Lecture 5: Slides - The International Biometric Society

Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Plan of the Session
Session V: Network Meta-Analysis
James
Carpenter1 ,
Ulrike
Krahn2,3 ,
Gerta
R¨
ucker4 ,
At the end of this session the aim is that you should understand
Guido
Schwarzer4
1 London
School of Hygiene and Tropical Medicine & MRC Clinical Trials Unit, London, UK
of Medical Biostatistics, Epidemiology and Informatics, Mainz, Germany
3 Institute of Medical Informatics, Biometry and Epidemiology, Duisburg-Essen, Germany
4 Institute for Medical Biometry and Statistics, Freiburg, Germany
2 Institute
[email protected]
1
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
the principal model and underlying assumptions;
I
how to assess heterogeneity/inconsistency;
I
how to assess the evidence flow.
I
carry out a network meta-analysis;
I
report and visualize results;
I
interpret the results taking diagnostic measures and graphics into
account.
Carpenter/Krahn/R¨
ucker/Schwarzer
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Locating Inconsistency
Florence, 6 July 2014
Flow of evidence
2
Summary
National Institute for Health and Care Excellence (NICE)
Classical meta-analysis
comparison between two treatments
Network meta-analysis/ Multiple-treatments comparison/ Mixed-treatment
comparison
I
I
Overview
Background: Evidence based healthcare decisions
I
the background for network meta-analysis;
The objectives are that you are able to:
IBC Short Course
Florence, 6 July 2014
Overview
I
I
”has a preference for data from head-to-head RCTs [...]”
I
”[...] evidence from mixed treatment analyses may be presented if it is
considered to add information [...].”
I
”If data from head-to-head RCTs are not available, indirect treatment
comparison methods should be used [...].”
comparison between a set of treatments
(see NICE: Guide to the methods of technology appraisal, 2008 and NICE Decision Support
Unit, Series of Technical Support Documents)
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
3
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
4
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Overview
Illustrative example: Network meta-analysis in diabetes
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Adjusted indirect comparison
HbA1c change in patients with type II
diabetes and a baseline therapy of sulfonylurea (by Senn et al., 2013)
Study
migl
●
10 treatments
benf
●
plac
plac
acar
●
●
MD= -1.150
26 RCTs: 25 two-armed, 1 three-armed
Study
28 assessed pairwise comparisons
SUal ●
●
rosi versus plac
vild
rosi
15 different designs
MD= -1.148
●
(defined by compared treatments,
●
●
e.g. plac:acar, plac:acar:metf)
?
sita
metf
rosi
metf versus plac
●
piog
metf
ind
dir
dir
ˆ
θrosi:metf
=ˆ
θrosi:plac
−ˆ
θmetf:plac
dir
ind
dir
Vrosi:metf
= Vrosi:plac
+ Vmetf:plac
15 observed / 45 possible edges
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Florence, 6 July 2014
Locating Inconsistency
Flow of evidence
5
Summary
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Combination of direct and indirect evidence
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Locating Inconsistency
Florence, 6 July 2014
Flow of evidence
6
Summary
Assumption
Study
I
plac
of transitivity
I
MD= -1.150
Study
rosi versus plac
I
rosi
MD= -1.148
?
I
Study
of consistency
metf versus plac
I
metf
MD= -0.073
I
ˆ
θdir
ˆ dir
V
1
V dir
+
+
ˆ
θind
V ind
1
V ind
,
V nma =
Carpenter/Krahn/R¨
ucker/Schwarzer
extension of transitivity: direct and indirect estimates are in agreement
can be tested statistically
(see Salanti, 2012)
metf versus rosi
ˆ
θnma =
an indirect comparison validly estimates an unobserved head-to-head
comparison
cannot be tested statistically, but can be evaluated conceptually and
contextually
1
1
V dir
+
1
V ind
Session V: Network Meta-Analysis
Florence, 6 July 2014
7
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
8
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Overview
Different models and estimation methods...
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Fixed effects model
Bayesian approaches (using WinBUGS, see e.g. Dias et al. 2011, Lu
& Ades 2004 or R package gemtc)
Frequentist approaches
I
I
using two-way linear mixed models with main effects for treatment and
trial (Piepho et al. 2012, SAS)
using multivariate meta-regression (White et al. 2012, R package
mvmeta)
two approaches using generalized least squares (Krahn et al. 2013,
R¨
ucker 2012) leading to identical estimates (R¨
ucker 2014) implemented
in R package netmeta :
I
I
I
each study
with p treatments
a) reducing dimensions
p − 1 comparisons
to a reference treatment
with corresponding SEs
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Decomposition of Q
with E() = 0 and Cov() =: V = diag(V1dir , . . . , Vkdir )
Example:
3 treatments, independent studies s = 1, · · · , 4, designs d = AB, AC, BC



Y1.AB 
 1
 1
Y2.AB 
 , X = 
Y = 
Y3.AC 
 0
Y4.BC
−1
b) reducing weights
p(p−1)
all 2 pairwise
comparisons
with adjusted SEs
Session V: Network Meta-Analysis
Model and Estimating
Y = X θnma + Florence, 6 July 2014
Locating Inconsistency
Flow of evidence
9
Summary
Two-stage fixed effects model fitting
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction

0
!

