Changelog

Source: NEWS.md

frmtmb.sample 0.27.0

• posterior_epred(), posterior_linpred() and posterior_predict() with allow_new_levels = TRUE answer at a me(gr = ) level the fit did not see: each draw takes the level’s latent value from normal(x, sdx), as brms does. Before, they refused it, where fitted() and brms answered. A new level of a group-level term stays refused. Needs frmtmb 0.80.0.

• The package code does not use the global assignment operator <<-, which rOpenSci’s pkgcheck refuses. A closure that keeps state between calls writes to an environment that it creates, and a condition handler returns its value. The results do not change. The tmbstan build check keeps its memo in memo, an environment in its enclosure, and the test helper with_tmbstan_hpp() resets it there.

frmtmb.sample 0.26.0

Needs frmtmb 0.79.0. The exact gp() field is non-centered in core, so the sampler has nothing to transform. It reads the sampling-API exports gp_tape_scope(), gp_tape_order(), gp_nugget_message(), and the new readers of a me() draw and of a latent residual.

Breaking changes

• An exact gp() is non-centered in core, as brms’s zgp, at a nugget of 1e-12 * sdgp^2. So frm_sample() samples the tape frm() fits, with the sampler’s priors added, and has nothing to transform. The block no longer has a factor for reparameterize =, and that argument does not reach it: an exact field is non-centered with or without it. The b[i] columns of its draws are z, over a farthest-first order of the positions, where they were the field. ranef() and every predictive method read the field through expand_b(), as before.

• frm_sample(fit) starts at random where the fit’s noise has reached the nugget of an exact gp(), and says so. Chains started at that mode stay pinned: at true sigma 1e-6, n 15, a median sigma of 1.83e-8 and Rhat 2.25. These chains do not converge at this noise either: sigma read 24 and 394 times the truth on two seeds at n 15. laplace = TRUE is the route here. init = "last.par.best" starts at the mode. At ordinary noise the mode stays the start.

• frm_sample(laplace = TRUE) on an exact gp() takes 3.2 to 3.5 times 0.78.0’s wall clock at 300 positions, 531 to 570 s against 165 s. The dense gradient causes this. Memory holds, 1.91 to 1.95 GB against 1.92 GB.

• brms’s class "sds" carries a smooth’s default, student_t(3, 0, <scale>), and a smooth’s prior on frm_sample(), with coef the term as written. Class "sd" no longer reaches a smooth, as in brms. The table lists the rows brms lists.

• A missing value’s draw is reported under brms’s name, Ymi_<resp>[k] (Ymi[k] with one response, Yl_ under mi(sd)), on the response’s own scale. It was miss_<k>, on the sampler’s chart: miss_1 under Gamma ranged from -3.05 to 3.01 where brms’s Ymi[1] ranges from 0.04 to 15.3. The sampler’s Jacobian does not change.

• A new row of a me() term draws a new latent value per posterior draw from normal(x, sdx), per row, or per level under me(gr = ). It does so in posterior_epred(), posterior_linpred() and posterior_predict(), as brms 2.23.0 does. Before, they read x: on a model with sdx = 0.3, the sd over draws of the expected value was 0.603 of brms’s. The 95 percent interval was 0.634 wide against 1.036. New data must hold sdx, and corme is not read, as in brms.

• One me() call is one latent value per row and draw, across the responses and dpars one method answers for. This diverges from brms 2.23.0, which draws one per predictor on new data, against its own Stan program, which has one Xme_1: at the training rows the residuals correlate at 0.734, one latent predicts 0.739 and brms’s replicates give -0.001.

• posterior_linpred() with no dpar on an ordinal family with a cs() term returns brms’s layout, draws x rows x thresholds, with every margin unnamed. Before, it returned the shared part, the predictor at no threshold of the model. dpar = "mu" and transform = TRUE read the shared part, as in brms.

• conditional_effects() of a multivariate model gives one panel set per response without resp, and takes several, each named <resp>.<resp>_<effect>, as brms does. A name given twice is refused, where brms repeats the panel. Before, it gave the first response alone and failed on several.

• A univariate model does not read resp, as in brms: posterior_epred(), posterior_linpred(), posterior_predict(), fitted(), predict(), residuals(), log_lik(), logLik(), bayes_R2(), loo_R2() and the three loo_*() predictions give the same answer for an empty vector. They give the same answer for a name given twice, an unknown name, NA and a number. Before, some gave internal errors.

• update(<draws>, prior = ) merges with the first call’s prior, as brms’s update() does: a new row takes its slot and the other rows stay. A prior with no rows keeps them. Before, the new prior replaced the first call’s: after an update() that added a sigma prior, b_x moved from 9.93 to 2.09. There is one exception: an edited row of a default_prior() table stays here, where brms loses it because the row’s source is still "default": b_x reads 9.52 here and 2.00 in brms.

• posterior_predict(newdata = ) answers under ar(), ma(), arma(), cosy() and unstr() with cov = TRUE, through core’s route over newdata’s own groups and times. Before, it refused. log_lik(newdata = ) still refuses where brms answers; its message now names posterior_predict(newdata = ) as a route on these models, and that is all that changes.

• A new row with no x or sdx of a me() term reads NA in posterior_epred(), posterior_linpred() and posterior_predict(), with a warning of class frmtmb_me_missing. The other rows answer.

• conditional_effects() of a me(gr = ) model holds the group at NA, as brms does. Panels are in brms’s order, with the effects outside and the responses inside.

New features

• bayes_R2() and loo_R2() read an empty resp as every response. Before, they stopped with an internal error.

• pp_mixture() of a multivariate mixture reads the one response named, with brms’s refusal of none or several. Before, it refused every such model, and read the first response when resp was NULL.

Bug fixes

• frm_sample() on nl = TRUE with cumulative(), sratio(), cratio(), acat() or hurdle_cumulative() samples the body. Before, it sampled the thresholds alone without a word, as frm() fitted them, and posterior_epred() on those draws stopped with a refusal.

• frm_sample() takes bf(..., nl = TRUE, autocor = ~ term), which was refused, and nl = TRUE on categorical(), multinomial(), the simplex families and a mixture(), with one body in every location, as brms.

• The drawn route for an exact gp() needs long chains at low noise, and the numbers are stated. At true sigma 1e-3, 15 rows, 4 chains of 12000 read sigma 5.9e-4, 1.05e-3, 2.3e-3 (Rhat 1.025, sdgp Rhat 1.07, 58 minutes). brms’s chains read 5.2e-4, 1.03e-3, 2.4e-3 (1.031, 1.15, 53 minutes, 20000 iterations). Neither reaches Rhat 1.01 at those lengths. frm_sample(laplace = TRUE) serves the outer parameters only: it refuses posterior_epred(), posterior_linpred(), posterior_predict(), fitted(), predict(), log_lik(), loo() and conditional_effects().

Other changes

• frm_sample() samples mi() under every family core takes, and under asym_laplace() and mixture() where frm() refuses a value another density reads. From a frm() fit it holds the missing values the fit left out of its objective, through core’s new mi_sampling().

• frm_sample(formula, laplace = TRUE) takes the route of frm_sample(fit, laplace = TRUE): a missing value no density reads leaves the objective. On an exgaussian with 200 rows and 60 missing the formula read sigma 0.910 and beta 1.590. It now reads 0.360 and 2.132, the fit 0.355 and 2.146, and the observed rows, fully sampled, 0.359 and 2.139. A value a density reads with no mode to expand about is refused on this route as on frm().

• frm_sample() samples core’s latent residuals: ar(), ma(), arma(), cosy() and unstr() on a family with no residual y - mu. Its defaults are brms’s: student_t(3, 0, <scale>) on sderr, flat on ar and ma over (-1, 1) and on cosy over (0, 1), and lkj(1) on cortime. On a poisson ar(cov = TRUE) with 150 rows, the posterior means of the intercept, slope, ar[1] and sderr read 0.797, 0.478, 0.456 and 0.692. In brms they are 0.796, 0.476, 0.459 and 0.690. variables() lists ar[1] and sderr. The latents are the columns latent_state[k], where brms has zerr[n] and err[n]. These are the chain’s own states: ma(), arma() and cosy() have one extra state per group, so there are 72 of them for 60 rows in 12 groups. They are not err on a chain with sderr other than 1.

• frm_sample() fits the forms of a latent residual that frm() refuses: cov = FALSE after the inverse link, and a mean less than 2 marginal sds of the latent from the edge of an identity or inverse link.

• posterior_epred(), log_lik() and posterior_predict() read each draw’s latents on the fitted rows, and posterior_predict() and posterior_epred() draw new latents on new rows under cov = TRUE. Under cov = FALSE new rows are refused in brms’s words.

frmtmb.sample 0.25.0

Needs frmtmb 0.78.0. Its set_prior() takes the classes meanme, sdme and corme, and its default_prior() prints the group rows of a sampling-route table as (flat). It exports predict_rescor_draw() and gives rescor_row_loglik() its resp.

Breaking changes

• A default_prior(route = "sample") table passed as prior = applies an edited class row to every group, as brms does: the groups of sd, cor and corme, and each level of thres(gr = ) for the Intercept and delta rows. Before, the table’s group rows carried the class row’s default and applied it over the edit: on two me(gr = g) terms, a class corme row edited to lkj(4) sampled lkj(1). Its Kolmogorov-Smirnov p-value against the lkj(4) correlation was 1.3e-32. The same held for cor and sd.

• For the brms classes, a bound on one coefficient (coef with lb or ub) is refused with brms’s message: “Prior argument ‘coef’ may not be specified when using boundaries.” brms 2.23.0 refuses it for b, sd, meanme and sdme. Before, the sampler took the bound and the prior draws ignored it: sdme of one term at ub = 0.5 had prior draws up to 10.7. A bound on a class row or on a group’s row is taken, as in brms, and frm() keeps its box on one coefficient. Three cases are not refused. Class theta, frmtmb’s own, takes a bound on one internal parameter. A row that restates lb = 0 on sd, sdgp, lscale or sdme bounds nothing. And set_prior(class = "Intercept", nlpar = "a") on a nonlinear parameter with a ~ 1 is brms’s class row class = "b", nlpar = "a", so its bound is a class bound. With a bound, that spelling is refused where a has more than an intercept.

• That spelling’s bound now reaches its prior draws. Before, it was a bound on one coefficient, and the prior draws ignored it: under lb = 0 on y ~ a * exp(-b * x), the prior draws of a went down to -1.67. Now they are the prior truncated at the bound, under brms’s names prior_b_a and prior_b_b.

• A me() model is sampled under brms 2.23.0’s default priors on its hyperparameters: flat on sdme itself, and lkj(1) on corme. Before, the sampler was flat on log(sdme) and on the internal chart of the correlation. We sampled two me(gr = g) terms with 10 levels and one row-level me() term (4 chains of 3000 draws). The posterior means of sdme_mexg and sdme_mezg were 12.6 and 13.6 Monte Carlo standard errors from brms’s, and the slopes up to 8.0. Now every mean, sd, 5 and 95 percent quantile of 14 variables is within 2.52 of brms’s. prior = "flat" stays flat on sdme, and is flat on the Cholesky factor of corme, which is brms’s (flat).

• get_prior(route = "sample") and prior_summary() of draws list these defaults in brms’s rows.

• posterior_predict() of a set_rescor(TRUE) model draws the responses of a row jointly, as brms does: per draw, a multivariate normal with covariance diag(sigma) R diag(sigma), or for student() the multivariate t with the shared nu. Before, no call returned that draw: without resp the method returned the first response only. Draws of two responses taken one call each correlated at -0.05 to 0.08, on a fit whose residual correlation was 0.82. On a fit of three responses (200 rows, 2000 draws), the correlations of the predictive residuals are now 0.750, 0.554 and 0.315. They are within 1.3 Monte Carlo standard errors of what the draws imply, and within 0.7 of brms’s own draw from the same parameter draws. Under sigma ~ x and under student(), the same holds within 1.4 and 0.8. A sigma that varies by row is drawn at that row’s value.

• Without resp, posterior_predict(), posterior_epred(), posterior_linpred() and predictive_error() of a multivariate model return every response, as brms does: a draws x observations x responses array with brms’s response names. fitted(), predict() and residuals() summarize it into brms’s observations x 4 x responses array. Before, they returned the first response only. Without rescor, each layer is the draw that resp = gives for its response.

