Session I: Basic Methods of Meta-Analysis
Short Introduction to R
James Carpenter1, Ulrike Krahn2,3, Gerta R¨ucker4, Guido Schwarzer4
1London School of Hygiene and Tropical Medicine & MRC Clinical Trials Unit, London, UK 2Institute of Medical Biostatistics, Epidemiology and Informatics, Mainz, Germany 3Institute of Medical Informatics, Biometry and Epidemiology, Duisburg-Essen, Germany
4Institute for Medical Biometry and Statistics, Freiburg, Germany
IBC Short Course Florence, 6 July 2014
1
Plan of the Session
At the end of this session theaim is that you should understand
I the basic principles of meta-analysis;
I how R works.
Theobjectivesare that you are able to:
I use the help system and read data into R;
I conduct a meta-analysis with binary outcomes;
I summarize the main result of a meta-analysis to a medical audience.
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 2
Introduction R in Action Meta-Analysis Binary Data Summary References
What is a Meta-Analysis?
See Egger and Smith (1997), BMJ Definitions:I Term “Meta”:
Implies occuring later, more comprehensive, new but related discipline which critically deals with original discipline
I Term “Meta-Analysis”:
Statistically combining and analysing data from separate studies
I Term “Systematic Review”:
Any type of review utilising strategies to avoid bias Meta-analysis:
I May or may not be part of a systematic review
I Medical Subject Heading (MeSH) in Medline
Introduction R in Action Meta-Analysis Binary Data Summary References
Why Conduct a Meta-Analysis?
Egger and Smith (1997), BMJ
I More objective appraisal of evidence which may lead to resolution of
uncertainty and disagreement
I Reduce probability of false negative results and thus prevent undue
delays in introduction of effective treatments into practice
I Heterogeneity between study results may be explored (and sometimes
explained)
I Allows testing of a priori hypotheses regarding treatment effects in
What is R?
I General purpose statistical package (http://www.r-project.org/)
I Based on statistical programming language S (→S-PLUS)
I Almost 20 years old, actively developed and maintained
I Available for Windows, Linux, Unix, Mac OS
I Released under the GNU General Public License (GPL) version 2
(or any later version)
I Licence costs:
0
e
/ 0$
I R can be used in regulated clinical trial environments
(http://www.r-project.org/doc/R-FDA.pdf)
I More than 5000 add-on packages available on CRAN
(http://cran.at.r-project.org/)
I Short introductions / reviews of add-on packages inThe R Journal
(http://journal.r-project.org/) – successor of R News
I Mailing lists: http://stat.ethz.ch/mailman/listinfo/(R-help)
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 5
R – Used as an overgrown calculator
> 2 + 2 [1] 4 > 4 -+ 1 [1] 3 > exp(1) [1] 2.718282 > x = 2.25 > x + x [1] 4.5Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 6
Introduction R in Action Meta-Analysis Binary Data Summary References
R – Working with vectors
> 1:10 [1] 1 2 3 4 5 6 7 8 9 10 > c(1:4, 6:10) [1] 1 2 3 4 6 7 8 9 10 > y = c(1,4,9,16) > sqrt(y) [1] 1 2 3 4 > sqrt(y)[4] [1] 4> y^2 # same result: y**2
[1] 1 16 81 256
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 7
Introduction R in Action Meta-Analysis Binary Data Summary References
R – Calculate mean and standard deviance
> x = 1:10 > sum(x) / length(x) [1] 5.5 > mean(x) [1] 5.5 > sqrt(sum((x-mean(x))^2)/(length(x)-1)) [1] 3.02765 > sd(x) [1] 3.02765R – Missing values