θnma
0 nma
 , θ
= AB
, V = diag(V1.AB , ..., V4.BC )
nma
1
θAC
1
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Florence, 6 July 2014
Locating Inconsistency
Flow of evidence
10
Summary
Two-stage fixed effects model
1. Aggregation per study design d by multivariate meta-analysis methods
1. Aggregation per study design d by multivariate meta-analysis methods
ˆ
θddir := Vddir
X
s∈Sd
Vs−1 Ys ,
Vddir
ˆ
θddir := Vddir
−1


 X

−1 
dir
ˆ

Vs 
:= Cov(θd ) = 
X
s∈Sd
s∈Sd
Vs−1 Ys ,
Vddir
−1


 X

−1 
dir
ˆ

Vs 
:= Cov(θd ) = 
s∈Sd
2. Model fitting
dir 0
θˆdir = (θˆ1dir , . . . , θˆD
) = Xa θnma + a
2. Model fitting
with E(a ) = 0 and Cov(a ) =: Va = diag(V1dir , . . . , VDdir )
dir 0
θˆdir = (θˆ1dir , . . . , θˆD
) = Xa θnma + a
Example:
dir
ˆdir
with one further study of design ABC with treatment effect estimates ˆ
θAB
ABC , θAC ABC , and
dir
covariance matrix VABC
with E(a ) = 0 and Cov(a ) =: Va = diag(V1dir , . . . , VDdir )
Example:
θ
ˆdir
ˆdir 

θAB 
 1
ˆdir 

= θAC  , Xa =  0
ˆdir 
−1
θBC

!
0

θnma
dir
dir
1 , θnma = AB
, Va = diag(VAB
, ..., VBC
)
nma

θAC
1
ˆ
θdir
 ˆdir 

 θAB 
 1
 ˆdir 
 0
 θAC 
 , X = −1
dir
=  ˆ
θBC
a


ˆdir

 1
θAB ABC 
ˆ
0
θdir
AC ABC
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
11
Carpenter/Krahn/R¨
ucker/Schwarzer

0

!
1

θnma
dir
dir
1 , θnma = AB
nma , Va = diag(VAB , ..., VABC )

