Fixes

• updating the C++ to compile on M1 Mac

Changes

• message about the effect size scale of parameter estimates is always shown
• compatibility with BayesTools 0.2.0+

Features

• three-level meta-analysis by specifying study_ids argument in RoBMA. However, note that this is (1) an experimental feature and (2) the computational expense of fitting selection models with clustering is extreme. As of now, it is almost impossible to have more than 2-3 estimates clustered within a single study).

Fixes

• adding Windows ucrt patch (thanks to Tomas Kalibera)

Fixes

• incorrectly formatted citations in vignettes and capitalization

Features

• adding informed_prior() function (from the BayesTools package) that allows specification of various informed prior distributions from the field of medicine and psychology
• adding a vignette reproducing the example of dentine sensitivity with the informed Bayesian model-averaged meta-analysis from Bartoš et al., 2021 (open-access),
• further reductions of fitted object size when setting save = "min"

Fixes

• correcting wrong PEESE transformation for the individual models summaries (issue #12)
• fixing error message for missing conditional PET-PEESE
• fixing incorrect lower bound check for log(OR)

Features

• adding interpret() function (issue #11)
• adding effect size transformation via output_scale argument to plot() and plot_models() functions
• better handling of effect size transformations and scaling - BayesTools style back-end functions with Jacobian transformations

Please notice that this is a major release that breaks backwards compatibility.

Changes

• naming of the arguments specifying prior distributions for the different parameters/components of the models changed (priors_mu -> priors_effect, priors_tau -> priors_heterogeneity, and priors_omega -> priors_bias),
• prior distributions for specifying weight functions now use a dedicated function (prior(distribution = "two.sided", parameters = ...) -> prior_weightfunction(distribution = "two.sided", parameters = ...)),
• new dedicated function for specifying no publication bias adjustment component / no heterogeneity component (prior_none()),
• new dedicated functions for specifying models with the PET and PEESE publication bias adjustments (prior_PET(distribution = "Cauchy", parameters = ...) and prior_PEESE(distribution = "Cauchy", parameters = ...)),
• new default prior distribution specification for the publication bias adjustment part of the models (corresponding to the RoBMA-PSMA model from Bartoš et al., 2021 preprint),
• new model_type argument allowing to specify different “pre-canned” models ("PSMA" = RoBMA-PSMA, "PP" = RoBMA-PP, "2w" = corresponding to Maier et al., in press , manuscript),
• combine_data function allows combination of different effect sizes / variability measures into a common effect size measure (also used from within the RoBMA function),
• better and improved automatic fitting procedure now enabled by default (can be turned of with autofit = FALSE)
• prior distributions can be specified on the different scale than the supplied effect sizes (the package fits the model on Fisher’s z scale and back transforms the results back to the scale that was used for prior distributions specification, Cohen’s d by default, but both of them can be overwritten with the prior_scale and transformation arguments),
• new prior distributions, e.g., beta or fixed weight functions,
• estimates from individual models are now plotted with the plot_models() function and the forest plot can be obtained with the forest() function,
• the posterior distribution plots for the individual weights are no able supported, however, the weightfunction and the PET-PEESE publication bias adjustments can be visualized with the plot.RoBMA() function and parameter = "weightfunction" and parameter = "PET-PEESE".

Fixes

• check_setup function not working at all

Changes

• the studies’s true effects are now marginalized out of the random effects models and are no longer estimated (see Appendix A of our prerint for more details). As a results, arguments referring to the true effects are now disabled.
• all models are now being estimated using the likelihood of effect sizes (instead of test-statistics as usually defined). We reproduced the simulation study that we used to evaluate the method performance and it achieved identical results (up to MCMC error, before marginalizing out the true effects). A big advantage of using the normal likelihood for effect sizes is a considerable speed up of the whole estimation process.
• as a results of these two changes, the results of the models will differ to those of pre 1.2.0 version

Fixes

• autofit being turn on if any control argument was specified

Fixes

• vdiffr not being used conditionally in unit tests

Fixes

• inability to fit a model without specifying a seed
• inability to produce individual model plots due to incompatibility with the newer versions of ggplot2

Features

• parallel within and between model fitting using the parallel package with ‘parallel = TRUE’ argument

Fixes:

• models being fitted automatically until reaching R-hat lower than 1.05 without setting max_rhat and autofit control parameters
• bug preventing to draw a bivariate plot of mu and tau
• range for parameter estimates from individual models no containing 0 (or 1 in case of OR measured effect sizes)
• inability to fit a model with only null mu distributions if correlation or OR measured effect sizes were specified
• ordering of the estimated and observed effects when both of them are requested simultaneously
• formatting of this file (NEWS.md)

Improvements:

• priors plot: parameter specification, default plotting range, clearer x-axis labels in cases when the parameter is defined on transformed scale
• parameters plots: probability scale always ends at the same spot as is the last tick on the density scale
• adding warnings if any of the specified models has Rhat higher than 1.05 or the specified value
• grouping the same warnings messages together

Fixes:

• inability to run models without the silent = TRUE control

Features:

• x-axis rescaling for the weight function plot (by setting ‘rescale_x = TRUE’ in the ‘plot.RoBMA’ function)
• setting expected direction of the effect in for RoBMA function

Fixes:

• marginal likelihood calculation for models with spike prior distribution on mean parameter which location was not set to 0

CRAM requested changes:

• changing information messages from ‘cat’ to ‘message’ from plot related functions
• saving and returning the ‘par’ settings to the user defined one in the base plot functions

Fixes:

• the summary and plot function now shows quantile based confidence intervals for individual models instead of the HPD provided before (this affects only ‘summary’/‘plot’ with ‘type = “individual”’, all other confidence intervals were quantile based before)

Fixes:

• summary function returning median instead of mean

Fixes:

• incorrectly weighted theta estimates
• models with non-zero point prior distribution incorrectly plotted using when “models” option in case that the mu parameter was transformed