Metainsight User Guide - (Including Bayesian Functionality) Version 0.1

NIHR Complex Review Support Unit (CRSU)
MetaInsight User Guide – (including

Bayesian functionality) Version 0.1
Xin Y, Cooper N, Owen RK, Freeman S & Sutton AJ
The UK NIHR Complex Review Support Unit (CRSU)
Funding information: National Institute for Health

Research, Grant/Award Number: 14/178/29
MetaInsight (including Bayesian estimates) V 3** Beta

available at:
https://crsu.shinyapps.io/metainsight
We delivered a webinar on MetaInsight as part of Cochrane network meta-analysis webinar

series through Cochrane Training on 3rd December 2019. Professor Alex Sutton and Dr. Yiqiao
Xin, presented the background of the MetaInsight development, a live demonstration
(25min), and discussed related issues and future plans for the app. A recording of the webinar
is available on the Cochrane Training website: .
https://training.cochrane.org/resource/metainsight-complex-review-support-unit-crsu-
network-meta-analysis-nma-web-based-app
Alternatively, they are directly available on Youtube: Part 1: Background and introduction
to MetaInsight
https://www.youtube.com/watch?v=RR_tkICQv_s&feature=youtu.be
Part 2: Demonstration of MetaInsight
https://www.youtube.com/watch?v=b-fYoUdksRo&feature=youtu.be
Part 3: Limitations and future plans
https://www.youtube.com/watch?v=g0n5yxQ4Z34&feature=youtu.be
1
We kindly ask you to cite MetaInsight whenever its output is used. MetaInsight should be cited
as:
Owen RK, Bradbury N, Xin Y, Cooper N, & Sutton AJ. MetaInsight: An interactive web-based
tool for analysing, interrogating and visualizing network meta-analyses using R-shiny and
netmeta. Research Synthesis Methods. 2019.
https://onlinelibrary.wiley.com/doi/full/10.1002/jrsm.1373
This manual can be cited as:
Xin Y, Cooper N, Owen RK, Freeman S & Sutton AJ. MetaInsight User Guide – (including
Bayesian functionality) Version 0.1. Accessed from: https://crsu.shinyapps.io/metainsight
/
May 2020
2
Contents
Introduction ....................................................................................................................... 4
The home page .................................................................................................................. 6
Using an inbuilt dataset ..................................................................................................... 7
Illustrative example ................................................................................................... 7
Loading your own data ...................................................................................................... 9
Network meta-analysis (NMA) ........................................................................................ 14
Choose data type and NMA analysis options .......................................................... 15
Data summary ......................................................................................................... 17
A plot of individual study results ..................................................................... 17
Network plot.................................................................................................... 19
Frequentist network meta-analysis ......................................................................... 21
Forest plot ....................................................................................................... 21
Comparison of all treatment pairs .................................................................. 23
Inconsistency ................................................................................................... 24
Bayesian network meta-analysis ............................................................................. 25
Forest plot ....................................................................................................... 25
Comparison of all treatment pairs .................................................................. 28
Ranking table ................................................................................................... 28
Nodesplit model .............................................................................................. 30
Bayesian result details ..................................................................................... 31
Deviance report ............................................................................................... 33
Model details ................................................................................................... 37
Perform sensitivity analysis ..................................................................................... 38
Sensitivity analysis: network plots .................................................................. 39
Sensitivity analysis: forest plots ...................................................................... 40
Sensitivity analysis: comparison of all treatment pairs ................................... 42
Sensitivity analysis: ranking ............................................................................. 43
Sensitivity analysis: Inconsistency ................................................................... 45
Sensitivity analysis: Bayesian details and deviance report ............................. 47
Output results.......................................................................................................... 49
Acknowledgement........................................................................................................... 51
Full list of packages.......................................................................................................... 51
References ....................................................................................................................... 51
3
Introduction
MetaInsight is an interactive web-based application for conducting network meta-analysis

and was developed by the NIHR Complex Review Support Unit (CRSU). In clinical settings,
there are often more than two available intervention options. Traditional meta-analysis, or
‘pairwise meta-analysis’ compares two interventions and therefore is limited in its ability to
answer questions when three or more interventions are compared. Network meta-analysis
(NMA) aims to evaluate multiple (n>2) interventions that may or may not have been directly
compared in the studies. This approach combines studies making different intervention
comparisons, that together form a connected network of evidence, to obtain relative
treatment effects for all interventions compared to one another. NMA has been increasingly
adopted in recent years by Cochrane reviews, guideline developers and decision-making
bodies such as National Institute of Heath and Care Excellence (NICE) in the UK to address the
questions such as ‘which intervention is the ‘best’ overall?’. Further information on NMA is
available elsewhere (1-4).
Currently, NMA is primarily conducted in statistical packages such as WinBUGs, R and STATA,
and depending on the packages chosen, the software coding can be difficult for non-
statisticians and hinders the progress of carrying out systematic reviews containing NMA. It
revealed a need for the development of a user-friendly interface to sustainably increase
capacity by empowering informed non-specialists to be able to conduct more clinically
relevant reviews. To this end, MetaInsight was created, although it has also been found to be
a user-friendly and time-saving tool for the more experienced analyst.
MetaInsight is freely available and requires no specialist software, beyond a modern web
browser, for the user to install but leverages established analysis routines. The tool is
interactive and uses an intuitive ‘point and click’ interface and presents results in visually
intuitive and appealing ways. It can also carry out sensitivity analysis on existing NMAs. It is
hoped that this tool will assist those in conducting NMA, and, in turn, increase the relevance
of published meta-analyses.
The tool can interactively conduct NMA for both binary and continuous outcomes for fixed
and random effects models [12] and facilitates sensitivity analyses via the inclusion and
exclusion of studies. It allows users to upload their own datasets into a webpage and provides
graphical representations of the treatment network and various aspects of the comparison
results. This includes forest plots of individual studies, forest plots of pooled analysis results,
comparison between pairwise results and network results for each of the comparisons,
ranking of interventions, and inconsistency tests.
MetaInsight has been built primarily using the Shiny (5), netmeta (6) and GeMTC (7) packages
for R. Shiny allows users to build interactive web applications and host them on a server. The
statistical analysis is conducted using the packages netmeta or GeMTC. A full list of all R
packages used in the building of MetaInsight can be found in Section 7.
4
We initially developed the app for conducting frequentist NMA only. In October 2019
MetaInsight version 2 was released which, in addition to the frequentist NMA, enables
Bayesian NMA to be conducted. Within each version, there were separate Apps for
continuous outcomes (e.g. weight change) and binary outcomes (e.g. success rate). In April
2020, we further integrated the two Bayesian apps together into one app.
The link to the MetaInsight (including Bayesian estimates) V3 is as below:

• https://crsu.shinyapps.io/metainsight/
In this User Guide we will demonstrate the features available within MetaInsight version 3**
Beta. We will start by showing you the home page and how you can upload your own dataset
to MetaInsight and then demonstrate the features of MetaInsight using one of the inbuilt
datasets allowing you to replicate the analysis if you wish. MetaInsight consists of five pages
(Home, Load Data, Data analysis, Full update history, and Privacy notice) which can be
accessed by clicking on the appropriate page name in the grey bar at the top of the web page.
This User Guide will guide you through each page in turn.
If you have any feedback or queries about MetaInsight, please feel free to contact Professor
Alex Sutton ajs22@leicester.ac.uk.
5
The home page
The MetaInsight (continuous outcome version) home page is shown in Figure 1. The name
indicates the version number v 3** Beta. A radiobutton is displayed underneath the title for
users to select the outcome of their analysis. On the right of the diagram, it highlights the
new features since last update. Below the diagram you can find the authors, contact email,
packages used, and disclaimers from the authors.
Navigate to a page using the

grey bar at the top of the page
Outcome selection. Please make

sure this is selected according to
the outcome of your interest.
Figure 1 Home page of MetaInsight V3
6
Using an inbuilt dataset
MetaInsight comes with one inbuilt dataset each for its continuous outcome example and
binary outcome example, which can be used to familiarise yourself with the features of
MetaInsight. The datasets are available to download in csv format and can be used to help
ensure that your own data is in the correct format for upload to MetaInsight.
Download the example

inbuilt dataset to assist
with formatting
Figure 2 Downloading inbuilt dataset
Illustrative example
Throughout the remainder of this User Guide we will use the inbuilt dataset. This dataset is
from a systematic review and NMA of pharmacological interventions for the treatment of
obesity. The example dataset evaluated change from baseline in body mass index (BMI) for
anti-obesity interventions (8). The dataset consists of 24 studies evaluating (in various
combinations) the effectiveness of 6 interventions of interest - placebo, orlistat, sibutramine,
metformin, orlistat + sibutramine, and rimonbant. Twenty of these studies were 2-arm trials,
3 studies were 3-arm trials, and one study was a 4-arm trial. All studies reported estimates of
mean difference from baseline in BMI, together with its corresponding uncertainty. The
7
dataset can be downloaded by clicking the download button and the resulting CSV file is
shown in Figure 3.
Figure 3 Screenshot of csv file resulting from downloading the dataset in long form.
8
Loading your own data
The ‘Load Data’ page consists of a grey box on the left-hand side of the page and three tabs:
Long format upload, Wide format upload, and View data (Figure 4). Users can import their
data easily by uploading a comma delimited (.csv) file in either long or wide format. The first
row of the data file should contain the column headings. Column headings are specified on
either ‘Long format upload' or ‘wide format upload’ tab. To assist with the labelling of column
headings inbuilt example datasets are available to download as previously mentioned.
Figure 4 Load data page of MetaInsight V3
For the continuous app, the long format data would require six columns containing the ‘Study
ID’ (as sequential and numerical code), ‘Study’ name, treatment code ‘T’ (as sequential and
numerical code), the number of participants ‘N’, the mean treatment effects ‘Mean’ and
corresponding standard deviations ‘SD’ for each arm of the study. The wide format data would
include ‘Study ID’, ‘Study’, and sets of ‘T’, ‘N’, ‘Mean’ and ‘SD’ depending on the number of
arms (i.e., ‘T.1’…‘T.6’, ‘N.1’… ‘N.6’, ‘Mean.1’…’Mean.6’, ‘SD.1’…’SD.6’).
For the binary outcomes, the long format data would require five columns containing the
‘Study ID’, ’Study’ name, treatment code ‘T’, the number of participants with the outcome of
interest in each arm of the study ‘R’, and the number of participants in each arm of the study
‘N’. Similar to the wide form with the continuous outcomes, the wide format data would
9
include ‘study ID’, ‘Study’ and sets of ‘T’, ‘R’ and ‘N’ depending on the number of arms (i.e.,
‘T.1’… ‘ T.6’, ‘R.1’…’R.6’, ‘N.1’…’N.6).
An example of long and wide format data entry is given in Figure 5 and Figure 6. Please note
that the reference treatment needs to be labelled as 1. The reference treatment is the
treatment option that all the other treatments will be compared to in the analysis. The
reference treatment is usually either a placebo or usual care group, or the most common
comparator used in the included studies. In this example, the reference treatment is placebo,
thus labelled as 1.
Figure 5 Upload data to the MetaInsight app - Long data format
10
Figure 6 Upload data to the MetaInsight app – Wide data format
To load your own dataset:

1. Formatting: Follow the instructions of ‘Step 1’ on either ‘long format upload’ or ‘wide
format upload’ subtabs to format your data. Please feel free to download the example
inbuilt dataset for each type of format to assist labelling.
2. Saving: Save your formatted data as ‘csv (Comma delimited) (*.csv) format.
3. Uploading: Click the `Select' button at the left of the grey box (Figure 4). A box will
appear which will allow you to navigate through your computers file system until you
find the file you wish to upload. Select the file you wish to upload and click open. Once
the upload is completed a blue bar will appear showing `Upload complete' (Figure 4).
4. Checking: Click on the `View data’ tab if you want to check your dataset has uploaded
in the format expected (Figure 5 or Figure 6 ).
These steps are shown in Figure 7.
Please note: The `View data’ tab will always display the dataset that is currently loaded and
used for analysis within MetaInsight. Prior to upload of your own dataset, this tab will
display the inbuilt default dataset.
11
1. Formatting 4. Checking
2. Saving
3. Uploading
Figure 7 Steps for uploading data
Users may enter their own text label for each intervention by either copying-and-pasting from
Microsoft Excel, or via any tab separated file (e.g. txt, or Excel). The latter is recommended as the users
can save the label file and copy-paste from it to the editor each time they re-visit the app. These text
labels will appear in data and outputs of results. Users are reminded to keep labels short to allow for
clear displays on visual outputs.
To load your own text label (Figure 8):

1. Download the example ‘treatment labels.txt’ text file template.
2. Adapt the template by entering the labels to match with the numerical treatment codes
in the data file. Labels should be short to allow for clear display on figures. Please note:
(1) keep the heading row which contains the ‘Number’ and ‘Label’, case sensitive; (2)
the Number column and Label column are separated by tab; (3) Treatment names may
only contain letters, digits, and underscore (_).
3. Copy the number and labels from the txt file and paste into the editor.
12
Figure 8 Load labels for treatment codes
Helpful Hints for uploading your own dataset to avoid error messages
• Excel spreadsheets must be saved in comma delimited (csv) format before uploading to
MetaInsight
• The labelling of column headings is case sensitive
• The ‘Study’ field should be unique for every study; in wide form, it should be unique for
every row.
• If the labels are not successfully shown in the results, try using Excel to enter the labels (two
columns: Number and Label), and copy and paste from Excel.
13
Network meta-analysis (NMA)
Once the data are uploaded, click on ‘Data analysis’ at the top of the page (Figure 9) and you
will see a page with a side bar on the left and a main section on the right. The side bar
contains the choices relevant to your data, and the NMA analysis options (see section 5.1).
The main section contains the following tabs for data summary (see section 5.2), frequentist
NMA analysis results (see section 5.3) and Bayesian NMA analysis results (see Section 5.4):
1. Data summary, which contains two subtabs:

1a. Study Results, and
1b. Network Plot, and
2. Frequentist network meta-analysis, which contains three subtabs:
2a. Forest plot, and
2b. Comparison of all treatment pairs, and
2c. Inconsistency
3. Bayesian network meta-analysis, which contains seven subtabs:
3a. Forest plot
3b. Comparison of all treatment pairs
3c. Ranking table
3d. Nodesplit model
3e. Bayesian result details
3f. Deviance report
3g. Model details
Figure 9 Data analysis tab of the MetaInsight V3
14
Choose data type and NMA analysis options

Firstly, select the outcome relevant for your data type. In this example, we have continuous
data and our outcome of interest is the Mean Difference in BMI, and therefore we leave the
radio button selected as the default, ‘Mean Difference (MD)’ (Figure 10). If your outcome of
interest is standardised mean difference, then ‘standardised mean difference (SMD)’ should
be selected. Figure 11 shows the interface of the ‘Data analysis’ tab with binary outcomes
where three outcomes can be selected from: odds ratio (OR), risk ratio (RR) and risk
difference (RD).
For the treatment rankings, the user must specify whether smaller outcome values are
desirable or undesirable. In this example, as the outcome is change in BMI, smaller
(negative) values indicate a reduction in BMI and thus we select ‘desirable’. Similarly, this
applies to other clinical outcomes such as blood pressure lowering drugs for hypertension or
cigarette per day for smoking cessation. Examples for the opposite situation where you
would need to select ‘undesirable’ includes outcomes such as steps per day for interventions
that aims to increase physical activities, quality of life measurement where usually a higher
score indicates better quality of life.
Finally, the user will need to select the type of NMA model to fit to the data – fixed or
random effects. Fixed effect model assumes that the true effect size does not vary between
studies, whereas random effect models assumes that the true effects in each individual
study are sampled from a distribution. (9) The default is random effect (RE).
Underneath the choices for data type and model options is the study selection section
(Figure 10). One of the best features of the MetaInsight app lies in its convenience for
conducting sensitivity analysis by excluding one or more studies from the main analysis from
the sidebar and results are displayed side by side with the all-inclusive primary results.
Studies can easily be excluded from the analysis using a simple checkbox interface. The user
can choose to exclude one or more studies and the NMA will be updated and displayed
alongside the complete case analysis which remains visible to facilitate comparison. There
are three caveats in terms of study exclusion:
(1) If the reference treatment is excluded completely from the analysis (i.e. all the
studies that contain the reference treatment are excluded), the user will need to re-
number the treatments, i.e. to remove treatment 1 on the treatment code editing
box shown in ‘step 2’ on page 11. If any other treatment is excluded, the results
should display with no need to relabel.
(2) If the network becomes disconnected due to exclusion of studies, both the network
plot and the frequentist results will show error messages. If this happens, users are
advised to check their network and continue excluding studies until only one
connected network exists and all other disconnected studies have been selected for
removal. The app will update each time a new study is excluded; this means the app
will automatically show results after the cause of the error (the disconnected
network in this case) has been addressed.
(3) If studies were excluded until only two treatments existing in the network, the
frequentist analysis is still able to perform the pairwise meta-analysis.
15
Choices for data type, direction

of preferences of outcome, and
NMA model. Note the two
options for a continuous
outcome: mean difference and
standardised mean difference
Study selection section: The user can

choose to exclude one or more studies
and the NMA will be updated and
displayed alongside the complete case
analysis which remains visible to
facilitate comparison.
Figure 10 Choices for data type and NMA analysis options with continuous outcomes
Choices for data type, direction of

preferences of outcome, and
NMA model. Note the three
options for binary outcomes: odds
ratio, risk ratio, and risk
difference.
Figure 11 Choices for data type and NMA analysis options with binary outcomes
16
Data summary
The ‘1. Data summary’ page contains two subtabs (Figure 12):
• ‘1a. study results’, and

• ‘1b. network plot’.
A plot of individual study results
‘1a. Study Results’ contains a plot of the individual study results as included in the NMA
grouped by each pairwise treatment comparison. These study results are displayed on a
common scale in a forest plot (Figure 12). Please note this forest plot does not display any
pooled results. It only provides a visualisation of the individual study results based on the
uploaded data.
There are two pieces of key information you can see from observing this plot. Firstly, it
informs how much evidence there is for each comparison. For this specific example, there
are a lot of studies comparing orlistat versus placebo, sibutramine versus placebo, and
sibutramine vs. orlistat, but not much evidence for the other comparisons.
Secondly, users can examine the results from each individual study within each of the
comparisons to visually examine the degree of heterogeneity between studies and identify
potential ‘outliers’ that the user may wish to examine further or exclude from the analysis as
part of the sensitivity analysis. A visual inspection of the treatments for obesity example
shows the study results for the sibutramine vs placebo comparison to be similar with almost
all studies showing a statistically significant reduction in BMI with sibutramine compared to
placebo. While for the orlistat vs placebo comparison, the majority of studies show orlistat
to be more effective than placebo (i.e. the mean effect estimate is negative showing a
reduction in BMI) but with many of the 95% confidence intervals crossing the line of no
effect.
17
Figure 12 Forest plot of individual study results grouped by treatment comparison.
18
Network plot
Click on the ‘1b. network plot’ and you will see two network diagrams; on the left is the
network plot of all studies and on the right is the plot for the sensitivity analysis (See section
5.5).
These network plots provide a visual display of where interventions have been compared in
head-to-head trials, and importantly whether there is a connected network of evidence; an
essential requirement for NMA. Each node on the plot represents an individual intervention
with connecting lines between nodes indicating where one or more of the trials have
evaluated both interventions on a head-to-head basis. The number along the lines indicates
the number of trials making each comparison.
Initially, the two network plots will be identical but if studies are excluded by clicking on
their study names on the left sidebar, the user will see the numbers on the network plot on
the right change as these depict the number of studies for each intervention comparison in
the NMA. The structure of the network plot may also change if by excluding studies data on
a particular treatment comparison is no longer available, as shown on Figure 13.
Study selection
Network plots
Figure 13 A network plot of all studies and network plot with studies excluded
19
Underneath each network plot is the plot information box (Figure 14). In this example, it
shows there are 24 studies, comparing 6 treatments, forming 35 pairwise comparisons and
there is one connected network.
The distance matrix shows how two treatments are connected in the network. The lower
triangle and the upper triangle display the same information, so users only need to look at
one of them. When the distance is 1, it indicates that direct evidence exists between that
comparison, or graphically, there exists an instance where one line connects the two
treatments (for example, metformin vs orlistat in the left network plot) (note, longer routes
between the treatments may also exist). When the distance is 2, it means the shortest
connection between the two treatments is through another comparator, or graphically,
there are two lines needed to connect the two treatments (for example, metformin vs.
orli_sibut in the left network plot). You can see that there is one 3 distance in the matrix,
which is between rimonbant and orli_sibut indicating that the shortest distance between
these two treatments is through another two treatments, or at least three lines are needed
to connect them. The distance matrix provides a good way of checking the network and
allows the users to assess the degree of the ‘indirectness’ between any two interventions.
Distance 3: the shortest distance

