New Features

• Add experimental support for the pathfinder and laplace algorithms in the cmdstanr backend. (#1591)
• Automatically recompute fit criteria previously stored in the model if potentially results-changing arguments are provided to the criterion method.
• Allow to turn off automatic broadcasting of constant priors.
• Allow for joint likelihood evaluation in kfold via argument joint.
• Use several Stan built-in functions implemented since version 2.26 to improve the efficiency of multiple model classes. (#1077)

Other Changes

• Change make_stancode and make_standata to be aliases of stancode and standata, respectively. Change get_prior to be an alias of a new generic method default_prior. This enable other packages to define new stancode, standata and default_prior methods to generate Stan code and data, and extract the default priors, for their own objects building on brms. Thanks to Ven Popov for helping with this. (#1604)
• Change the default prior of the shape parameter of negbinomial models to inv_gamma(0.4, 0.3) thanks to Aki Vehtari. (#1614)
• No longer automatically canonicalize the Stan code if cmdstanr is used as backend. (#1544)
• Export read_csv_as_stanfit thanks to Ven Popov. (#1619)
• Make installation of shinystan optional. This means that the package has to be loaded, via library(shinystan), before launch_shinystan can be used. (#1595)
• Improve parameter class names in the summary output.
• Show histograms rather than densities in the plot method by default.
• Deprecate argument N in the plot method in favor of argument nvariables.
• Remove deprecated argument exact_loo in method kfold.

Bug Fixes

• Remove some remaining uses of Stan's old array syntax.
• Fix a bug in formula parsing of missing values terms with interactions thank to Guido Biele. (#1608)
• Ensure compatibility of combine_models with moment matching. (#1603)
• Ensure compatibility with the latest splines2 package version. (#1580)
• Fix output of rmulti_normal thanks to Ven Popov. (#1588)
• Prevent memory leaks when executing kfold or reloo in parallel.

New Features

• Apply the horseshoe and R2D2 priors globally, that is, for all additive predictor terms specified in the same formula. (#1492)
• Use as.brmsprior to transform objects into a brmsprior. (#1491)
• Use matrix data as non-linear covariates. (#1488)

Other Changes

• Switch to the new array syntax of Stan. This increases the version requirements of Stan to >= 2.26.
• No longer support the lasso prior as it is not a good shrinkage prior and incompatible with the newly implemented global shrinkage prior framework.
• No longer support multiple deprecated prior options for categorical and multivariate models after around 3 years of deprecation. (#1420)
• Deprecate argument newdata of get_refmodel.brmsfit(). (#1502)
• Disallow binomial models without trials argument after several years of deprecation. (#1501)

Bug Fixes

• Fix a long-standing bug in the post-processing of spline models that could lead to non-sensible results if predictions were performed on a different machine than where the model was originally fitted. Old spline models can be repaired via restructure. Special thanks to Simon Wood, Ruben Arslan, Marta Kołczyńska, Patrick Hogan, and Urs Kalbitzer. (#1465)
• Fix a bunch of minor issues occurring for rare feature combinations.

New Features

• Model unstructured autocorrelation matrices via the unstr term thanks to the help of Sebastian Weber. (#1435)
• Model ordinal data with an extra category (non-response or similar) via the hurdle_cumulative family thanks to Stephen Wild. (#1448)
• Improve user control over model recompilation via argument recompile in post-processing methods that require a compiled Stan model.
• Extend control over the point_estimate feature in prepare_predictions via the new argument ndraws_point_estimate.
• Add support for the latent projection available in projpred versions >= 2.4.0. (#1451)

Bug Fixes

• Fix a Stan syntax error in threaded models with lasso priors. (#1427)
• Fix Stan compilation issues for some of the more special link functions such as cauchit or softplus.
• Fix a bug for predictions in projpred, previously requiring more variables in newdata than necessary. (#1457, #1459, #1460)

New Features

• Support regression splines with fixed degrees of freedom specified via s(..., fx = TRUE).
• Reuse user-specified control arguments originally passed to the Stan backend in update and related methods. (#1373, #1378)
• Allow to retain unused factors levels via drop_unused_levels = FALSE in brm and related functions. (#1346)
• Automatically update old default priors based on new input when when updating models via update.brmsfit. (#1380)
• Allow to use dirichlet priors for more parameter types. (#1165)

Other Changes

• Improve efficiency of converting models fitted with backend = "cmdstanr" to stanfit objects thanks to Simon Mills and Jacob Socolar. (#1331)
• Allow for more O1 optimization of brms-generated Stan models thanks to Aki Vehtari. (#1382)