I NA: Not Available > x = c(1:5, NA) > mean(x) [1] NA > # Getting help:> help(mean) # Show documentation on R command mean > ?mean # Show documentation on R command mean > help.start() # Show documentation in webbrowser > help("+") # Show help on arithmetic operator + > ?"+" # Show help on arithmetic operator + > mean(x, na.rm=TRUE)
[1] 3
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 9
Example: Aggressive Non-Hodgkin Lymphoma
Greb et al. (2008), Cochrane Database Syst Rev1, CD004024:
I Cochrane Review including 15 randomised controlled trials (RCTs)
I Adult patients with aggressive non-Hodgkin lymphoma
I First line treatment with high-dose chemotherapy (HDCT) versus
conventional chemotherapy
I Primary outcome:
Overall survival (14 RCTs, 2444 patients)
I Secondary outcome:
Complete response (14 RCTs, 2126 patients)
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 10
Introduction R in Action Meta-Analysis Binary Data Summary References
Aggressive Non-Hodgkin Lymphoma – Forestplot
Study De Souza Gianni Gisselbrecht Haioun Intragumtornchai Kaiser Kluin−Nelemans Martelli Martelli 2003 Milpied Rodriguez 2003 Santini Verdonck Vitolo 0.5 1 2 Hazard Ratio
Favours HDCT Favours control
HR 0.92 0.52 1.45 0.96 0.64 1.08 1.23 0.69 1.01 0.64 1.34 0.81 1.40 1.41 95%−CI [0.45; 1.89] [0.24; 1.11] [1.08; 1.93] [0.71; 1.30] [0.30; 1.36] [0.75; 1.55] [0.72; 2.08] [0.29; 1.65] [0.59; 1.73] [0.40; 1.05] [0.68; 2.65] [0.48; 1.37] [0.73; 2.67] [0.82; 2.41]
Introduction R in Action Meta-Analysis Binary Data Summary References
Meta-Analysis – Calculate a Weighted Mean
Weighted mean of estimated treatment effects in individual studies (Fleiss, 1993): ˆ θ= PK k=1wk·ˆθk PK k=1wk
I Estimated treatment effectθˆk in studyk (k = 1, ...,K)
I Weight wk correspond to information of study k
I Methods of meta-analysis differ in definition of weights
(especially fixed effect and random effects model)
> args(weighted.mean) function (x, w, ...) NULL
> weighted.mean(1:2, c(0.2, 0.8)) [1] 1.8
Fixed effect model – Inverse Variance Method
Fixed effect model:ˆ
θk =θ+σˆkk, ki.i.d.∼ N(0,1), k = 1, . . . ,K (1)
Maximum likelihood estimate under model (1) for given (ˆθk,σˆk): ˆ θF = PK k=1θˆk/σˆ2k PK k=11/σˆ2k = PK k=1wkˆθk PK k=1wk with weights wk = 1/σˆ2 k. Estimated variance of ˆθF: d Var (ˆθF) = 1 PK k=1wk = 1 PK k=11/σˆ2k
(1-α) confidence interval for ˆθF:
ˆ
θF ± z1−α2 S.E.(ˆθF) with S.E.(ˆθF) =
q d
Var(ˆθF)
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 13
Fixed Effect Model – Graphical Presentation
Odds ratio 0.1 0.2 0.5 1 2 5 10 Study 1 Study 2 Study 3 Study 4 Study 5 True effect
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 14
Introduction R in Action Meta-Analysis Binary Data Summary References
Aggressive Non-Hodgkin Lymphoma – Overall Survival
Hazard log(HR) SE(log(HR)) wk = Study Ratio (HR) 95%-CI (=ˆθk) (=σˆk) 1/σˆ2k De Souza 0.92 0.45 - 1.89 -0.0807 0.3672 7.42 Gianni 0.52 0.24 - 1.11 -0.6527 0.3850 6.75 Gisselbrecht 1.45 1.08 - 1.93 0.3683 0.1487 45.22 Intragumtornchai 0.96 0.71 - 1.30 -0.0387 0.1529 42.77 Kaiser 0.64 0.30 - 1.36 -0.4480 0.3852 6.74 Kluin-Nelemans 1.08 0.75 - 1.55 0.0761 0.1834 29.73 Martelli 1996 1.23 0.72 - 2.08 0.2031 0.2697 13.75 Martelli 2003 0.69 0.29 - 1.65 -0.3773 0.4473 5.00 Milpied 1.01 0.59 - 1.73 0.0087 0.2748 13.24 Rodriguez 2003 0.64 0.40 - 1.05 -0.4402 0.2481 16.25 Santini 1998 1.34 0.68 - 2.65 0.2921 0.3482 8.25 Santini-2 0.81 0.48 - 1.37 -0.2141 0.2697 13.75 Verdonck 1.40 0.73 - 2.67 0.3368 0.3290 9.24 Vitolo 1.41 0.82 - 2.41 0.3412 0.2749 13.23 Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 15