between rimonbant and
orli_sibut is through another
two treatments.
Figure 14 network plot information box
In some cases, if the network is disconnected (i.e., two or more disconnected subnetworks
exist), the network plot will not be produced, and the app will give error messages. If this
happens, then the network information box will show the ‘number of networks’ to equal 2
instead of 1This is shown in Figure 15. If disconnection happens, the users could try
excluding all the studies in one network, then the plot and the results for the other network
should appear. One caveat to this is that if the reference treatment is contained in a
disconnected subnetwork that has been completely removed, the user will need to
renumber the treatments to make one of the remaining treatments the ‘new’ reference
treatment to enable the analysis to run.
20
Error message is produced for

the network plot since there are
two disconnected networks.
Figure 15 Using network plot information to detect disconnected networks
Frequentist network meta-analysis

After examining the forest plot of individual study results and the network plots, click on the
‘2. Frequentist network meta-analysis’ tab where three subtabs are displayed:
2a. Forest plot
2b. Comparison of all treatment pairs
2c. Inconsistency
These tabs display key results produced from the R netmeta package. On each tab, the
results for all studies and for sensitivity analysis are displayed side by side, which enables an
immediate comparison to examine the impact of excluded studies. The result is immediately
shown once you click each tab name. The functions of each tab are described as below.
Forest plot
The ‘2a. forest plot’ tab displays the forest plot of the pooled effect estimates, and their
associated uncertainty (95% confidence intervals), for all interventions compared to the
reference treatment (coded as treatment 1 on the ‘Load data’ page – placebo in this
example).
21
Figure 16 shows the plot displayed by the default options, i.e. mean difference (as displayed
‘MD’) in the heading for the mean pooled estimates, and random effects model (as
displayed under the title).
Key information is displayed below the forest plot, including the between-study standard
deviation (‘tau’, indicating how much heterogeneity there is), the number of studies, and the
number of treatments in the meta-analysis. The results for this example show rimonbant has
the largest pooled mean difference as compared to placebo (the reference treatment);
rimonbant appeared to decrease BMI by approximately -3.76 kg/m2. Despite having the
largest mean difference, the confidence interval is relatively wide (95%CI: -5.52, -1.99). This
is likely to be caused by the lack of evidence for rimonbant, which can be seen from the
network plot in Figure 13 that there is only one study (only one line linking the node
rimonbant to the network, with one study for that comparison) evaluated rimonbant
(Rimonbant vs. Metformin). If we remove that study (by checking the treatment code from
the ‘data analysis’ tab on the ‘load data’ page), we will get a new NMA with only five
treatment options in the network (see the forest plot on the right of Figure 16). This,
however, does not change the pooled estimates for other interventions versus placebo as
the connection with rimonbant is not in any of the loops within the network. In this way, the
MetaInsight produces a good platform for interrogating the analysis and checking the impact
of individual studies on the results.
Figure 16 Forest plots
22
Comparison of all treatment pairs
Clicking on the tab ‘2b. comparison of all treatment pairs’ produces a matrix of all the
comparisons in a network meta-analysis. As for the previous tabs, the top matrix displays
results with all studies included, and the bottom league table displays results with studies
excluded. Above the leading diagonal, in the upper triangle, are the treatment comparisons
obtained from pairwise meta-analyses only (i.e., with only direct head to head evidence).
These are calculated as the treatment in the row versus the treatment in the column. For
example, the pooled mean difference from pairwise analysis of rimonbant (row) versus.
Metformin (column) is -1.73 (95%Crl -3.24, -0.22). Below the leading diagonal, in the lower
triangle, are the treatment comparisons obtained from the network meta-analysis. These
are calculated as the treatment in the column versus the treatment in the row. Both
pairwise and network estimates are presented as point estimates and corresponding 95%
confidence intervals.
The interventions are ordered from the most effective intervention to the least effective
intervention along the leading diagonal (Figure 17). The ranking is dependent on the choice
buttons on the side bar; for this example, a smaller BMI means the treatment is more
effective, and therefore we leave the selection as default, ‘smaller outcome values are
desirable’. If we change the selection to ‘undesirable’, the order would change to the
opposite on the leading diagonal with placebo being the best treatment.
Figure 17 Comparison of all treatment pairs
23
Inconsistency
Clicking on ‘2c. inconsistency’ provides the inconsistency tests results (obtained using the
netmeta package) (Figure 18). Consistencies between treatment effect estimates obtained
from direct and indirect evidence are one of the key assumptions underpinning NMA. The
inconsistency table displays the agreement between the direct evidence and indirect
evidence including: the number of studies directly comparing treatments of interest (‘No.
Studies’), NMA treatment effect estimates (‘NMA’), treatment effect estimates obtained
from direct (i.e. head to head) comparisons (‘Direct’), and treatment effect estimates
obtained from indirect information (‘Indirect’). For treatments that belong to a closed loop
in the network of evidence (i.e. there exists both direct and indirect information), the
difference between the direct and indirect estimates is calculated together with the lower
and upper limit of the 95% confidence interval (‘difference’, ‘Diff_95CI_lower’ and
‘Diff_95CI_upper’). Differences between direct and indirect information are further
quantified using p-values where a low p-value can be used to indicate conflicting evidence
unlikely to be attributable to chance alone (‘pValue’). In the example (Figure 18, we can see
that there does not appear to be any important estimated differences between direct and
indirect information (although power is often limited in such assessments).
Figure 18 Assessment of inconsistency
24
Bayesian network meta-analysis

The Bayesian NMA is conducted with the R package GEMTC that itself calls the Bayesian
simulation analysis software JAGS (http://mcmc-jags.sourceforge.net/).
Under the ‘Data analysis’ tab, click on ‘3. Bayesian network meta-analysis’. The tab names
and their functions under the Bayesian NMA are summarised in Table 1 below.
Table 1 Bayesian NMA functions
Tab names functions

