosteller F, eds. Medical uses of statistics. 2nd
`15.
`ed. Waltham, MA: NEJM Books, 1992.
`Lonn E, Yusuf S, Arnold MJ, et al. Homocysteine lowering
`16.
`with folic acid and B vitamins in vascular disease. N Engl J Med
`2006;354:1567-77. [Erratum, N Engl J Med 2006;355:746.]
`Lees KR, Zivin JA, Ashwood T, et al. NXY-059 for acute ische-
`17.
`mic stroke. N Engl J Med 2006;354:588-600.
`Al-Marzouki S, Roberts I, Marshall T, Evans S. The effect of
`18.
`scientific misconduct on the results of clinical trials: a Delphi
`survey. Contemp Clin Trials 2005;26:331-7.
`Moher D, Schulz KF, Altman DG, et al. The CONSORT State-
`19.
`ment: revised recommendations for improving the quality of
`reports of parallel-group randomized trials. (Accessed Novem-
`ber 1, 2007, at http://www.consort-statement.org/.)
`International Conference on Harmonisation (ICH). Guid-
`20.
`ance for industry: E9 statistical principles for clinical trials. Rock-
`ville, MD: Food and Drug Administration, September 1998. (Ac-
`cessed November 1, 2007, at http://www.fda.gov/cder/guidance/
`ICH_E9-fnl.PDF.)
`Cuzick J. Forest plots and the interpretation of subgroups.
`21.
`Lancet 2005;365:1308.
`Wactawski-Wende J, Kotchen JM, Anderson GL, et al. Calci-
`22.
`um plus vitamin D supplementation and the risk of colorectal
`cancer. N Engl J Med 2006;354:684-96.
`Copyright © 2007 Massachusetts Medical Society.
`
`full text of all journal articles on the world wide web
`Access to the complete text of the Journal on the Internet is free to all subscribers. To use this Web site, subscribers should go
`to the Journal’s home page (www.nejm.org) and register by entering their names and subscriber numbers as they appear on
`their mailing labels. After this one-time registration, subscribers can use their passwords to log on for electronic access to the
`entire Journal from any computer that is connected to the Internet. Features include a library of all issues since January 1993
`and abstracts since January 1975, a full-text search capacity, and a personal archive for saving articles and search results of
`interest. All articles can be printed in a format that is virtually identical to that of the typeset pages. Beginning 6 months after
`publication, the full text of all Original Articles and Special Articles is available free to nonsubscribers who have completed a
`brief registration.
`
`2194
`
`n engl j med 357;21 www.nejm.org november 22, 2007
`
`The New England Journal of Medicine
`
`Downloaded from nejm.org on April 16, 2019. For personal use only. No other uses without permission.
`
` Copyright © 2007 Massachusetts Medical Society. All rights reserved.
`
`Page 6 of 6
`
`