• resp takes several responses, in the order you name them, in these methods: posterior_predict(), posterior_epred(), posterior_linpred(), predictive_error(), fitted(), predict(), residuals(), log_lik(), loo(), waic(), loo_predict(), loo_epred(), loo_linpred() and loo_predictive_interval(). Each takes brms’s spelling of a name (ya for y_a) as well as frmtmb’s. As in brms, each refuses a name given twice and reads an empty resp as every response. Under set_rescor(TRUE), a resp that is not in the model’s order is answered, with the correlations in the order named. brms 2.23.0 stops there, because it looks up the correlations by the names in the order given.

• A response can predict more than one value per row: the category probabilities of an ordinal or categorical response. The stacked layout has no place for it. posterior_epred() and fitted() without resp refuse such a model; brms binds the category columns beside the other responses there. predict() without resp refuses to summarize it, as brms stops; with summary = FALSE it returns the draws. Name the response with resp =. predictive_interval() without resp is refused, as brms refuses it.

• log_lik(resp =) of a set_rescor(TRUE) model is brms’s: one response gives its own density, and several give their joint density. Before, every resp gave the joint density of all the responses, so loo(resp = "y1") on three responses reported -724.7 where the density of y1 gives -280.8. loo(), waic(), loo_predict(), loo_R2() and the other loo_*() functions read the corrected matrix through their resp. log_lik(combine = FALSE) returns the joint matrix, as in brms, where it was refused.

• One response of a student() model with set_rescor(TRUE) is drawn, and its density taken, as a t with the shared nu. That t is the margin of the multivariate t. brms 2.23.0 reads nu = 1 there (its one-response preparation finds no nu of its own), which is not the margin of its own model. This is a deliberate divergence from brms.

• Where a ~ 1, set_prior(class = "Intercept", nlpar = "a") and brms’s class row set_prior(class = "b", nlpar = "a") are one prior slot. With both, the Intercept spelling’s density stands, in either order, as frm() and brms resolve them. A bound on either row is the class bound, the Intercept spelling’s where both set one. The fit’s own prior and the call’s prior = combine in the same way. prior_summary() of the draws lists both rows as they were written, as brms does.

New features

• Priors on brms’s classes meanme, sdme and corme apply when sampling. sample_prior = "yes" draws them under brms’s names: prior_meanme_1, prior_sdme_1 and prior_Lme_1, numbered by the group of terms that share a gr =. Where a term has a prior of its own, there is one column per term, prior_meanme_1_meanme_mex, as brms’s renaming spells it. With brms’s defaults only prior_Lme_1 is drawn, as in brms.

• sample_prior = "only" samples a me() model once meanme and sdme have proper priors. Before, it refused every such model. With a flat one it is refused with brms’s words, naming 'meanme_1' or 'sdme_1'.

Bug fixes

• The draws of a model with correlated me(gr = ) terms carry brms’s corme_<gr>__<coef1>__<coef2> column. Before, only the ungrouped corme__ column was derived, so the grouped correlation was absent from the draws.

• The default-prior message prints brms’s (flat) on a bounded parameter with its bound, delta (flat), lb = 0 for the distance of equidistant ordinal thresholds. Before, it printed delta flat(). The new sdme line reads (flat), lb = 0 too.

• bayes_R2() and loo_R2() refuse a response named twice in resp, as the other methods do. Before, bayes_R2() gave two identical rows, and loo_R2() gave two rows with different values.

frmtmb.sample 0.24.0

Needs frmtmb 0.77.0. That release reads brms’s (flat) as a prior density and gives build_objective() its prior_only field. Its sigma is the innovation sd under the covariance form of ar(), ma() and arma().

Breaking changes

• Under the covariance form of ar(), ma() and arma(), the draws of sigma are of the innovation sd, as brms’s are. Before, they were of the marginal sd. On an AR(1) with ar near 0.92, the posterior mean of sigma was 1.154 against brms’s 0.432. It is now 0.433. summary(), posterior_summary(), variables() and the draws read the same parameter as brms’s.

• The default student_t(3, 0, s) on sigma and a prior of your own are densities on the innovation sd itself. A lb or ub on sigma is taken, as in brms. Under sigma ~ x, the default and a prior of your own on the intercept of sigma are read at the log innovation sd, as brms’s Intercept_sigma. With normal(-1, 0.5) there, the posterior mean of ar was 7.95 Monte Carlo standard errors from brms’s and is now 0.55.

• We compared the draws with brms’s own (4 chains of 4000 on each side): ar(), ma() and arma() under the defaults, ar() under user priors, and ar() under sigma ~ x. Every posterior mean is within 1.8 Monte Carlo standard errors of brms’s. Every sd ratio is within 0.969 to 1.037. The bulk effective sample size of sigma is 2.0, 0.94 and 1.2 times what it was under ar(), ma() and arma().

• conditional_effects(method = "posterior_predict") on draws draws its band at the draws’ sigma. Under these terms that is now the innovation sd. brms does the same, because its conditional_effects() drops every autocorrelation term. The band is narrower than before by sqrt(1 - ar^2) for an AR(1).

• Draws saved with frmtmb.sample 0.23.0 for a model with one of these terms still report the old marginal sd as sigma. summary(), posterior_summary(), as.matrix() and variables() show it. Their log_lik(), loo() and posterior_predict() are refused. Sample such a model again.

• The message of the default priors no longer marks sigma with “on brms’s innovation sd”: it is brms’s sigma.

• prior = "flat" is brms’s (flat) in every slot: no density on the parameter brms declares, inside the bounds brms declares. That is the model brms fits with "" in every row of its prior table; brms has no opt-out of its priors. Until now the opt-out added no density at all, which is flat on frmtmb’s internal parameters: on a log standard deviation, improper at sd = 0, and on a first-order correlation, (1 - rho^2)^(-3/2). A standard deviation, sigma and the other positive distributional parameters are now flat on their own scale. A correlation matrix (cor, rescor, cortime, lncor) is flat on its Cholesky factor, and the ordered thresholds and delta are flat on themselves. The defaults that are already flat on brms’s parameter stay: ar, ma, cosy, lagsar, and dirichlet(1) on a simplex, a mixture’s weights included. Against brms 2.23.0 with "" in every slot, 4 chains of 3000 draws each. On y ~ 1 + (1 | g) with 6 groups the posterior mean of sd_g__Intercept is 1.458 in brms, 1.455 now and was 1.197. On y ~ x + (1 + x | g) with 8 groups the correlation’s mean is -0.283 in brms, -0.265 now, and was -1.000 (sd below 0.0001, 529 divergent transitions). On ar(cov = TRUE) with 6 series of 5 the mean of ar[1] is 0.481 in brms, 0.480 now and was 0.541. A random-effect block now non-centers under prior = "flat", since flat on the standard deviation closes the tail at sd = 0, and prior_summary() lists the (flat) rows.

• A bounds-only specification that replaces a default, set_prior("", class = "sd", lb = 0.1), is brms’s (flat) on that slot: flat on the standard deviation above 0.1, where it was flat on its log. So is set_prior("", class = ...) with no bound, and a table row with an empty prior: on cor it is flat on the Cholesky factor, on the intercept no density. The prior_summary() of a prior = "flat" run lists the intercepts as (flat) too. Passed back as prior =, as the object or as its table through as.brmsprior(), it gives the same density: log density and gradient to 1e-12 relative at 40 points.

• The named list of prior objects (prior = list(theta_1 = prior_normal(0, 1))) is refused, as in frmtmb. It took over the internal parameters it named after the class priors. A name inside a correlation block removed the default lkj(1) from the whole block: list(theta_4 = prior_normal(0, 1)) on (1 + x + z | g) left two correlations with no density. The draws’ prior_summary() no longer carries an overrides attribute.

• frmtmb.spatial’s car() dependence parameter is the one parameter brms has a class for that is not yet brms’s (flat): brms declares it on (0, 1) and leaves it flat. Here it carries no density, which is flat on its logit and not integrable at 0. Under prior = "flat" such a posterior is improper; give it a prior with class "theta" on the coef default_prior() lists.

• prior = refuses a default_prior() or validate_prior() table of route = "fit", the default route, edited, cut down to some rows or columns, or not. It refuses a priorlist made from one with as.brmsprior(), + or c() too. Its (flat) rows are frm()’s defaults. Read as brms’s (flat), they replaced brms’s defaults on the intercept, sigma, the standard deviations and the correlations without a word. The message points to route = "sample", whose table is brms’s defaults, and to prior = "flat". as.data.frame() drops a table’s route. So as.brmsprior() marks a data frame that holds frmtmb’s printed (flat). The priorlist is refused where its (flat) rows would replace one of brms’s proper defaults. A route = "sample" table never does that. A table built by hand in brms’s spelling, with "" for (flat), is read as written, and so is brms’s own.

• delta, the distance between equidistant ordinal thresholds, is flat on delta by default, brms’s (flat) with lb = 0. frmtmb holds log(delta) for an ordered family, and the default was flat on that.

• prior_summary() and the default-prior message print brms’s (flat) with lb and ub for cosy, for lagsar and for a first-order cov = TRUE ar or ma. They printed uniform(-1, 1). The density is unchanged.

• set_prior(class = "theta") on a standard deviation, a correlation or rescor is refused, as in frmtmb (?set_prior). One such prior used to retire the default lkj(1) on every correlation of its block.

• The evidence ratio of a point hypothesis is brms’s. hypothesis() divides the kernel density of the posterior draws at the point by the kernel density of the prior draws there, as brms’s evidence_ratio() does. It returns the prior draws of each hypothesis as prior_samples. Without prior draws for every parameter a hypothesis names, its Evid.Ratio is NA with no message, as in brms. Before, the denominator was the prior density read off the prior specification. So draws made without prior draws gave a number where brms gives NA. And sigma, a standard deviation, a correlation, a nonlinear function and the intercept were refused with a warning. brms gives a ratio for all but the intercept. We compared with brms 2.23.0 on y ~ x + (1 | g) (4 chains of 6000 a side, seed 11). The ratios of x = 0, x = 0.2, sigma = 1 and sd_g__Intercept = 0.5 are within 1.7 Monte Carlo standard errors of brms’s.

• as_draws*(inc_warmup = TRUE) and as.mcmc(inc_warmup = TRUE) return the warmup instead of refusing, and update() of draws samples again instead of refusing.

New features

• frm_sample() samples a mo() term on two values, as brms 2.23.0 does (Jmo = 1). The draws report the weight simo_<term>1[1] as a constant 1, as brms’s do. On 200 rows, y ~ mo(x) + z with x in 1 and 2 agrees with brms’s own Stan program (2 chains of 2000 each): every posterior mean within 1.6 Monte Carlo standard errors, every posterior sd within 3.3 percent.

• frm_sample() refuses a mo() term on one value. That term is 0 on every row, so its coefficient has an improper posterior under the default flat prior. brms 2.23.0 cannot sample it either: its Stan program rejects Jmo = 0.

• frm_sample(sample_prior = ) takes brms’s "no" (the default), "yes" and "only", and TRUE and FALSE.

  • "yes" adds draws from the priors to the draws, under brms’s names, as brm(sample_prior = "yes") does. That is one column per prior statement of brms’s Stan program for the model: prior_b for a class-wide prior, prior_b_x for each coefficient of a vector with a coefficient-specific prior, prior_Intercept, prior_sigma, prior_sd_g, prior_cor_g, prior_simo_mox1[k], and so on. A flat prior has no column. A bounded parameter is drawn truncated to its bounds. We checked 19 models. They cover population-level and monotonic coefficients, intercepts of the location and of a distributional parameter, nonlinear parameters, sigma, nu, phi and hu. They also cover smooth standard deviations in two predictors, group-level standard deviations, correlations, rescor and mo() simplexes. And they cover mixture weights, ordinal thresholds, ar of two responses, and multivariate and bounded priors. On each, the names are brms 2.23.0’s, in brms’s order. Each column is drawn directly from its prior, independent of the likelihood, as brms draws it in generated quantities. On 39 columns of 9 models the one-sample Kolmogorov-Smirnov p-value against the exact (truncated) prior is between 0.125 and 0.986. Against brms’s own draws of two models (14 columns, 4000 draws a side), it is between 0.033 and 0.962. An ordinal model with a threshold prior of its own gets brms’s prior_Intercept__1, prior_Intercept__2, …, one per threshold. A gp() term’s sdgp and lscale get no prior draws, and neither do brms’s meanme and sdme, which frmtmb has no prior slot for.
  • "only" samples the prior of the whole model with the likelihood left out, as brms does. So posterior_predict(), pp_check() and conditional_effects() give the prior predictive distribution. Every parameter needs a proper prior: a flat one is refused with brms’s message (“Sampling from priors is not possible as some parameters have no proper priors. Error occurred for parameter ‘b’.”). Unlike brms, a vector with some coefficients flat is refused too. So is a model with missing values of an mi() term, which have no density without the likelihood. A mi(sdy) response keeps its measurement density, as in brms. A model with me() terms is refused, naming brms’s classes meanme and sdme: brms needs proper priors there, and frmtmb has no prior slot for them yet. We checked y ~ x + (1 | g) with a normal(0, 1) prior on b (4 chains of 4000). We compared each mean and each 5, 50 and 95 percent quantile of b_Intercept, b_x, sigma, sd_g__Intercept, r_g[1,Intercept] and of a prior predictive draw. Each is within 3.0 Monte Carlo standard errors of brms’s. At two more seeds every quantile of b_x, sigma and sd_g__Intercept is within 1.3 of the exact prior’s (brms’s: 2.4).