θAC
0
1
Session V: Network Meta-Analysis
Florence, 6 July 2014
12
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
R package netmeta
>
>
>
>
>
>
>
>
seTE treat1 treat2
0.1414
metf
plac
0.0992
metf
plac
0.3579
metf
acar
0.1435
rosi
plac
0.3758
0.4669
metf
acar
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
R package netmeta: Object netmeta
# 1. Install R package netmeta
install.packages("netmeta")
# 2. Load R package netmeta
library(netmeta)
# 3. Load data set Senn2013
data(Senn2013)
# 4. Print data Set
Senn2013
TE
1 -1.90
2 -0.82
3 -0.20
4 -1.34
...
27 -1.20
28 -1.00
Overview
>
>
>
>
>
>
+
studlab
DeFronzo1995
Lewin2007
Willms1999
Davidson2007
plac
plac
Fixed effects model (default)
The netmeta function generates an object of class
netmeta with corresponding functions print, summary,
forest, netgraph, netheat, decomp.design, and netmeasures
nma <- netmeta(TE, seTE, treat1, treat2, studlab,
data=Senn2013, sm="MD", reference="plac")
Warning messages:
1: In netmeta(TE, seTE, treat1, treat2, studlab, data = Senn2013,
sm = "MD", : Treatments within a comparison have been
re-sorted in increasing order.
Willms1999
Willms1999
Session V: Network Meta-Analysis
Model and Estimating
#
#
#
#
Decomposition of Q
Florence, 6 July 2014
Locating Inconsistency
Flow of evidence
13
Summary
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Florence, 6 July 2014
Locating Inconsistency
Flow of evidence
R package netmeta: Network graph
R package netmeta: Estimating treatment effects
> netgraph(nma, seq=c("plac", "benf", "migl", "acar", "sulf",
+
"metf", "rosi", "piog", "sita", "vild"))
> print(nma,digits=2)
14
Summary
Original data (with adjusted standard errors for multi-arm studies):
migl
●
benf
plac
acar
●
●
SUal ●
●
●
vild
●
sita
metf
●
rosi
Carpenter/Krahn/R¨
ucker/Schwarzer
DeFronzo1995
Lewin2007
Willms1999
Davidson2007
...
Willms1999
Willms1999
●
treat1 treat2
TE seTE seTE.adj narms multiarm
metf
plac -1.90 0.14
0.14
2
metf
plac -0.82 0.10
0.10
2
acar
metf 0.20 0.36
0.39
3
*
plac
rosi 1.34 0.14
0.14
2
metf
acar
plac -1.20 0.38
plac -1.00 0.47
0.41
0.82
3
3
*
*
●
piog
Session V: Network Meta-Analysis
...
Florence, 6 July 2014
15
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
16
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
R package netmeta: Estimating treatment effects
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
R package netmeta: Estimating treatment effects
Data utilised in network meta-analysis (fixed effect model):
treat1 treat2
DeFronzo1995 metf
plac
Lewin2007
metf
plac
Willms1999
acar
metf
Davidson2007 plac
rosi
...
Willms1999
metf
plac
Willms1999
acar
plac
Number of treatment arms (by study):
narms
Alex1998
2
...
Willms1999
3
Wolffenbuttel1999
2
Yang2003
2
Zhu2003
2
...
> #
...
95%-CI
[-1.23; -1.00]
[-1.23; -1.00]
[ 0.06; 0.51]
[ 1.11; 1.30]
-1.11
-0.83
[-1.23; -1.00]
[-1.04; -0.61]
Q leverage
30.89 0.18
8.79 0.36
0.05 0.09
0.93 0.11
0.04
0.04
0.02
0.02
each comparison's contribution to the heterogeneity
statistic Q_total
> # leverage: small leverage means that the precision of a contrast
benefits largely from the indirect inference;
a leverage of 1 means that there is no gain at all
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
MD
-1.11
-1.11
0.29
1.20
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Locating Inconsistency
Florence, 6 July 2014
Flow of evidence
17
Summary
R package netmeta: Estimating treatment effects
Q:
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Locating Inconsistency
Florence, 6 July 2014
Flow of evidence
18
Summary
R package netmeta: Estimating treatment effects
Number of studies: k=26
Number of treatments: n=10
Number of pairwise comparisons: m=28
Without a specification of argument reference.group in function netmeta
matrices of the TE estimates, lower and upper 95%-confidence limits for all
possible contrasts are shown, e.g.:
Fixed effect model
Fixed effect model
Treatment estimate (sm='MD', reference.group='plac'):
MD
95%-CI
acar -0.83 [-1.04; -0.61]
benf -0.91 [-1.15; -0.66]
metf -1.11 [-1.23; -1.00]
migl -0.94 [-1.19; -0.70]
piog -1.07 [-1.22; -0.92]
plac 0.00 [ 0.00; 0.00]
rosi -1.20 [-1.30; -1.11]
sita -0.57 [-0.82; -0.32]
sulf -0.44 [-0.62; -0.26]
vild -0.70 [-0.95; -0.45]
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
Treatment estimate (sm='MD'):
acar benf metf migl piog
acar 0.00 0.08 0.29 0.12 0.24
benf -0.08 0.00 0.21 0.04 0.16
metf -0.29 -0.21 0.00 -0.17 -0.05
migl -0.12 -0.04 0.17 0.00 0.12
piog -0.24 -0.16 0.05 -0.12 0.00
plac 0.83 0.91 1.11 0.94 1.07
...
19
Carpenter/Krahn/R¨
ucker/Schwarzer
plac...
-0.83
-0.91
-1.11
-0.94
-1.07
0.00
Session V: Network Meta-Analysis
Florence, 6 July 2014
20
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Generalization to a random effects model
Y = X θnma + b + with Cov(Y ) =: Vτ = V + Vhet (τ)
Example: studies of design AB, AC, BC, and ABC
0
Var(YAC )
0
0
0
0
0
1
0
0
0
0
0
1
1/2
Introduction
0
0
0
0
0
0
Var(YAB ABC )
Cov(YAB ABC , YAC ABC )




