NEWS.md
stan_jm() is not available for 32bit Windows
Some improvements to prior distributions, as described in detail in the vignette Prior Distributions for rstanarm Models and book Regression and Other Stories. These changes shouldn’t cause any existing code to error, but default priors have changed in some cases:
autoscale argument to functions like normal(), student_t(), etc., now defaults to FALSE except when used by default priors (default priors still do autoscalinng). This makes it simpler to specify non-default priors. (#432)kfold() for stan_gamm4() models that used random argument (#435)posterior_predict() and posterior_linpred() when using newdata with family = mgcv::betar (#406, #407)singular.ok now rules out singular design matrices in stan_lm() (#402)data is a data.table object (#434, @danschrage)New method posterior_epred() returns the posterior distribution of the conditional expectation, which is equivalent to (and may eventually entirely replace) setting argument transform=TRUE with posterior_linpred(). (#432)
Added convenience functions logit() and invlogit() that are just wrappers for qlogis() and plogis(). These were previously provided by the arm package. (#432)
src/Makevars{.win} now uses a more robust way to find StanHeaders
Fixed bug where ranef() and coef() methods for glmer-style models printed the wrong output for certain combinations of varying intercepts and slopes.
Fixed a bug where posterior_predict() failed for stan_glmer() models estimated with family = mgcv::betar.
Fixed bug in bayes_R2() for bernoulli models. (Thanks to @mcol)
loo_R2() can now be called on the same fitted model object multiple times with identical (not just up to rng noise) results. (Thanks to @mcol)
New vignette on doing MRP using rstanarm. (Thanks to @lauken13)
4x speedup for most GLMs (stan_glm()) and GAMs (stan_gamm4() without random argument). This comes from using Stan’s new compound _glm functions (normal_id_glm, bernoulli_logit_glm, poisson_log_glm, neg_binomial_2_log_glm) under the hood whenever possible. (Thanks to @avehtari and @VMatthijs)
compare_models() is deprecated in favor of loo_compare() to keep up with the loo package (loo::loo_compare())
The kfold() method now has a cores argument and parallelizes by fold rather than by Markov chain (unless otherwise specified), which should be much more efficient when many cores are available.
For stan_glm() with algorithm='optimizing', Pareto smoothed importance sampling (arxiv.org/abs/1507.02646, mc-stan.org/loo/reference/psis.html) is now used to diagnose and improve inference (see https://avehtari.github.io/RAOS-Examples/BigData/bigdata.html). This also now means that we can use PSIS-LOO also when algorithm='optimizing'. (Thanks to @avehtari)
For stan_glm() the "meanfield" and "fullrank" ADVI algorithms also include the PSIS diagnostics and adjustments, but so far we have not seen any example where these would be better than optimzation or MCMC.
stan_clogit() now works even when there are no common predictorsprior.info() works better with models produced by stan_jm() and stan_mvmer()
stan_glm() (only) gets a mean_PPD argument that when FALSE avoids drawing from the posterior predictive distribution in the Stan codeposterior_linpred() now works even if the model was estimated with algorithm = "optimizing"
stan_jm() and stan_mvmer() now correctly include the intercept in the longitudinal submodelCompatible with loo package version >= 2.0
QR = TRUE no longer ignores the autoscale argument and has better behavior when autoscale = FALSE
posterior_linpred() now has a draws argument like for posterior_predict()
Dynamic predictions are now supported in posterior_traj() for stan_jm models.
More options for K-fold CV, including manually specifying the folds or using helper functions to create them for particular model/data combinations.
Lots of good stuff in this release.
stan_polr() and stan_lm() handle the K = 1 case betterThe prior_aux arguments now defaults to exponential rather than Cauchy. This should be a safer default.