• New prior_draws(), brms’s, with brms’s variable and its deprecated pars; brms’s deprecated prior_samples() reaches it while brms is loaded. A name gets the prior column whose name is the longest match at its start, in random order (b_x reads prior_b). The overall intercept b_Intercept is left out with brms’s warning.

• The warmup iterations rstan saves (save_warmup = TRUE, the default of rstan and of brms) stay in the draws’ stanfit, as in a brmsfit. inc_warmup = TRUE on as_draws(), as_draws_array(), as_draws_df(), as_draws_list(), as_draws_matrix(), as_draws_rvars(), as.mcmc(), as.matrix(), as.array(), as.data.frame(), rhat() and neff_ratio() converts them to the draws’ variables. It puts them before each chain’s draws, as brms does. as.mcmc() refuses it with combine_chains = TRUE and on thinned draws, with brms’s messages, and every accessor refuses a warmup that no longer fits the draws. Every other method reads the post-warmup draws alone, as before.

• update() of draws samples the same model again with changed arguments of frm_sample(), as brms’s update() does for sample_prior, prior and the sampler’s options: update(ds, sample_prior = "yes") adds prior draws. A new formula or new data is refused, as before.

frmtmb.sample 0.23.0

Needs frmtmb 0.76.0, which exports autocor_flat_priors(), autocor_bound_chart(), autocor_innov_sd() and autocor_row_loglik() and codes mo() from its minimum.

Breaking changes

• With frmtmb 0.76.0, which codes an integer mo() variable from its minimum as brms 2.23.0 does, frm_sample() samples brms’s model on such data. On 300 rows with values 1 to 4, the draws have the 3 weights simo_mok1[1] to simo_mok1[3], where they had 4. Every posterior mean agrees with brms’s within 2 Monte Carlo standard errors, where the Intercept was 12 of them away. The code of this package did not change; its tests need frmtmb 0.76.0.

• The default priors of a residual-correlation term are brms’s. Before, an ar(), ma() or cosy() coefficient was flat on frmtmb’s internal parameter. For a first-order cov = TRUE ar or ma that is (1 - rho^2)^(-3/2) on rho, with its mass at -1 and 1. For cosy it is a density flat on the logit of its whole window. And the default on sigma was on frmtmb’s marginal sd, where brms’s sigma under a first-order cov = TRUE ar(), ma() or arma() is the innovation sd. Now each coefficient is flat on itself, as brms’s (flat) is: on (-1, 1) for ar and ma under cov = TRUE, on brms’s [0, 1] for cosy. The default student_t() on sigma is read on brms’s innovation sd, sigma sqrt(1 - ar^2), sigma / sqrt(1 + ma^2) or sigma sqrt((1 - ar^2) / (1 + ma^2 + 2 ar ma)), with the Jacobian of that factor. The reference is brms 2.23.0’s own draws, 4 chains of 4000 on each side (seed 3). Against them, the means of ar[1] were 42.3 Monte Carlo standard errors apart, on an AR(1) of 0.9 in 8 groups of 6 time points. For sigma they were 19.6 apart. An MA(1) of 0.6 had every ma[1] draw at 1. Now every parameter of the ar, ma and arma cases is within 1.6 standard errors of brms’s, the sds within 4 percent. For cosy (0.03 in 30 groups of 4), 48 percent of the draws were below 0, where brms has none. The mean was 0.011 against brms’s 0.076. Now there are none, and the means are 0.075 to 0.078 against brms’s 0.075 to 0.078 on three seeds. A cov = TRUE term of order two or more, which brms does not have, is flat on its stationary or invertible region. Its sigma default stays on the marginal sd. The cov = FALSE coefficients, brms’s default form, are flat as before, as in brms. frm_sample()’s message and default_prior(route = "sample") list the new rows.

• sigma under a first-order cov = TRUE ar(), ma() or arma() is the marginal sd. Every prior on class sigma is read at the innovation sd, as brms’s is. The draws’ sigma column, summary() and variables() report frmtmb’s marginal sd: on the AR(1) above, a posterior mean of 1.16. brms’s sigma is the innovation sd (0.43); the model is the same. A prior of your own on class "sigma" is read where the default is, at the innovation sd. So a default_prior(route = "sample") table passed back as prior =, or the default’s text typed, gives the default’s density exactly. Before, a user prior there was read on the marginal sd. That density was up to 0.72 nats from brms’s with the same text (three points on the AR(1) data). frm() with the same prior is the mode of the density frm_sample() samples. A lb above 0 or a ub on sigma there is refused.

• A prior of your own keeps the bound brms declares for its slot. brms keeps a declared bound under a user prior: real<lower=1> nu, real<lower=0> sigma under an identity link, real<lower=0,upper=min_Y> ndt, real<lower=0,upper=1> cosy. Before, a user prior of the same slot dropped the default’s bound: with set_prior("normal(2, 2)", class = "nu") 3.55 percent of the draws were below 1. Now the bound stays on each side the prior leaves unset, and a lb or ub of your own replaces that side, as in brms. With set_prior("normal(0, 0.1)", class = "cosy") the posterior mean is 0.054 against brms’s 0.053 (z 1.5); it was 0.009, with draws below 0. A bounds-only prior on a residual-correlation class keeps the flat density as well: set_prior("", class = "ar", lb = 0.5) is flat on [0.5, 1] in brms. Here it was flat on frmtmb’s internal parameter (mean 0.946 against brms’s 0.916, z 43.8); now the mean is 0.916 (z 0.5).

• A prior on class = "theta", coef = "thetaac_1" is refused where a class reaches that parameter (ar, ma, cosy, cortime); see frmtmb’s NEWS. The message for the retired lower/upper arguments names coef = "theta_1" as its example.

New features

• log_lik() of draws no longer refuses a residual covariance term (ar(cov = TRUE), ma() and arma() with cov = TRUE, cosy(), unstr(), fcor(), frmtmb.spatial’s sar() and other registered terms). A column is brms 2.23.0’s leave-one-out conditional density of the row given the other rows of its group, normal or Student-t. It equals brms’s log_lik() at the same draws to 2.0e-15 relative on 18 cases. loo(), waic(), psis() and the loo_*() functions take these models. The values do not sum to the log-likelihood, as in brms. On a group with a missing time point, the conditional is the fitted model’s. Its ar(), ma() and arma() lags count time levels, where brms’s count positions.

Bug fixes

• A bounded residual-correlation parameter is sampled through one logistic onto its interval, as Stan’s own bound is in brms. Stan’s log bound on frmtmb’s internal parameter, stacked on the term’s own map, gave 5 to 9 divergent transitions per 4 chains of 2000. That was on a cosy near its bound at 0 (seeds 3 to 5, mode and random starts), where brms has none. Now there are none.

• When the ML estimate of cosy is below brms’s bound at 0, the warning says that the bound is the default, brms’s declaration. It names set_prior("", class = "cosy", lb = -1/(d - 1)) to sample the whole window frm() fits. It said “the requested bound(s) on thetaac_1”.

frmtmb.sample 0.22.0

Needs frmtmb 0.75.0, which exports frmtmb_covstruct_info() and frmtmb_autocor_pars() and gives coefficients brms’s names.

Breaking changes

• The draws of a model with an exp() term carry brms’s names: b_expx, sd_g__expx and cor_g__Intercept__expx, as brms 2.23.0 names them, where they were b_.frm_expx, sd_g__.frm_expx and cor_g__Intercept__.frm_expx. fixef() of the draws names the row expx, and hypothesis(draws, "expx > 0") works. The change is in frmtmb, which builds the names, and the new test of it needs frmtmb 0.75.0. The same holds for a nonlinear parameter of a multivariate model, b_yn_a_Intercept where the draws held b_a_yn_Intercept.

• fixef() of the draws lists the rows in brms’s order, which is the order of fixef() of the fit: every centered intercept first. For bf(y ~ x, sigma ~ exp(z)) brms 2.23.0 gives Intercept, sigma_Intercept, x, sigma_expz. The draws’ fixef() listed Intercept, x, sigma_Intercept, sigma_expz, the order of the draws’ columns, which is unchanged.

New features

• The draws of a model with frmtmb.spatial’s sar() carry brms’s lagsar or errorsar column, derived from the sampled internal parameter, as ar[1] and cosy are. The names come from core’s registry (frmtmb::frmtmb_autocor_pars()), so a term another package registers gets its own column the same way.

• The default prior on lagsar and errorsar is brms’s: flat on rho over its interval, which frm_sample() announces as uniform(lower, upper), the ends shown to 6 digits. The same holds for a bounded parameter of a term another package registers. set_prior(class = "lagsar") and class = "errorsar" replace it.

Bug fixes

• The derived sd_ and cor_ columns asked core’s frmtmb_covstruct_info() which structures are fields, where they kept their own list of five names. A field another package registers, beside a block on the same grouping factor, was mapped onto that block’s names. It failed the map’s own check and sent every draw through the slow path: on one model, 0.70 ms a draw against 0.053 ms on the fast path now. The values were right either way. Needs frmtmb 0.75.0, which exports frmtmb_covstruct_info().

frmtmb.sample 0.21.1

Needs frmtmb 0.74.0. The package itself loads on frmtmb 0.73.0, but four of its test files need what frmtmb 0.74.0 and frmtmb.latent 0.7.0 add: the covariance structure registry, frmtmb_family(simplex_pars =) and mixture_mvn() in frmtmb.latent.

Bug fixes

• The note of default priors from frm_sample() names every simplex that a family declares with frmtmb_family(simplex_pars =), by its brms name. It named only the sbhaz simplex of cox(), so the dirichlet(1) default on the simplex of another family was not announced. For cox() the note now reads “sbhaz: dirichlet(1), brms’s default, on the simplex the cox() family declares”.

Other changes

• The test of mixture_mvn() draws reads the family from frmtmb.latent, where it moved, and skips without it. The help of posterior_predict() and pp_mixture() links it there.

• frmtmb.survival is a suggested package. The tests of cox() draws read cox() from it, and skip without it.

frmtmb.sample 0.21.0

Needs frmtmb 0.73.0. It registers its draws methods on the loo_*(), kfold(), reloo(), loo_moment_match() and loo_subsample() generics that frmtmb now defines, and re-exports them. It calls the sampling-API export check_weights_unbounded().

Breaking changes

• On a subset of the draws (ndraws, draw_ids), loo(), psis() and the weights of every loo_*() function take the relative efficiency on one chain. brms 2.23.0 does the same, and so did the loo_* types of pp_check(). They used r_eff = 1. The same draws are now weighted the same way everywhere. On 60-row gaussian and random-intercept models with 4000 draws, loo(ndraws =) at 500, 1000 and 3000 draws moved by at most 0.0065 in elpd_loo. Its standard error was 5.0. On every draw nothing changes, and frm_sample(thin =) draws keep their chains.

• posterior_average(weights = "stacking") and "pseudobma" take their weights from loo_model_weights(): one log_lik() per model and its relative efficiency, as brms’s model_weights() does. Before, they read a loo criterion that add_criterion() stored, which can come from other draws. Without a stored criterion the weights are identical to the old ones.

• The kfold(), reloo(), loo_moment_match() and loo_subsample() generics, and those of the six loo_*() functions under New features, are frmtmb’s, re-exported here. On a maximum-likelihood fit each refuses, with frmtmb alone attached too: the six name frm_sample(), the other four what to use instead. loo_moment_match() and loo_subsample() had no method for a fit, even with this package loaded.