Introduction R in Action Meta-Analysis Binary Data Summary References
R packages for meta-analysis on CRAN
I rmeta (Lumley, 2012)
I Fixed and random effects meta-analysis (Mantel-Haenszel, Peto, DerSimonian-Laird)
I metafor (Viechtbauer, 2010)
I Tests for funnel plot asymmetry / Trim and fill method
I General linear (mixed-effects) model approach for meta-regression I Multivariate meta-analysis
I meta (Schwarzer, 2007)
I Tests for funnel plot asymmetry / Trim and fill method I Import data from RevMan 5 / Link to R package metafor
I mvmeta (Gasparrini, 2014)
I Multivariate meta-analysis and meta-regression on multiple outcomes
I metasens (Schwarzer et al., 2014; Carpenter et al., 2009)
I Advanced methods to model and adjust for bias in meta-analysis I Add-on package to R packagemeta/ replaces R packagecopas
I netmeta (R¨ucker et al., 2014)
I Network meta-analysis
I Add-on package to R packagemeta
R package meta
Function Comment
metabin Meta-analysis of binary outcome data
metacont Meta-analysis of continuous outcome data
metagen Generic inverse variance meta-analysis
metacor Meta-analysis of correlations
metainc Meta-analysis of incidence rates
metaprop Meta-analysis of single proportions
read.rm5 Import RevMan 5 data files (csv-files with special format)
metacr Meta-analysis of outcome data from Cochrane review
forest Forest plot
funnel Plot to assess funnel plot asymmetry
metabias Test for funnel plot asymmetry
trimfill Trim and fill method for meta-analysis
metareg Meta-regression (wrapper function to R package metafor)
... Cumulative meta-analysis / Influence analysis in meta-analysis
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 17
Meta-Analysis of Overall Survival
> os = read.csv("hd-os.txt", as.is=TRUE) > library(meta) > m1 = metagen(logHR, selogHR, + sm="HR", data=os, studlab=study, + comb.random=FALSE) > str(m1) List of 62 $ TE : num [1:14] -0.0807 -0.6527 0.3683 -0.0387 -0.448 ... $ seTE : num [1:14] 0.367 0.385 0.149 0.153 0.385 ...$ studlab : chr [1:14] "De Souza" "Gianni" "Gisselbrecht" "Haioun" ... $ w.fixed : num [1:14] 7.42 6.75 45.22 42.77 6.74 ... $ w.random : num [1:14] 6.18 5.71 20.34 19.83 5.7 ... $ TE.fixed : num 0.0441 $ seTE.fixed : num 0.0657 $ lower.fixed : num -0.0848 $ upper.fixed : num 0.173 $ zval.fixed : num 0.671 $ pval.fixed : num 0.502
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 18
Introduction R in Action Meta-Analysis Binary Data Summary References
Meta-Analysis of Survival Data – Print Function
> class(m1)[1] "metagen" "meta"
> m1 # Calls R function print.meta
HR 95%-CI %W(fixed) De Souza 0.9225 [0.4491; 1.8946] 3.21 Gianni 0.5206 [0.2448; 1.1073] 2.92 Gisselbrecht 1.4453 [1.0799; 1.9343] 19.55 Haioun 0.9620 [0.7129; 1.2982] 18.49 Intragumtornchai 0.6389 [0.3003; 1.3593] 2.91 Kaiser 1.0791 [0.7533; 1.5458] 12.85 Kluin-Nelemans 1.2252 [0.7222; 2.0786] 5.94 Martelli 0.6857 [0.2854; 1.6477] 2.16 Martelli 2003 1.0087 [0.5887; 1.7286] 5.72 Milpied 0.6439 [0.3960; 1.0471] 7.02 Rodriguez 2003 1.3392 [0.6768; 2.6500] 3.57 Santini 0.8073 [0.4758; 1.3696] 5.94 Verdonck 1.4005 [0.7349; 2.6688] 3.99
Introduction R in Action Meta-Analysis Binary Data Summary References
Meta-Analysis of Overall Survival – Summary Function
> summary(m1) # Calls summary.meta and print.summary.metaNumber of studies combined: k=14