Bug Fixes

• Fix problems with missing boundaries of sdme parameters in models with known response standard errors thanks to Solomon Kurz. (#1348)
• Fix Stan code of gamma models with softplus link.
• Allow for more flexible data inputs to brm_multiple. (#1383)
• Ensure that control_params returns the right values for models fitted with the cmdstanr backend. (#1390)
• Fix problems in multivariate spline models when using the subset addition term. (#1385)

brms 2.17.0

New Features

• Add full user control for boundaries of most parameters via the lb and ub arguments of set_prior and related functions. (#878, #1094)
• Add family logistic_normal for simplex responses. (#1274)
• Add argument future_args to kfold and reloo for additional control over parallel execution via futures.
• Add families beta_binomial & zero_inflated_beta_binomial for potentially over-dispersed and zero-inflated binomial response models thanks to Hayden Rabel. (#1319 & #1311)
• Display ppd_* plots in pp_check via argument prefix. (#1313)
• Support the log link in binomial and beta type families. (#1316)

Other changes

• Argument brms_seed has been added to get_refmodel.brmsfit(). (#1287)
• Deprecate argument inits in favor of init for consistency with the Stan backends.
• Improve speed of the summary method for high-dimensional models. (#1330)

Bug Fixes

• Fix Stan code of threaded multivariate models thanks to Anirban Mukherjee. (#1277)
• Fix usage of int_conditions in conditional_smooths thanks to Urs Kalbitzer. (#1280)
• Fix an error sometimes occurring for multilevel (reference) models in projpred's K-fold CV. (#1286)
• Fix response values in make_standata for bernoulli families when only 1s are present thanks to Facundo Munoz. (#1298)
• Fix pp_check for censored responses to work for all plot types thanks to Hayden Rabel. (#1327)
• Ensure that argument overwrite in add_criterion works as expected for all criteria thanks to Andrew Milne. (#1323)
• Fix a problem in launch_shinystan occurring when warmup draws were saved thanks to Frank Weber. (#1257, #1329)
• Fix numerical stability problems in log_lik for ordinal models. (#1192)

Bug Fixes

• Fix a bug causing problems during post-processing of models fitted with older versions of brms and the cmdstanr backend thanks to Riccardo Fusaroli. (#1218)

New Features

• Support several methods of the posterior package. (#1204)
• Substantially extend compatibility of brms models with emmeans thanks to Mattan S. Ben-Shachar. (#907, #1134)
• Combine missing value (mi) terms with subset addition terms. (#1063)
• Expose function get_dpar for use in the post-processing of custom families thank to Martin Modrak. (#1131)
• Support the squareplus link function in all families and distributional parameters that also allow for the log link function.
• Add argument incl_thres to posterior_linpred.brmsfit() allowing to subtract the threshold-excluding linear predictor from the thresholds in case of an ordinal family. (#1137)
• Add a "mock" backend option to facilitate testing thanks to Martin Modrak. (#1116)
• Add option file_refit = "always" to always overwrite models stored via the file argument. (#1151)
• Initial GPU support via OpenCL thanks to the help Rok Češnovar. (#1166)
• Support argument robust in method hypothesis. (#1170)
• Vectorize the Stan code of custom likelihoods via argument loop of custom_family. (#1084)
• Experimentally allow category specific effects for ordinal cumulative models. (#1060)
• Regenerate Stan code of an existing model via argument regenerate of method stancode.
• Support expose_functions for models fitted with the cmdstanr backend thanks to Sebastian Weber. (#1176)
• Support log_prob and related functionality in models fitted with the cmdstanr backend via function add_rstan_model. (#1184)

Other Changes

• Remove use of cbind to express multivariate models after over two years of deprecation (please use mvbind instead).
• Method posterior_linpred(transform = TRUE) is now equal to posterior_epred(dpar = "mu") and no longer deprecated.
• Refactor and extend internal post-processing functions for ordinal and categorical models thanks to Frank Weber. (#1159)
• Ignore NA values in interval censored boundaries as long as they are unused. (#1070)
• Take offsets into account when deriving default priors for overall intercept parameters. (#923)
• Soft deprecate measurement error (me) terms in favor of the more general and consistent missing value (mi) terms. (#698)

Bug Fixes

• Fix an issue in the post-processing of non-normal ARMA models thanks to Thomas Buehrens. (#1149)
• Fix an issue with default baseline hazard knots in cox models thanks to Malcolm Gillies. (#1143)
• Fix a bug in non-linear models caused by accidental merging of operators in the non-linear formula thanks to Fernando Miguez. (#1142)
• Correctly trigger a refit for file_refit = "on_change" if factor level names have changed thanks to Martin Modrak. (#1128)
• Validate factors in validate_newdata even when they are simultaneously used as predictors and grouping variables thanks to Martin Modrak. (#1141)
• Fix a bug in the Stan code generation of threaded mixture models with predicted mixture probabilities thanks to Riccardo Fusaroli. (#1150)
• Remove duplicated Stan code related to the horseshoe prior thanks to Max Joseph. (#1167)
• Fix an issue in the post-processing of non-looped non-linear parameters thanks to Sebastian Weber.
• Fix an issue in the Stan code of threaded non-looped non-linear models thanks to Sebastian Weber. (#1175)
• Fix problems in the post-processing of multivariate meta-analytic models that could lead to incorrect handling of known standard errors.