• frm_sample() from a formula refuses a posterior that does not integrate. Suppose that every row with a positive weight has a random effect of its own in mu. The likelihood grows like sigma^-(sum(w) - n) as sigma shrinks to 0. A prior that is finite at 0, as brms’s default student_t(3, 0, s) on sigma is, then leaves the posterior improper once sum(w) - n reaches 1. Such a run used to return chains that had not mixed (largest Rhat 2.5 on 2 chains of 1000, sigma near 1e-4). The check is core’s, against the priors the chains sample under. A prior on sigma that vanishes at 0 fast enough makes the posterior proper but not informative: lognormal(0, 1) gave a posterior median sigma of 7.2e-05 and a largest Rhat of 2.83. A usable model has weights that sum to the number of rows or fewer, or no observation-level term.

New features

• brms’s leave-one-out functions on draws: loo_epred(), loo_linpred(), loo_predict(), loo_predictive_interval(), loo_R2() and loo_model_weights(), with brms 2.23.0’s arguments, defaults and output shapes. A multivariate model without resp gives one layer per response, named as brms names them, and a category margin stays third. Each computes one log_lik() and one loo::psis() per call, on the same draws as the prediction (per response for loo_R2(), per model for loo_model_weights()). The weights are those of psis() and of loo(save_psis = TRUE). psis_object = reuses weights and computes no log_lik(). On the same posterior draws as a brms 2.23.0 fit, every loo_epred() and loo_linpred() value is identical to brms’s under R’s reference BLAS. So is every loo_model_weights() weight (a stacking pair of 0.155 and 0.845). The loo_R2() bootstrap draws at one seed agree to rounding (largest relative difference 1.1e-14). Under OpenBLAS, where brms’s and frmtmb’s predictors round apart in the last bits, they agree to that rounding.

• add_criterion(x, "loo_R2") stores the loo_R2() bootstrap draws, as brms stores them, and loo_R2() reads them. It was refused. add_criterion() gives each criterion the arguments it takes, as brms does: add_criterion(x, c("loo", "loo_R2"), seed = 1) seeds the bootstrap and no longer stops in loo().

• psis() honors draw_ids, which it ignored, and refuses an argument it does not read, which loo::psis() swallowed.

• loo() and the loo_*() functions accept brms’s default values by name (allow_new_levels = FALSE, sample_new_levels = "uncertainty", newdata = NULL), which loo() refused, and loo() and loo_model_weights() take draw_ids.

• resp takes brms’s spelling (ya for a column y_a) in all six functions, and the layers of a multivariate result carry it. loo_epred(), loo_linpred(), loo_predict(), loo_predictive_interval(), loo_model_weights() and psis() also take frmtmb’s y_a, which log_lik() and posterior_epred() take. loo_R2() takes brms’s spelling only, as bayes_R2() and brms do.

• brmsfamily("negbinomial2") and dirichlet_multinomial() fits sample under brms 2.23.0’s default priors: sigma of negbinomial2 gets gamma(0.4, 0.3) from brms’s family hook (where negbinomial’s shape gets inv_gamma(0.4, 0.3)), and phi gets gamma(0.01, 0.01). On the same data the posterior means agree with brms’s to 0.05 posterior sd, with and without a group term.

• A dirichlet_multinomial() fit’s draws take the multinomial() paths: posterior_linpred() with no dpar stacks the mu<cat> predictors into brms’s draws x observations x (K - 1) array. posterior_epred() and posterior_predict() return draws x observations x K arrays that read trials() from newdata. predictive_error() and pp_check(type = "error_binned") refuse the family as polytomous.

Divergences from brms

• newdata, re_formula and the new-level arguments, given a value other than brms’s default, are refused by the loo_*() functions, as loo() refuses them: a weight comes from the observation’s own column of log_lik(), so it exists only for a row the model was fitted on. brms computes the weights on the new rows’ own response.

• ndraws and draw_ids choose the draws once, for the weights and the prediction. brms draws a separate random subsample for each, so its weights belong to other draws: on a 60-row gaussian model with ndraws = 100 the two subsamples shared 2 draws.

• An argument the loo_*() functions have no use for is an error that names it. loo_linpred(scale = "response"), which brms 2.23.0 reads nowhere (its result is the default’s, on a logit link too), is one.

• loo_R2() refuses an ordinal or categorical family, as bayes_R2() does. brms refuses the categorical families and computes the ordinal ones with a warning.

frmtmb.sample 0.20.0

Needs frmtmb 0.72.0. It calls the sampling-API export ord_lp_blocks() and writes uniform() and lognormal() priors. It reads the per-response lncor_<resp> names and the per-coefficient sd priors of that release.

Breaking changes

• posterior_linpred() with no dpar on a family with several location predictors returns brms’s draws x observations x P array, every margin unnamed: every mu<cat> of categorical(), multinomial() and the simplex families (K - 1 layers with a reference category, K without), and every component’s mu<k> of a mixture. It returned the first predictor’s matrix. On these families transform = TRUE without a dpar is refused in brms’s words, a two-category categorical() with its one predictor included.

• A default steps aside only for a prior on its own slot, as in brms. A narrower prior keeps the defaults on the parameters it does not reach. set_prior(class = "sd", group = "g") replaces the default on the standard deviations of g. It keeps student_t(3, 0, 2.5) on the other groups. Adding coef = "x" narrows the prior to sd_g__x. Before, the defaults of the whole class stepped aside, and the other standard deviations were flat.

• A class-wide prior does not replace a default written per coefficient. For example, set_prior("inv_gamma(2, 1)", class = "lscale") on a gp(x) model keeps brms’s tuned default on lscale_gpx, as brms keeps it. Before, the class-wide prior replaced it. The default-prior message now names a specification that applies to no parameter. That happens for this reason, or when a more specific specification of your own covers its parameters. The message names the coef to write.

• frm_sample() puts brms 2.23.0’s default prior on every distributional parameter that brms gives one. They were flat on the link scale: shape (gamma(0.01, 0.01), and inv_gamma(0.4, 0.3) for the negative binomials), phi (gamma(0.01, 0.01)), nu (gamma(2, 0.1)) and kappa. So were the exgaussian beta, alpha, zi, hu, zoi, coi and quantile (beta(1, 1)), and frmtmb.eam’s bs, bias and ndt. A parameter with a formula has brms’s default on its intercept, on the link scale. On small data the posteriors move. A beta phi near 300 on 10 rows moved by 30 Monte Carlo standard errors. A negative binomial shape and a Student-t nu had no proper posterior before.

• The ordinal thresholds take brms’s student_t(3, 0, 2.5), class Intercept, and a mixture’s weights without a formula take brms’s dirichlet(1) on the simplex. Both were flat.

• sigma ~ 1 takes brms’s default on sigma’s intercept. It got a natural-scale default that addressed a slot the model does not have. sigma without a formula now reads the response on the identity scale, as brms does. Under a log mu link it read it on the log scale.

• An unstr() residual correlation takes brms’s lkj(1), class cortime.

• frm_sample(seed = ) fixes every chain. The init jitter of chains 2 and up came from the session’s random stream, so two calls with seed = 3 differed unless set.seed() came first. The jitter is now drawn from seed and no longer from the session’s stream. rstan’s sampler still advances the session’s stream, as under brm().

• A natural-scale default carries brms’s bound on its class where frmtmb’s link does not hold the parameter inside it. A log-linked ndt is bounded above by the smallest response. An identity- or log-linked nu is bounded below by 1. brms declares them so, and Stan then transforms the draws inside rather than meeting the density’s edge. Where the link holds ndt inside a bound of its own, frmtmb.eam’s, the uniform is over the link’s image and no bound is written: uniform(0, min_Y) for an ungrouped wiener(), lba() or rdm(), and uniform(0, 1) for the fraction ndt_group() makes of ndt.

• posterior_predict(newdata =) reads every addition term from newdata. It read only trunc(), rate() and thres(gr = ), and gave the simulator no other term, from the method’s first version on. A binomial with y | trials(N) and newdata N = 1000 drew 0 or 1 (column mean 0.595 where posterior_epred() gave 576.3). A gaussian with se(s) drew at sd 0.033 where newdata had s = 5. predict(), predictive_error() and residuals() of draws, which call it, change with it. Newdata without the trials() or se() variable is now refused, naming the term, as brms refuses it.

New features

• Draws of a logistic_normal() response inside mvbf() report its correlations as lncor_<resp>__b__c, with <resp> brms’s Stan name of the response. The sampler’s default puts brms’s lkj(1) on each such response.

• The draws carry brms’s sd_<group>__<coef> and cor_<group>__<coef1>__<coef2>, and a smooth’s sds_, rescor__, ar[1], cosy and cortime__, as columns of the draws matrix. So as_draws_df(draws)$sd_g__Intercept works as on a brms fit. They are derived from each draw’s theta_ columns, which stay. The names are core’s, the ones variables() of the fit lists.

Bug fixes

• posterior_predict(newdata =) and posterior_predict(re_formula =) hand the simulator each distributional parameter on its natural scale. They handed it the reporting scale, which for a mixture’s theta is already the softmax probability. The simulator applied the softmax again, so the mixing weights were wrong. The defect is in every release from 0.2.0 (frmtmb 0.52.0, whose afd00a43 gave a mixture’s weights their softmax reporting scale). With frmtmb’s fix of the same release, a mixture also draws each class at its own rows’ trials().

frmtmb.sample 0.19.0

Needs frmtmb 0.71.0: it reads the sampling-API exports ce_method(), ce_predict_family_check(), ce_sim_one() and pp_check_resp(), ce_plan_eval(one = ), and the gen_extreme_value() chart gev_xi_chart_frame().

Breaking changes

• loo() and waic() of draws take brms’s arguments in brms’s order, (x, ..., compare, resp, ndraws, model_names). Draws given without a name are further models, which are compared as brms compares them. So ndraws must be named, as it is in brms.

• loo() and waic() of draws refuse by name the arguments they dropped. These are newdata, re_formula and the other arguments log_lik() refuses, which brms reads. loo(newdata = ) returned the in-sample elpd in silence (lane remeasure, R2). moment_match = TRUE and reloo = TRUE refit, so they are refused too. pointwise and k_threshold change nothing and are accepted.

New features

• The draws of core’s dirichlet() and logistic_normal() fits take brms’s shapes. posterior_epred() on a dirichlet() fit is the draws x observations x categories array of category means. posterior_predict() draws rows on the simplex for both families, with layers named by category, as brms’s are.

• The correlations of a logistic_normal() fit are the draws column lncor__b__c, brms’s name. They take brms’s lkj(1) default, and each sigma<k> takes the default that brms gives sigma.

• As in brms, posterior_epred() refuses logistic_normal(). pp_check() and predictive_error() refuse both families.

• frm_sample() samples a gen_extreme_value() xi without a formula as brms samples it, through tmp_xi. It maps tmp_xi into the support that each draw’s mu and sigma leave, and gives it brms’s default normal(0, 2.5). On its log1p scale the support’s edge was a wall in the sampled space. The review fitted n = 40 data sets at xi -0.45, 0.20 and -0.85 (4 chains x 2000). There 1187, 263 and 3856 of 4000 transitions diverged, with Rhat 2.30 on the last, against 0, 1 and 0 in brms. Through tmp_xi none diverge, and the posterior mean of xi is within 0.005 of brms’s at the first two. One difference from brms remains: frmtmb holds xi above -1 on every route, where brms’s range goes on (dev/famuni-out/tmpxi.txt, tmpxi-sample.txt). The hold is material where the data put mass there. On the xi -0.85 data brms puts 29 percent of its posterior below -1. There frm_sample()’s posterior is brms’s conditioned on xi > -1 (median of xi -0.876 against -0.877; the review’s round 2). A prior on class xi is then a prior on tmp_xi, as in brms. A cens() or trunc() response, a second GEV response, or GEV components of a mixture() keep the log1p scale, and frm_sample() warns that they do.

• frm_sample() puts brms’s default normal(0, 4) on the intercept of a predicted xi, xi ~ 1 included (normal(0, 2.5) under link_xi = "identity").

• loo() and waic() of several draws objects return brms’s comparison. It holds the criterion of each model (loos), loo::loo_compare() of them (diffs) and brms’s pairwise ic_diffs__, printed as brms prints its loolist. compare = and model_names = are brms’s. Models with different numbers of observations are refused in brms’s words. Models of different responses get loo’s warning, through brms’s yhash attribute.

• logLik() of draws is log_lik(), the draws by observations matrix, as brms’s logLik() of a fit is. It had no method.

• bayes_R2() of draws takes newdata and re_formula, which brms hands to posterior_epred(). With newdata it is the R-squared of newdata’s rows, against newdata’s own response column. A newdata without that column is refused in brms’s words.

• model.frame() of draws is the model frame of their fit. stats’ default rebuilt a frame from the formula’s environment, which holds none of the data. It stopped at “object ‘count’ not found” on brms’s ported row brmsfit-methods:120.