 +

Cov(YAB ABC , YAC ABC )
Var(YAC ABC )
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Locating Inconsistency
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Florence, 6 July 2014
Flow of evidence
> # In order to explicitely conduct a random effects model
> nma_re <- netmeta(TE, seTE, treat1, treat2, studlab,
+
data=Senn2013, sm="MD", reference="plac",
+
comb.random=TRUE)
Warning messages:
1: In netmeta(TE, seTE, treat1, treat2, studlab, data = Senn2013,
sm = "MD", : Treatments within a comparison have been
re-sorted in increasing order.
Note: Even though object nma has been generated without argument
comb.random all necessary information on the random effects network
meta-analysis is also part of object nma.
0 
0 

0 

1/2
1
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
0
0
Var(YBC )
0
0
Introduction
R package netmeta: Estimating treatment effects
Estimation of τ2 e.g. by a generalized DerSimonian-Laird method with the
assumptions:
I τ2 is the same for all designs/comparisons
I random effects of multi-arm studies have correlation 1/2
Var(Y
AB )


0

0
Vτ = 

0

0
1
0

0
1

2
0
τ 0

0
0

0
0
Overview
21
Summary
R package netmeta: Estimating treatment effects
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Florence, 6 July 2014
Locating Inconsistency
Flow of evidence
22
Summary
R package netmeta: Forest plot
> forest(nma,xlab="HbA1c mean difference", xlim=c(-1.5,1.5))
Treatment
Fixed Effect Model
acar
benf
metf
migl
piog
plac
rosi
sita
sulf
vild
> # with object nma or nma_re we obtain
Quantifying heterogeneity/inconsistency:
tau^2 = 0.1087; I^2 = 81.4%
Test of heterogeneity/inconsistency:
Q d.f. p.value
96.99
18 < 0.0001
−1.5 −1 −0.5
0
0.5
1
MD
95%−CI
−0.83
−0.91
−1.11
−0.94
−1.07
0.00
−1.20
−0.57
−0.44
−0.70
[−1.04; −0.61]
[−1.15; −0.66]
[−1.23; −1.00]
[−1.19; −0.70]
[−1.22; −0.92]
[−1.30; −1.11]
[−0.82; −0.32]
[−0.62; −0.26]
[−0.95; −0.45]
1.5
HbA1c mean difference
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
23
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
24
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
R package netmeta: Forest plot
Decomposition of Cochran’s Q in network meta-analysis
> forest(nma_re,xlab="HbA1c mean difference")
Heterogeneity of the whole network
XX
(Ys − Xs θˆnma )0 Vs−1 (Ys − Xs θˆnma )
Q nma :=
Treatment
Random Effects Model
acar
benf
metf
migl
piog
plac
rosi
sita
sulf
vild
−1.5 −1 −0.5
0
0.5
1
MD
95%−CI
−0.84
−0.73
−1.13
−0.95
−1.13
0.00
−1.23
−0.57
−0.42
−0.70
[−1.32; −0.36]
[−1.29; −0.17]
[−1.43; −0.82]
[−1.40; −0.50]
[−1.56; −0.70]
d s∈Sd
P
(χ2 distributed with s (as − 1) − (n − 1) degrees of freedom
n: number of treatments, as : number of arms in study s)
[−1.48; −0.98]
[−1.26; 0.12]
[−0.89; 0.06]
[−1.39; −0.01]
Q nma = Q within + Q inc
1.5
HbA1c mean difference
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Locating Inconsistency
Florence, 6 July 2014
Flow of evidence
25
Summary
I
Q within : Heterogeneity within pairwise comparisons OR within designs
I
Q inc : Inconsistency/Heterogeneity between pairwise comparisons OR
between designs
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Locating Inconsistency
Florence, 6 July 2014
Flow of evidence
Decomposition of Q nma in pairwise comparisons
Decomposition of Q nma in pairwise comparisons
> #Heterogeneity of the whole network
> round(nma$Q, 1)
>
>
>
>
>
97
> #Heterogeneity within pairwise comparisons
> round(nma$Q.heterogeneity, 1)
> #Heterogeneity between pairwise comparisons
> round(nma$Q.inconsistency, 1)
22.5
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
27
Summary
#Decomposition of the heterogeneity within pairwise comparisons
Qd <- nma$Q.decomp
