MetaAnalysis

1
Meta-Analysis

• Brian F. Gage, MD, MSc
• 10/28/09
• Thanks to Elna Nagasako, MD, PhD for several
  slides

2
Types of Overviews

• Literature reviews
• Systematic reviews
• Meta-analyses

? todays focus
3
Goals

• Why do a meta-analysis?
• Who should do a meta-analysis?
• When to do a meta-analysis
• How to do a meta-analysis
• Successful examples

4
Why do a Meta-Analysis?

• Quantitatively, to combine the results of
  multiple studies

5
Meta-Analysis of Hemorrhage
Ximelagatran (Xi) reduced risk of major
hemorrhage by 27 compared with warfarin (95 CI
6-44)
6
Who should do a meta-analysis?

• Your team should have both domain and statistical
  expertise
• One of these areas should be very strong
• Domain expertise is needed
• to interpret the results of the trial,
• to write the discussion, and
• to network w/ investigators who have unpublished
  data
• Stats expertise is needed
• to do and understand the math
• to provide credibility to your results and
  methods

7
When to do a Meta-analysis?

• The field needs to be ripe
• Not too many studies and not too few
• Concurrent w/ the results of the latest large
  trial
• Typically in collaboration w/ the PI of the trial
• Provides domain expertise and early access to
  data
• Concurrent with a literature review for another
  reason (e.g. a grant)
• Otherwise, its not worth undertaking

8
E.g. Warfarin and Fractures Motivated us to
Examine Role of Vitamin K on Bone Health

• Overall adjusted OR 1.25 (95 CI, 1.06 1.48)
• N7587 without previous warfarin exposure, N
  4461 with long term warfarin exposure (gt 1 year)

(Gage et al. 2006)
9
Meta-analysis of Vitamin K2 Supplementation and
Fracture

• Searching for Articles
• Vitamin K2, menatetrenone, menaquinones And
  Postmenopausal, osteoporosis, fracture
• Criteria for Inclusion of Study
• Double-blind randomized-controlled trial
• 45 mg vitamin K2 vs. placebo
• Postmenopausal osteoporotic women
• Outcome variable fractures

10
Meta-analysis of Vitamin K2 Supplementation and
Fracture
Summary of Selected Studies
11
Meta-analysis Vitamin K2 Supplementation and
Fracture
Statistical Analysis Random-effects model
(DerSimonian R et al. 1986)
12
Steps to Complete a Meta-analysis

• Clear question
• Comprehensive, unbiased lit. review
• Explicit inclusion/exclusion criteria
• Uniform and unbiased abstraction of results AND
  characteristics of each study
• Clear presentation of each studys findings
• Calculation of a summary estimate (95 CI)
• Assessment of heterogeneity
• Assessment of publication bias
• Subgroup and sensitivity analysis

S. B. Hulley et al., chapter 13
13
Heart Attack Risk Seen in Drug for Diabetes
The Food and Drug Administration is trying to
estimate the number of heart attacks that may be
linked to GlaxoSmithKlines Avandia. (NYT May 22,
2007)
14
Nissen SE, Wolski K.Clear Question Effect of
rosiglitazone on the risk of myocardial
infarction and death from cardiovascular causes.
NEJM 2007 356(24)2457-71.
15
Rosiglitazone

• Thiazolidinedione (PPAR-? agonist)
• PPAR-? receptors ligand-activated nuclear
  transcription factors
• Affect blood glucose by increasing insulin
  sensitivity
• FDA approval based on reduction in blood glucose
  and glycated hemoglobin levels
• Extent of cardiovascular effect not known

16
Comprehensive identification of studies

• Meta-analysis of 42 trials
• Trials drawn from FDA approval submission,
  clinical trial registry, recent large-scale RCTs
• Subjects
• Mean age 56
• Mean Hgb A1c 8.2

17
Clear Clinical question

• Does rosiglitazone reduce cardiovascular
  morbidity and mortality in patients with type 2
  diabetes mellitus?

18
Explicit Criteria for Inclusion

• Randomized comparator group not receiving
  rosiglitazone
• Similar duration of treatment in all groups
• Drug exposure gt 24 weeks
• Available outcome data for MI and death from CV
  causes

19
Study selection three sources

• Studies submitted to the FDA for rosiglitazone
  approval hearing (N5)
• GlaxoSmithKline clinical-trial registry (N35
  Phase 2-4)
• DREAM ADOPT studies (N2)

Inclusion of unpublished studies via a clinical
trial registry helps mitigate publication bias.
20
Study validity

• Not assessed beyond requiring randomized
  comparator group
• Trials not intended to assess CV outcomes

21
Statistical heterogeneity

• Are the differences between studies significant?
• More significant observed differences in
  results are real
• Less significant observed differences are due
  to chance

Cochrans Q statistic is one method for assessing
study heterogeneity in a meta-analysis.
22
Fixed or Random Treatment Effects

• Fixed if all studies were infinitely large
  theyd give identical results.
• Random Each study is evaluating a different
  true effect.