• conditional_effects(method = "predict") on draws (brms’s "posterior_predict") draws one response per grid row and posterior draw, at the parameters of that draw. That is the posterior predictive band brms draws. Under the default robust = TRUE, estimate__ is its median and se__ its MAD. A category-valued response is refused, as the fit method refuses it.

• transform = applies to the predicted draws, in conditional_effects(method = "predict") and posterior_predict(), with brms’s warning that the argument is deprecated.

• frm_sample() of a frm_multiple() result samples the fit of every imputation and combines the draws, as brms’s brm_multiple() combines its chains. nchains() is the number of imputations times chains. A seed gives imputation j the seed seed + j - 1. Methods that read the data in sample read the first imputed data set, as brms’s do.

• The diagnostics of combined imputations are brms’s. summary() leaves out Rhat, Bulk_ESS and Tail_ESS, as brms’s does for a brm_multiple() fit. loo() takes a relative efficiency of 1, as brms does. Chains of different imputations sample different posteriors, so their R-hat measures the imputations.

Bug fixes

• pp_check() of multivariate draws refuses a missing, plural or unknown resp in brms 2.23.0’s words, as the fit method now does. A missing one was “pp_check() is not supported yet for multivariate fits”.

• summary(draws, waic = TRUE) is refused by name, naming waic(x). Old brms printed WAIC there, and brms 2.23.0 ignores the argument. It got the generic “has no argument waic”. waic = FALSE is accepted.

frmtmb.sample 0.18.0

Needs frmtmb 0.70.0: it reads simplex_frame_terms() and the Dirichlet prior entries frmtmb now resolves, and a cox() fit’s baseline through mo_chart_frame().

Breaking changes

• A cox() baseline is sampled as brms samples it: in softmax coordinates under brms’s default dirichlet(1) (class sbhaz), and the draws carry the weights as sbhaz[k], brms’s names, in place of the sbhaz_raw_k coordinates. With no density on the coordinates two of them ran to -3e11 and -9e12 on a right-censored model, with 239 divergences and R-hat 3.07; now the weights’ means and sds agree with brms 2.23.0’s within 2.1 Monte Carlo standard errors on two data sets, with 0 divergences (dev/moparity-cox-sample.R).

• An exact gp() field is sampled non-centered, as brms’s sdgp * L zgp, when its sd and length scales have priors (the defaults do). The centered field is a funnel: on lane surface’s y ~ gp(x) the formula route ran 1286 of 2000 transitions into the maximum tree depth with R-hat 1.52 and a bulk ESS of 7; non-centered, none, R-hat at most 1.016, bulk ESS 312 or more, in a third of the time. The factor is a Cholesky factorization of the kernel on the tape, n^3 / 3 per gradient for n positions. reparameterize = FALSE keeps the centered field.

New features

• A simo or sbhaz prior row replaces dirichlet(1) on its simplex, with the softmax’s Jacobian. With dirichlet(c(5, 1, 1)) the draws of simo_ agree with brms 2.23.0’s on a weakly and a strongly identified simplex: the means within 1.71 Monte Carlo standard errors and the sds within 2.51 (dev/moparity-dir-sample.R).

• mo() crossed with a factor and group-level mo() terms sample, with brms’s names (simo_mox:fb1[k], bsp_mox:fb, r_g[, mox]). On mo(x) * f2, with a two-level f2, and on mo(x) + (mo(x) | g) the means and sds agree with brms 2.23.0’s within 2.12 Monte Carlo standard errors (dev/moparity-dir-sample.R).

• as_tmbstan() names only the simplexes with no density on the fit’s tape: a MAP fit with a simo row has its Dirichlet there. The function still samples the fit’s own tape, as documented; it adds no prior.

Bug fixes

• parnames() on draws warns once whichever generic dispatched it, as on a fit.

frmtmb.sample 0.17.0

Needs frmtmb 0.69.0: it imports and re-exports frmtmb’s stancode(), standata(), pp_mixture() and add_criterion() generics, and it holds a mo() simplex through frmtmb’s mo_simplex(), mo_coords(), mo_chart_frame(), mo_frame_terms() and summary_mo_frame().

Breaking changes

• stancode(), standata() and pp_mixture() are frmtmb’s generics, re-exported here; this package keeps its methods for draws. frmtmb defines them because it now has methods for a fit.

• pp_mixture()’s summary names its components P(K = k | Y) and numbers its rows, as brms does and as frmtmb’s method for a fit does. They were class1, class2. On draws of a model that is not a mixture it stops in brms’s words, “Method ‘pp_mixture’ can only be applied to mixture models.”, as on a fit.

• frm_sample() puts brms’s default dirichlet(1) on every mo() simplex and reports brms’s simo_ weights. The sampler put no density on a simplex, so the fit’s coordinates were the prior. In frmtmb’s old softmax coordinates that prior is 1 / prod(w), and a weakly identified simplex sat at a vertex (44 divergences); in the coordinates frmtmb now fits in it is not integrable at the barycenter, and the draws sat there with sd 0. The simplex is now sampled in softmax coordinates under dirichlet(1), which is the flat density on the simplex (so prior = "flat" keeps it), and the draws carry the weights as simo_<term>1[k], brms’s names, in place of the zeta<j>_k coordinates. Against brms 2.23.0 on a weakly identified simplex and on brms_monotonic’s fit1 (4 chains of 2000 each side), every weight’s mean and sd agree within 1.7 Monte Carlo standard errors, with 0 divergences and R-hat at most 1.0033. Models without mo() terms sample exactly as before. A simo row in prior = is still refused: set_prior() has no Dirichlet density.

• With more than one mo() term, each term’s weights are adjacent columns, term by term, as brms orders them.

• check_laplace() on a mo() fit compares each simplex on the weight scale: the ML weights and their delta-method errors (as summary() gives them) against the posterior of the simo_ weights. The coordinates are not comparable, since the fit’s sheet and the draws’ need not be the same. The simplex is sampled under dirichlet(1) there too, the flat density on the simplex, under which the weights’ posterior is the fit’s likelihood; a MAP fit keeps its own prior. A weight at 0 or 1 is reported in a message, with no Wald error.

• as_tmbstan() says in a message, on a fit with mo() terms, that the simplex has no density on that route (it is sampled flat in the fit’s own coordinates, which a weakly identified simplex collapses in) and that frm_sample() samples it under dirichlet(1).

New features

• plot() of draws is brms’s display: a histogram and a trace per parameter through bayesplot::mcmc_combo(), nvariables to a page, with brms’s pars, variable, regex, fixed, combo, bins, theme, plot, ask and newpage, and N as brms’s deprecated alias with brms’s warning. It refused every call before.

• add_criterion() on draws stores loo, waic and bayes_R2 in x$criteria as brms does, and loo() and waic() with no further argument return the stored object, brms’s use_stored. overwrite = TRUE clears the stored criterion before it computes, as brms does, so it recomputes. kfold, loo_subsample, loo_R2 and marglik are refused with their own reasons.

Bug fixes

• posterior_samples() on draws warns once that it is deprecated. While brms was loaded it warned twice, once from brms’s generic and once from the method.

• The compatibility pre-flight of frm_sample() no longer loads R’s base-priority packages to ask whether a name is one of their functions; it reads them without loading. On a headless Linux machine, loading tcltk warned “no DISPLAY variable so Tk is not available” into an unrelated call.

• check_laplace() leaves a parameter whose z_shift or sd_ratio is not finite out of its “Laplace/Wald approximation questionable” message. The test passed NA to if (any(...)), so such a parameter stopped the call (“missing value where TRUE/FALSE needed”) when no other parameter was flagged, and was listed as NA when one was. The returned table is unchanged.

frmtmb.sample 0.16.0

Needs frmtmb 0.68.0: for ord_thres_linpred(); for gp_brms_natural(), the prior classes "sdgp" and "lscale" and the kriging draw; and for ordinal mixtures, hurdle_cumulative() with thres(gr = ) and cs(), and the names of the draws of a model with no location column.

Breaking changes

• Through frmtmb, a smooth’s null-space draws (bs_sx_1) are on brms’s scale, and a prior with coef = "sx_1" is about brms’s parameter: frmtmb now builds s() with brms’s diagonal.penalty = TRUE basis.

• A gp() term’s draws carry brms’s names on brms’s scales: sdgp_<term> and lscale_<term> where they were theta_1 and theta_2 on the internal log scale, as brms names them (sdgp_gpxfa for level a of gp(x, by = f)). The lengthscale is on the inputs brms divides by their largest distance.

• The default priors include brms’s for a GP: student_t(3, 0, s) on class "sdgp" per prefix, and on class "lscale" an inverse gamma per sub-GP, tuned to that sub-GP’s own distances as brms’s def_lscale_prior() tunes it. get_prior(route = "sample") lists them; on nine gp() designs every density row equals brms default_prior()’s (63 rows, dev/gpby-priors.R). A Hilbert-space gp(k = ) block, whose lengthscales the defaults left flat, is now non-centered.

New features

• posterior_linpred(incl_thres = TRUE) returns brms’s linear predictor with the thresholds included on every ordinal family: a draws by observations by thresholds array with brms’s dimnames, layer k being disc * (thres_k - mu) for cumulative() and sratio() and disc * (mu - thres_k) for cratio() and acat(), under every threshold structure, with cs() terms, and NA past the thresholds of a thres(gr = ) level that has fewer. As in brms it is ignored with dpar, nlpar or transform = TRUE and on a family that is not ordinal. hurdle_cumulative() is refused, a divergence the user decided on 2026-10-06: brms returns the hurdle probability beside the threshold predictors times 1 - hu there, which is the predictor of nothing (dev/upstream-bugs.md, brms-21). An ordinal mixture is refused in brms’s words, “‘incl_thres’ is not supported for mixture models.”

• Draws of an ordinal mixture carry brms’s names, each component’s thresholds b_mu1_Intercept[k] and b_mu2_Intercept[k]. Under order = "mu" both come from the one shared vector, which is read back through the first component that reports it uncentered. posterior_epred() and log_lik() agree with the theta-weighted sum of brms’s R-side category probabilities at the stored columns to the double epsilon.

• brms’s default normal(0, 1) on the intercept of a modeled disc<k> of an ordinal mixture, as on disc.

• Through frmtmb 0.68.0, draws of a cs() model on cumulative(), as brms samples it: the per-threshold coefficients are bcs_x[k], and a draw’s posterior_epred() equals brms’s own R-side density at that draw to 1.9e-16 (dev/rel068-cs-probe.R). A row whose offsets cross two thresholds has NaN probabilities and an NA draw, as in core.

Bug fixes

• A prediction at a position an exact gp() did not see is a draw of the field there, from its conditional law given the draw’s own values at the fitted positions, as brms’s posterior_epred() draws it. It was that law’s mean, so posterior_epred(newdata = ) past the data had no spread from the field itself. At the same 1500 parameter draws of a brms fit, the spread of the predicted field at six new positions is 0.978 to 1.001 of brms’s, where the mean alone reached 0.861 to 1.000 (dev/gpby-brms-epred2.R, final build in dev/gpby-p1-nugget-brms-1e-6.txt). The law carries frmtmb’s nugget, 1e-6 of the variance, as white noise per distinct position, as the fit’s latent values do; two rows at one position take one value exactly.

• The draw’s factor costs under 1 ms per posterior draw at 300 new positions: a pivoted Cholesky factor that builds only the kernel columns it pivots on (rank 7 to 15 there) and stops where what is left is below 1e-12 of the prior variance, which joins the white part (dev/gpby-p1-krigfactor.R). On posterior_epred() at 300 new positions over 300 draws, CPU time is 0.78 s, against 5.76 s for a dense Cholesky of the same law and 0.63 s for the previous release’s conditional mean, on one machine in one session (dev/gpby-p1-krigdraw.R). # frmtmb.sample 0.15.0

Needs frmtmb 0.67.0: for ord_delta_info(), the ordinal families’ disc and threshold structures and hurdle_cumulative()’s inverse threshold map; for the parnames() generic, which it imports, and the conditional-effects engine that draws crossed and multi-membership new levels; and for arma_cond_fill_dpars(), arma_cond_fill_epred() and response_codes_newdata().

New features

• Draws of an ordinal fit with equidistant or sum-to-zero thresholds carry brms’s names: every threshold b_Intercept[k] and, for equidistant thresholds, delta. The sampler holds fewer parameters than that (two per vector, or one fewer than the thresholds); the columns past them are added after the sampled ones, before lp__, as a mixture’s last weight is, and dropped again whenever a draw is handed back to the model.