HR 95%-CI z p.value
Fixed effect model 1.0451 [0.9187; 1.1888] 0.6709 0.5023 Quantifying heterogeneity:
tau^2 = 0.0270; H = 1.19 [1; 1.64]; I^2 = 29.9% [0%; 63%] Test of heterogeneity:
Q d.f. p.value
18.55 13 0.1379
Details on meta-analytical method: - Inverse variance method
Overall Survival – Forestplot
> forest(m1, hetstat=FALSE) # Calls function forest.meta Study
Fixed effect model
De Souza Gianni Gisselbrecht Haioun Intragumtornchai Kaiser Kluin−Nelemans Martelli Martelli 2003 Milpied Rodriguez 2003 Santini Verdonck Vitolo TE −0.08 −0.65 0.37 −0.04 −0.45 0.08 0.20 −0.38 0.01 −0.44 0.29 −0.21 0.34 0.34 seTE 0.3672 0.3850 0.1487 0.1529 0.3852 0.1834 0.2697 0.4473 0.2748 0.2481 0.3482 0.2697 0.3290 0.2749 0.5 1 2 Hazard Ratio HR 1.05 0.92 0.52 1.45 0.96 0.64 1.08 1.23 0.69 1.01 0.64 1.34 0.81 1.40 1.41 95%−CI [0.92; 1.19] [0.45; 1.89] [0.24; 1.11] [1.08; 1.93] [0.71; 1.30] [0.30; 1.36] [0.75; 1.55] [0.72; 2.08] [0.29; 1.65] [0.59; 1.73] [0.40; 1.05] [0.68; 2.65] [0.48; 1.37] [0.73; 2.67] [0.82; 2.41] W(fixed) 100% 3.2% 2.9% 19.5% 18.5% 2.9% 12.9% 5.9% 2.2% 5.7% 7.0% 3.6% 5.9% 4.0% 5.7%
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 21
Example: Aggressive Non-Hodgkin Lymphoma
Greb et al. (2008), Cochrane Database Syst Rev1, CD004024:
I Cochrane Review including 15 randomised controlled trials (RCTs)
I Adult patients with aggressive non-Hodgkin lymphoma
I First line treatment with high-dose chemotherapy (HDCT) versus
conventional chemotherapy
I Primary outcome:
Overall survival (14 RCTs, 2444 patients)
I Secondary outcome:
Complete response (14 RCTs, 2126 patients)
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 22
Introduction R in Action Meta-Analysis Binary Data Summary References
Aggressive Non-Hodgkin Lymphoma – Complete Response
HDCT Control Study
Events Total Events Total
De Souza 14 28 10 26 Gianni 46 48 35 50 Gisselbrecht 119 189 116 181 Intragumtornchai 10 23 9 25 Kaiser 110 158 97 154 Kluin-Nelemans 67 98 56 96 Martelli 1996 3 22 4 27 Martelli 2003 57 75 51 75 Milpied 74 98 56 99 Rodriguez 2003 39 55 30 53 Santini 1998 46 63 34 61 Santini-2 80 117 71 106 Verdonck 25 38 26 35 Vitolo 35 60 46 66
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 23
Introduction R in Action Meta-Analysis Binary Data Summary References
Milpied Study – Complete Response (CR)
CR no CR
HDCT 74 (a) 24 (b) 98 (a+b =nT)
Control 56 (c) 43 (d) 99 (c+d =nC)
130 (a+c) 67 (b+d) 197 (n)
Binary Data – Effect Measures
LetI pT: Experimental event rate ˆpT =a/(a+b)
I pC: Control event rate pˆC =c/(c +d)
Risk Ratio φ: φ=pT pC ˆ φ= pˆT ˆ pC Odds ratio ψ: ψ= pT 1−pT ! pC 1−pC ! =φ× 1−pC 1−pT ˆ ψ= a d b c (2) Risk Difference η: η=pT −pC ˆη= ˆpT−pˆC
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 25
Binary Data – Effect Measures
> cr = read.csv("hd-cr.txt", as.is=TRUE) > library(meta)
> mil = metabin(crHDCT, nHDCT, crControl, nControl,
+ sm="OR", data=cr, studlab=study,
+ subset=study=="Milpied")
> round(exp(mil$TE), 2)
[1] 2.37
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 26
Introduction R in Action Meta-Analysis Binary Data Summary References
Binary Effect Measures – Confidence Interval
Large sample variance estimates (Fleiss, 1993):d Var(log ˆφ) = 1 a+ 1 c − 1 a+b − 1 c +d d Var(log ˆψ) = 1 a+ 1 b + 1 c + 1 d (3) d Var(ηˆ) = a b (a+b)3 + c d (c+d)3
(1−α)-confidence interval (on log scale for risk ratio and odds ratio):
ˆ
θ ± z1−α2 S.E.(ˆθ)
with standard error S.E.(ˆθ) =
q d
Var(ˆθ).
