[email protected] Value and Limitations of Meta-Analysis in the Era of Evidence-Based Medicine Giuseppe Biondi-Zoccai, MD Division

[email protected] www.metcardio.org

Value and Limitations of Meta-Analysis in the Era of Evidence-Based Medicine

Giuseppe Biondi-Zoccai, MD

Division of Cardiology, Department of Internal Medicine, University of Turin, Turin, Italy

Meta-analysis and Evidence-based medicine Training in Cardiology (METCARDIO), Turin, Italy


Index

• How to define meta-analyses? Key concepts

• What comes first? Scientific hierarchy and The

Cochrane Collaboration

• Where’s the beef? Strenghts of meta-analyses

• Any toxic asset? Weaknesses of meta-analyses

• See one, do one, teach one. Structured approach

to systematic reviews


Why should you trust me?Meta-analyses or manuscript pertinent to meta-

analyses that I have co-authored since graduation

Total = 51


Why are meta-analysis important: exponential increase in worldwide PubMed citations

PubMed search strategy: ("2001"[PDAT] : "2005"[PDAT]) AND (("systematic"[title/abstract] AND "review"[title/abstract]) OR ("systematic"[title/abstract] AND "overview"[title/abstract]) OR ("meta-analysis"[title/abstract] OR "meta-analyses"[title/abstract]))


Index









Famous quotes

“If I have seen further it is by standing on the shoulders of giants”

Isaac Newton

“The great advances in science usually result from new tools rather than from new doctrines”

Freeman Dyson


Famous quotes

“I like to think of the meta-analytic process as similar to being in a helicopter.

On the ground individual trees are visible with high resolution.

This resolution diminishes as the helicopter rises, and in its place we begin to see patterns not visible from the ground”

Ingram Olkin


Baby steps of meta-analysis• 1904 - Karl Pearson (UK): correlation between inoculation of

vaccine for typhoid fever and mortality across apparently conflicting studies

• 1931 – Leonard Tippet (UK): comparison of differences between and within farming techniques on agricultural yield adjusting for sample size across several studies

• 1937 – William Cochran (UK): combination of effect sizes across different studies of medical treatments

• 1970s – Robert Rosenthal and Gene Glass (USA), Archie Cochrane (UK): combination of effect sizes across different studies of, respectively, educational and psychological treatments

• 1980s – exponential development/use of meta-analytic methods


Minimal glossary• Review: viewpoint on a subject quoting different primary authors

• Overview: as above

• Qualitative review: deliberately avoids a systematic approach

• Systematic review: deliberately uses a systematic approach to study search,

selection, abstraction, appraisal and pooling

• Quantitative review: uses quantitative methods to appraise or synthesize

data

• Meta-analysis: uses specific statistical methods for data pooling and/or

exploratory analysis

• Individual patient data meta-analysis: uses specific stastistical methods

for data pooling or exploration exploiting individual patient data

→ Our goal: systematic review (± meta-analysis)


Qualitative review

Tung et al, Ann Intern Med 2006


Systematic review and meta-analyses

• What is a systematic review?

– A systematic appraisal of the methodological quality,

clinical relevance and consistency of published

evidence on a specific clinical topic in order to provide

clear suggestions for a specific healthcare problem

• What is a meta-analysis?

– A quantitative synthesis that, preserving the identity of

individual studies, tries to provide an estimate of the

overall effect of an intervention, exposure, or diagnostic

strategy


Systematic review (w/o meta-analysis)

Hackan et al, JAMA 2003


Systematic review and meta-analysis

Agostoni et al, J Am Coll Cardiol 2004


Index









EBM hierarchy of evidence1. N of 1 randomized controlled trial

2. Systematic reviews of homogeneous randomized trials

3. Single (large) randomized trial

4. Systematic review of homogeneous observational studies addressing patient-important outcomes

5. Single observational study addressing patient-important outcomes

6. Physiologic studies (eg blood pressure, cardiac output, exercise capacity, bone density, and so forth)

7. Unsystematic clinical observations

Guyatt and Rennie, Users’ guide to the medical literature, 2002


Parallel hierarchy of scientific studies in cardiovascular medicine

Biondi-Zoccai, Ital Heart J 2003

Qualitative reviews

Systematic reviews

Meta-analyses from individual studies

Meta-analyses from individual patient data

Case reports and series

Observational studies

Observational controlled studies

Randomized controlled trials

Multicenter randomized controlled trials




The Cochrane Collaboration

Mission Statement:

The Cochrane Collaboration is an world-wide organization that aims to help people make well informed decisions about healthcare by preparing, maintaining and promoting the accessibility of systematic reviews of the effects of healthcare interventions


The Cochrane Collaboration

• Over 6000 contributors• 50 Collaborative Review Groups (CRGs)• 12 centers throughout the world• 9 fields• 11 Methods Groups• 1 Consumer Network• The Campbell Collaboration (focusing on

education/social sciences)


Cochrane resources• Cochrane Database of Systematic Reviews (CDSR) – contains

Cochrane systematic reviews

• Database of Abstracts of Reviews of Effectiveness (DARE) – contains abstracts of non-Cochrane reviews

• Cochrane Central Controlled Trials Register (CENTRAL) – contains titles or abstracts of RCTs from multiple sources

• Cochrane Database of Methodology Reviews – contains Cochrane reviews of methods papers

• Cochrane Methodology Register (CMR) – contains abstracts of non-Cochrane methods papers

• Health Technology Assessment Database (HTA) – contains abstracts of HTA papers

• NHS Economic Evaluation Database (NHS EED) – contains abstracts of economic analysis papers


Index









Pros• Application to any clinical research question

• Systematic searches for clinical evidence

• Explicit and standardized methods for search and selection

of evidence sources

• Thorough appraisal of the internal validity of primary studies

• Quantitative synthesis with increased statistical power

• Increased external validity by appraising the effect of an

intervention (exposure) across different settings

• Test subgroup hypotheses

• Explore clinical and statistical heterogeneity

Lau et al, Lancet 1998


Any application feasible: meta-analysis of intervention studies

Landoni et al, Am J Kidney Dis 2006


Any application feasible: meta-analysis of diagnostic studies

Hamon et al, JACC 2006


Any application feasible: meta-analysis of prognostic studies


Thorough appraisal of internal validity and quality of selected studies

Landoni et al, J Cardiothorac Vasc Anesth 2007


Increasing statistical power and external validity

De Luca et al, EHJ 2009


Test subgroup analyses

ATC, BMJ 2002


Explore statistical and clinical heterogeneity

Biondi-Zoccai et al, Am Heart J 2005


Explore small study effects

Abbate et al, J Am Coll Cardiol 2008

Review: Late percutaneous coronary intervention for infarct-related artery occlusionComparison: 01 Late percutaneous coronary intervention vs best medical therapy for infarct-related artery occlusion Outcome: 01 Death

0.1 0.2 0.5 1 2 5 10

0.0

0.4

0.8

1.2

1.6

SE(log OR)

OR (fixed)


Arguably the most important meta-analysis ever….

Antman et al, JAMA 1992


…showing discrepancies among evidence and experts


Index









Cons• “Exercise in mega-silliness”

• “Mixing apples with oranges”

• Not original research

• Big RCTs definitely better

• Pertinent studies might not be found, or may be of low

quality or internal validity

• Publication and small study bias

• Average effect largely unapplicable to individuals

• Duplicate efforts may lead to discordant resultsLau et al, Lancet 1998


What if I mix apples and oranges…

Hooper et al, BMJ 2006


What if I mix apples and oranges…


What if only few/low quality studies are found?

Biondi-Zoccai et al, J Endovasc Ther 2009


What if small positive studies are selectively published?

Pre

cisi

on

(sta

ndar

d e

rror

of

log

rel

ativ

e ris

k)

Effect(relative risk)

P<0.001 at Egger testP<0.001 at Peters test

0.01 0.1 1 10 100

0.0

0.4

0.8

1.2

1.6

Favours cilostazol Favours control



What if meta-analyses disagree?