New Features

• Turn off normalization in the Stan model via argument normalize to increase sampling efficiency thanks to Andrew Johnson. (#1017, #1053)
• Enable posterior_predict for truncated continuous models even if the required CDF or quantile functions are unavailable.
• Update and export validate_prior to validate priors supplied by the user.
• Add support for within-chain threading with rstan (Stan >= 2.25) backend.
• Apply the R2-D2 shrinkage prior to population-level coefficients via function R2D2 to be used in set_prior.
• Extend support for arma correlation structures in non-normal families.
• Extend scope of variables passed via data2 for use in the evaluation of most model terms.
• Refit models previously stored on disc only when necessary thanks to Martin Modrak. The behavior can be controlled via file_refit. (#1058)
• Allow for a finer tuning of informational messages printed in brm via the silent argument. (#1076)
• Allow stanvars to alter distributional parameters. (#1061)
• Allow stanvars to be used inside threaded likelihoods. (#1111)

Other Changes

• Improve numerical stability of ordinal sequential models (families sratio and cratio) thanks to Andrew Johnson. (#1087)

Bug Fixes

• Allow fitting multinomial models with the cmdstanr backend thanks to Andrew Johnson. (#1033)
• Allow user-defined Stan functions in threaded models. (#1034)
• Allow usage of the : operator in autocorrelation terms.
• Fix Stan code generation when specifying coefficient-level priors on spline terms.
• Fix numerical issues occurring in edge cases during post-processing of Gaussian processes thanks to Marta Kołczyńska.
• Fix an error during post-processing of new levels in multi-membership terms thanks to Guilherme Mohor.
• Fix a bug in the Stan code of threaded wiener drift diffusion models thanks to the GitHub user yanivabir. (#1085)
• Fix a bug in the threaded Stan code for GPs with categorical by variables thanks to Reece Willoughby. (#1081)
• Fix a bug in the threaded Stan code when using QR decomposition thanks to Steve Bronder. (#1086)
• Include offsets in emmeans related methods thanks to Russell V. Lenth. (#1096)

New Features

• Turn off normalization in the Stan model via argument normalize. to increase sampling efficiency thanks to Andrew Johnson. (#1017, #1053)
• Enable posterior_predict for truncated continuous models even if the required CDF or quantile functions are unavailable.
• Update and export validate_prior to validate priors supplied by the user.
• Add support for within-chain threading with rstan (Stan >= 2.25) backend.
• Apply the R2-D2 shrinkage prior to population-level coefficients via function R2D2 to be used in set_prior.
• Extend support for arma correlation structures in non-normal families.
• Extend scope of variables passed via data2 for use in the evaluation of most model terms.

Other Changes

• Improve numerical stability of ordinal sequential models (families sratio and cratio) thanks to Andrew Johnson. (#1087)

Bug Fixes

• Allow fitting multinomial models with the cmdstanr backend thanks to Andrew Johnson. (#1033)
• Allow user-defined Stan functions in threaded models. (#1034)
• Allow usage of the : operator in autocorrelation terms.
• Fix Stan code generation when specifying coefficient-level priors on spline terms.
• Fix numerical issues occurring in edge cases during post-processing of Gaussian processes thanks to Marta Kołczyńska.
• Fix an error during post-processing of new levels in multi-membership terms thanks to Guilherme Mohor.
• Fix a bug in the Stan code of threaded wiener drift diffusion models thanks to the GitHub user yanivabir. (#1085)
• Fix a bug in the threaded Stan code for GPs with categorical by variables thanks to Reece Willoughby. (#1081)
• Fix a bug in the threaded Stan code when using QR decomposition thanks to Steve Bronder. (#1086)

New Features

• Experimentally support within-chain parallelizaion via reduce_sum using argument threads in brm thanks to Sebastian Weber. (#892)
• Add algorithm fixed_param to sample from fixed parameter values. (#973)
• No longer remove NA values in data if there are unused because of the subset addition argument. (#895)
• Combine by variables and within-group correlation matrices in group-level terms. (#674)
• Add argument robust to the summary method. (#976)
• Parallelize evaluation of the posterior_predict and log_lik methods via argument cores. (#819)
• Compute effective number of parameters in kfold.
• Show prior sources and vectorization in the print output of brmsprior objects. (#761)
• Store unused variables in the model's data frame via argument unused of function brmsformula.
• Support posterior mean predictions in emmeans via dpar = "mean" thanks to Russell V. Lenth. (#993)
• Improve control of which parameters should be saved via function save_pars and corresponding argument in brm. (#746)
• Add method posterior_smooths to computing predictions of individual smooth terms. (#738)
• Allow to display grouping variables in conditional_effects using the effects argument. (#1012)

Other Changes

• Improve sampling efficiency for a lot of models by using Stan's GLM-primitives even in non-GLM cases. (#984)
• Improve sampling efficiency of multilevel models with within-group covariances thanks to David Westergaard. (#977)
• Deprecate argument probs in the conditional_effects method in favor of argument prob.

Bug Fixes

• Fix a problem in pp_check inducing wronger observation orders in time series models thanks to Fiona Seaton. (#1007)
• Fix multiple problems with loo_moment_match that prevented it from working for some more complex models.