Title: Analysis and Interpretation of Subgroups David L. DeMets, Ph.D. University of Wisconsin Madison, WI
1Analysis and Interpretationof SubgroupsDavid
L. DeMets, Ph.D.University of WisconsinMadison,
WI
2Disclosure Information
The following relationships exist related to this
presentation Trials Sponsorship Consultant/
DSMB Member NIH Astra-Zeneca Merck Myog
en Bristol Myers Squibb Guidant Eli Lilly
Actelion Pfizer GlaxoSmith Klne Wyeth-Ayerst
3Subgroup Motivation
- Response of average patient is not response of
individual patient
- (Yusuf, et al., JAMA, 1991)
- (Bernard, et al., 1957)
- Move RCT results closer to individual patient
- (Julian, Current Cardiovascular Trials, 2000)
4Primary and Secondary Questions
- Primary
- most important, central question
- ideally, only one
- stated in advance
- basis for design and sample size
- Secondary
- related to primary
- stated in advance
- limited in number
5Subgroup Questions
- Questions about effect of therapy in a
sub-population of subjects entered into the
trial
- Primary outcome
- Secondary outcome
- Rationale
- Consistency of primary and secondary outcomes
- Confirmation
- Exploratory
6Subgroup Issues
- Properly Defined
- False Positives
- Degree of Consistency
- Precision of Treatment Effect Estimate
7Appropriate/Proper Subgroups
- Defined in advance
- Defined by baseline measurements only
- Based on external definitions, no slicing and
dicing
8Inappropriate Subgroups
- Defined by
- Unblinded post hoc review of eligibility
- Compliance to treatment
- Intermediate outcomes (e.g. blood pressure
change)
- Produce biased or nonsense analysis
9False Positive Rates
- The greater the number of subgroups analyzed
separately, the larger the probability of making
false positive conclusions.
- No. of Subgroups False Positive Rate
- 1 .05
- 2 .08
- 3 .11
- 4 .13
- 5 .14
- 10 .19
10Example - Subgroup Concern
- Second International Study of Infarct Survival
(ISIS 2) (Lancet, 1988)
- 2 x 2 factorial design
- (aspirin vs. placebo and streptokinase vs.
placebo)
- vascular and total mortality in patients with an
acute myocardial infarction (MI)
- Gemini or Libra astrological birth signs did
somewhat worse on aspirin while all other signs
and overall results impressive and highly
significant benefit from aspirin
11Consistency of Results
- Absolute consistency not realistic
- Variation across properly defined subgroups to be
expected
- Degree of variation depends on
- True effect size
- Size of trial and subgroups
- Look cautiously for qualitative differences
12MERIT-HF Study Design
- Chronic heart failure patients
- Randomized placebo controlled
- Metoprolol vs. placebo
- Two-week placebo run in (compliance)
- Entered 3991 patients
- Terminated early
- Mean follow-up approximately one year
- The International Steering Committee on Behalf of
the MERIT-HF Study Group,
- Am J Cardiol 1997 80(9B)54J-58J. The MERIT-HF
Study Group, ACC, March 1999.
13MERIT Total Mortality
14MERIT
15MERIT(AHJ, 2001)
16BHAT Hazard Ratios for All-Cause Mortality
17SubgroupsRef NEJM, 1996
- Not all subgroup results are valid
- False positive claims
- Classic Example in cardiology
- Major trial identified possible subgroup result
18Praise IRef NEJM, 1996
- Amlodipine vs. placebo
- NYHA class II-III
- Randomized double-blind
- Mortality/hospitalization outcomes
- Stratified by etiology (ischemic/non-ischemic)
- 1153 patients
19Subgroup or Treatment Interaction
- PRAISE I (NEJM, 1996)
- Patients with ischemic and non-ischemic
- Predefined stratified randomization
- Mortality, a secondary outcome
- Group Relative Risk P-value
- Total .84 .07
- Ischemic 1.02
- Non-Ischemic .54
- Interaction Test P .004
20PRAISE I
21PRAISE I - Interaction
- Overall P 0.07
- Etiology by Trt Interaction
- P 0.004
- Ischemic P NS
- Non-Ischemic P
22PRAISE I - Ischemic
23PRAISE I Non- Ischemic
24PRAISE II
- Repeated non-ischemic strata
- Amlodipine vs. placebo
- Randomized double-blind
- 1653 patients
- Mortality outcome
- RR ? 1.0
25Subgroup by Treatment Interaction
- Trial Relative Risk
- PRAISE I
- Non-Ischemic Subgroup .54
- PRAISE II
- All Non-Ischemic 1.05
- Subgroup replication still required
26Estimate of Treatment Effect
- Overall power/precision of trial barely adequate
- Each subgroup has limited power and precision
- Best estimate of treatment effect for any
subgroup is overall estimate
27Three Views
- Ignore subgroups and analyze only by treatment
groups. (Sleight 2000)
- Do subgroup analyses --- However view all results
with caution.
- Plan for subgroup analyses in advance. Do not
mine data. Be cautious
- Distinguish confirmatory vs. exporatory
- Consistency
28Subgroup Analysis
-
- Guidelines for Subgroups
- 1. Defined by baseline measurements
- 2. Stated in advance (in protocol)
- 3. Limited number of confirmatory
- 4. Label exploratory subgroups
- 5. Interpreted cautiously, qualitatively
- 6. Look for general consistency of results
- Formal adjustments for multiplicity problematic
if more than a few are done