3a. Forest plot Forest plot of the Bayesian NMA results (each treatment compared
to the reference treatment), and model fit statistics
3b. Comparison of Estimates of comparison of all treatment pairs from the NMA
all treatment pairs
3c. Ranking table Ranking table and median rank chart for each treatment to be the
best
3d. Nodesplit model Inconsistency tests with a “node-splitting” model (assesses
whether the direct and indirect evidence is inconsistent)
3e. Bayesian result Direct simulation result output from R including mean, SD and
details quantiles from the iterations for each treatment effect, and
Gelman convergence assessment plot for checking model
convergence.
3f. Deviance report Three plots are provided for checking model fit of individual data
points: a. residual deviance from NMA model and UME
inconsistency model (please see section 5.4.6 for details), b. per-
arm residual deviance for all studies, c. leverage plot. Deviance
data output are displayed at the end of the page. Studies with
outlying results can be identified using these tools
3g. Model details 3g.1 – model codes (adds transparency for what the app is doing
“behind the scenes”).
3g.2 – initial values: initial values used in each of the four
simulation chains for the analysis.
3g.3 – Download simulations: simulated result data are available to
download (e.g. to import into other packages).
Now we will illustrate these functions below.
Forest plot
The Bayesian method uses an approach where it runs Markov chain Monte Carlo simulations
and calculates the estimated model parameter distributions empirically. Because the
simulations take much longer time and memory to run than the frequentist approach
(currently it runs 25,000 sample iterations in total), we added the running buttons as seen in
Figure 19. Whenever the users change the radio button selections on the side bar, no
matter which tab the users are at, it is necessary to go back to the ‘forest plot’ page and click
this button to re-run the analysis. This is different from the frequentist analysis, where for
example, if you select a few studies to exclude, the app will immediately start running after
each additional study is selected, but in the Bayesian analysis, you will need to click the
25
button to tell the app that now the selection for the exclusion has finished and the analysis
can start running (Figure 20).
Figure 19 Running buttons for Bayesian analysis
Always click here if you

change the model
specification and need to
re-run the model.
Figure 20 Running button – always click to re-run whenever you change the selections on side bar
Once the button to run the analysis has been clicked you will see the status bar on the right
corner showing the progress of the analysis (Figure 21). Depending on the complexity of the
model and the speed of your computer, the results usually take about 10 – 20 seconds to
run, but can be up to 1 min, so please be patient and please do not re-click the button if the
model is still running.
Figure 21 Status bar indicating the model is running
26
The Bayesian pages have the same layout as the frequentist pages, where the results from
all studies are displayed side by side with the results from sensitivity analysis. Depending on
the complexity of the network, it usually takes up to 1 minute for the results to display.
Similar to the frequentist result (Figure 16), the Bayesian result also shows that rimonbant
has the largest pooled estimate compared with placebo but with large uncertainty depicted
by the wider 95% credible interval. Frequentist results and Bayesian results do not
necessarily agree with each other, for example, in some extreme situations where a lot of
zero events exist in the binary data.
Model fit parameters
Figure 22 forest plot results of Bayesian NMA
Underneath the forest plot are the model fit parameters (Figure 22). Dbar indicates the
posterior mean of the deviance, and the pD is the effective number of parameters. The DIC
is the sum of Dbar and pD. Lower values of the DIC suggest a more parsimonious model and
therefore the DIC value could be used to select the most appropriate model(s). A rule of
thumb is that differences in DIC over five are important, whereas a small differences (less
than 3) suggests there was little difference between the models (10).
27
Comparison of all treatment pairs

Clicking the ‘3b. comparison of all treatment pairs’ tab will provide a table (Figure 23). This table shows the
estimate from network meta-analysis only. This is different from the frequentist analysis where the pairwise
estimates are displayed in the upper triangle. In the Bayesian comparison, the table displays the results for all
treatment pairs in both the upper triangle and lower triangle, but with the comparison switched over; that is A vs.
B or B vs. A. For both above and below the leading diagonal, the results are for the treatment at the top of the
same column vs. treatment at the left hand side of the same row.
-3.76: rimonbant (treatment at

the same column) vs. placebo
(treatment at the same row)
Figure 23 Comparison of all treatment pairs – Bayesian network meta-analysis
Please note that this table of the Bayesian NMA is not ordered based on the ranking of the
treatment effect. The ranking is shown on the tab ‘3c. ranking table’ (please see next
section).
Ranking table
Clicking on ‘3c. Ranking table’, we can see a ranking table on the left, and a median rank
chart on the right. The ranking table shows the probability for each treatment to be ranked
as the most effective treatment, ‘rank 1’, then second most effective treatment, ‘rank 2’,
until the least effective treatment. For example, the probability metformin is ranked 1 is
0.7%, ranked 2 46.0%, ranked 3 33.9%, etc. With this table, you could calculate the Surface
Under the Cumulative Ranking (SUCRA) (11), median rank, or mean rank.
28
The median rank for

metformin
Figure 24 Ranking probability table: Probability for each treatment at each rank
On the right is the median rank chart which visualises the median rank for each treatment.
The median rank is the rank of the middle iteration when every simulated estimate of the
rank is ordered. In this case, we run the simulation for 20000 iterations and monitor rank[4]
(for metformin) and order the 20000 rank values obtained, then we look at the rank of the
10000th and 10001st iteration. In this case, it will be rank 3 for both 10000st and 10001st
iterations for metformin, and therefore the median rank for metformin will be rank 3.
Actually, in majority of the cases, the rank at 10000th and 10001st would be the same which
will be used as the median rank. Very occasionally, when the rank for 10000th and 10001st
are not the same, it will take an average of the two as the median rank.
Users can easily calculate the median rank themselves based on the table. To do this, you
just need to add up the probabilities from the probability of being rank 1, rank 2, rank 3,
until the sum is over 0.5, then the last rank you added will be the median rank.
29
Figure 25 Median rank chart
Nodesplit model
Next tab is the ‘nodesplit model’ where it runs the nodesplit model to check the
inconsistency between direct evidence and indirect evidence. (12) This is a separate model
from the main NMA model so no results will be displayed here until you click the button on
this tab to run the nodesplit analysis. Clicking the button will start the analysis simulation
and, same as the main analysis, you can see a status bar showing the model is running. It
takes a much longer time to run than the initial NMA analysis as it fits a distinct model for
each treatment pair in the network that has both direct and indirect evidence. The results
are displayed in a very similar structure as the inconsistency checking for the frequentist
analysis (Section 5.3.3) as shown in Figure 26. For each treatment comparison that has both
direct and indirect estimates, the analysis will provide the mean and credible intervals (the
value at 2.5% and 97.5% quantiles) for the direct, indirect and network estimates together
with a Bayesian p-value for the related test of inconsistency between the direct and indirect
evidence for each treatment comparison.
30
Figure 26 Nodesplit model results
Bayesian result details
Clicking on the ‘Bayesian result details’ tab provides the direct R output from running the
analysis based on the GeMTC package. Here, you can find more detail about the analysis and
results used for the forest plot on tab ‘3a. forest plot’ as shown in Figure 27.
31
Burn-in for 5000, and