Biondi-Zoccai et al, BMJ 2006


Appraisal tools: QUOROM

Moher et al, Lancet 1999


Appraisal tools: Oxman and Guyatt’sEvaluates the internal validity of a review on 9 separate questions for

which 3 distinct anwers are eligible (“yes”, “partially/can’t tell”, “no”):

1. Where the search methods used to find evidence stated?2. Was the search for evidence reasonably comprehensive?3. Were the criteria for deciding which studies to include in the overview reported4. Was bias in the selection of studies avoided5. Were the criteria used for assessing the validity of the included studies reported?6. Was the validity of all studies referred to in the text assessed using appropriate

criteria7. Were the methods used to combine the findings of the relevant studies reported?8. Were the findings of the relevant studies combined appropriately relative to the

primary question the overview addresses?9. Were the conclusions made by the author(s) supported by the data and/or

analysis reported in the overview?Question 10 summarizes the previous ones and, specifically, asks to rate the

scientific quality of the review from 1 (being extensively flawed) to 3 (carrying major flaws) to 5 (carrying minor flaws) to 7 (minimally flawed). The developers of the index specify that if the “partially/can’t tell” answer is used one or more times in questions 2, 4, 6, or 8, a review is likely to have minor flaws at best and is difficult to rule out major flaws (ie a score≤4). If the “no” option is used on question 2, 4, 6 or 8, the review is likely to have major flaws (ie a score≤3).

Oxman et al, J Clin Epidemiol 1991


Index









Algorithm for systematic reviews

• Definition of question and hypothetical solution

• Prospective design of the systematic review

• Problem formulation (population, intervention or

exposure, comparison, outcome [PICO])

• Data search

• Data abstraction and appraisal

• Data analysis ± quantitative synthesis

• Result interpretation and dissemination

Biondi-Zoccai et al, Ital Heart J 2004

FE

ED

-BA

CK

ON

HY

PO

TH

ES

IS


Definition of question and prospective design

• The clinical question should be clearly

stated, being as much explicit as possible

• The review should be designed in as much

details as possible, and yet with a limited a

priori knowledge of the subject



Problem formulation according to the PICO approach

• Population of interest – eg elderly male >2 weeks after

myocardial infarction)

• Intervention (or exposure) – eg intracoronary

infusion of progenitor blood cells

• Comparison – eg patients treated with progenitor cells vs

standard therapy

• Outcome(s) – eg change in echocardiographic left ventricular

ejection fraction from discharge to 6-month control



Data search

• After definition of question according to

PICO approach, the appropriate key-words

are used to search several databases

• Useful resources: BioMedCentral, CENTRAL,

clinicaltrials.gov, EMBASE/Scopus, LILACS, and

PubMed

• Conference proceedings

• Cross-referencing (snowballing)

• Contact with experts


Example of search strategies

Biondi-Zoccai et al, Int J Epidemiol 2005 Biondi-Zoccai et al, Am Heart J 2008


A simple PubMed strategy for clinical studies on percutaneous coronary intervention for left main coronary artery disease: left AND main AND coronary AND stent* NOT case reports [pt] NOT review [pt] NOT editorial [pt]

A complex PubMed strategy for randomized clinical trials on invasive vs conservative strategies in acute coronary syndromes: (randomized controlled trial[pt] OR controlled clinical trial[pt] OR randomized controlled trials[mh] OR random allocation[mh] OR double-blind method[mh] OR single-blind method[mh] OR clinical trial[pt] OR clinical trials[mh] OR (clinical trial[tw] OR ((singl*[tw] OR doubl*[tw] OR trebl*[tw] OR tripl*[tw]) AND (mask*[tw] OR blind[tw])) OR (latin square[tw]) OR placebos[mh] OR placebo*[tw] OR random*[tw] OR research design[mh:noexp] OR comparative study[mh] OR evaluation studies[mh] OR follow-up studies[mh] OR prospective studies[mh] OR cross-over studies[mh] OR control*[tw] OR prospectiv*[tw] OR volunteer*[tw]) NOT (animal[mh] NOT human[mh]) NOT (comment[pt] OR editorial[pt] OR meta-analysis[pt] OR practice-guideline[pt] OR review[pt])) AND ((invasive OR conservative AND (coronary OR unstable angina OR acute coronary syndrome* OR unstable coronary syndrome* OR myocardial infarction)))