Different statistical methods are used depending
on whether treatment effects are assumed to be
fixed or random.
23
Doses, Baseline Demographic Characteristics,
Study Periods, and Glycated Hemoglobin Levels

• Intervention Rosiglitazone alone in
  combination
• Control placebo, glyburide, metformin, insulin,
  lifestyle changes
• Daily dose 2 mg, 2 mg BID, 4 mg, 4 mg BID, 8 mg
• Population Type 2 DM /- poor control on another
  agent, with CHF, recently diagnoses, elderly,
  Korean, Chinese mild-mod Alzheimers disease,
  chronic psoriasis, IGT
• Baseline A1c 6.3-9.9

24
Clear Presentation of Individual Trial Results

• Small trials
• MI 0-5 (rsg), 0-3 (ctrl)
• CV death 0-3 (rsg), 0-2 (ctrl)
• DREAM
• MI 15 (rsg), 9 (ctrl)
• CV death 12 (rsg), 10 (ctrl)
• ADOPT
• MI 27 (rsg), 41 (ctrl)
• CV death 2 (rsg), 5 (ctrl)

Small total numbers of events affects ability to
detect heterogeneity.
25
What are the results?

• What are the overall results of the overview?
• How precise were the results?

Oxman AD, Cook DJ, Guyatt GH. Users Guide to the
Medical Literature How to Use an Overview. JAMA
1994 272 1367-1371.
26
Combining results

• Meta-analysis the use of statistical methods to
  combine results of individual studies
• Vote counting
• Are there more positive or negative studies?
• Treat-as-one-trial method
• Simpsons paradox
• Alters effect of randomization
• Direct average of risk differences
• Minimizes influence of large studies, magnifies
  effect of small studies

Cochrane Collaboration open learning material for
reviewers V1.1, Nov 2002 (http//www.cochrane-net.
org/openlearning/index.htm)
27
Which studies should be given more weight?

• Large studies?
• More events?
• Better quality?

Variance is a measure that incorporates both
study size and event rate.
Cochrane Collaboration open learning material for
reviewers V1.1, Nov 2002 (http//www.cochrane-net.
org/openlearning/index.htm)
28
Calculation of Summary Estimate

• Many trials had few cardiovascular events, so
  the odds ratios and 95 confidence intervals were
  calculated with the use of the Peto method.
  Because all trials had similar durations of
  follow-up for all treatment groups, the use of
  odds ratios represents a valid approach to
  assessing the risk associated with the use of
  rosiglitazone. Trials in which patients had no
  adverse cardiovascular events in either group
  were excluded from analyses. All reported P
  values are two-sided.
  Statistical heterogeneity across the various
  trials was tested with the use of Cochrans Q
  statistic. A P value of more than the nominal
  level of 0.10 for the Q statistic indicated a
  lack of heterogeneity across trials, allowing for
  the use of a fixed effects model.

29
Statistical methods

• Peto (fixed effect, dichotomous)
• Odds ratios only
• Used for trials with small treatment effects
• Best method for rare outcomes
• Less useful for small trials with large treatment
  effects
• DerSimonian Laird (random effect, either
  continuous or dichotomous)
• Mantel-Haenszel (fixed effect, either continuous
  or dichotomous)

Cochrane Collaboration open learning material for
reviewers V1.1, Nov 2002 (http//www.cochrane-net.
org/openlearning/index.htm)
30
Rates of MI death from CV causes
31
(No Transcript)
32
Rosiglitazone Cardiovascular Risk

• Strengths of trial
• Inclusion of unpublished studies
• Use of major CV events as primary outcome
• Comparison with placebo
• Weaknesses
• Summary data rather than patient-level data
• No dose-response analyses
• No standard method for identifying or validating
  outcomes
• Small total numbers of events

Although the study has significant limitations,
it suggests that rosiglitazone is unlikely to
have cardiovascular benefits.
Psaty BM Furberg CD. NEJM 2007 3562522-2524.
33
Limitations of the linear physiologic argument

• If increased glycated hemoglobin levels increase
  the risk of adverse health outcomes, do
  reductions in glycated hemoglobin improve health
  outcomes?
• Peroxisome-proliferator-activated receptor
  agonists activate multiple glucose and lipid
  metabolism genes.
• Beneficial rosiglitazone effects
• Reduction of glycated hemoglobin
• Reduction in fasting glucose
• Adverse rosiglitazone effects
• Increase in body weight
• Adverse effects on lipids
• Fluid retention

Psaty BM Furberg CD. NEJM 2007 3562522-2524.
34
A major failure of the drug use and drug
approval process in the US.

• Surrogate endpoints may neglect other effects
  that influence health outcomes.
• For drugs approved based on changes in surrogate
  endpoints, timely completion of post-marketing
  studies is important for the assessment of
  clinically meaningful outcomes.

Psaty BM Furberg CD. NEJM 2007 3562522-2524.
35
Summary

• Why do a meta-analysis?
• Who should do a meta-analysis?
• When to do a meta-analysis
• How to do a meta-analysis
• Successful examples
• Helpful software Cochrane Review Manager 5

36
Next Week

• Defer Communicating your research
• Gage Organizing the study, questionnaires,
  database, team
• Do Problem Set 5
• Read Hulley chapters 13, 15, 16