insight

• is_converged() now supports CmdStanMCMC models from *cmdstanr*, with the same checks and thresholds as for other Stan models. The values come from the $diagnostic_summary() and $summary() methods of the model. Other *cmdstanr* models, for example from $variational() or $optimize(), return FALSE with a message. stanfit models from rstan::read_stan_csv() store the maximum treedepth outside control. For these models, is_converged() counts treedepth hits against this stored maximum, not against the default of 10.
• get_parameters() for *lavaan* and *blavaan* models gets a type argument (in ...). If standardize = TRUE and type is given, it is passed to lavaan::standardizedSolution() or blavaan::standardizedPosterior(), for example type = "std.lv" or type = "std.nox". The default is still "std.all".
• get_variance() for nlme::lme() models now returns cor.slopes, the correlations between random slopes, as for *lme4* models. Before, it was missing. This works for nested and non-nested grouping factors with general covariance blocks (pdSymm and the default pdLogChol). Blocks of class pdDiag or pdIdent do not estimate correlations, so they add no entry to cor.slopes.
• export_table(format = "html") passes arguments of gt::gt() from ... to gt::gt(), for example id or rowname_col (#631).
• is_converged() now supports models fitted with Stan (stanfit, brmsfit and stanreg). Before, it did not work for these models. It checks divergent transitions, transitions at the maximum treedepth, E-BFMI, R-hat, and bulk and tail effective sample size, with the thresholds of the warnings that *rstan* gives after sampling. E-BFMI is computed as in rstan::check_hmc_diagnostics(). The values of the checks are returned in the diagnostics attribute. If convergence cannot be assessed, for example for models fitted with variational inference or with brms::brm_multiple(), FALSE is returned with a message (#619).
• is_converged() now supports coxph models from *survival*. Before, it did not work for these models. It checks whether the model ran out of iterations and whether a coefficient is possibly infinite, as *survival* does when it fits the model. If a check fails, it returns FALSE. The attribute diagnostics shows the result of each check and the coefficients that are possibly infinite. For penalized models, the checks do not apply. For models with ties = "exact" or tt() terms, and in the other cases that the help page lists, is_converged() cannot recompute them. In these cases, it returns FALSE. Objects of other classes that inherit from coxph, such as survival::clogit() and survey::svycoxph(), are not supported (#1090).

Bug fixes

• find_parameters() and clean_parameters() now work for multivariate *brms* models with addition terms in the response, such as resp_weights() or trials(). Before, find_parameters() returned empty elements with wrong response names, and clean_parameters() gave an error (easystats/bayestestR#700).
• If you use the hidden variable argument with summary = TRUE, get_parameters() for *brms* models no longer fails with the error object 'parms' not found. The summary then has the columns Parameter and Estimate.
• get_predicted() for zero-inflated *glmmTMB* models no longer fails for population-level predictions (include_random = FALSE, or a model without random effects) when the data object of the model call was changed or removed after the fit. Also, get_predicted() for zero-inflated *glmmTMB* and *pscl* models now accepts new data in the newdata argument. Before, it gave the error formal argument "newdata" matched by multiple actual arguments.
• null_model() now works when the model was fitted inside a function and its call uses a local object other than the data, for example family = fam. Before, it returned NULL for mixed models and gave an error for other models (easystats/performance#145).
• find_random_slopes() no longer reports a grouping factor or a *brms* correlation id as a random slope because its name is part of a slope name. An example is p in (0 + Intercept | p | g). For *brms* models, the explicit Intercept is no longer a random slope. Before, get_variance() warned that random slopes were not present as fixed effects for these models.
• get_modelmatrix() with new data for *nlme* (lme, gls), clmm and *brms* models returns one row per data row. A row with a missing value or a factor level that the model data does not have holds NA values. Before, it was dropped. For *brms* models, an ordered factor given as a character vector in data gets the polynomial contrasts of the model.
• get_predicted() with new data for lme and gls models returns NA for rows with a missing value or a factor level that the model data does not have in a fixed-effect predictor, with a warning. If no row can be predicted, it returns NULL with a warning. Before, such a row made get_predicted() return the predictions for the model data. If data has no grouping column, lme models give population-level predictions, as for *lme4* models.
• get_predicted() for lme and gls models uses the ci argument, with or without new data. As for other models, the default ci = NULL returns no intervals or standard errors. Before, the interval level did not follow ci. For lme models, an explicit level argument is passed to predict().
• The confidence intervals of get_predicted() for lme models, with or without new data, use one degree of freedom value for all rows, the smallest of the values per coefficient. Before, the values per coefficient were recycled against the predictions.
• get_predicted() with new data for clmm models returns NULL with a warning. Before, it returned the fitted values of the model data, transformed with the inverse link. If data has the rows of the model data and all of its predictor columns with the same values, the fitted values are still returned.
• find_parameters() for *brms* models again returns the standard deviations of smooth terms (sds_*) in the smooth_terms element. It also returns the simplex parameters of monotonic effects (simo_*) in the simplex element. For multivariate models, each response gets only its own parameters (easystats/parameters#1191)
• get_data() for merModList objects (package *merTools*) gets a verbose argument. If verbose = FALSE, it gives no warning that the data cannot be accessed.
• get_variance() for nlme::lme() models with a random-effect block that has no intercept, for example random = ~ 0 + day | Dog / Side:
• For nested grouping factors, get_variance() returned NULL.
• Without nested grouping factors, var.slope left out the first slope of the block, and cor.slope_intercept returned the correlation of two slopes as a slope-intercept correlation.