Qd$Q <- round(Qd$Q, 1)
Qd$pval.Q <- round(Qd$pval.Q, 3)
Qd[Qd$df!=0,]
2
4
6
7
9
12
74.5
26
treat1 treat2
Q df pval.Q
acar
plac 0.1 1 0.811
benf
plac 4.4 1 0.036
metf
plac 42.2 3 0.000
metf
rosi 0.2 1 0.665
migl
plac 6.4 2 0.040
plac
rosi 21.3 5 0.001
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
28
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Decomposition of Q nma in design components
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Decomposition of Q nma in design components
Design-specific decomposition of within-designs Q statistic
Design
Q df p.value
acar:sulf 0.00 0
-metf:piog 0.00 0
-metf:rosi 0.19 1
0.6655
metf:sulf 0.00 0
-piog:rosi 0.00 0
-plac:acar 0.00 0
-plac:benf 4.38 1
0.0363
plac:metf 42.16 2 < 0.0001
plac:migl 6.45 2
0.0398
plac:piog 0.00 0
-plac:rosi 21.27 5
0.0007
plac:sita 0.00 0
-plac:vild 0.00 0
-rosi:sulf 0.00 0
-plac:acar:metf 0.00 0
--
> decomp.design(nma)
Q statistics to assess homogeneity / consistency
Q df p.value
Whole network
96.99 18 < 0.0001
Within designs 74.46 11 < 0.0001
Between designs 22.53 7
0.0021
...
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Florence, 6 July 2014
Locating Inconsistency
Flow of evidence
29
Summary
> netheat(nma)
Introduction
Model and Estimating
Q inc :=
Decomposition of Q
Locating Inconsistency
Florence, 6 July 2014
Flow of evidence
30
Summary
X
Qdinc with
d
Qdinc := (θˆddir − Xd θˆnma )0 Vd−1 (θˆddir − Xd θˆnma )
8
metf:sulf
rosi:sulf
metf:piog
plac:piog
plac:rosi
plac:metf
metf:rosi
piog:rosi
plac:acar_plac:acar:metf
plac:metf_plac:acar:metf
acar:sulf
plac:acar
Overview
Session V: Network Meta-Analysis
Generalized Cochran’s Q for design inconsistency
metf:sulf
rosi:sulf
metf:piog
plac:piog
plac:rosi
plac:metf
metf:rosi
piog:rosi
plac:acar_plac:acar:metf
plac:metf_plac:acar:metf
acar:sulf
plac:acar
Locating inconsistency: The net heat plot
Carpenter/Krahn/R¨
ucker/Schwarzer
P
(χ2 distributed with d (ad − 1) − (n − 1) degrees of freedom
n: number of treatments, ad : number of arms in design d)
6
4
Network meta-analysis example in diabetes:
2
Q inc =0.04 + 0.00 + 0.20 + · · · = 22.53
0
with df = 16 − 9 = 7 (p value: 0.002)
-2
(see Krahn et al., 2013)
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
31
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
32
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Detaching of single designs
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Detaching of single designs in the diabetes example
1. Model fitting that allows for a deviating effect in one design d
2. Recalculation of the Q statistic
I
inc
=
Q(d)
P
d0
Qdinc
0 (d)
migl
3. Investigation of the change in inconsistency for each summand
I Q inc
d0
−
benf
acar
plac
0
Qdinc
0 (d) ∀ d = 1, . . . , D
SUal
vild
Visualization ∀d = 1, . . . , D in the nma heat plot:
metf
rosi
Summary
Decomposition of Q
Locating Inconsistency
Flow of evidence
34
Summary
migl
plac
benf
acar
plac:metf *
Net heat plot in the diabetes example
benf
acar
Model and Estimating
plac:acar *
migl
Introduction
rosi:SUal
Detaching of single designs in the diabetes example
Overview
Florence, 6 July 2014
piog:rosi
Flow of evidence
Session V: Network Meta-Analysis
metf:SUal
Locating Inconsistency
Carpenter/Krahn/R¨
ucker/Schwarzer
metf:rosi
Decomposition of Q
33
metf:piog
Model and Estimating
Florence, 6 July 2014
acar:SUal
Introduction
Session V: Network Meta-Analysis
= 0.04 + 0.00 + 0.20 + · · · = 22.53 with dfQ inc = 16 − 9 = 7 (p value: 0.002)
plac:rosi
Overview
piog





Q inc
Carpenter/Krahn/R¨
ucker/Schwarzer
sita
plac:piog
inc
Qdinc
0 < Qd 0 (d)
increase after detaching/
inconsistent evidence
decrease after detaching/
supporting evidence