• brms’s default normal(0, 1) on the intercept of an ordinal family’s disc, which is what places that intercept, since the likelihood cannot tell it from the scale of the thresholds. Under link_disc = "identity" brms uses lognormal(0, 1), which set_prior() does not carry, and the default-prior message names the gap.

• nsamples(incl_warmup = TRUE) counts the iterations the sampler saved, warmup included, from the stanfit that frm_sample() keeps, as brms counts them. It was refused. Draws with no stanfit behind them refuse it by name. nsamples(subset = ) takes brms’s check that the subset fits the count.

• posterior_samples(pars = ) returns the coefficients in brms’s order, as brms’s variables() lists them: the intercept of every distributional parameter first (b_Intercept, b_sigma_Intercept, b_x, b_sigma_x), a nonlinear parameter’s intercept staying with its own coefficients (b_a_Intercept, b_a_z, b_b_Intercept). Without pars the columns are those of variables(), as before. A pars that matches nothing returns NULL, as in brms.

• parnames() is frmtmb’s generic, re-exported here. With brms loaded it warns once, as brms’s does; brms’s generic and the method each warned before.

• conditional_effects() on draws answers crossed (1 | g) + (1 | h) + (1 | g:h) at an unseen combination of observed levels, and a new member of an mm() term with a by variable.

Bug fixes

• A hurdle_cumulative() fit’s draws are named b_Intercept[k], as the other ordinal families’ are. They were tau_raw_k.

• posterior_predict(newdata = ) of a model with ar(), ma() or arma() and cov = FALSE was refused as a structured draw, the same refusal as a cov = TRUE block, though core’s predict() answers it. It answers now, and with brms’s treatment of a missing response: a row of newdata whose response is NA, or every row when the column is absent, is filled with a draw at that draw’s parameters before the rows after it read its residual. re_formula = is no longer refused on such a model either.

• posterior_epred(newdata = ) under cov = FALSE fills a missing response with a draw, as brms’s posterior_epred() does, so the expected-response draws on the rows after it carry the spread of the unobserved past. Core’s fitted() on a maximum likelihood fit fills with the expected value instead, because it has no draws to carry the fill through.

• predictive_error(newdata = ) of a bernoulli() model codes newdata’s response as the fit coded its own, so a response held as two values other than 0 and 1 (brms’s level-order coding, new in frmtmb) is compared with the draws on the same scale.

frmtmb.sample 0.14.0

Needs frmtmb 0.66.0: for subset_resp_check() and subset_newdata(), for cs_build(), ce_level_plan() and the other engine functions of conditional_smooths() and conditional_effects(), for the ordinal families’ inverse threshold maps that name the draws, and for equated parameters.

Breaking changes

• An ordinal fit’s draws carry brms’s names. The thresholds are b_Intercept[1], b_Intercept[2] (per group under thres(gr = ), per response in a multivariate model) and hold the thresholds, not cumulative()’s first threshold and log increments; a cs() column is bcs_<column>[k]. They were tau_raw_k and bcs<j>_k. variables(ds) now lists what variables(fit) lists. A draws object saved with the old names still works. Code that selected tau_raw_1 from the draws must select b_Intercept[1] now. The names come from each ordinal family’s inverse threshold map, which frmtmb 0.66.0 declares.

Bug fixes

• posterior_predict(), posterior_epred() and posterior_linpred() on frm_sample(laplace = TRUE) draws returned silently wrong numbers. Such draws hold no random effects, and every draws method read them in the layout of full draws: sigma and theta landed in the random effects and NA past the last column. On y ~ x + ar(t, g) + (1 | g) all 4500 cells of posterior_epred() were NaN; at newdata the values were finite and wrong, each group effect read off theta_1 or lp__. fitted(), predict(), residuals() and bayes_R2() returned the same kind of numbers; pp_check(), predictive_interval(), pp_mixture() and hypothesis() on sd_ quantities stopped in an internal error from quantile() or bayesplot. Laplace draws are now read in their own layout. A quantity that reads what was integrated out (the group-level coefficients, a smooth’s coefficients, mi() values) is refused, naming the function you called, as log_lik() already refused; one that does not, such as re_formula = NA on a model with group-level terms alone, is computed exactly and equals the full draws’ answer bit for bit. The new “Laplace draws” section of ?frm_sample lists both, and says why the missing values are not filled in (user decision, 2026-09-29).

• conditional_effects() on draws under re_formula = NULL reads each grid row’s groups by brms’s rule (found by lane sampfix). With conditions = data.frame(g = <a level>), every draw’s effects for a NEW group were drawn into the first level and predicted there: moving r_g[1,] or r_g[2,] by 10 left the curve where it was, and levels 1 and 2 drew the same curve. The rule is now core’s, per row and per group-level term: a term reads its draws where the row sets every one of its grouping variables to an observed level, in conditions or as an effect ("x:g"), and draws a new level otherwise, from each draw’s own covariance: an unset variable, an unseen level ("99"), a row of a mixed conditions column, or a nested (1 | g / h) with g set and h unset. At levels 1, 2 and 3 every column and the spaghetti frame equal brms’s at the same five draws, with a difference of 0 (dev/postfit2-celevel-brms.R in the frmtmb repository), and "x:g" equals brms’s to 1.6e-8.

  On a gr(g, by = f) term a row reads only the term of its own f level; every re_formula = NULL display of such a model was refused before, even at an observed g. An mm(g1, g2) term’s unset members are new levels instead of the first observed one. At the same draws and seed, the gr() band at an unseen g and the mm() band with members unset equal brms’s sample_new_levels = "gaussian" bands to 8.9e-16 (dev/postfit2-p2-brms.R in the frmtmb repository). Two different unseen members are two new levels here; brms 2.23.0 draws them once, an artifact of how it numbers unseen values. mm(g1, g2) beside (1 | g1) with g1 unset or unseen now draws both new levels, where it was refused. One new level is drawn once per draw and shared by every panel that names it. The core NEWS has the details.

• predictive_error(), pp_check() and bayes_R2() of draws no longer read a missing mi() response as an observation. The frame holds the placeholder 0 there, and the predictive error was 0 - yrep; it is NA now, as brms’s is, bayes_R2() is NA, as brms’s is, and pp_check() leaves the row out with brms’s warning. A pp_check() loo type on such a model is refused by name, since log_lik() is not defined there; it died inside bayesplot.

• frm_sample(laplace = TRUE) on a model with nothing to integrate out died in tmbstan with “invalid argument to unary operator”. It now says so and samples the model itself; the draws equal those of the call without laplace. On a REML fit sampled as it stands (prior = "flat") it is refused: that objective integrates the coefficients out too, and the draws carried b_Intercept over other parameters’ values.

• frm_sample() on a model with one outer parameter (y ~ 0 + x) died in rstan with “no more scalars to read”: a length-one init was read as a scalar. Each init is now passed as a one-dimensional array.

• A draws object with stanfit = NULL works with every method that reads the chain count (nchains(), summary(), VarCorr(), as_draws_*(), posterior_summary(), mcmc_plot() and more), as one chain. nuts_params() and log_posterior() refuse it by name.

New features

• pp_check()’s loo_pit_overlay, loo_pit_qq, loo_intervals and loo_ribbon types work. They failed on every model, because nothing built the PSIS weights bayesplot needs. They are built as brms builds them, loo::psis() on log_lik() of the same draws as the predictions, and match brms 2.23.0 bitwise on brms’s own draws. Where log_lik() refuses, these types refuse and say so. A type that bayesplot::available_ppc() does not list, such as the deprecated loo_pit, is refused with brms’s message.
• posterior_average(), brms’s model-averaged draws: draws from several frm_sample() objects in proportion to their weights, given as numbers or computed by "stacking" (the default), "pseudobma", "loo" or "waic". "kfold" and "bma" are refused, as kfold() and post_prob() are. On the same draws, weights and seed its output equals brms’s to 1.4e-17, with the same rows, names and attributes (dev/postfit2-brms-compare.R in the frmtmb repository).
• conditional_smooths() on draws: brms’s median, MAD and quantiles of each smooth term’s curves, and brms’s spaghetti. On ten draws that a fixed_param brms fit shares, every column and the spaghetti frame equal brms’s to 5.3e-15.
• conditional_effects() on draws takes spaghetti, surface, too_far and select_points, which it refused before. The spaghetti frame equals brms’s to 5.4e-15 on the same ten draws, for a one- and a two-variable effect. On laplace draws it draws its default curves instead of refusing every model with a group-level term.
• log_lik(), posterior_epred(), posterior_linpred(), posterior_predict() and predictive_error() on a multivariate model whose responses use subset() need one resp =, as brms’s do, and on newdata keep the rows where the response’s subset is TRUE. log_lik(resp = ) returns one column per row of that response.
• posterior_predict(newdata = ) reads the exposure of a rate() response from newdata.
• nobs() passes brms’s resp = to the fit’s method, which counts that response’s rows.
• The draws of a model with an equated parameter, bf(y ~ x, sigma1 = "sigma2"), carry a sigma1 column that is a copy of sigma2, as brms’s draws do, and the default priors skip the equated parameter, which owns no coefficient.
• as.array(), as.matrix() and as.data.frame() of draws take brms’s chain and iteration, which brms hands to posterior::subset_draws(): as.array(ds, chain = 1) is the first chain. They were refused by name. Any other name is still refused.
• conditional_effects(categorical = FALSE) on a hurdle_cumulative() fit scores the expected category by the category codes, so the hurdle scores 0, as core’s does. Every other ordinal family is unchanged.
• The default-prior disclosure names kappa, which brms gives a gamma prior, beside shape, phi and nu: xbeta() and von_mises() carry it.

frmtmb.sample 0.13.0

Needs frmtmb 0.65.0, for arma_cond_resp() and arma_cond_dpars(). With an older frmtmb the package installs and loads, and then log_lik(), loo() and posterior_predict() stop at their first call with “could not find function”.

• log_lik() and loo() now work on brms’s default cov = FALSE form of ar(), ma() and arma(), which was refused before. A row’s log-density is taken at brms’s one-step conditional mean, the regression on the group’s earlier OBSERVED residuals, so a column is an observation and the columns multiply to the whole likelihood. It matches brms::log_lik() cell for cell at the same parameter values (gaussian ar(1) bitwise; student arma(1,1) and arma(1,1) + (1 | g) to 4.4e-15 absolute), and loo(), waic() and psis() follow. loo()’s help says what that makes the estimate: PSIS-LOO over these columns leaves out one CONDITIONAL density, the one-step-ahead density of a row given its observed past, and not the series. The cov = TRUE covariance form, cosy() and unstr() stay refused, because there the smallest independent unit is a whole group; the refusal now names the form that does factor.

• A draws object with no stanfit behind it no longer errors in log_lik(): x$stanfit@sim is an error and not a NULL on such an object, so the chain identifier could not be read at all.

frmtmb.sample 0.12.0

Needs frmtmb 0.64.0, for rescor_row_loglik() and fit_extras(resp =).

• log_lik() on a Student-t rescor model used the gaussian joint density. It now uses the multivariate t through frmtmb’s rescor_row_loglik(); on the test fixture the old value was 38.6 log units off. On a multivariate model with an ordinal response, each response reads its own thresholds.

• Draws of the models frmtmb 0.64.0 adds: thres(gr = ) thresholds per group, gr(g, by = f) blocks merged in ranef(), coef() and VarCorr() as brms merges them, and posterior_predict() for brms’s cov = FALSE ARMA drawn around each row’s one-step mean.

• An intercept written 0 + Intercept, or under bf(center = FALSE), gets no default prior, as brms leaves class "b" flat.

• The flat-prior advice no longer reads class = "Intercept", = "" for a univariate intercept.

frmtmb.sample 0.11.0

• The pp_check(type = "error_binned") refusal on draws of a multinomial fit calls the response a set of counts over categories, not a category, as the fit method now does.

• Two test files no longer leave Rplots.pdf in the tests directory: test-draws-methods.R, whose device guard covered one block and not the pairs() call, and the generated test-brms-suite-methods.R, whose generator now opens a null device for the file.

• posterior_predict() on draws now draws a cs() term. On an sratio(), cratio() or acat() model with a category-specific effect it drew every row as if the term were absent, in sample and at newdata, with no warning. predict() and pp_check() on draws go through it and change with it. posterior_epred() and log_lik() were right and are unchanged. Needs the frmtmb that exports cs_offsets_add().