Introduction R in Action Meta-Analysis Binary Data Summary References
Binary Effect Measures – Confidence Interval
Large sample variance estimates (Fleiss, 1993):
d Var(log ˆφ) = 1 a+ 0.5 + 1 c + 0.5− 1 a+b+ 0.5 − 1 c+d + 0.5 d Var(log ˆψ) = 1 a+ 0.5 + 1 b+ 0.5 + 1 c + 0.5 + 1 d + 0.5 d Var(ˆη) = (a+ 0.5) (b+ 0.5) (a+b+ 1)3 + (c + 0.5) (d + 0.5) (c+d + 1)3
Add 0.5 if any cell counts are zero (Gart and Zweifel, 1967; Pettigrew et al., 1986)
Binary Effect Measures – Confidence Interval
> mil = metabin(crHDCT, nHDCT, crControl, nControl,+ sm="OR", data=cr, studlab=study,
+ subset=study=="Milpied")
> # Print variance estimate > mil$seTE^2
[1] 0.09629314
> ## Print confidence interval > print(mil, digits=2)
OR 95%-CI z p.value
2.37 [1.29; 4.35] 2.78 0.0055 Details:
- Inverse variance method
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 29
Aggressive Non-Hodgkin Lymphoma – Forestplot
Study De Souza Gianni Gisselbrecht Intragumtornchai Kaiser Kluin−Nelemans Martelli Martelli 2003 Milpied Rodriguez 2003 Santini Santini−2 Verdonck Vitolo Events 14 46 119 10 110 67 3 57 74 39 46 80 25 35 Total 28 48 189 23 158 98 22 75 98 55 63 117 38 60 HDCT Events 10 35 116 9 97 56 4 51 56 30 34 71 26 46 Total 26 50 181 25 154 96 27 75 99 53 61 106 35 66 Control 0.1 0.5 1 2 10 Odds Ratio
Favours control Favours HDCT OR 1.60 9.86 0.95 1.37 1.35 1.54 0.91 1.49 2.37 1.87 2.15 1.07 0.67 0.61 95%−CI [0.54; 4.73] [2.11; 45.96] [0.62; 1.45] [0.43; 4.36] [0.84; 2.16] [0.86; 2.78] [0.18; 4.57] [0.73; 3.06] [1.29; 4.35] [0.84; 4.14] [1.01; 4.56] [0.61; 1.87] [0.24; 1.83] [0.29; 1.27]
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 30
Introduction R in Action Meta-Analysis Binary Data Summary References
Naive Pooling – Fictitious Example
CR no CR pˆT ˆpC RR [95%-CI]c HDCT 4 56 Study 1 Control 11 139 6.7% 7.3% 0.91 [0.30; 2.74] HDCT 40 140 Study 2 Control 12 38 22.2% 24.0% 0.93 [0.53; 1.63] HDCT 44 196 Study 1&2 Control 23 177 18.3% 11.5% 1.59 [1.00; 2.55] Appropriate meta-analysis 0.92 [0.56; 1.52]
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 31
Introduction R in Action Meta-Analysis Binary Data Summary References
Inverse Variance Method – Odds ratio – Definition
Overall odds ratio ˆψIV (Fleiss, 1993):ˆ ψIV = exp K P k=1 wk ·log ˆψk K P k=1 wk (4) I Study index: k= 1, ...,K I Weights: wk = 1 .