plac:metf
>
Qdinc
0 (d)
plac:acar
Qdinc
0
plac:acar
plac
8
plac:metf
plac:piog
SUal
vild
SUal
vild
6
plac:rosi
acar:SUal
4
metf:piog
metf
sita
metf
metf:rosi
sita
2
metf:SUal
rosi
piog
rosi
piog
piog:rosi
rosi:SUal
0
plac:acar *
plac:metf *
-2
Q inc
= 0.04 + 0.00 + 0.20 + · · · = 22.53 with dfQ inc = 16 − 9 = 7 (p value: 0.002)
inc
Q(metf:SUal)
= 0.48 + 0.00 + 0.23 + · · · = 7.52
Q inc
= 0.04 + 0.00 + 0.20 + · · · = 22.53 with dfQ inc = 16 − 9 = 7 (p value: 0.002)
inc
Q(metf:SUal)
= 0.48 + 0.00 + 0.23 + · · · = 7.52
diff
Q.,metf:SUal
= (−0.45, 0.00, −0.03, . . . )0
diff
Q.,metf:SUal
= (−0.45, 0.00, −0.03, . . . )0
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
34
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
34
Locating Inconsistency
Flow of evidence
Summary
plac:metf *
plac:acar *
rosi:SUal
piog:rosi
metf:SUal
metf:rosi
plac:acar
plac:metf *
plac:acar *
piog:rosi
metf:rosi
plac:metf
plac:rosi
metf:piog
metf:piog
6
plac:rosi
vild
acar:SUal
metf:rosi
piog
Summary
8
plac:piog
SUal
4
4
metf:piog
2
piog:rosi
rosi
Flow of evidence
plac:metf
metf:piog
plac:metf
sita
Locating Inconsistency
plac:acar
plac
6
plac:rosi
metf
benf
acar
plac:piog
vild
Decomposition of Q
acar:SUal
migl
8
rosi:SUal
SUal
Model and Estimating
Net heat plot in the diabetes example
metf:SUal
plac
Introduction
acar:SUal
acar
plac:piog
benf
rosi:SUal
migl
metf:SUal
Net heat plot in the diabetes example
Overview
plac:rosi
Decomposition of Q
plac:piog
Model and Estimating
plac:metf
Introduction
plac:acar
Overview
metf
plac:acar *
metf:rosi
sita
2
metf:SUal
plac:metf *
0
rosi
piog
piog:rosi
plac:acar
rosi:SUal
acar:SUal
0
plac:acar *
-2
* three-armed study
Q inc
= 0.04 + 0.00 + 0.20 + · · · = 22.53 with dfQ inc = 16 − 9 = 7 (p value: 0.002)
inc
Q(metf:SUal)
= 0.48 + 0.00 + 0.23 + · · · = 7.52
plac:metf *
-2
* three-armed study
diff
Q.,metf:SUal
= (−0.45, 0.00, −0.03, . . . )0
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Florence, 6 July 2014
Locating Inconsistency
Flow of evidence
34
Summary
acar
Introduction
Model and Estimating
plac:piog
vild
piog
H := Xa (Xa0 Va−1 Xa )−1 Xa0 Va−1 .
metf:rosi
2
piog:rosi
rosi
Summary
Xa θˆnma =H θˆdir
4
plac:metf
sita
Flow of evidence
6
metf:piog
plac:rosi
metf
Locating Inconsistency
35
θˆdir = Xa θnma + a
8
rosi:SUal
SUal
Decomposition of Q
Florence, 6 July 2014
acar:SUal
plac:acar
plac:metf *
plac:acar *
piog:rosi
metf:rosi
plac:metf
plac:rosi
metf:SUal
plac
Overview
Session V: Network Meta-Analysis
Contributions of direct estimates
metf:piog
plac:piog
benf
rosi:SUal
migl
metf:SUal
Net heat plot in the diabetes example
Carpenter/Krahn/R¨
ucker/Schwarzer
plac:acar *
plac:metf *
0
plac:acar
acar:SUal
* three-armed study
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
-2
Florence, 6 July 2014
36
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
37
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Contributions of direct estimates
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
SUal
vild
metf
sita
rosi
piog
metf:SUal
rosi:SUal
plac:piog
metf:piog
plac:rosi
plac:metf
metf:rosi
piog:rosi
plac:acar *
plac:metf *
plac:acar
acar:SUal
56 28 -5 9 17 -24 8
4
3 -4 13 16
42 41 4 -10 -43 22 -10 -6 2
acar
4
5
0
plac
3
5 -4 22 68 2 -21 5 -10 1 -1 -2 -1
1 83 10 6
benf
0 16 17
-6 3 36 49 24 35 -6 -15 -1 2
3 -6 4
metf:SUal
rosi:SUal
plac:piog
metf:piog
plac:rosi
plac:metf
metf:rosi
piog:rosi
plac:acar *
plac:metf *
plac:acar
acar:SUal
migl
0
2
2
-11 7 14 -18 23 56 -11 -5 -2 4
5
4
SUal
14 -13 -10 19 60 -46 18 10 2 -3 -3 -2
vild
metf
sita
9 -9 -32 -48 59 -25 12 20 1 -2 -1 0
rosi
18 15 4 -5 18 17 -2 -1 9 -6 57 -34
-11 7 14 -18 23 56 -11 -5 -2 4
5
piog
