- Faster code: all models should now run by up to 50% faster
- For standardised data, where mean in control group is by definition
0, you can now say
`pooling_control = "remove"`

when calling`baggr()`

. This will avoid estimating parameters which are known to be 0. - Predicting effects for new data: for models with covariates you can
use
`effect_draw(object, newdata = ...)`

or (equivalently)`predict(object, newdata = ...)`

to generate predictions for any number of new samples - I updated the calculation of the pooling metric so that it’s comparable 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 run
`baggr()`

without any extra steps like`prepare_ma()`

, by just defining`effect`

when running baggr (or it will default to log OR). - I added alias
`posterior_predict()`

for drawing from posterior sample. This is more consistent with regression modeling and RStan ecosystem.

Bugs:

- Transforms of samples on
`baggr_compare`

plots previously didn’t work for some plots. This is now fixed. - Fixed a bug when covariates didn’t work for some types of summary-level binary data.

Misc: * Printing `baggr`

and `baggr_compare`

objects is now better at showing intervals and you can also change their
widths with arguments passed to `print.baggr()`

or directly
to `baggr_compare()`

* Added `student_t()`

and
`lognormal()`

priors and updated some prior documentation *
Removed some cases where input data would be reordered (previously this
could happen to either individual-level continuous data or summary data
of binary events) * More warning prompts at various stages of model
fitting * Faster installation and package checks.

- You can add numerical values to
`plot.baggr_compare`

and`baggr_plot`

graphics (a la forest plot) - You don’t need to convert summary data to individual-level data
before running
`model="logit"`

, call to`baggr()`

should detect it automatically now `pooling()`

includes extra metrics, including study weights calculation (and better documentation)- You can now plot the objects returned by
`loocv()`

to understand out-of-sample performance graphically - Risk difference models are now easy to fit, you only need to
transform your binary input data with
`prepare_ma(..., effect = "RD")`

Misc:

- You can plot hyperparameter values only (without group-specific
estimates) in
`baggr_compare()`

now - Removed an unnecessary dependency on the
`quantreg`

package - Rare event corrections (
`prepare_ma()`

) can now be applied either to particular studies or all data (the literature sometimes recommends the latter) - Clearer prompts about priors and pooling in control arms when working with individual-level data models.
- Can now set priors for error terms in linear regression models
(
`prior_sigma`

) - Added
`lognormal()`

prior and updated some prior documentation

Bug fixes:

- Print errors when examining LOO CV results
- LOO CV with full pooling and binary outcomes now works again after being broken in 0.6. Some of the results in 0.5 and 0.6 releases may have been wrong
- Individual-level Rubin model with covariates was also broken in 0.6
- Fixed a calculation of default beta prior
- No more confusing warnings about setting
`prior_control`

for`"logit"`

model. `binary_to_individual`

with non-integer number of events warns user and throws an error now- Confusing results in
`baggr_binary`

vignette (rare events section) - Fixes crashes for elpd calculations with unusual binary input data

- Mu & tau models now also print correlations between effects, via
a new function
`mutau_cor`

- You can now change type of visual comparison
(
`baggr_compare`

) on the fly (between`"effects"`

and`"groups"`

). Printing comparisons also returns posterior predictive draws. - Upgraded forest plots to work with
`forestplot`

2.0

Minor bug fixes:

- Fixed errors that could happen when using multiple factor
covariates, or various covariate models with
`loocv()`

- Fixed a bug with reporting wrong SD’s for effect in the v0.6
`mutau`

model when using`plot.baggr_compare`

- Fixed ordering of groups in
`baggr_compare()`

- Various small changes to reduce amount of persistent messages triggered by normal user behaviour.
- Fixed a bug where priors for meta-regressions were set even though there were no covariates.

- Various documentation fixes for re-submission of v0.6 to CRAN (first one since v0.4).
- Added
`summary`

option for`effect_draw`

. - Factor covariates will work (better) now.
- Removed some non-essential code for faster compilation on CRAN.

New

`"mutau_full"`

model is a generalisation of the`"mutau"`

model into individual-level data. The idea is similar as for the recent`"rubin_full"`

changes, see version 0.6.0.I also reparameterised the

`mutau`

model. It should be faster and have fewer divergent transition warnings.Some of the code around the mu and tau model has also been rewritten on the back end.

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

- Spike and slab model can be called via
`model="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 via
`model="rubin_full"`

rather than`"full"`

. Old syntax will still work, however. Made some documentation and code improvements around this issue. - Leave-one-out cross-validation works for
`model="rubin_full"`

now. It works the same way as for`model="logit"`

. See`?baggr`

for more information on how to use it. - It’s now possible to use
`model="rubin"`

with the same inputs as`model="mutau"`

. Some data columns are removed automatically in that case.

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

- Fixed a few issues with formatting data for individual-level data models.
- Fixed a major bug with distributions of baselines in the
`rubin_full`

(`full`

) model. - Fixed glitchy display for some
`baggr_compare`

plots.

- Fixed and random effects for
`baggr`

models now have their own separate functions,`fixed_effects`

and`random_effects`

, in addition to`group_effects`

- LOO CV works for the logistic model (as does general cross-validation).
- Vignette for binary data analysis has been rewritten in parts.
- L’Abbe plots for binary data, see
`labbe()`

. - There is now more automatic conversion between summary-level and
individual-level data for binary data (e.g. you can run
`baggr()`

with summary data and`model="logit"`

for automatic conversion) - For logistic model, priors can be specified for rates of events in
the control arm, see arguments
`prior_control`

and`prior_control_sd`

in`baggr()`

- There are experimental features for working with models of quantiles. We advise against fitting such models using the package until these features have been fully tested and documented.

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

- Covariates can now be used in all baggr() models: in “rubin” model they 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 through fixed_effects(),
- you will also see them when printing baggr objects;
- when using forest_plot() you can request
`show = "covariates"`

- Prototype of pp_check() function now works for Rubin model (thanks to Brice Green) you can apply it to generate new datasets according to posterior distribution of treatment effect and contrast them with the observed quantities as part of model checking
- baggr_compare() function now has standard output which you can print() or plot(), thanks to Brice Green
- Vignettes and documentation were updated to better describe binary data analysis
- We now give more warnings when plugging in stupid inputs.

- Messages for setting priors were accidentally given when e.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 next release.
- LOO CV is also disabled for some models. Prompts will be given.

- Binary data models for both summary and individual-level data.
- New vignette for working with binary data; see
`vignette("baggr_binary")`

. - Expanded helper functions (esp.
`prepare_ma`

), esp. for prepping binary data. - Added forest plot functionality for all types of models.
- Various outputs can now be transformed (main use case is
`exp`

, but any transform is allowed). - Reworked vignette sections for pooling and cross-validation.
- Pooling statistics are now calculated for the whole model and better documented.
- More consistent theming, similar to bayesplot (thanks to Brice Green)
- Comparison of leave-one-out cross-validations with
`loo_compare`

(thanks to Brice Green)

- Re-enabled missing Cauchy priors

- Users can now define priors in
`baggr()`

using a syntax similar to`rstanarm`

. Extra priors are available `baggr()`

outputs prior predictive distributions; they can be examined using`baggr_compare`

and`effect_plot`

,`effect_draw`

– 2 new functions- More types of model comparisons are possible
- LOO CV function has been reworked
- Full pooling and no pooling models have been reworked to avoid divergent transitions.

First package version for CRAN.