dVar(log ˆψk) (→fixed effect model) I See formulae (2) and (3) for definition of ˆψk anddVar(log ˆψk) I Analogous for risk ratio as effect measure: log ˆφk
I For risk difference: ˆηk (without exp function in equation (4))
Meta-Analysis of CR – Inverse Variance Method
> cr = read.csv("hd-cr.txt", as.is=TRUE)> library(meta)
> m2 = metabin(crHDCT, nHDCT, crControl, nControl,
+ sm="OR", data=cr, studlab=study,
+ comb.random=FALSE, method="Inverse")
> summary(m2)
Number of studies combined: k=14
OR 95%-CI z p.value
Fixed effect model 1.3228 [1.0999; 1.5909] 2.9713 0.003 Quantifying heterogeneity:
tau^2 = 0.0897; H = 1.3 [1; 1.78]; I^2 = 41% [0%; 68.6%] Test of heterogeneity:
Q d.f. p.value
22.03 13 0.0549
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 33
Forest Plot – CR – Inverse Variance Method
> forest(m2, hetstat=FALSE, text.fixed="IV estimate")Study IV estimate De Souza Gianni Gisselbrecht Intragumtornchai Kaiser Kluin−Nelemans Martelli Martelli 2003 Milpied Rodriguez 2003 Santini Santini−2 Verdonck Vitolo Events 14 46 119 10 110 67 3 57 74 39 46 80 25 35 Total 1072 28 48 189 23 158 98 22 75 98 55 63 117 38 60 Experimental Events 10 35 116 9 97 56 4 51 56 30 34 71 26 46 Total 1054 26 50 181 25 154 96 27 75 99 53 61 106 35 66 Control 0.1 0.5 1 2 10 Odds Ratio OR 1.32 1.60 9.86 0.95 1.37 1.35 1.54 0.91 1.49 2.37 1.87 2.15 1.07 0.67 0.61 95%−CI [1.10; 1.59] [0.54; 4.73] [2.11; 45.96] [0.62; 1.45] [0.43; 4.36] [0.84; 2.16] [0.86; 2.78] [0.18; 4.57] [0.73; 3.06] [1.29; 4.35] [0.84; 4.14] [1.01; 4.56] [0.61; 1.87] [0.24; 1.83] [0.29; 1.27] W(fixed) 100% 2.9% 1.4% 19.0% 2.5% 15.3% 9.8% 1.3% 6.6% 9.2% 5.4% 6.0% 10.8% 3.3% 6.3%
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 34
Introduction R in Action Meta-Analysis Binary Data Summary References
Mantel-Haenszel Method – Odds ratio – Definition
Mantel and Haenszel (1959):I Estimator for common odds ratio in stratified case-control study
I Can be used in meta-analysis of RCTs
I Fixed effect method
Mantel-Haenszel odds ratio ˆψMH:
ˆ ψMH = k P k=1 wk ·ψˆk k P k=1 wk (5) I Weights: wk = bkck nk
Introduction R in Action Meta-Analysis Binary Data Summary References
Meta-Analysis of CR – Mantel-Haenszel Method
> cr = read.csv("hd-cr.txt", as.is=TRUE)> library(meta)
> m3 = metabin(crHDCT, nHDCT, crControl, nControl,
+ sm="OR", data=cr, studlab=study,
+ comb.random=FALSE, method="MH")
> # Same result (Mantel-Haenszel method is default) > m3 = metabin(crHDCT, nHDCT, crControl, nControl,
+ sm="OR", data=cr, studlab=study,
+ comb.random=FALSE)
> # Same result (use of R function update.meta) > m3 = update(m2, method="MH")
Forest Plot – CR – Mantel-Haenszel Method
> forest(m3, hetstat=FALSE, text.fixed="MH estimate")Study MH estimate De Souza Gianni Gisselbrecht Intragumtornchai Kaiser Kluin−Nelemans Martelli Martelli 2003 Milpied Rodriguez 2003 Santini Santini−2 Verdonck Vitolo Events 14 46 119 10 110 67 3 57 74 39 46 80 25 35 Total 1072 28 48 189 23 158 98 22 75 98 55 63 117 38 60 Experimental Events 10 35 116 9 97 56 4 51 56 30 34 71 26 46 Total 1054 26 50 181 25 154 96 27 75 99 53 61 106 35 66 Control 0.1 0.5 1 2 10 Odds Ratio OR 1.35 1.60 9.86 0.95 1.37 1.35 1.54 0.91 1.49 2.37 1.87 2.15 1.07 0.67 0.61 95%−CI [1.12; 1.61] [0.54; 4.73] [2.11; 45.96] [0.62; 1.45] [0.43; 4.36] [0.84; 2.16] [0.86; 2.78] [0.18; 4.57] [0.73; 3.06] [1.29; 4.35] [0.84; 4.14] [1.01; 4.56] [0.61; 1.87] [0.24; 1.83] [0.29; 1.27] W(fixed) 100% 2.6% 0.7% 21.8% 2.4% 14.8% 8.9% 1.5% 6.1% 6.8% 4.4% 4.6% 11.7% 4.6% 9.1%
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 37
Summary
Meta-analysisI Central to evidence based medicine (see, e.g. website of Cochrane
Collaboration http://www.cochrane.org/).
