SlideShare a Scribd company logo
1 of 39
Meta-analysis II

Adrian V. Hernandez, M.D., Ph.D.
     Assistant Professor of Medicine
      Quantitative Health Sciences


           October 21, 2010
OUTLINE SECOND PART

SECOND PART: 50 MINUTES
• Analysis (models, methods, heterogeneity, publication
 bias, quality, subgroup analysis)
• Reporting of meta-analysis (PRISMA, MOOSE guidelines)




                                         Meta-analysis II l
                                         October 10, 2012 l
MODELS

 • Fixed effects
        -Assumption: Effect is the same across
              studies and differences due to chance
        -Common effect unknown
        -Objective: Estimation of common effect with
              more precision
        -Pool studies using weights ↔ sample size




                                    Meta-analysis II l
                                    October 10, 2012 l
MODELS (2)

 • Random effects
       -Assumption: Effect is different across
             studies and there is an average effect
       -Average effect unknown
       -Objective: Estimation of average effect of
             studies
       -Pool studies using similar weights




                                    Meta-analysis II l
                                    October 10, 2012 l
When are effects similar between models?
  Large effect
  Balanced arms
  Study sizes similar
  Low heterogeneity of effects

           HF Fixed MH       1.59 (1.34-1.89)
               Random        1.56 (1.32-1.86)

           Edema Fixed MH    2.04 (1.85-2.26)
                   Random    2.41 (1.91-3.04)


                            Hernandez AV et al. 2010 (submitted)
                                     Meta-analysis II l
                                     October 10, 2012 l
METHODS TO COMBINE STUDY EFFECTS

 • Inverse Variance (IV)
       Common, flexible
       Binary/continuous data
       Log OR, Log RR, log HR, standardized ratios




                                  Meta-analysis II l
                                  October 10, 2012 l
METHODS TO COMBINE STUDY EFFECTS (2)

 • Mantel-Haenzel (MH)
       Binary outcomes only
       Special cases: sparse outcomes, unbalanced
              arms
       Correction for zeros in arms
       Other situations: Effects similar to IV


                                      Meta-analysis II l
                                      October 10, 2012 l
METHODS TO COMBINE STUDY EFFECTS (3)

 • Peto
          Binary outcomes only, a few outcomes
          Small effect
          Balanced arms
          No correction for zeros in arms

                            HF Fixed MH     1.59 (1.34-1.89)
                                Peto        1.59 (1.34-1.88)


                                       Meta-analysis II l
                                       October 10, 2012 l
Meta-analysis II l
October 10, 2012 l
Meta-analysis II l
October 10, 2012 l
Meta-analysis II l
October 10, 2012 l
METHODS TO COMBINE STUDY EFFECTS (4)

 • DerSimonian and Laird
       Any type of effect measures
       Random model
       Larger CI of the pooled effect
       More weight to smaller studies




                                        Meta-analysis II l
                                        October 10, 2012 l
HETEROGENEITY

• Degree of dissimilarity in effects of individual studies
• Why?
       Participants
       Interventions
       Co-interventions
       Outcomes
       Biases of studies (according to hierarchy), etc.

                                        Meta-analysis II l
                                        October 10, 2012 l
Meta-analysis II l
October 10, 2012 l
HETEROGENEITY: Pseudo-tests

    • Eyeballing forest plots
    • Point estimates
    • Significance level
    • Confidence intervals




                                Meta-analysis II l
                                October 10, 2012 l
HETEROGENEITY: Test it!
• Cochrane Q test (Х2, p)
      AND
• I2: Amount of heterogeneity (0-100%)
    It needs 95% CI (<25% Low; 25-50% Mod; >50 High)
                       (Ioannidis JP. J Eval Clin Pract 2008:14:951-7)



 Bad news for both: Low power
 Bad news for I2: No software for CIs

                                         Meta-analysis II l
                                         October 10, 2012 l
Meta-analysis II l
October 10, 2012 l
HETEROGENEITY: Address it!
    • Check data again
    • Do not perform a meta-analysis
    • Explore vs. ignore
    • Use random-effects models
    • Change effect measure (e.g. MD →SMD)
    • Exclude studies




                                  Meta-analysis II l
                                  October 10, 2012 l
EXPLORING HETEROGENEITY

• Subgroup analysis
    Exploratory only
    Low power to detect significant effects
    Better pre-specify in protocol
    Generates hypotheses