results are counted from
the 5001st interaction until
the 25000th .
These are the results

displayed on ‘3a. forest
pot’ tab.
Figure 27 Results details of the Bayesian NMA primary analysis
You can also obtain the Gelman-Rubin convergence assessment plots (Figure 28) for the
Bayesian NMA model. This plot will be useful for checking the simulated model has
converged – a requirement for valid analysis results. Broadly the red line on the plot should
reach stability around the value 1 (as in Figure 27). Further details about this plot can be
found elsewhere. (13)
The current version of the app uses default setting for the number of simulated iterations
for burn-in (5000) and for the model results (20000). While we have observed evidence of
convergence using these defaults for datasets we have tested, we hope to allow more
flexibility in the future so that users can adjust the burn-in period and model results period
(and perhaps the number of simulated chains run and their initial values) based on the
Gelman convergence assessment plot.
32
Figure 28 Gelman convergence assessment plot
Deviance report
Next tab is the ‘3f. Deviance report’ which provides three interactive deviance checking
plots, and deviance data output at the bottom. Once you enter this page, an unrelated mean
effect (UME) inconsistency model (10) will immediately start running which will produce part
of the results shown on the first plot. These plots are designed to be interactive so users can
move cursors on the point to identify which study these outliers were from. The
corresponding specific study and arm information will be displayed beside the point.
We will start by introducing the second and third plots as they are only contributed by the
data from the primary model, and introducing the first (i.e. top) plot in the end, as it plotted
the results from both the primary model and the inconsistency model.
The second plot (Figure 29) is the contribution for each study arm to the residual deviance.
Ideally each point should contribute around one, with bigger values indicating poorer fit and
thus higher residuals. Move your cursor to the three points that have a residual deviance
larger than 2 and the corresponding study which contributed to the points will show. For the
obesity example, study data points with high residual deviances are both study arms of
Ozcelik 2004, and the 2nd arm of Borges 2007. Please note that arms from the same study
are ordered from left to right, for example, the first data point on the left of this chart comes
from the first arm of Aydin 2004, then towards the right, the second arm of Aydin 2004, and
then the third arm of Aydin 2004.
33
Circled points are

the high residual
deviance.
Figure 29 Per-arm residual deviance for all studies
It may be sensible to explore further the studies identified as having high residual deviances,
checking no mistakes were made with data entry, exploring whether the studies differ from
the others studies included in the NMA, whether the studies were poor quality and
therefore at risk of bias. It may also be sensible to do a sensitivity analysis excluding these
studies (note taking whole studies out not just individual arms) to assess the impact their
omission has on the overall NMA results.
The contour plot in Figure 30 takes this a step further by simultaneously looking at residual
deviance and leverage (which is a statistical measure of influence of a data point on model
estimation) simultaneously. The text suggests leverage values outside the contour of 3
(which is a data points contribution to the DIC) are considered poorly fitting. In the obesity
example two points are (just) outside 3. The points leverage values are not extreme, but it is
the (square root) of the residual deviance which are - i.e. the 2 highest values - Ozcelik and
Borges, which we can see by hovering cursor over the points.
34
Circled points have a

leverage over 3, indicating
they have large influence.
Figure 30 Leverage plot
Now, we come back to the first plot (Figure 31). The first plot visualizes the level of
inconsistency for each study arm separately in the analysis. It displays the residual deviance
from an inconsistency model, against the residual deviance from the NMA model. For both
models, the larger the residual deviance, the poorer the fit of that study arm data to the
respective model. When the points are on the equality line (the dotted line on the plot), it
means that relaxing the consistency assumption, by fitting an inconsistency model cannot
improve the model fit and thus the deviance cannot be attributed to conflict between this
and other study’s results when consistency is assumed. When points are below the line, it
means, for that data point, the residual deviance from the NMA model is larger than for the
inconsistency model, indicating that there may be some evidence of inconsistency.
35
Equality line
These are the points

with larger residual
deviance (>2).
Figure 31 Residual deviance plot from NMA model and UME inconsistency model for all studies, on the
‘3f.deviance report’ page.
For the inbuilt example (displayed in Figure 31), there are three points with residual
deviances greater than 2. Two points are on the line indicating that the NMA model and
inconsistency model both provide a similarly poor fit to these data points. The third point is
in the very top right corner again showing poor model fit but it is below the equality line
indicating the inconsistency model provides a better fit. Moving cursor to the points, we can
see that the point in the very top right corner is from the study Borges 2007, and the other
two points (with deviance >2) on the equality line are from Ozcelik 2004.
Users can also use the data table at ’3g. Model details’ – ‘3g-4. Deviance details’ to view the
complete deviance data., where ‘dev.ab’ in the top table is the residual deviance for each
trial arm from the NMA model (Figure 32) and ‘dev.ab’ in the bottom table is from the
inconsistency UME model. Further investigation of these studies may be appropriate.
36
These are the points

with larger residual
deviance (>2).
Figure 32 Deviance data from the NMA model (please note the data may change due to the differences in each
simulation)
Model details
Clicking on the last tab, ‘model details’ will provide additional details about the model
(Figure 33). This page contains the information behind all the summary output, including
model codes and initial values, and you can download the simulated output data (as
produced by the JAGS package). These are primarily included to add transparency to the app
and provide code that ensures any analysis is reproducible (good practice for data analysis),
i.e. you could run this code directly in R to reproduce the results the app presents.
37
Figure 33 Model details
Perform sensitivity analysis