4
18 15 4 -5 18 17 -2 -1 9 -6 57 -34
27 20 4 -4 21 15 -1 0 -7 6 -40 53
Network Estimate
plac
Network Estimate
acar
Direct Estimate
metf:SUal
rosi:SUal
plac:piog
metf:piog
plac:rosi
plac:metf
metf:rosi
piog:rosi
plac:acar *
plac:metf *
plac:acar
acar:SUal
benf
metf:SUal
rosi:SUal
plac:piog
metf:piog
plac:rosi
plac:metf
metf:rosi
piog:rosi
plac:acar *
plac:metf *
plac:acar
acar:SUal
* three-armed study
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Summary
Contributions of direct estimates
Direct Estimate
migl
Flow of evidence
* three-armed study
Florence, 6 July 2014
Locating Inconsistency
Flow of evidence
38
Summary
Contributions of direct estimates
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Florence, 6 July 2014
Locating Inconsistency
Flow of evidence
39
Summary
Net heat plot in the diabetes example
plac
SUal
vild
metf
sita
rosi
piog
Network Estimate
acar
metf:SUal
rosi:SUal
plac:piog
metf:piog
plac:rosi
plac:metf
metf:rosi
piog:rosi
plac:acar *
plac:metf *
plac:acar
acar:SUal
acar
acar:SUal
plac:acar
plac:metf *
plac:acar *
piog:rosi
metf:rosi
plac:metf
plac:rosi
metf:SUal
plac
8
rosi:SUal
plac:piog
SUal
vild
6
metf:piog
plac:rosi
4
plac:metf
metf
sita
metf:rosi
2
piog:rosi
rosi
piog
plac:acar *
plac:metf *
0
plac:acar
acar:SUal
* three-armed study
Carpenter/Krahn/R¨
ucker/Schwarzer
metf:piog
benf
plac:piog
migl
rosi:SUal
benf
metf:SUal
migl
metf:SUal
rosi:SUal
plac:piog
metf:piog
plac:rosi
plac:metf
metf:rosi
piog:rosi
plac:acar *
plac:metf *
plac:acar
acar:SUal
Direct Estimate
Session V: Network Meta-Analysis
* three-armed study
Florence, 6 July 2014
40
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
-2
Florence, 6 July 2014
41
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
acar
Decomposition of Q
dir
wtu
:=
8
rosi:SUal
4
plac:metf
piog
3. Minimal parallelism (large is nice)
2
plac:acar *
plac:metf *
πtu =
0
plac:acar
acar:SUal
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Locating Inconsistency
˜ with
(where hi,j is one element of H. In the case of multi-arm studies, H
design specific net flows is used instead; see K¨onig et al., 2013)
Florence, 6 July 2014
Flow of evidence
42
Summary
Examples
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Florence, 6 July 2014
Locating Inconsistency
Flow of evidence
43
Summary
Characterizing measures in the diabetes example
3
3
2
2
3
> m <- netmeasures(nma)
> # Direct evidence proportion
> m$proportion[order(m$proportion)]
Direct evidence proportion 2/3
Mean path length
2/3 × 1 + 1/3 × 2 = 4/3
Minimal parallelism
1/2 × 2 + 1/2 × 1 = 3/2
1
1
1
1
maxd htu,d
-2
* three-armed study
Introduction
htu,d
d
piog:rosi
Overview
X
metf:rosi
sita
rosi
Summary
nma )
var(θˆtu
= htu,tu
dir )
var(θˆtu
ηtu =
plac:rosi
metf
Flow of evidence
6
metf:piog
vild
Locating Inconsistency
2. Mean path length (short is nice)
plac:piog
SUal
Model and Estimating
In the case of only two-arm studies:
1. Direct evidence proportion
acar:SUal
plac:acar
plac:metf *
plac:acar *
piog:rosi
metf:rosi
plac:metf
plac:rosi
metf:SUal
plac
Introduction
Characterizing the flow of evidence
metf:piog
plac:piog
benf
rosi:SUal
migl
metf:SUal
Net heat plot in the diabetes example
Overview
... sulf:vild acar:metf metf:rosi piog:rosi plac:piog rosi:sulf ...
0.0000
0.1025
0.1750
0.1995
0.3578
0.4106
plac:rosi plac:benf plac:migl plac:sita plac:vild
0.8317
1.0000
1.0000
1.0000
1.0000
migl
●
benf
●
plac
acar
●
●
1
1
3
3
1
½
½
2
SUal
Direct evidence proportion 0
Mean path length 4
Minimal parallelism 2
●
●
●
●
sita
metf
●
rosi
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
44
Carpenter/Krahn/R¨
ucker/Schwarzer
vild
●
piog
Session V: Network Meta-Analysis