• On draws, a newdata with a grouping level the fit did not see is refused by the draws methods themselves, with or without allow_new_levels, and the refusal says that allow_new_levels = TRUE is refused too. Without the flag, core’s message reached the caller and recommended “Use allow_new_levels = TRUE”, which the draws methods then refused. The same holds for sample_new_levels alone and for a newdata without the grouping column.

• re_formula = NA with allow_new_levels = TRUE and an unseen level now answers, as brms does, instead of being refused: the check asked about the level without the re_formula that drops it.

• The default priors of a categorical, multinomial, mixture or latent-class model carry the dpar of each location (mub, mu1), as brms writes them. They used to be written without it, which frmtmb now refuses, so this is what keeps those models sampling. The resolved defaults are the same parameters with the same densities. With no prior of your own, a categorical or mixture fit samples exactly as before. With a dpar prior of your own on a mixture, the default for that slot now steps aside instead of being written and then overridden, which changes the order the priors are applied in: the draws at a fixed seed differ (by up to 8.74 on one measured fit), and prior_summary() no longer lists the overridden defaults. The resp-, dpar- and nlpar-qualified defaults of the other families with several location dpars (lba(), rdm(), hmm(), lca(), mixture_mvn()) follow the same rule.

• The default-prior announcement names each slot with its resp, dpar and nlpar, so every line is a spelling set_prior() takes. On a multivariate model it used to print Intercept with no response.

• hypothesis()’s advice for a coefficient with no proper prior gives a full set_prior() call, with resp, dpar, nlpar and coef as the model needs them. It used to say set_prior(class = "b"), which a multivariate model refuses.

frmtmb.sample 0.10.0

• The default priors are brms’s on a multivariate model: each response gets its own Intercept, sigma and sd defaults, read off that response, and set_rescor(TRUE) gets lkj(1) on the residual correlation. Until now a multivariate model was sampled with no default priors at all. This moves the draws of every multivariate model sampled without prior = "flat".

• The default intercept location takes off the mean offset(), as brms does: y ~ 1 + offset(off) with y near 2 and off = 10 gets student_t(3, -8, 2.5), where it got student_t(3, 2, 2.5). A location is now written to 15 significant digits, as brms writes it. This moves the draws of a model with an offset in its location formula.

• The sd default is written once per response, distributional parameter and nonlinear parameter, to follow frmtmb’s brms scoping of class "sd". With no prior =, every standard deviation gets the density it got before. With prior = set_prior(..., class = "sd"), that prior now replaces the location parameter’s default only, and a phi or sigma block keeps its own default, as in brms. A SMOOTH’s smoothing standard deviation is keyed the same way, so sigma ~ s(z) gets its own default and its own sd row. This moves the draws of a distributional model sampled with a class-wide sd prior.

• BREAKING: predictive_error() and residuals() on draws refuse an ordinal, categorical or multinomial model, as brms refuses them.

• pp_check() on draws refuses type = "error_binned" for a category response in the same words the fit method uses, and ignores resp on a model with one response, which is what brms does with it. A multivariate model still selects with resp and still refuses a name none of its responses has.

• A one-sided re_formula on draws keeps the terms it names, in posterior_epred(), posterior_linpred(), posterior_predict(), fitted() and predict(). Each draw is evaluated through frmtmb::frm_linpred(), which kept every term for any formula; the fix is in frmtmb’s development version, so it needs that frmtmb. On a (1 + x | g) + (1 | h) fit, re_formula = ~ (1 | g) used to be off the draw’s own intercept-only construction by up to 2.98; it now equals it. A term the fit does not have is an error that names it. The draws already carry each draw’s sampled group effects, so the uncertainty at a known level was in them before and nothing is added.

frmtmb.sample 0.9.0

• log_lik() is frmtmb’s generic now, re-exported here, rather than a second generic of the same name defined in this package. frmtmb owns the name because it has a method for it: log_lik(fit) on a maximum-likelihood fit refuses and names the route to draws, as loo() and waic() do, where it used to be “could not find function”. log_lik(draws) is unchanged, and the frmtmb_draws method is still registered on rstantools’ generic as well. The method page is ?log_lik.frmtmb_draws; ?frmtmb::log_lik documents the generic.

• posterior_epred() and posterior_predict() on draws refuse every name in ... they do not read, where they used to ignore it, as the other methods here already do. The new-level arguments pass through to the refusal described below.

• Requires frmtmb 0.61.0, for log_lik(), frm_linpred(), fixef_by_dpar(), brms_fixef_rows() and the brms shape helpers.

• fitted(), predict() and residuals() on draws exist (item 2.6f). They are brms’s three summarizing methods, each a summary of the draws method beside it: posterior_epred(), posterior_predict() and predictive_error(). Before this there was no method at all, so fitted(ds) reached stats::fitted.default(), read ds$fitted.values and returned NULL, a wrong answer with nothing said. summary = FALSE gives the draws, robust and probs are brms’s, and predict() on an ordinal or categorical response returns brms’s P(Y = k) proportions.

• BREAKING: nsamples(), posterior_samples() and parnames() answer instead of refusing. All three were refused as “the deprecated brms spelling”. brms keeps all three live with a deprecation warning, and the refusal stopped a ported script dead where brms’s warning does not. Each now does what brms’s does and warns as brms warns:

  • posterior_samples(x) is as.data.frame(x), with brms’s pars (a regular expression unless fixed = TRUE), add_chain, subset, as.matrix and as.array;
  • nsamples(x) is ndraws(x), with brms’s subset;
  • parnames(x) is variables(x).

  nsamples(incl_warmup = TRUE) is still refused: frm_sample() discards the warmup rather than storing it, so there is nothing to count.

• fixef() on an ordinal draws object reports the thresholds, Intercept[1], Intercept[2], then the coefficients and any cs() terms, in brms’s order and on the model’s own scale, which is what the fit’s fixef() reports. It reported the slopes alone, because the sampler stores the thresholds as tau_raw.

• posterior_epred(), posterior_linpred(), posterior_predict(), fitted(), predict() and residuals() on draws refuse allow_new_levels = TRUE when newdata holds a grouping level the fit did not see, with a message that names the function called and says why. That case used to fail in the design builder with “Use allow_new_levels = TRUE”, which the caller had just done. Predicting an unseen level from draws is not implemented; predict(fit, allow_new_levels = TRUE) on the maximum-likelihood fit does it. Every other use answers as before: allow_new_levels = FALSE, brms’s default, and TRUE with no newdata or with levels the fit saw.

• BREAKING: hypothesis() on draws no longer returns an object of class brmshypothesis. The class is "frmtmb_hypothesis" alone; the shape and every element are unchanged, and what stops working is is(x, "brmshypothesis") and inherits(x, "brmshypothesis") in a ported script. frmtmb owns print() and plot() for its own class, so neither changes.

• point_estimate and ndraws_point_estimate are honored by posterior_epred(), posterior_predict(), posterior_linpred() and log_lik(). They were accepted and ignored, so posterior_epred(ds, point_estimate = "median", ndraws_point_estimate = 2) returned all 25 draws where brms returns

  2. As in brms it is a PARAMETER-space operation: the draws are collapsed to their mean or median first and the method runs once at that one vector, repeated ndraws_point_estimate times.

• posterior_epred(), posterior_linpred() and posterior_predict() build each draw’s prediction with frmtmb::frm_linpred() rather than predict(), which is brms’s predictive summary in frmtmb 0.61.0.

frmtmb.sample 0.8.0

• Requires frmtmb 0.60.0, for frm_stop() and the other condition helpers.

• tmbstan (>= 1.2.1) in Suggests, and check_tmbstan_build() now names that version as the remedy. The CI step that pinned a dated snapshot of the Stan trio is removed.

• Requires the frmtmb release that exports frm_stop(); the frmtmb (>= 0.59.0) floor must move to it.

• BREAKING: every error, warning and message that frmtmb.sample raises is classed. An error has the class c("frmtmb_sample_error", "frmtmb_error", "error", "condition"), and warnings and messages follow the same pattern, so tryCatch(frmtmb_error = ) catches any refusal. The class vector no longer contains simpleError, simpleWarning or simpleMessage. See ?frmtmb::frmtmb-conditions.

• BREAKING: a value that matches none of the choices of loo_compare(criterion =), hypothesis(scope =), pp_check(prefix =) and predictive_error(method =) is refused with a frmtmb_sample_error that names the argument, the value and the choices. The old text was 'arg' should be one of .... A partial value still matches.

• posterior_interval(), as.matrix(), rhat(), neff_ratio() and the other accessors refuse an exact variable name that the draws do not carry as a frmtmb_sample_error. The posterior package used to raise it as a plain error. The text keeps posterior’s words, missing in the draws object.

frmtmb.sample 0.7.0

• Requires frmtmb 0.59.0, for brms_par_labels(), brms_coef_table(), brms_stan_name(), expand_b(), varcorr_layout(), the hypothesis() helpers (hyp_eval_in(), hyp_expr_vars()) and check_prior_slots() on ?frmtmb::frmtmb-sampling-api.

• BREAKING: the draws carry brms’s names. b_Intercept, b_x, b_sigma_Intercept for the coefficients and r_g[1,Intercept] for a group-level coefficient (r_gs[lev.1,Intercept] for a level with a space, r_g:h[1_p,Intercept] for an interaction group), in ds$draws, variables(), every as_draws_*(), as.matrix(), as.array(), as.data.frame(), rhat(), neff_ratio(), summary() and posterior_summary(). They were Intercept, x, sigma_Intercept and b[1]. A covariance parameter brms does not sample keeps the name confint() gives it on the fit, theta_1; a block with no brms counterpart (reduced rank, smooth, GP, CAR, SPDE) keeps b[i]. What stops working: ds$draws[, "x"], and anything that selected b[ columns.

• BREAKING: a distributional parameter nobody wrote a formula for is stored on its natural scale, under brms’s name. The column is sigma (or shape, nu, and sigma_ya for response y_a of a multivariate model) and holds sigma, as brms’s does. It used to be b_sigma_Intercept holding log sigma, which brms reports only when sigma ~ 1 is written out, so fixef(ds) had a row, coef(ds)$g a slice, and posterior_summary(ds) and as_draws_df(ds) a column that brms does not have. They are gone. What stops working: ds$draws[, "b_sigma_Intercept"] on such a model; use sigma, or log(ds$draws[, "sigma"]) for the old values.

  A mixture’s weights are stored as brms’s simplex: theta1 ... theta<K-1> hold the mixing probabilities, and brms’s thetaK is a derived column before lp__. They used to be the log ratio under the name theta1_Intercept. Every reader that hands a draw to the model maps all K back to the log ratios together.

• Draws labels a level repeats are suffixed as brms suffixes them. Levels lvl 1 and lvl.1 are both r_gd[lvl.1,Intercept] in brms, which names the later one r_gd[lvl.1,Intercept]__1. Without the suffix every accessor failed on “Duplicate variable names”. hypothesis() on draws refuses a name that matches two columns instead of reading the first.

• BREAKING: the accessors return brms’s objects. as.matrix(), as.array() and as.data.frame() return brms’s unclassed draws objects and take brms’s pars, variable, draw and subset, with brms’s deprecation warnings; as_draws_*() take variable, regex and inc_warmup; as_draws() is brms’s as_draws_list(). posterior_summary(x, pars, variable, probs, robust) and posterior_interval() cover every variable by default, as brms’s do. Compared against brms’s own installed methods on the same draws: 88 of 90 calls identical(), the other 2 rvars objects that differ only by a cache environment (dev/brmsnames-findings.md).

• BREAKING: fixef(), ranef(), coef() and VarCorr() are brms’s. summary = FALSE returns the raw draws, as in brms, where it used to return a summary; robust, probs, pars and groups are brms’s. coef() broadcasts every population-level coefficient, as brms’s does, and VarCorr() has brms’s sd/cor/cov/residual__ structure, with one residual__ row per response on a multivariate model. ranef() on a model with no group-level effects is refused with brms’s message. ranef() and coef() on a reduced-rank block, whose draws are factor scores rather than coefficients, compute the coefficients per draw; on frm_sample(laplace = TRUE) draws, which carry no group-level draws, they are refused by name rather than returned as NA.

• BREAKING: hypothesis() returns brms’s brmshypothesis object, takes class = "b" by default and brms’s argument order, supports robust, seed and scope = "ranef" or "coef", and reports Evid.Ratio and Post.Prob in brms’s columns. evid_ratio, post_prob, z, p and attr(h, "draws") are gone; the draws are h$samples. The string is read with brms’s renaming, so x:fe is the interaction coefficient and not R’s : operator, and a stored column such as r_g[1,Intercept] can be named with class = NULL, as in brms.