The purpose of the inbuilt sensitivity analysis is to investigate the influence of the different
studies on the NMA results. For instance, you may want to explore the impact on the results
of excluding one or more studies with higher risk of bias, small sample size, with zero events
(for binary outcomes), with slightly different population characteristics, intervention
characteristics, high deviance or leverage, etc. to assess the robustness of the analysis
results to such issues. Studies can easily be excluded from the analysis using a simple
checkbox interface (Figure 34). The user can choose to exclude one or more studies and the
NMA will be updated (for Bayesian NMA, the running button needs to be clicked to update
the results) and displayed alongside the complete case analysis which remains visible to
facilitate comparison. For illustrative purposes only, in this particular example, we have
chosen to exclude 2 studies (Kaya 2004, and Sathuapalan 2008).
38
Figure 34 Checkbox input for sensitivity analyses
Sensitivity analysis: network plots
Now, clicking on the ‘1. Data summary’ – ‘1b. Network plot’, as shown in Figure 35, you can
see that the network plot with studies excluded on the right only contains five treatments,
as the only study that contains rimonbant (Sathuapalan 2008) has been removed from the
analysis and thus rimonbant is no longer connected to the network of evidence and has
therefore disappeared from the network plot.
39
Kaya 2004 is a four-arm study comparing placebo, orlistat,

sibutramine, and orli_sibut. After removing Kaya 2004, the
number of trials on the line connecting each of these
treatment nodes was reduced by 1. After removing
Sathuapaian 2008 which compares rimonbant and
metformin, rimonbant disappeared from the network.
Figure 35 Network plots illustrating the network plot including all studies to the right and the network plot of
treatment comparisons after excluding studies to the right
Sensitivity analysis: forest plots
Similarly, summary forest plots (together with the number of studies included, number of
treatments and the estimated between-study standard deviation) can be displayed for all
studies and with selected studies excluded (Figure 36). Now with the exclusion of the
Sathuapalan 2008 study, rimonbant as a treatment has disappeared from the forest plot
(displayed in the right). With the exclusion of the Kaya 2004 study, the incremental
effectiveness of ‘orli_sibut’ versus placebo became almost zero, with very large uncertainty
(i.e. very large confidence intervals). Click on the ‘1. Data summary’ – ‘1a. Study results’,
then uncheck the Kaya 2004 study, you will see that it provides very certain positive
evidence for the comparison ‘Orli_sibut’ versus placebo whereas the other study (Borges
40
2007) provides an indifferent result. This confirms that after excluding Kaya 2004, the pooled
result for this comparison becomes uncertain.
There were two trials (Kaya 2004 and Borges 2007) providing
direct evidence for the comparison between orli_sibut and
placebo. Kaya 2004 shows a strong evidence that orli_sibut was
more effective than the placebo while the other (Borges 2007)
shows no difference. This can be observed from the ‘1a. study
results’ tab. After excluding Kaya 2004 in the sensitivity
analysis, the estimate for this comparison became uncertain.
Figure 36 Forest plots from frequentist NMA illustrating the results of treatment comparisons including all studies
to the left, and results with studies excluded to the right
Click the ‘3. Bayesian network meta-analysis’ tab, and under the ‘3a. forest plot’ tab, click
the button to run the sensitivity analysis (Figure 37).
Click this button to run sensitivity

41
analysis of Bayesian NMA
Figure 37 Running button for Bayesian sensitivity analysis
The Bayesian NMA sensitivity analysis provided similar results as the frequentist analysis as
shown in Figure 38, where the Orli_sibut is no longer more effective compared to placebo.
Similar results as the frequentist

analysis shown in Figure 36.
Figure 38 Forest plots from Bayesian NMA illustrating the results of treatment comparisons including all studies to
the left, and results with studies excluded to the right
Sensitivity analysis: comparison of all treatment pairs
Coming back to the frequentist analysis, with the two studies excluded (checked on the
‘select studies to exclude’ list), click on ‘2b. comparison of all treatment pairs’, you will see
the comparison table as presented in Figure 39. Kaya 2004 provides a direct comparison
between orli_sibut and orlistat, which closes the loop for ‘orlistat – placebo - orli_sibut’
(Figure 35 to the left). After excluding Kaya 2004, Orli_sibut was no longer in the loop (Figure
35 to the right), and therefore there was no longer indirect evidence contributing to the
42
estimate for orli_sibut versus placebo; the network estimate for orli_sibut versus placebo
becomes the same as the direct estimate (0.10, 95% CI -1.87, 2.07).
After excluding
Kaya 2004,
orli_sib vs.
placebo was no
longer in a
closed loop and
thus indirect
evidence no
longer exists and
the network
pooled estimate
is the same as
the estimate
from direct
evidence only.
Figure 39 Comparison table illustrating the all treatment pairs comparison including all studies at the top, and
results with studies excluded at the bottom
Sensitivity analysis: ranking
The treatment rankings also changed after removing these two studies. It can be observed
that after removing rimonbant from the ranking, the most effective treatment becomes
metformin (table below), compared to the previous orli_sibut in the table above. Orli_Sibut
moves from the second most effective treatment to the second least effective treatment,
with very a similar treatment effect as placebo. The order for the other treatments did not
change, i.e. metformin, sibutramine, orlistat and placebo.
43
Excluded from the

analysis Original ranking for orli_sibut: second
most effective treatment
Ranking changed for

orli_sibut: second least
effective treatment
Figure 40 Comparison table illustrating the all treatment pairs comparison including all studies at the top, and
results with studies excluded at the bottom: ranking changed
This change in the rankings is also reflected in the Bayesian ‘3c. Ranking table’. Clicking ‘3.
Bayesian network meta-analysis’, then ‘3c. Ranking table’ will bring us directly to the results
of the rankings. As shown in Figure 41, the ranking is very slightly different from the
frequentist results where the orli_sibut is ranked in last place, i.e. least effective, compared
to the frequentist results where orli_sibut is ranked before placebo. This is not so surprising
given the estimate is so close to 0 with a very wide confidence interval.
44
Figure 41 Sensitivity analysis of Bayesian NMA: ranking table
Sensitivity analysis: Inconsistency
Assessment of inconsistency between direct and indirect evidence could also be evaluated
after excluding studies for both frequentist NMA and Bayesian NMA. Figure 42 provides the
inconsistency assessment results for all studies at the top and with studies excluded at the
bottom for frequentist NMA and Figure 43 provides the inconsistency results for all studies
on the left and with studies excluded on the right Note, the number of the rows in the table
was reduced after excluding the two studies. This is because one treatment (Rimonbant) was
removed from the network. There was no evidence of inconsistency observed for both the
full network and the reduced network, for both frequentist and Bayesian NMA.
45
Figure 42 Inconsistency table illustrating the estimates including all studies at the top, and results with studies
excluded at the bottom
46
Figure 43 Sensitivity analysis of Bayesian NMA: nodesplit model.
Sensitivity analysis: Bayesian details and deviance report
Additionally, the Bayesian NMA also provides Bayesian details (Figure 44) and deviance
report for the sensitivity analysis with studies excluded.
47
Figure 44 Sensitivity analysis: Bayesian result details
Model fit did not change very much but it can be observed that in the sensitivity analysis the
data points from Borge 2007 no longer produce large residual deviances. This could be
explained by looking at the treatments trialled in Borge 2007 and the excluded trials. It
suggests that the excluded trials (Kaya 2004, and Sathuapalan 2008) had data on the same
treatment comparison or in a loop with the treatment comparison with Borge and the data
were in conflict to some degree, causing the large deviance. Please note that although
excluding studies may reduce the deviance of the remaining studies in the network, this
does not necessarily mean the excluded studies are ‘biased’, or ‘wrong’, it just means they
are in conflict with the remaining studies.
48
These outliers were no