The Stan programs do not drop any constants and should now be safe to use with the bridgesampling package
hs() and hs_plus() priors have new defaults based on a new paper by Aki Vehtari and Juho Piironen
stan_gamm4() is now more closely based on mgcv::jagam(), which may affect some estimates but the options remain largely the same
The product_normal() prior permits df = 1, which is a product of … one normal variate
The build system is more conventional now. It should require less RAM to build from source but it is slower unless you utilize parallel make and LTO
stan_jm() and stan_mvmer() contributed by Sam Brilleman
bayes_R2() method to calculate a quantity similar to R2
stan_nlmer(), which is similar to lme4::nlmer but watch out for multimodal posterior distributions
stan_clogit(), which is similar to survival::clogit but accepts lme4-style group-specific terms
The mgcv::betar family is supported for the lme4-like modeling functions, allowing for beta regressions with lme4-style group terms and / or smooth nonlinear functions of predictors
Fix to stan_glmer() Bernoulli models with multiple group-specific intercept terms that could result in draws from the wrong posterior distribution
Fix bug with contrasts in stan_aov() (thanks to Henrik Singmann)
Fix bug with na.action in stan_glmer() (thanks to Henrik Singmann)
Fix for intercept with identity or square root link functions for the auxiliary parameter of a beta regression
Fix for special case where only the intercepts vary by group and a non-default prior is specified for their standard deviation
Fix for off-by-one error in some lme4-style models with multiple grouping terms
New methods loo_linpred(), loo_pit(), loo_predict(), and loo_predictive_interval()
Support for many more plotfuns in pp_check() that are implemented in the bayesplot package
Option to compute latent residuals in stan_polr() (Thanks to Nate Sanders)
The pairs plot now uses the ggplot2 package
VarCorr() could return duplicates in cases where a stan_{g}lmer model used grouping factor level names with spaces
The pairs() function now works with group-specific parameters
The stan_gamm4() function works better now
Fix a problem with factor levels after estimating a model via stan_lm()
New model-fitting function(s) stan_betareg() (and stan_betareg.fit()) that uses the same likelihoods as those supported by the betareg() function in the betareg package (Thanks to Imad Ali)
New choices for priors on coefficients: laplace(), lasso(), product_normal()
The hs() and hs_plus() priors now have new global_df and global_scale arguments
stan_{g}lmer() models that only have group-specific intercept shifts are considerably faster now
Models with Student t priors and low degrees of freedom (that are not 1, 2, or 4) may work better now due to Cornish-Fisher transformations
Many functions for priors have gained an autoscale argument that defaults to TRUE and indicates that rstanarm should make internal changes to the prior based on the scales of the variables so that they default priors are weakly informative
The new compare_models() function does more extensive checking that the models being compared are compatible
reloo() if data was not specifiedpp_validate() that was only introduced on GitHubUses the new bayesplot and rstantools R packages
The new prior_summary() function can be used to figure out what priors were actually used
stan_gamm4() is better implemented, can be followed by plot_nonlinear(), posterior_predict() (with newdata), etc.
Hyperparameters (i.e. covariance matrices in general) for lme4 style models are now returned by as.matrix() and as.data.frame()
pp_validate() can now be used if optimization or variational Bayesian inference was used to estimate the original model
Fix for bad bug in posterior_predict() when factor labels have spaces in lme4-style models
Fix when weights are used in Poisson models
posterior_linpred() gains an XZ argument to output the design matrixstan_biglm() function that somewhat supports biglm::biglm
as.array() method for stanreg objects
posterior_predict() with newdata now works correctly for ordinal models
stan_lm() now works when intercept is omitted
stan_glmer.fit() no longer permit models with duplicative group-specific terms since they don’t make sense and are usually a mistake on the user’s part
posterior_predict() with lme4-style models no longer fails if there are spaces or colons in the levels of the grouping variables
posterior_predict() with ordinal models outputs a character matrix now
pp_validate() function based on the BayesValidate package by Sam Cook
posterior_vs_prior() function to visualize the effect of conditioning on the data
Works (again) with R versions back to 3.0.2 (untested though)
Fix problem with models that had group-specific coefficients, which were mislabled. Although the parameters were estimated correctly, users of previous versions of rstanarm should run such models again to obtain correct summaries and posterior predictions. Thanks to someone named Luke for pointing this problem out on stan-users.
Vignettes now view correctly on the CRAN webiste thanks to Yihui Xie
Fix problem with models without intercepts thanks to Paul-Christian Buerkner
Fix problem with specifying binomial ‘size’ for posterior_predict using newdata
Fix problem with lme4-style formulas that use the same grouping factor multiple times
Fix conclusion in rstanarm vignette thanks to someone named Michael
Group-specific design matrices are kept sparse throughout to reduce memory consumption
The log_lik() function now has a newdata argument
New vignette on hierarchical partial pooling