   → Editors and reviewers like subgroup analysis

                                      Meta-analysis II l
                                      October 10, 2012 l
EXPLORING HETEROGENEITY (2)


• How to perform subgroup analysis?
   → By baseline characteristics (e.g. age, gender)
   → By quality
   → By sample size
   → By follow-up time




                                    Meta-analysis II l
                                    October 10, 2012 l
Meta-analysis II l
October 10, 2012 l
Meta-analysis II l
October 10, 2012 l
Use of thiazolidinediones and risk of heart failure and
     peripheral edema in patients at high risk of diabetes and
                         type 2 diabetes:
        A systematic review and meta-analysis of placebo-
                   controlled randomized trials

Follow-up                             OR (95% CI)          RR (95% CI)
     ≥12 months
              MH                     1.57 [1.31-1.87]    1.50 [1.28-1.76]
              Random                 1.67 [1.16-2.40]    1.66 [1.10-2.50]
     <12 months
              MH                     2.71 [0.94-7.79]    2.68 [0.93-7.67]
              Random                 2.56 [0.88-7.44]    2.52 [0.88-7.25]

                                     Hernandez AV et al. 2010 (Submitted)




                                              Meta-analysis II l
                                              October 10, 2012 l
EXPLORING HETEROGENEITY (3)


• Meta-regression
   →Evaluates factors that explain heterogeneity of


        effects


   →Bad news: Low power
                  Lack of data

                                   Meta-analysis II l
                                   October 10, 2012 l
Meta-analysis II l
October 10, 2012 l
EVALUATION OF PUBLICATION BIAS
 Funnel Plot: Size effect vs. SE/SS ; Asymmetry?




                      De Luca G et al. Am Heart J 2007; 153:343-53
                                        Meta-analysis II l
                                        October 10, 2012 l
EVALUATION OF PUBLICATION BIAS (2)

 Asymmetrical: Only due to publication bias?




                    Dentali F et al. Ann Intern Med 2007; 146: 278-88
                                          Meta-analysis II l
                                          October 10, 2012 l
EVALUATION OF PUBLICATION BIAS (3)

 Pseudo-test: Visual inspection of funnel plot


 Test!: Begg-Mazumdar test
       Asymmetry regression test
       Kendall test, etc
       → Bad news: All have low power



                                    Meta-analysis II l
                                    October 10, 2012 l
EVALUATION OF PUBLICATION BIAS (4)




 In 95% of MAs, the use of asymmetry regression
 tests is inappropriate:
       → Highly heterogeneous (I2 > 50%)
       → <10 studies


                                 Meta-analysis II l
                                 October 10, 2012 l
QUALITY OF STUDIES: Observational

    • Design: Prospective cohort
              Retrospective cohort
              Case-control
    • Quality of measurement of factors
    • Patient enrollment (consecutive vs no)




                                     Meta-analysis II l
                                     October 10, 2012 l
QUALITY OF STUDIES: RCTs
    • Difficult to define
           Design/conduct/analysis?
           Clinical relevance?
           Reporting?


    • Several scales: 39
    Egger M et al. Systematic reviews in health care. Meta-analysis in
                    context. 2nd Edition, BMJ London 2001. pp87-108.



                                                   Meta-analysis II l
                                                   October 10, 2012 l
QUALITY OF STUDIES: RCTs (2)

 TC Chalmers et al. Control Clin Trials 1981; 2: 31-49
    30 items, complex
       - Internal validity (R, Blinding, Attrition,
               stat analysis)
       - External validity
       - Data presentation/Organizational aspects


 → Low weight to internal validity

                                        Meta-analysis II l
                                        October 10, 2012 l
QUALITY OF STUDIES: RCTs (3)
 AR Jadad et al. Control Clin Trials 1996; 17: 1-12
  5 items, 5 points, ≥3 high quality
    - Randomization: Description of method? 1
                      Appropriate? 1
    - Double blinding: Description of method? 1
                       Appropriate? 1
    - Description of withdrawal/dropouts? 1


 → More weight to reporting than methodology
                                  Meta-analysis II l
                                  October 10, 2012 l
REPORTING


• RCTs: PRISMA (Preferred Reporting Items for
       Systematic reviews and Meta-Analyses)
• Observational: MOOSE (Meta-analysis Of
       Observational Studies in Epidemiology)




                                  Meta-analysis II l
                                  October 10, 2012 l
PRISMA




Replace and improve the old QUOROM (1999) guidelines
27 items
Title, Abstract, Introduction, Methods, Results, Discussion
  and Funding.