Florence, 6 July 2014
45
Overview
Introduction
Model and Estimating
Decomposition of Q
Locating Inconsistency
Flow of evidence
Summary
Characterizing measures in the diabetes example
Overview
I
> # Minimal parallelism
> m$minpar[order(m$minpar)]
I
●
benf
Summary
●
●
R package netmeta provides some methods: effect estimating,
decomposition of a generalized Q statistic, net heat plot, measures for
characterizing the evidence flow
vild
●
sita
metf
●
rosi
●
piog
Session V: Network Meta-Analysis
Decomposition of Q
I
●
●
Locating Inconsistency
Florence, 6 July 2014
Flow of evidence
46
Summary
Assumption of consistency
Issues known from classical meta-analysis when evaluating the validity
of results (e.g. heterogeneity, limited power for assessing
heterogeneity, selection bias, extent to which results rest on a few
studies)
plac
●
Model and Estimating
Flow of evidence
I
●
acar
SUal
Introduction
Locating Inconsistency
Estimating the effects of all pairwise treatment comparisons in a trial
network
Indirect evidence is taken into account
I
migl
Overview
Decomposition of Q
Network meta-analysis
plac:benf ... sita:sulf sulf:vild
1.0000
3.1383
3.1383
Carpenter/Krahn/R¨
ucker/Schwarzer
Model and Estimating
Summary
> # Mean path length
> m$meanpath[order(m$meanpath)]
plac:benf ... plac:sulf rosi:sulf
1.0000
2.2156
2.3054
Introduction
Carpenter/Krahn/R¨
ucker/Schwarzer
Overview
Introduction
Session V: Network Meta-Analysis
Model and Estimating
Decomposition of Q
Locating Inconsistency
Florence, 6 July 2014
Flow of evidence
47
Summary
I R¨
ucker G, Schwarzer G, Krahn U, K¨
onig J (2014). netmeta: Network
References
I Dias S, Welton NJ, Sutton AJ, Ades AE (2011). NICE DSU Technical Support
Document 2: A Generalised Linear Modelling Framework for Pairwise and Network
Meta-Analysis of Randomised Controlled Trials. http://www.nicedsu.org.uk.
meta-analysis with R. http://www.CRAN.R-project.org/package=netmeta. R
package, version 0.5-0.
I Salanti G (2012). Indirect and mixed-treatment comparison, network, or
multiple-treatments meta-analysis: many names, many benefits, many concerns for
the next generation evidence synthesis tool. Research Synthesis Methods 2012.
I Krahn U, Binder H, K¨
onig J (2013). A graphical tool for locating inconsistency in
network meta-analyses. BMC Med Res Methodol.
I Senn S, Gavini F, Magrez D, Scheen A (2013). Issues in performing a network
I K¨
onig J, Krahn U, Binder H (2013). Visualizing the flow of evidence in network
meta-analysis. Stat Methods Med Res.
meta-analysis and characterizing mixed treatment comparisons. Stat Med.
I White IR, Barrett JK, Jackson D, Higgins JPT (2012). Consistency and
I Lu G, Ades AE (2004). Combination of direct and indirect evidence in mixed
inconsistency in network metaanalysis: model estimation using multivariate
meta-regression. Res Syn Meth.
treatment comparisons. Stat Med.
I NICE (2008). Evidence synthesis. Guide to the Methods of Technology Appraisal.
http://www.nicedsu.org.uk/Evidence-Synthesis-TSDseries%282391675%29.htm.
I NICE DSU Technical Support Documents (2011). http://www.nicedsu.org.uk.
I Piepho HP, Williams ER, Madden LV (2012). The use of two-way linear mixed
models in multitreatment meta-analysis. Biometrics.
I R¨
ucker G (2012). Network meta-analysis, electrical networks and graph theory.
Res SynMeth.
I R¨
ucker G, Schwarzer G (2012). Reduce dimension or reduce weights? Comparing
two approaches to multi-arm studies in network meta-analysis. Stat Med.
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
47
Carpenter/Krahn/R¨
ucker/Schwarzer
Session V: Network Meta-Analysis
Florence, 6 July 2014
47