I Only as good as the evidence that it relies on.
I Some important issues are:
I Heterogeneity of study results. The more you can explain, the better →subgroup-analysis / meta-regression.
I Bias in contributing studies / non-representativeness of studies (e.g. due to publication bias).
I Binary data raises some special issues, especially if event rates are low. R
I Modern statistical packages for data analysis, management & graphics
I Use of additional software easily possible, e.g. for meta-analysis
I Use of command line necessary
I Extended documentation available (Online, Use-R! books, ...)
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 38
Introduction R in Action Meta-Analysis Binary Data Summary References References
Carpenter, J., R¨ucker, G., and Schwarzer, G. (2009). copas: An R package for fitting the Copas selection model. The R Journal, 1(2):31–36.
Egger, M. and Smith, G. D. (1997). Meta-analysis: Potentials and promise. British Medical Journal, 315:1371–1374.
Fleiss, J. L. (1993). The statistical basis of meta-analysis. Statistical Methods in Medical Research, 2:121–145.
Gart, J. J. and Zweifel, J. R. (1967). On the bias of various estimators of the logit and its variance with application to quantal bioassay. Biometrika, 54:181–187.
Gasparrini, A. (2014). mvmeta: Multivariate and univariate meta-analysis and meta-regression. R package version 0.4.3.
Greb, A., Bohlius, J., Schiefer, D., Schwarzer, G., Schulz, H., and Engert, A. (2008). High-dose chemotherapy with autologous stem cell transplantation in the first line treatment of aggressive non-hodgkin lymphoma (nhl) in adults. Cochrane Database Syst Rev, 1:CD004024. DOI: 10.1002/14651858.CD004024.pub2.
Lumley, T. (2012). rmeta: Meta-analysis. R package version 2.16.
Mantel, N. and Haenszel, W. (1959). Statistical aspects of the analysis of data from retrospective studies of disease. Journal of the National Cancer Institute, 22(4):719–748.
Pettigrew, H. M., Gart, J. J., and Thomas, D. G. (1986). The bias and higher cumulants of the logarithm of a binomial variate.Biometrika, 73:425–435.
Carpenter/Krahn/R¨ucker/Schwarzer Session I: Introduction to Meta-Analysis Florence, 6 July 2014 38
Introduction R in Action Meta-Analysis Binary Data Summary References
R Core Team (2014). R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria.
R¨ucker, G., Schwarzer, G., Krahn, U., and K¨onig, J. (2014). netmeta: Network meta-Analysis with R. R package version 0.5-0.
Schwarzer, G. (2007). meta: An R package for meta-analysis.R News, 7(3):40–45. Schwarzer, G., Carpenter, J., and R¨ucker, G. (2014).metasens: Advanced statistical
methods to model and adjust for bias in meta-analysis. R package version 0.1-0. Viechtbauer, W. (2010). Conducting meta-analyses in R with the metafor package.
Journal of Statistical Software, 36(3):1–48.