                                     Meta-analysis II l
                                     October 10, 2012 l
PRISMA flow chart




                    Meta-analysis II l
                    October 10, 2012 l
PRISMA guidelines: Improvements

    • Clear description of objective (PICOS)
    • Improve description of selection of studies
    (search strategy). Publish at least one.
    • Improve evaluation of risk of bias within studies
    (quality)
    • Improve description and evaluation of
    publication bias.
    • Suggest publishing the protocol of the MA



                                        Meta-analysis II l
                                        October 10, 2012 l
MOOSE




35 items, 1 point to each
Background, Search Strategy, Methods, Results,
      Discussion & Conclusion


                                  Meta-analysis II l
                                  October 10, 2012 l
Meta-analysis II l
October 10, 2012 l

More Related Content

Similar to Second Part.MA.Oct202010

Research methods (variables & hypothesis)
Research methods (variables & hypothesis)Research methods (variables & hypothesis)
Research methods (variables & hypothesis)Crystal Delosa
 
Experimental, Quasi experimental, Single-Case, and Internet-based Researches ...
Experimental, Quasi experimental, Single-Case, and Internet-based Researches ...Experimental, Quasi experimental, Single-Case, and Internet-based Researches ...
Experimental, Quasi experimental, Single-Case, and Internet-based Researches ...Hatice Çilsalar
 
httpwww.jmde.com Articles Journal of MultiDiscipli.docx
 httpwww.jmde.com  Articles Journal of MultiDiscipli.docx httpwww.jmde.com  Articles Journal of MultiDiscipli.docx
httpwww.jmde.com Articles Journal of MultiDiscipli.docxMARRY7
 
Causal comparative research
Causal comparative researchCausal comparative research
Causal comparative researchNoor Hasmida
 
No support for declining effect sizes over time - SSSP 2014
No support for declining effect sizes over time  - SSSP 2014No support for declining effect sizes over time  - SSSP 2014
No support for declining effect sizes over time - SSSP 2014Chris Martin
 
No support for declining effect sizes over time - Chris C Martin and Gregory ...
No support for declining effect sizes over time - Chris C Martin and Gregory ...No support for declining effect sizes over time - Chris C Martin and Gregory ...
No support for declining effect sizes over time - Chris C Martin and Gregory ...Chris Martin
 

Similar to Second Part.MA.Oct202010 (6)

Research methods (variables & hypothesis)
Research methods (variables & hypothesis)Research methods (variables & hypothesis)
Research methods (variables & hypothesis)
 
Experimental, Quasi experimental, Single-Case, and Internet-based Researches ...
Experimental, Quasi experimental, Single-Case, and Internet-based Researches ...Experimental, Quasi experimental, Single-Case, and Internet-based Researches ...
Experimental, Quasi experimental, Single-Case, and Internet-based Researches ...
 
httpwww.jmde.com Articles Journal of MultiDiscipli.docx
 httpwww.jmde.com  Articles Journal of MultiDiscipli.docx httpwww.jmde.com  Articles Journal of MultiDiscipli.docx
httpwww.jmde.com Articles Journal of MultiDiscipli.docx
 
Causal comparative research
Causal comparative researchCausal comparative research
Causal comparative research
 
No support for declining effect sizes over time - SSSP 2014
No support for declining effect sizes over time  - SSSP 2014No support for declining effect sizes over time  - SSSP 2014
No support for declining effect sizes over time - SSSP 2014
 
No support for declining effect sizes over time - Chris C Martin and Gregory ...
No support for declining effect sizes over time - Chris C Martin and Gregory ...No support for declining effect sizes over time - Chris C Martin and Gregory ...
No support for declining effect sizes over time - Chris C Martin and Gregory ...
 