Study selection

• 1st - screening of titles and abstracts

• 2nd – potentially pertinent citations are then

retrieved as full reports and appraised

according to prespecified and explicit

inclusion/exclusion criteria

• 3rd – studies fullfilling both inclusion and

exclusion criteria, are then included in the

systematic review


Andreotti et al,

Eur Heart J 2005


Data abstraction and appraisal

• Abstraction of outcomes and moderator

variables, possibly on prespecified data form

• Appraisal of the internal validity of primary

studies (eg the risk of selection, performance,

adjudication and attrition bias)

• Performed by single vs multiple reviewers, with

divergences resolved by consensus (possibly

after formal tests for agreement)


Internal validity of primary studies

• Many scales for the quality of included studies have been reported, but none is reliable or robust

• The recommended approach is to individually appraise the potential risk of the 4 biases (eg A-low, B-moderate, C-high, D-unclear from reported data):

– Selection bias (one group is different than the other)

– Performance bias (treatment is systematically different)

– Adjudication bias (outcome adjudication is selectively

different)

– Attrition bias (follow-up duration or completeness is

different)


Another common classification scheme for bias


Data synthesis

• Quantitative data synthesis is central to

the practice of meta-analysis, and is based

on a major assumptio:

individual studies that are going to be

pooled are relatively homogeneous, both

clinically and statistically, to provide a

meaningful central tendency effect

estimate


Effect sizes and p valuesForms of research findings suitable to meta-analysis:• Central tendency research:

– incidence or prevalence rates– mean (standard error)

• Pre-post contrasts:– changes in continuous or categorical variables

• Group contrasts:– experimentally created groups:

• comparison of outcomes between experimental and control groups

– naturally or non-experimentally occurring groups• treatment, prognostic or diagnostic features

• Association between variables:– correlation coefficients– regression coefficients


Effect sizes and p values• The effect size makes meta-analysis possible:

– it is the “dependent variable”– it standardizes findings across studies such that they can be

directly compared

• Any standardized index can be an “effect size” as long as it meets the following:– is comparable across studies (generally requires

standardization)– represents the magnitude and direction of the relationship of

interest– is independent of sample size

• We identify p values (for effect) for measuring alpha error for hypothesis testing and corresponding confidence intervals


Continous variables

• Continous variables can be pooled with

– Weighted mean differences (WMD), if the

same variable is used across studies

– Standardized mean differences (SMD), if

similar but not identical variables are used

– Inverse variance weighting, if only point

estimates and standard errors are available


Relative risks• Relative risks (RR) are defined as the ratio

of incidence rates, and are thus used for dichotomic variables)

• What is the meaning of RR:– RR=1 means no difference in risk– RR<1 means reduced risk in group 1 vs 2– RR>1 means increased risk in group 1 vs 2

• RRs are easier to interpret but are less userfriendly from a statistical point of view (RRAvsB≠1/RRBvsA) and may appear over-optimistic


Odds ratios• Odds ratios (OR) are defined as the

ratio of the odds (P/[1-P]) and also used for dichotomic variables

• When prevalences are low, they are a good approximation of RR

• They behave similarly to RR (OR=1 means no difference in risk, …)

• ORs are less easy to interpret but more flexible from a statistical point of view (ORAvsB=1/ORBvsA), yet also overoptimistic


Risk differences and number needed to treat/harm

• The risk difference (RD), ie absolute risk difference, is the difference between the incidence of events in the experimental vs control groups

• The RD is theoretically the most clinically relevant statistics, but changes too much with disease prevalence

• The number to treat (NNT), defined as 1/RD, identifies the number of patients that we need to treat with the experimental therapy to avoid one event*

• The NNT is the most clinically meaningful parameter to express the impact of a treatment on a dichotomic outcome (eg death), but has the same limits of RD

*Numbers needed to harm (NNH) similarly express the number of patients that we have to treat with the experimental therapy to cause one adverse event


RR, OR or RD/NNT?

