Network Meta-Analysis Using Bayesian Methods
Network meta-analyses (mixed treatment comparisons) in the Bayesian
framework using JAGS. Includes methods to assess heterogeneity and
inconsistency, and a number of standard visualizations.
van Valkenhoef et al. (2012) <10.1002>;
van Valkenhoef et al. (2015) <10.1002>.10.1002>10.1002>
- mtc.deviance() for convenient access to deviance statistics and
plots. Additional deviance statistics calculated.
Bugfixes and improvements:
- Depend on rjags >= 4-0, as JAGS 4 features are being used.
- Fix node-splitting models with down-weighted studies.
- Additional sanity checking of input data.
- More appropriate default limits on plotCovariateEffect.
- Corrections to atrialFibrillation dataset.
- Network meta-regression using study-level covariates. Shared,
per-class, exchangeable, or unrelated coefficients.
- Allow adjusting relative effects and rank probabilities to a specific
covariate level. Plot treatment effects versus covariate levels.
- Down-weighting (variance inflation or likelihood adjustment) of
studies (e.g. for inclusion of lower quality evidence).
- Add example datasets for new regression features, as well as examples
for contrast-based data and rate data.
- Improved model fit diagnostics including per-arm or per-study
- Removed "write.mtc.network" function to write the deprecated GeMTC
XML format, as most networks can't be saved in that format anyway.
- Fixed degrees of freedom in ANOHE I^2 computations.
- Update for compatibility with JAGS 4.0
- Support for WinBUGS and OpenBUGS was removed due to their various
problems (lack of portability, lack of support for selecting blocks
from arrays, bugs in setting monitors, and the constant need to track
minor differences in syntax between JAGS and BUGS).
- Added binom/log (risk ratio) support.
- Ability to generate n * n relative effects table, and to generate
plots directly from the table, bypassing the time-consuming
computation of the quantiles.
- Added residual deviance statistics (replaces and improves upon the
DIC previously computed by JAGS).
- Example datasets are now standard R data, and can be accessed using
- Forest plots: allow treatment descriptions to be displayed instead of
- Network plot: make proportional line width optional.
- Additional validations to catch obvious data entry mistakes.
- Better safeguards for starting values, and alternative approach to
baseline priors for scales without infinite support.
- The XML package is no longer required to install gemtc, and will be
loaded only if needed to read/write legacy .gemtc XML files.
- Network plot: line width is now proportional to the number of
- Forest plots: new options 'draw.no.effect' and 'center.label'. See
?blobbogram for documentation.
- Better validation of network properties and more helpful reporting of
- Prevent zero prior heterogeneity variance (because it causes JAGS
- Clear error when asked to parse non-existant XML file.
- Use the '.Call' interface for C code
- Better detect invalid treatment names
- Better detect the likelihood / link
- Fix node-splitting models in sparse networks (#31)
- Customizable prior distribution for heterogeneity. The prior can be
set on the standard deviation, variance, or precision. It can use any
distribution supported by BUGS/JAGS. Informative priors for the log
odds ratio (based on Turner et al. 2012) are also available.
- Variance of the vague normal distributions for all other parameters
can be controlled using 'om.scale'. By default it is still
- Change handling of 4+-arm trials in node-splitting models and ANOHE
- Enable linearModel='fixed' for mtc.anohe()
- Add literature references for various methods to gemtc-package
- Add mtc.data.studyrow() to convert the one-study-per-row format often
used in BUGS models to the one-arm-per-row format used by GeMTC
- Upgrade unit testing infrastructure to testthat >= 0.8
- Move validation tests into testthat infrastructure
- Add regression test suite that tries all summaries and plots on all
models to check whether they crash.
- General clean up of example and test data files
- Models with relative-effect data from trials with 4 or more arms
can't be run using WinBUGS or OpenBUGS (#28). This seems to be due to
a BUGS-bug. A warning is produced, saying to use JAGS instead.
- Fix empty first page on PDF output of some plots (#12)
- Fix ordering of generated data / initial values in models for
relative effect data (#21)
- Fix generating node-splitting models for relative effect data (#22)
- Fix mtc.nodesplit.comparisons getting confused when both relative
effect and arm-based data are present (#25)
- Fix duplicated parameters in UME models (#26)
- Fix incompatibility with igraph 0.7
- Fix node-splitting model generation bug
- Remove Matrix dependency by not using the Matrix-dependent parts of
- Correctly use C99 "inline" facility
- Converted several plots to use the standard R mechanism of asking for
the next page, rather than our own hack.
- The plot() for the summary() of mtc.anohe() now has a nicer default
- Import Matrix package because of igraph changes.
- mtc.anohe() mixed up treatments or studies in some extraordinary
circumstances, due to automatic conversions to/from factor.
- blobbogram() would crash when it could not find reasonable scales and
none were specified.
- blobbogram() created an empty first page when plotting to off-screen
- mtc.anohe() called sd() on a matrix, which is deprecated.
- mtc.nodesplit() used "A vs B" for d.A.B instead of "B vs A" like the
- node-splitting models
- mtc.nodesplit() wrapper to run multiple node-splitting models
- calculate DIC model fit (JAGS only)
- add preferredDirection to rank.probability()
- default plot() for rank.probability()
- fixed effect models
- poisson/log likelihood/link for survival data
- allow std.err instead of std.dev + sampleSize in continuous data sets
- refuse to generate models when treatments are duplicated in
multiple arms of the same study, since the code does not take this
- network$data is now network$data.ab for consistency with
- likelihood/link implementations can now define their data
- export ll.call for programmatic access to likelihood/link specific
- multi-arm trial decomposition sometimes resulted in NA standard
- WinBUGS and OpenBUGS had complaints about the relative effect code
- safer matrix indexing: explicit 'drop='
- safer data frame indexing: use [['x']] instead of $x
Many additional features:
- support for relative effect data
- support for mixed arm-based and relative effect data
- binom/cloglog likelihood/link for rate data
- unrelated mean effects (UME) model
- unrelated study effects (USE) model
- analysis of heterogeneity with heterogeneity plot (EXPERIMENTAL)
- full access to generated code, data structures, and parameters
- guard against "impossible" initial values
Also, bugfixes and documentation updates.
Dropped Java dependency by rewriting core algorithms in R.
- Remove dependency on rJava
- Use own forest plot methods instead of meta package
- Fix bug in rank.probabilities when only two treatments in network
- Various bugfixes
- Compatibility note: inconsistency models are no longer supported by
- Compatibility note: the argument order of mtc.network has changed
- Compatibility note: the YADAS sampler is no longer available
Bugfix release. Addresses the following issues:
- Correct the initial values generation for the random effects standard
- Do not fail on relative.effect when saving and loading samples
- Make Java open tab-completed paths
- Use correct column name for continuous data
- When only YADAS available, run YADAS by default
- Fix compatibility issue with Java on 32bit Mac OS X.
First official release of the GeMTC R package. The package enables
Bayesian network meta-analysis (also known as MTC, Mixed Treatment
Comparisons) in R. Network meta-analysis models can be generated and
then run using MCMC software: JAGS (using the rjags package), OpenBUGS
(using the BRugs package), WinBUGS (using the R2WinBUGS package) or
YADAS (provided by GeMTC). The GeMTC GUI can be used instead of or in
conjunction with the R package.