Movatterモバイル変換

baggr 0.7.11 (late 2024)

Two big additions:

1. You can run models with individual-level data (continuous andlogit) with arbitrary number of treatment effects. Instead of definingcolumntreatment in data as binary, just use a factor. Thiswas designed especially for multi-arm experiments. There may be ageneralisation into network meta-analysis later.

2. You can add random effects for clusters (for now only of therandom intercepts variety, u ~ normal(0, sigma_cluster)) and definepriors for them.

Bugs:

• Better checks for wrongtest_data inbaggr() andloocv()

baggr 0.7.8 (late 2023)

New functionality: - You can run meta-analyses with just one row ofdata, but must specify priors -baggr_plot can be made tolook more like forest plot withbaggr_plot(bg, style = "forest") - Plotting baggr andbaggr_compare objects now has more powerfuladd_valuesfunctionality. - You can customise colour of hypermean,e.g. baggr_plot(bg, hyper = "red") - For meta-regressions,you can draw abubble()

Back end and minor changes:

• Various CRAN and rstan bugfixes. Now that rstan 2.26 is on CRAN,this package requires users to update to it.
• A few “quality of life”-type and presentation upgrades.
• You can print hyperparameters of a model usinghypermean() andhypersd() (defaults tosummaries)
• Printing baggr models now also includes correlations and theposterior predictive distribution.
• loo_compare now has better annotation

Bugfixes:

• Forest plots now display correctly even if you don’t labelstudies/groups

baggr 0.7.6 (March 2022)

• Various CRAN and rstan bugfixes. A few quality-of-life typeupgrades.
• You can run meta-analyses with just one row of data, but mustspecify names for

baggr 0.7.4 (late 2022)

• Faster code: all models should now run by up to 50% faster
• For standardised data, where mean in control group is by definition0, you can now saypooling_control = "remove" when callingbaggr(). This will avoid estimating parameters which areknown to be 0.
• Predicting effects for new data: for models with covariates you canuseeffect_draw(object, newdata = ...) or (equivalently)predict(object, newdata = ...) to generate predictions forany number of new samples
• I updated the calculation of the pooling metric so that it’scomparable with frequentist packages. See the help file.

Misc:

• More information when printing models.
• Baggr automatically checks for a grouping column.
• For binary data, you can runbaggr() without any extrasteps likeprepare_ma(), by just definingeffect when running baggr (or it will default to logOR).
• I added aliasposterior_predict() for drawing fromposterior sample. This is more consistent with regression modeling andRStan ecosystem.

Bugs:

• Transforms of samples onbaggr_compare plots previouslydidn’t work for some plots. This is now fixed.
• Fixed a bug when covariates didn’t work for some types ofsummary-level binary data.

baggr 0.6.21 (January-March2022)

Misc: * Printingbaggr andbaggr_compareobjects is now better at showing intervals and you can also change theirwidths with arguments passed toprint.baggr() or directlytobaggr_compare() * Addedstudent_t() andlognormal() priors and updated some prior documentation *Removed some cases where input data would be reordered (previously thiscould happen to either individual-level continuous data or summary dataof binary events) * More warning prompts at various stages of modelfitting * Faster installation and package checks.

baggr 0.6.10-0.6.18 (Sept-Dec2021)

• You can add numerical values toplot.baggr_compare andbaggr_plot graphics (a la forest plot)
• You don’t need to convert summary data to individual-level databefore runningmodel="logit", call tobaggr()should detect it automatically now
• pooling() includes extra metrics, including studyweights calculation (and better documentation)
• You can now plot the objects returned byloocv() tounderstand out-of-sample performance graphically
• Risk difference models are now easy to fit, you only need totransform your binary input data withprepare_ma(..., effect = "RD")

Misc:

• You can plot hyperparameter values only (without group-specificestimates) inbaggr_compare() now
• Removed an unnecessary dependency on thequantregpackage
• Rare event corrections (prepare_ma()) can now beapplied either to particular studies or all data (the literaturesometimes recommends the latter)
• Clearer prompts about priors and pooling in control arms whenworking with individual-level data models.
• Can now set priors for error terms in linear regression models(prior_sigma)
• Addedlognormal() prior and updated some priordocumentation

Bug fixes:

• Print errors when examining LOO CV results
• LOO CV with full pooling and binary outcomes now works again afterbeing broken in 0.6. Some of the results in 0.5 and 0.6 releases mayhave been wrong
• Individual-level Rubin model with covariates was also broken in0.6
• Fixed a calculation of default beta prior
• No more confusing warnings about settingprior_controlfor"logit" model.
• binary_to_individual with non-integer number of eventswarns user and throws an error now
• Confusing results inbaggr_binary vignette (rare eventssection)
• Fixes crashes for elpd calculations with unusual binary inputdata

baggr 0.6.5-0.6.9(June-August 2021)

• Mu & tau models now also print correlations between effects, viaa new functionmutau_cor
• You can now change type of visual comparison(baggr_compare) on the fly (between"effects"and"groups"). Printing comparisons also returns posteriorpredictive draws.
• Upgraded forest plots to work withforestplot 2.0

Minor bug fixes:

• Fixed errors that could happen when using multiple factorcovariates, or various covariate models withloocv()
• Fixed a bug with reporting wrong SD’s for effect in the v0.6mutau model when usingplot.baggr_compare
• Fixed ordering of groups inbaggr_compare()
• Various small changes to reduce amount of persistent messagestriggered by normal user behaviour.
• Fixed a bug where priors for meta-regressions were set even thoughthere were no covariates.

baggr 0.6.3-0.6.4 (May 2021)

• Various documentation fixes for re-submission of v0.6 to CRAN (firstone since v0.4).
• Addedsummary option foreffect_draw.
• Factor covariates will work (better) now.
• Removed some non-essential code for faster compilation on CRAN.

baggr 0.6.2 (April 2021)

• New"mutau_full" model is a generalisation of the"mutau" model into individual-level data. The idea issimilar as for the recent"rubin_full" changes, see version0.6.0.