OR RR RD/NNT

Communication - + ++

Consistency + ++ -

Mathematics ++ - -


Our advice• Both RR and OR can be your first choice statistics for

uncommon events

• For common events, the OR is clearly less informative than the RR for the busy reader

• Complete your analyses by reporting RD and/or NNT for the sake of clarity

• Fixed effect methods are quite fine for homogeneous/ consistent data

• Random effect methods may be more appropriate for heterogeneous/inconsistent data, but often meta-regression (or even refraining from meta-analysis at all) might be the best option


Small study bias• Publication bias (eg the lower likelihood of

being published for studies with negative findings, or those originating in non-English speaking countries) may bias the results of a meta-analysis

• Other types of small study bias may undermine the validity of a meta-analysis

• A number of tests, analogical (eg the funnel plot) or analytical (eg Egger’s or Peter’s) have been proposed to appraise the likelihood of such small study bias

Peters et al, JAMA 2006


Statistical heterogeneity

• Statistical heterogeneity may be suspected

by inspecting tables (summary estimates/SE)

and forest plots, or analytically

• Chi-square, Breslow, or Cochran tests are

most commonly used

• While a 2-tailed p=0.05 is used for cut-off for

hypothesis testing of effect, a 2-tailed p=0.10

is conventionally chosen for heterogeneity


Statistical inconsistency

• Statistical inconsistency (I2) has been recently introduced to overcome the risk of alpha and beta error of standard tests for statistical heterogeneity

• It is computed as [(Q – df)/Q] x 100%, where Q is the chi-squared statistic and df is its degrees of freedom

• I2 values of 25% suggest low inconsistency, 50% moderate inconsistency, and 75% severe inconsistency

Higgins et al, BMJ 2003


Statistical packages• EasyMA (http://www.spc.univ-lyon1.fr/easyma.net/)

• RevMan (http://www.cochrane.org)

» For meta-analyses of medical interventions

• Meta-Test ([email protected])

• Meta-DiSc (http://www.hrc.es/investigacion/metadisc.html)

» For meta-analyses of diagnostic tests

• U of Pittsburgh (http://www.pitt.edu/~super1/lecture/lec1171/index.htm)

• FastPro• NCSS• SAS• SPSS• Stata• WEasyMA

FR

EE

WA

RE

S!

No

t fo

r fr

ee


Typical Revman outputReview: Late percutaneous coronary intervention for infarct-related artery occlusionComparison: 01 Late percutaneous coronary intervention vs best medical therapy for infarct-related artery occlusion Outcome: 01 Death

Study PCI Medical Rx OR (random) OR (random)or sub-category n/N n/N 95% CI 95% CI O - E Variance

TOPS 0/42 0/45 Not estimable 0.00 0.00 TOMIIS 1/25 1/19 0.75 [0.04, 12.82] 0.00 2.10 Horie 1/44 5/39 0.16 [0.02, 1.42] 0.00 1.25 TOAT 2/32 1/34 2.20 [0.19, 25.52] 0.00 1.56 Zeymer et al 6/145 17/151 0.34 [0.13, 0.89] 0.00 0.24 DECOPI 6/109 7/103 0.80 [0.26, 2.46] 0.00 0.33 BRAVE-2 4/182 8/183 0.49 [0.15, 1.66] 0.00 0.39 Silva et al 0/18 2/18 0.18 [0.01, 3.99] 0.00 2.51

Total (95% CI) 597 592 0.48 [0.28, 0.85]Total events: 20 (PCI), 41 (Medical Rx)Test for heterogeneity: Chi² = 4.25, df = 6 (P = 0.64), I² = 0%Test for overall effect: Z = 2.53 (P = 0.01)

0.1 0.2 0.5 1 2 5 10

Favours PCI Favours medical Rx

Review: Late percutaneous coronary intervention for infarct-related artery occlusionComparison: 01 Late percutaneous coronary intervention vs best medical therapy for infarct-related artery occlusion Outcome: 01 Death

Study PCI Medical Rx OR (fixed) OR (fixed)or sub-category n/N n/N 95% CI 95% CI O - E Variance