longer shown in the
sensitivity analysis.
Figure 45 Sensitivity analysis: Deviance report
Output results
Each part of the output can be individually downloaded from the app. Depending on the
type of the result, below the output, the user has the option to download results as a
portable document format (PDF), portable network graphic (PNG), for study results, a scalar
vector graphic (SVG), or for tables, a comma-separated values (CSV) file (Figure 46). The
ability to output results allows the user to include the produced figures directly in their NMA
report.
49
Figure 46 Download buttons to output results: an example
50
Acknowledgement
This project was funded by the UK National Institute for Health Research (Grant/Award
Number: 14/178/29). We are thankful to our colleagues, Naomi Bradbury, who was involved
in the development of the app. We are also grateful to Dikshyanta Rana for providing
feedback on a draft of the user manual.
Full list of packages
Package name Version

dplyr 0.8.5
knitr 1.28
plyr 1.8.6
data.table 1.12.8
ggplot2 3.3.0
metafor 2.1.0
gemtc 0.8.4
netmeta 1.2.1
BUGSnet 1.0.3
shiny 1.4.0
shinydashboard 0.7.1
shinyAce 0.4.1
rmarkdown 2.1
shinyalert 1.0
plotly 4.9.1
shinyjs 1.1
shinyWidgets 0.5.1
shinyBS 0.5.1
References
1. Caldwell DM, Ades AE, Higgins JP. Simultaneous comparison of multiple treatments:
combining direct and indirect evidence. Bmj. 2005;331(7521):897-900.
2. Hoaglin DC, Hawkins N, Jansen JP, Scott DA, Itzler R, Cappelleri JC, et al. Conducting
Indirect-Treatment-Comparison and Network-Meta-Analysis Studies: Report of the ISPOR
Task Force on Indirect Treatment Comparisons Good Research Practices: Part 2. Value in
Health. 2011;14(4):429-37.
3. Lu G, Ades AE. Combination of direct and indirect evidence in mixed treatment
comparisons. Stat Med. 2004;23(20):3105-24.
4. Lumley T. Network meta-analysis for indirect treatment comparisons. Statistics in
Medicine. 2002;21(16):2313-24.
51
5. Chang W, Cheng J, Allaire J, J., Xie Y, McPherson J, RStudio., et al. shiny: Web
Application Framework for R 2019 [Available from: https://cran.r-
project.org/web/packages/shiny/index.html.
6. Rücker GS, G.; Krahn, U.; König, J. netmeta: Network Meta-Analysis using
Frequentist Methods. 2017.
7. van Valkenhoef G, Kuiper J. Package ‘gemtc’. Network Meta-Analysis Using Bayesian
Methods. Version 0.8 - 2. 2016 [Available from: https://cran.r-
project.org/web/packages/gemtc/gemtc.pdf.
8. Gray LJ, Cooper N, Dunkley A, Warren FC, Ara R, Abrams K, et al. A systematic review
and mixed treatment comparison of pharmacological interventions for the treatment of
obesity. Obesity reviews. 2012;13(6):483-98.
9. Riley RD, Higgins JPT, Deeks JJ. Interpretation of random effects meta-analyses. BMJ.
2011;342:d549.
10. Dias S, Ades AE, Welton NJ, Jansen JP, Sutton AJ. Chapter 3. Model Fit, Model
Comparison and Outlier Detection. Network Meta‐Analysis for Decision Making. 2018:59-91.
11. Mbuagbaw L, Rochwerg B, Jaeschke R, Heels-Andsell D, Alhazzani W, Thabane L, et
al. Approaches to interpreting and choosing the best treatments in network meta-analyses.
Systematic Reviews. 2017;6(1):79.
12. Dias S, Welton NJ, Caldwell DM, Ades AE. Checking consistency in mixed treatment
comparison meta-analysis. Stat Med. 2010;29(7-8):932-44.
13. Welton NJ, Sutton AJ, Cooper NJ, Abrams KR, Ades AE. Bayesian Methods and
WinBUGS. Evidence Synthesis for Decision Making in Healthcare. Wiley Online Books2012.
p. 17-42.
52

Metainsight User Guide - (Including Bayesian Functionality) Version 0.1

Uploaded by

Copyright:

Available Formats

Metainsight User Guide - (Including Bayesian Functionality) Version 0.1

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Metainsight User Guide - (Including Bayesian Functionality) Version 0.1

Uploaded by

Copyright:

Available Formats

NIHR Complex Review Support Unit (CRSU)

MetaInsight User Guide – (including

The UK NIHR Complex Review Support Unit (CRSU)

Funding information: National Institute for Health

MetaInsight (including Bayesian estimates) V 3** Beta

We delivered a webinar on MetaInsight as part of Cochrane network meta-analysis webinar

This manual can be cited as:

MetaInsight is an interactive web-based application for conducting network meta-analysis

The link to the MetaInsight (including Bayesian estimates) V3 is as below:

The home page

Navigate to a page using the

Outcome selection. Please make

Figure 1 Home page of MetaInsight V3

Using an inbuilt dataset

Download the example

Figure 2 Downloading inbuilt dataset

Loading your own data

Figure 4 Load data page of MetaInsight V3

Figure 5 Upload data to the MetaInsight app - Long data format

Figure 6 Upload data to the MetaInsight app – Wide data format

To load your own dataset:

These steps are shown in Figure 7.

Figure 7 Steps for uploading data

To load your own text label (Figure 8):

Figure 8 Load labels for treatment codes

Network meta-analysis (NMA)

1. Data summary, which contains two subtabs:

Figure 9 Data analysis tab of the MetaInsight V3

Choose data type and NMA analysis options

Choices for data type, direction

Study selection section: The user can

Choices for data type, direction of

• ‘1a. study results’, and

A plot of individual study results

Figure 12 Forest plot of individual study results grouped by treatment comparison.

Distance 3: the shortest distance

Figure 14 network plot information box

Error message is produced for

Figure 15 Using network plot information to detect disconnected networks

Frequentist network meta-analysis

2a. Forest plot

2b. Comparison of all treatment pairs

Figure 16 Forest plots

Comparison of all treatment pairs

Figure 17 Comparison of all treatment pairs

Figure 18 Assessment of inconsistency

Bayesian network meta-analysis

Table 1 Bayesian NMA functions

Tab names functions

Now we will illustrate these functions below.

Figure 19 Running buttons for Bayesian analysis

Always click here if you

Figure 21 Status bar indicating the model is running

Model fit parameters

Figure 22 forest plot results of Bayesian NMA

Comparison of all treatment pairs

-3.76: rimonbant (treatment at

Figure 23 Comparison of all treatment pairs – Bayesian network meta-analysis

The median rank for