• I also reparameterised themutau model. It should befaster and have fewer divergent transition warnings.Some of the codearound the mu and tau model has also been rewritten on the backend.

On the back end the package now follows the rstantools recommendedway of compiling models. The user experience should be exactly the same,but this may avoid some problems when installing the package from GitHubor otherwise compiling it locally.

baggr 0.6.0 (February 2021)

New features

• Spike and slab model can be called viamodel="sslab".See?baggr for basics of working with this type of a model.A vignette will be added soon.
• Rubin model with full data is now called viamodel="rubin_full" rather than"full". Oldsyntax will still work, however. Made some documentation and codeimprovements around this issue.
• Leave-one-out cross-validation works formodel="rubin_full" now. It works the same way as formodel="logit". See?baggr for more informationon how to use it.
• It’s now possible to usemodel="rubin" with the sameinputs asmodel="mutau". Some data columns are removedautomatically in that case.

For v0.6 we added more generic code around plotting, printing,grabbing treatment effects etc. While there are no differences on thefront-end, this means that for the next versions we will be able toconsider some new models and have more homogeneous syntax for allmodels.

Bugs

• Fixed a few issues with formatting data for individual-level datamodels.
• Fixed a major bug with distributions of baselines in therubin_full (full) model.
• Fixed glitchy display for somebaggr_compareplots.

baggr 0.5.0 (June 2020)

New features

• Fixed and random effects forbaggr models now havetheir own separate functions,fixed_effects andrandom_effects, in addition togroup_effects
• LOO CV works for the logistic model (as does generalcross-validation).
• Vignette for binary data analysis has been rewritten in parts.
• L’Abbe plots for binary data, seelabbe().
• There is now more automatic conversion between summary-level andindividual-level data for binary data (e.g. you can runbaggr() with summary data andmodel="logit"for automatic conversion)
• For logistic model, priors can be specified for rates of events inthe control arm, see argumentsprior_control andprior_control_sd inbaggr()
• There are experimental features for working with models ofquantiles. We advise against fitting such models using the package untilthese features have been fully tested and documented.

Bug fixes

• Fixed some issues with printing of coefficients in meta-regressions,where wrong values were given for some models.

baggr 0.4.0 (February 2020)

New features

• Covariates can now be used in all baggr() models: in “rubin” modelthey give meta-regression (group-level covariates), while in “full” and“logit” models they can be used for “regular” regression(individual-level covariates)
• Priors for covariates are set through the argument prior_beta
• You can work with regression coefficients for covariates
  • you can access and summarise coefficients throughfixed_effects(),
  • you will also see them when printing baggr objects;
  • when using forest_plot() you can requestshow = "covariates"
• Prototype of pp_check() function now works for Rubin model (thanksto Brice Green) you can apply it to generate new datasets according toposterior distribution of treatment effect and contrast them with theobserved quantities as part of model checking
• baggr_compare() function now has standard output which you canprint() or plot(), thanks to Brice Green
• Vignettes and documentation were updated to better describe binarydata analysis
• We now give more warnings when plugging in stupid inputs.

Bug fixes

• Messages for setting priors were accidentally given whene.g. running full pooling models
• All models were re-written to standardise our approach and syntax.
  • “Full” model might now behave differently.
  • “Mutau” model will be re-written and generalised for nextrelease.
  • LOO CV is also disabled for some models. Prompts will be given.

baggr 0.3.0

New features

• Binary data models for both summary and individual-level data.
• New vignette for working with binary data; seevignette("baggr_binary").
• Expanded helper functions (esp. prepare_ma), esp. forprepping binary data.
• Added forest plot functionality for all types of models.
• Various outputs can now be transformed (main use case isexp, but any transform is allowed).
• Reworked vignette sections for pooling and cross-validation.
• Pooling statistics are now calculated for the whole model and betterdocumented.
• More consistent theming, similar to bayesplot (thanks to BriceGreen)
• Comparison of leave-one-out cross-validations withloo_compare (thanks to Brice Green)

Bug fixes

• Re-enabled missing Cauchy priors

baggr 0.2.0

New features

• Users can now define priors inbaggr() using a syntaxsimilar torstanarm. Extra priors are available
• baggr() outputs prior predictive distributions; theycan be examined usingbaggr_compare andeffect_plot,effect_draw – 2 newfunctions
• More types of model comparisons are possible
• LOO CV function has been reworked
• Full pooling and no pooling models have been reworked to avoiddivergent transitions.

baggr 0.1.0

First package version for CRAN.