var.slope of nlme::lme() models now leaves out the term named

(Intercept), not the first term of a block. So a block whose intercept is

not its first term no longer reports the intercept as a slope. Such a block,

for example from nlme::pdBlocked(), now also has its intercept variance in

var.intercept.

• get_variance() for *lme4* models with uncorrelated random effects, for example (1 + Days || Subject), now includes the random-slope variance in var.random. Before, var.random included only the random-intercept variance. For nested grouping factors, for example (1 + Days || grp / subgrp), var.random was NULL. This also changes performance::r2_nakagawa() and performance::icc() for these models (#304, easystats/performance#428).
• export_table(format = "html") for a list of data frames now shows the caption of each data frame as the label of its row group, and the footer of each data frame as a source note. Before, a list footer or a mix of captioned and uncaptioned tables gave an error. Of the footers stored as attributes, only the first one was kept, and a list caption became one title. If any data frame in the list has a Component, Effects, Group or Response column, the old output stays, but a list footer no longer gives an error (#631).
• export_table(format = "html") for a single data frame with a list footer (a multi-line, colored footer) gave an error. The parts of the footer are now joined into one source note, as in text output.
• export_table(format = "tt") for a list of data frames now shows the caption of each data frame as the label of its row group, and the footer of each data frame as a note. Before, the captions were shown as an extra group column, the first caption became the title, and only the footer of the first data frame, or only footer if given, was kept. A list caption, or a mix of table_title and table_caption attributes, gave an error. An unnamed column_names vector names the shown columns. Before, it also needed a name for the group column. The old behavior stays if a data frame has one column, no rows, row groups (indent_rows) or a Component, Effects, Group or Response column, if the data frames have different column names, or if by, row_groups or column_groups is given. For these lists, each element of a list footer still gives one note. Each footer changes as for a single data frame. The color is not shown as a note, new lines at the start and end are removed, and an empty footer gives no note.
• export_table(format = "tt") for a single data frame now shows a footer as one note. A character vector gives its first string, because the second string is the color, and a list gives its strings joined. Before, each string was a note, so the color of a colored footer was shown as a second note, and a list footer with colors (a multi-line, colored footer) gave an error.
• get_sigma(), get_variance() and get_variance_residual() now support *brms* models where sigma has an intercept-only formula, for example bf(y ~ x, sigma ~ 1). Sigma is the mean of the posterior draws of b_sigma_Intercept, back-transformed with the link function of sigma. Before, get_sigma() returned NULL and get_variance() returned a residual variance of 1. If sigma varies, for example with sigma ~ (1 | g), get_variance() now returns no residual variance with a message, as it already did for sigma ~ x (#967).
• find_parameters(), get_parameters() and clean_parameters() now work for *blavaan* models with multiple groups and equality constraints, for example group.equal = "loadings". Before, find_parameters() and get_parameters() failed with the error "replacement has length zero". get_parameters(standardize = TRUE) now gives the correct names for the posterior draws of all *blavaan* models. Before, the names were shifted, because the standardized draws also include the fixed parameters. get_parameters() now also works for *blavaan* models with user-defined parameters (:=) and returns their posterior draws. Before, it failed with the error "Arguments must be mcmc objects". get_data(source = "mf") now works for *lavaan* and *blavaan* models with multiple groups and returns the grouping variable as a column. Together, these fixes let parameters::model_parameters() work for multiple-group *blavaan* models (easystats/parameters#735).