More from Adrian V. Hernandez

More from Adrian V. Hernandez (7)

Clase V.27 Oct2012
Clase V.27 Oct2012Clase V.27 Oct2012
Clase V.27 Oct2012
 
Clase IV.20 Oct2012
Clase IV.20 Oct2012Clase IV.20 Oct2012
Clase IV.20 Oct2012
 
Clase III.13 Oct2012
Clase III.13 Oct2012Clase III.13 Oct2012
Clase III.13 Oct2012
 
Diseño y Conducción de Meta-análisis
Diseño y Conducción de Meta-análisisDiseño y Conducción de Meta-análisis
Diseño y Conducción de Meta-análisis
 
Clase II.Oct62012
Clase II.Oct62012Clase II.Oct62012
Clase II.Oct62012
 
Clase I.Sept292012
Clase I.Sept292012Clase I.Sept292012
Clase I.Sept292012
 
Analisis critico de ensayos de no-inferioridad en pacientes HIV virgenes a tt...
Analisis critico de ensayos de no-inferioridad en pacientes HIV virgenes a tt...Analisis critico de ensayos de no-inferioridad en pacientes HIV virgenes a tt...
Analisis critico de ensayos de no-inferioridad en pacientes HIV virgenes a tt...
 

Second Part.MA.Oct202010

  • 1. Meta-analysis II Adrian V. Hernandez, M.D., Ph.D. Assistant Professor of Medicine Quantitative Health Sciences October 21, 2010
  • 2. OUTLINE SECOND PART SECOND PART: 50 MINUTES • Analysis (models, methods, heterogeneity, publication bias, quality, subgroup analysis) • Reporting of meta-analysis (PRISMA, MOOSE guidelines) Meta-analysis II l October 10, 2012 l
  • 3. MODELS • Fixed effects -Assumption: Effect is the same across studies and differences due to chance -Common effect unknown -Objective: Estimation of common effect with more precision -Pool studies using weights ↔ sample size Meta-analysis II l October 10, 2012 l
  • 4. MODELS (2) • Random effects -Assumption: Effect is different across studies and there is an average effect -Average effect unknown -Objective: Estimation of average effect of studies -Pool studies using similar weights Meta-analysis II l October 10, 2012 l
  • 5. When are effects similar between models?  Large effect  Balanced arms  Study sizes similar  Low heterogeneity of effects HF Fixed MH 1.59 (1.34-1.89) Random 1.56 (1.32-1.86) Edema Fixed MH 2.04 (1.85-2.26) Random 2.41 (1.91-3.04) Hernandez AV et al. 2010 (submitted) Meta-analysis II l October 10, 2012 l
  • 6. METHODS TO COMBINE STUDY EFFECTS • Inverse Variance (IV) Common, flexible Binary/continuous data Log OR, Log RR, log HR, standardized ratios Meta-analysis II l October 10, 2012 l
  • 7. METHODS TO COMBINE STUDY EFFECTS (2) • Mantel-Haenzel (MH) Binary outcomes only Special cases: sparse outcomes, unbalanced arms Correction for zeros in arms Other situations: Effects similar to IV Meta-analysis II l October 10, 2012 l
  • 8. METHODS TO COMBINE STUDY EFFECTS (3) • Peto Binary outcomes only, a few outcomes Small effect Balanced arms No correction for zeros in arms HF Fixed MH 1.59 (1.34-1.89) Peto 1.59 (1.34-1.88) Meta-analysis II l October 10, 2012 l
  • 12. METHODS TO COMBINE STUDY EFFECTS (4) • DerSimonian and Laird Any type of effect measures Random model Larger CI of the pooled effect More weight to smaller studies Meta-analysis II l October 10, 2012 l
  • 13. HETEROGENEITY • Degree of dissimilarity in effects of individual studies • Why? Participants Interventions Co-interventions Outcomes Biases of studies (according to hierarchy), etc. Meta-analysis II l October 10, 2012 l
  • 15. HETEROGENEITY: Pseudo-tests • Eyeballing forest plots • Point estimates • Significance level • Confidence intervals Meta-analysis II l October 10, 2012 l
  • 16. HETEROGENEITY: Test it! • Cochrane Q test (Х2, p) AND • I2: Amount of heterogeneity (0-100%) It needs 95% CI (<25% Low; 25-50% Mod; >50 High) (Ioannidis JP. J Eval Clin Pract 2008:14:951-7)  Bad news for both: Low power  Bad news for I2: No software for CIs Meta-analysis II l October 10, 2012 l
  • 18. HETEROGENEITY: Address it! • Check data again • Do not perform a meta-analysis • Explore vs. ignore • Use random-effects models • Change effect measure (e.g. MD →SMD) • Exclude studies Meta-analysis II l October 10, 2012 l