TOPS 0/42 0/45 Not estimable 0.00 0.00 TOMIIS 1/25 1/19 0.75 [0.04, 12.82] 0.00 2.10 Horie 1/44 5/39 0.16 [0.02, 1.42] 0.00 1.25 TOAT 2/32 1/34 2.20 [0.19, 25.52] 0.00 1.56 Zeymer et al 6/145 17/151 0.34 [0.13, 0.89] 0.00 0.24 DECOPI 6/109 7/103 0.80 [0.26, 2.46] 0.00 0.33 BRAVE-2 4/182 8/183 0.49 [0.15, 1.66] 0.00 0.39 Silva et al 0/18 2/18 0.18 [0.01, 3.99] 0.00 2.51

Total (95% CI) 597 592 0.47 [0.27, 0.80]Total events: 20 (PCI), 41 (Medical Rx)Test for heterogeneity: Chi² = 4.25, df = 6 (P = 0.64), I² = 0%Test for overall effect: Z = 2.75 (P = 0.006)

0.1 0.2 0.5 1 2 5 10

Favours PCI Favours medical Rx


A few references• Biondi-Zoccai GGL et al. Parallel hierarchy of scientific studies in cardiovascular medicine. Ital Heart J 2003; 4: 819-20• Biondi-Zoccai GGL et al. Compliance with QUOROM and quality of reporting of overlapping meta-analyses on the role of

acetylcysteine in the prevention of contrast associated nephropathy: case study. BMJ 2006;332:202-209• Biondi-Zoccai GGL et al. A practical algorithm for systematic reviews in cardiovascular medicine. Ital Heart J 2004;5:486 -7• Bucher HC et al. The results of direct and indirect treatment comparisons in meta-analysis of randomized controlled trials. J

Clin Epidemiol 1997;50:683– 9• Cappelleri JC et al. Large trials vs meta-analysis of smaller trials: how do their results compare? JAMA 1996; 276: 1332-8• Clarke M et al, eds. Cochrane reviewers’ handbook 4.2.0. (www.cochrane.org/resources/handbook/handbook.pdf)• Cooper H et al, eds. The handbook of research synthesis. New York, NY: Russell Sage Foundation, 1994• Cucherat M et al. EasyMA: a program for the meta-analysis of clinical trials. Comput Methods Programs Biomed

1997;53:187- 90• Egger M et al, eds. Systematic reviews in health care: meta-analysis in context. 2nd ed. London: BMJ Publishing Group,

2001• Glass G. Primary, secondary and meta-analysis of research. Educ Res 1976;5:3-8• Glasziou P et al. Systematic reviews in health care. A practical guide. Cambridge: Cambridge University Press, 2001• Guyatt G et al, eds. Users’ guides to the medical literature. A manual for evidence-based clinical practice. Chicago, IL: AMA

Press, 2002• Higgins JPT et al. Measuring inconsistency in meta-analyses. BMJ 2003;327:557 – 60• Lau J et al. Summing up evidence: one answer is not always enough. Lancet 1998;351:123 -7• Moher D et al. Improving the quality of reports of meta-analyses of randomised controlled trials: the QUORUM statement.

Lancet 1999; 354: 1896-900• Petitti DB. Meta-analysis, decision analysis, and cost-effectiveness analysis: methods for quantitative synthesis in medicine.

New York, NY: Oxford University Press, 2000• Song F et al. Validity of indirect comparison for estimating efficacy of competing interventions: empirical evidence from published meta-analysis. BMJ 2003;326:472• Thompson SG et al. How should meta-regression analyses undertaken and interpreted? Stat Med 2002;21:1559-73


Take home messages• The validity of a meta-analysis refers to the

soundness of the original studies and the

procedures used to combine them (if appropriate)

• Dozens of potential validity threats have been

identified, and should always be borne in mind

• Given its current pivotal role in the hierarchy of

clinical evidence, all clinical decision-makers should

have a working knowledge of how to appraise

and/or conduct a systematic review/meta-analysis


Thank you for your attention!

For any correspondence: [email protected]

For further slides on these topics feel free to visit the metcardio.org website: http://www.metcardio.org/slides.html

Documents

[email protected] Value and Limitations of Meta-Analysis in the Era of Evidence-Based Medicine Giuseppe Biondi-Zoccai, MD Division