• get_varcov() for *fixest* models now supports vcov = "HC0". Before, it gave an error. The result is the same as fixest::vcov_hetero(ssc = fixest::ssc(K.adj = FALSE)), the heteroskedasticity-consistent matrix without the small-sample adjustment. For "HC", "HC4", "HC4m" and "HC5", the error now names the supported types (#1251).
• find_predictors() and find_variables() no longer return the matrix of the *brms* autocorrelation terms sar(), car() and fcor(), for example W in y ~ x + sar(W). This matrix is an object in data2, not a variable in the data. The grouping variable of car() is still returned.
• If *performance* is not installed, get_variance() now gives an error that asks you to install it. Before, it returned NULL without a message (#928).
• n_obs(disaggregate = TRUE) now returns the total number of trials for *brms* models with a trials() term in the response, for example y | trials(size) ~ x. Before, it ignored disaggregate and returned the number of data rows (#581).
• get_variance() and get_variance_distribution() returned a wrong distribution-specific variance for beta-binomial models with a probit or cloglog link. The cause was the back-transformation of the null-model mean, which always used the inverse logit. The mean now uses the inverse link of the model.
• get_residuals() for *gamlss* models now returns response residuals (observed minus expected values) for type = "response". Before, it returned the normalized quantile residuals without a warning. The mean() function of the family gives the expected values. Thus, the result is also correct for families where mu is not the mean, for example zero-inflated families. For families without a mean() function, the residuals use mu, with a warning (#327).
• For a list of data frames, caption = "" (or title = "") now makes export_table() remove the captions of all tables. This includes captions stored as attributes. subtitle = "" and footer = "" do the same for subtitles and footers. In a list of captions or footers, "" removes the caption or footer of that table. Before this fix, export_table() still printed the attributes (#930).
• For univariate non-linear *brms* models (nl = TRUE), find_parameters() now returns the coefficients of the non-linear parameters of mu as conditional parameters, and their group-level terms as random parameters. Before, they were returned as auxiliary parameters. find_auxiliary() no longer returns the names of these non-linear parameters. clean_parameters() labels their coefficients like summary() from *brms*, for example ult_Intercept. The effects, component and parameters arguments of get_parameters() now select the coefficients of non-linear parameters. get_varcov(), n_parameters(component = "conditional") and find_parameters(component = "location") now include these coefficients, where they included none before. This also applies to non-linear parameters that are nested with nlf() (#1076).
• get_parameters() for *brms* models now returns NULL when no parameter matches the selection of effects, component and parameters. Before, depending on the selection, it returned the posterior draws of all parameters, an empty data frame, or an error (#1076).
• get_modelmatrix() for clmm models now uses the contrasts stored in the model, so get_variance() works for clmm models fitted with, for example, contrasts = list(x = "contr.sum"). For clmm models, a data argument is now used. For clmm and brmsfit models, factor variables in data get the levels of the model data, and the contrasts of the model are used. Thus, new data with only some of the levels gives the correct columns. A contrasts.arg list is now accepted for both model classes.
• If variables used i