  • 19. EXPLORING HETEROGENEITY • Subgroup analysis  Exploratory only  Low power to detect significant effects  Better pre-specify in protocol  Generates hypotheses → Editors and reviewers like subgroup analysis Meta-analysis II l October 10, 2012 l
  • 20. EXPLORING HETEROGENEITY (2) • How to perform subgroup analysis? → By baseline characteristics (e.g. age, gender) → By quality → By sample size → By follow-up time Meta-analysis II l October 10, 2012 l
  • 23. Use of thiazolidinediones and risk of heart failure and peripheral edema in patients at high risk of diabetes and type 2 diabetes: A systematic review and meta-analysis of placebo- controlled randomized trials Follow-up OR (95% CI) RR (95% CI) ≥12 months MH 1.57 [1.31-1.87] 1.50 [1.28-1.76] Random 1.67 [1.16-2.40] 1.66 [1.10-2.50] <12 months MH 2.71 [0.94-7.79] 2.68 [0.93-7.67] Random 2.56 [0.88-7.44] 2.52 [0.88-7.25] Hernandez AV et al. 2010 (Submitted) Meta-analysis II l October 10, 2012 l
  • 24. EXPLORING HETEROGENEITY (3) • Meta-regression →Evaluates factors that explain heterogeneity of effects →Bad news: Low power Lack of data Meta-analysis II l October 10, 2012 l
  • 26. EVALUATION OF PUBLICATION BIAS Funnel Plot: Size effect vs. SE/SS ; Asymmetry? De Luca G et al. Am Heart J 2007; 153:343-53 Meta-analysis II l October 10, 2012 l
  • 27. EVALUATION OF PUBLICATION BIAS (2) Asymmetrical: Only due to publication bias? Dentali F et al. Ann Intern Med 2007; 146: 278-88 Meta-analysis II l October 10, 2012 l
  • 28. EVALUATION OF PUBLICATION BIAS (3) Pseudo-test: Visual inspection of funnel plot Test!: Begg-Mazumdar test Asymmetry regression test Kendall test, etc → Bad news: All have low power Meta-analysis II l October 10, 2012 l
  • 29. EVALUATION OF PUBLICATION BIAS (4) In 95% of MAs, the use of asymmetry regression tests is inappropriate: → Highly heterogeneous (I2 > 50%) → <10 studies Meta-analysis II l October 10, 2012 l
  • 30. QUALITY OF STUDIES: Observational • Design: Prospective cohort Retrospective cohort Case-control • Quality of measurement of factors • Patient enrollment (consecutive vs no) Meta-analysis II l October 10, 2012 l
  • 31. QUALITY OF STUDIES: RCTs • Difficult to define Design/conduct/analysis? Clinical relevance? Reporting? • Several scales: 39 Egger M et al. Systematic reviews in health care. Meta-analysis in context. 2nd Edition, BMJ London 2001. pp87-108. Meta-analysis II l October 10, 2012 l
  • 32. QUALITY OF STUDIES: RCTs (2) TC Chalmers et al. Control Clin Trials 1981; 2: 31-49 30 items, complex - Internal validity (R, Blinding, Attrition, stat analysis) - External validity - Data presentation/Organizational aspects → Low weight to internal validity Meta-analysis II l October 10, 2012 l
  • 33. QUALITY OF STUDIES: RCTs (3) AR Jadad et al. Control Clin Trials 1996; 17: 1-12 5 items, 5 points, ≥3 high quality - Randomization: Description of method? 1 Appropriate? 1 - Double blinding: Description of method? 1 Appropriate? 1 - Description of withdrawal/dropouts? 1 → More weight to reporting than methodology Meta-analysis II l October 10, 2012 l
  • 34. REPORTING • RCTs: PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses) • Observational: MOOSE (Meta-analysis Of Observational Studies in Epidemiology) Meta-analysis II l October 10, 2012 l
  • 35. PRISMA Replace and improve the old QUOROM (1999) guidelines 27 items Title, Abstract, Introduction, Methods, Results, Discussion and Funding. Meta-analysis II l October 10, 2012 l
  • 36. PRISMA flow chart Meta-analysis II l October 10, 2012 l
  • 37. PRISMA guidelines: Improvements • Clear description of objective (PICOS) • Improve description of selection of studies (search strategy). Publish at least one. • Improve evaluation of risk of bias within studies (quality) • Improve description and evaluation of publication bias. • Suggest publishing the protocol of the MA Meta-analysis II l October 10, 2012 l
  • 38. MOOSE 35 items, 1 point to each Background, Search Strategy, Methods, Results, Discussion & Conclusion Meta-analysis II l October 10, 2012 l