• BREAKING: bayes_R2() on a multivariate model returns one row per response, R2ya and R2y2, as brms does. It returned the first response alone, named R2y_a. resp takes brms’s spelling of a response (ya).

• BREAKING: brms’s positional slots on six more methods. bayes_R2(ds, NULL, TRUE, TRUE) now asks for brms’s robust summary; the fourth slot used to be probs and returned a Q100 column. posterior_summary()’s second slot is pars, hypothesis()’s third is class, pairs()’s second is pars, conditional_effects()’s third is conditions and pp_check()’s fourth is prefix. The formals audit of dev/brmsmatch-beyond.R now finds 0 of 68 methods diverging, from 6.

• BREAKING: summary() has brms’s columns and slots: Estimate, Est.Error, l-95% CI, u-95% CI, Rhat, Bulk_ESS, Tail_ESS, with prob, robust and mc_se.

• BREAKING: conditional_effects() defaults to brms’s robust = TRUE: estimate__ is the median of the drawn curves and se__ their MAD. robust = FALSE gives the mean and SD it used to.

• pp_check() takes brms’s prefix, group, x, newdata, resp and draw_ids, and brms’s default draw count with brms’s message. log_lik() takes brms’s pointwise, combine, add_point_estimate and cores, and refuses an argument it does not have, which it used to accept and ignore.

frmtmb.sample 0.6.0

• BREAKING, and a published number moves. rhat() and neff_ratio() now answer brms’s question. They called bayesplot::rhat() and bayesplot::neff_ratio() on the stanfit, which is rstan::summary()’s classic split-R-hat and rstan’s n_eff. brms computes neither. rhat.brmsfit is posterior::summarise_draws(rhat = posterior::rhat), the rank-normalized split-R-hat, and neff_ratio.brmsfit is min(ess_bulk, ess_tail) / ndraws. Both now do the same, computed on the draws array rather than on the stanfit.

  Measured on 4 chains of 500 (dev/brmsmatch-findings.md): the old rhat() differed from brms’s by 0.00563108 at most, against a whole signal of max|rhat - 1| = 0.0100318, so by 1.15 times the entire excess over 1 that the diagnostic reports. neff_ratio() differed by 0.43165, which is 77 times larger and is the number a user reads to decide whether to sample longer.

  These two no longer agree with ds$stanfit, which carries a different definition of the same idea. rstan::summary(ds$stanfit)$summary still has the sampler’s own numbers for anyone who wants them.

• BREAKING. summary()’s convergence columns move with them. It reported rstan’s n_eff and classic split-R-hat; it now reports Rhat, Bulk_ESS and Tail_ESS from posterior, which is what brms:::summary.brmsfit() reports, under brms’s own column names.

  The n_eff column is gone. summary(ds)[, "Rhat"] is now exactly rhat(ds) on those rows and neff_ratio(ds) is exactly pmin(Bulk_ESS, Tail_ESS) / ndraws(ds), so the package no longer disagrees with itself. It did: on the fit measured in dev/reviews/20260915-brmsmatch.md, sigma_Intercept read 0.99990396 in the Rhat column against 1.0037623 from rhat(ds), BELOW 1 in one place and above it in the other, and x read 1600.37 effective draws against 1028.60, a 55.6 percent overstatement of the number a user reads to decide whether to sample longer.

• rhat() and neff_ratio() now report frmtmb’s parameter names. They reported Stan’s, because they read the stanfit: 4 of 11 entries on the measured fit (beta[1], beta[2], betad, theta) were not addressable by the names variables(ds) lists, so rhat(ds)["x"] was NA. It is now a number. This falls out of the change above rather than being separate work.

• Both take brms’s pars, and it is brms’s pars FOR THESE TWO, which is not the pars that as.mcmc(), mcmc_plot() and posterior_interval() take. brms:::rhat.brmsfit() passes variable = pars to as_draws_array() rather than through extract_pars(), so NULL is the default and means every variable, a string is an EXACT variable name, regex = TRUE makes it a regular expression, and a name that is not there is an error. The other three keep the extract_pars rule, where NA means every variable and a string is a regular expression. ?draws-diagnostics documents both under Two different pars rules, both brms’s.

• BREAKING. Ten methods take brms’s arguments in brms’s POSITIONS. A positional call ported from brms used to answer a different question, twice without saying so. Every method below now matches brms’s method name for name, as far as this package’s own arguments go, and tests/testthat/test-draws-spellings.R asserts that against the installed brms.

  • as.mcmc(x, pars, fixed, combine_chains, inc_warmup, ...), was (x, combine_chains, ...). as.mcmc(ds, TRUE) used to return the pooled draws; brms reads that slot as pars and refuses a pars that is neither NA nor character, and so does this now.
  • posterior_interval(object, pars, variable, prob, regex, fixed, ...), was (object, prob, variable, ...). posterior_interval(ds, 0.9) used to return a 90% interval; it is now the same refusal brms gives.
  • log_lik(object, newdata, re_formula, resp, ndraws, draw_ids, ...), was (object, ndraws, resp, ...).
  • mcmc_plot(object, pars, type, variable, regex, fixed, ...), was (object, type, variable, ...).
  • posterior_epred(object, newdata, re_formula, re.form, resp, dpar, nlpar, ndraws, draw_ids, ...), was (object, newdata, resp, re_formula, re.form, ndraws, ...).
  • posterior_linpred(object, transform, newdata, re_formula, re.form, resp, dpar, nlpar, incl_thres, ndraws, draw_ids, ...).
  • posterior_predict(object, newdata, re_formula, re.form, transform, resp, negative_rt, ndraws, draw_ids, ...).
  • pp_mixture(x, newdata, re_formula, resp, ndraws, draw_ids, log, summary, robust, probs, ...), was (x, summary, ndraws, ...).
  • predictive_error(object, newdata, re_formula, re.form, method, resp, ndraws, draw_ids, ...), was (object, resp, re_formula, re.form, ndraws, ...).
  • psis(log_ratios, newdata, resp, model_name, ndraws, ...), was (log_ratios, ndraws, resp, ...).

  A call that named its arguments is unaffected. A call that passed them positionally must be re-read.

• New along the way, because brms puts them in those positions: draw_ids on the predictive methods, on log_lik() and on pp_mixture() names the draws to use by row index; dpar on posterior_epred(); log, robust and probs on pp_mixture(); method on predictive_error(), which also takes newdata now and re-evaluates the response term on it.

• Eight brms arguments are carried in brms’s position and refused, each with the reason and the replacement: log_lik(newdata =), log_lik(re_formula =), psis(newdata =), pp_mixture(newdata =), pp_mixture(re_formula =), as.mcmc(inc_warmup = TRUE), posterior_linpred(incl_thres = TRUE) and posterior_predict(negative_rt = TRUE). re_formula = NULL is brms’s own default and is refused nowhere: it is the no-op these methods already do.

• Two spellings of the random-effect switch survive on the five methods where brms itself ACCEPTS both, which is not the same as the four that declare both. posterior_epred(), posterior_linpred(), posterior_predict() and predictive_error() declare re_formula and lme4’s re.form on brmsfit. predictive_interval() declares neither and honors both anyway, because its whole brms body is posterior_predict(object, ...), so the alias reaches a formal one frame down. tests/testthat/test-draws-spellings.R derives “brms accepts it” from brms’s own code rather than restating a list.

• BREAKING, one method. pp_check() no longer takes re.form. Note that this is the one place the package is knowingly narrower than brms: pp_check.brmsfit() forwards its dots to posterior_predict() through do_call(), so brms accepts the alias there too. dev/argspell-brms-accepts.R in the monorepo has the measurement, and the decision is open rather than settled.

• BREAKING. An argument that lands in a method’s ... is an error naming it, rather than being swallowed. Every method of this package that took a ... it never read now refuses one; dev/argspell-report.R in the monorepo reports 0 still swallowing. See the frmtmb NEWS entry for why the refusal is the part that prevents the class.

frmtmb.sample 0.5.0

• frmtmb.sample no longer breaks brms, rstantools, loo, bridgesampling, bayesplot, posterior, coda or gratia. This package defined its own generic for 28 names those packages own, such as log_lik(), posterior_epred(), psis(), rhat() and as.mcmc(). UseMethod() reads the method table of the namespace where the generic it reached was defined, so after library(brms); library(frmtmb.sample), log_lik(fit) on a brmsfit found no method. Measured on the previous release: all 28 lost after brms, all 28 with brms only loaded, all 28 inside a package that imports frmtmb.sample, and every gratia method for posterior_samples() after gratia. It is now 0 in each of those orders. The package now calls frmtmb::frm_install_generics() from its .onLoad(), so each name resolves to the owner’s generic while the owner is loaded and to this package’s own while it is not.

• Two methods reached a foreign .default in silence and now reach their own. With posterior loaded after this package, rhat() on draws went to posterior’s rhat.default; with gratia loaded after it, posterior_samples() went to gratia’s. Both are now also registered on those owners’ generics.

• BREAKING. Three generics take their owner’s first argument, so a call that NAMED the old one fails:

  • bridge_sampler(samples, ...), was (x, ...), as bridgesampling.
  • bayes_factor(x1, x2, log = FALSE, ...), was (x, ...), as bridgesampling.
  • rhat(x, ...), was (object, ...). posterior and bayesplot both define rhat with different first arguments; this follows posterior, which is the one brms uses.

  Two only GAIN arguments and break nothing: post_prob() gains prior_prob and model_names from bridgesampling, and posterior_samples() gains pars from brms. posterior_linpred()’s generic gains transform, which its method already had.

  If you write an S3 method on any of these names, match the owner’s formals.

• While brms is loaded, parnames() and posterior_samples() on draws give brms’s own deprecation warning before this package’s refusal, because the generic that runs is brms’s.

• gratia (>= 0.9.0) joins Suggests, the first gratia with posterior_samples(), and posterior gains the floor (>= 1.0.0), whose rhat() this package now registers a method on. As in frmtmb, a partial install of either on the library path that lacks the generic, loaded before this package, stops it loading.

• The frmtmb floor moves to the release that exports frm_install_generics().

frmtmb.sample 0.4.2

• Requires frmtmb 0.56.0, and the requirement is a hard one. hypothesis() on a frmtmb_draws object now arms the note about a covariate that shadows a distributional parameter itself, through frmtmb::hyp_shadow_arm() and hyp_shadow_disarm(), which no earlier frmtmb exports. Before this release the draws method relied on core’s generic to arm that note; once frmtmb 0.56.0 stopped owning the hypothesis generic, the note fired on a fitted model and never on draws, in every session and whether or not brms was loaded.

• The generics this package re-exports from frmtmb no longer break brms’s methods when brms is loaded. With brms loaded but not attached, 26 of 26 re-exports lost their brms method in 0.4.1 and 0 of 26 do now. That repair comes from frmtmb 0.56.0 and reaches this package for free.

• Three method signatures now carry their generics’ full formals, because an S3 method must accept every argument its generic declares and the generics are now the owners’ rather than frmtmb’s.

• This package still DEFINES 28 generics of its own that break brms’s methods in the same way frmtmb’s did. That is not fixed here.

frmtmb.sample 0.4.1

• A broken tmbstan now skips the sampler tests instead of erroring them, and skip_sampler() asks the package’s own detector rather than only whether tmbstan is installed. 111 test blocks reach a sampler; 47 of them had been using a file-local copy of skip_sampler(), so searching for the helper’s name looked like coverage. All 111 go through it now.

• The \donttest examples carried the same weak guard. Fifteen topics called a sampler behind requireNamespace(), which a broken build satisfies. An example cannot skip, so it must not run.

• A check that does not depend on the marker string. On an affected build rstan::grad_log_prob() disagrees with -fit$obj$gr(), since the patched log-density overload is require_not_st_var and the missed one is require_st_var. With infinite bounds the two are the same number bitwise, 0 ulp over four model shapes and five points, so this catches an upstream that renames the placeholder, which the marker cannot.

• The check workflow pins the Stan trio to a dated snapshot and fails if model.hpp is still unpatched, reporting the state it measured rather than a boolean. Public RSPM rebuilt tmbstan the day after StanHeaders 2.39.1 landed, and four consecutive green runs of this package’s checks sampled a standard normal instead of the model.

frmtmb.sample 0.4.0

BREAKING. frm_sample(control =) is the SAMPLER’s control list now, brms’s spelling and brms’s meaning, passed to tmbstan unchanged. The fit-time options move to fit_control =. as_tmbstan() already meant control that way, so the two entry points agree rather than offering a third convention, and brms code that tightens adapt_delta ports across unchanged.

• That closes a silent discard. On the fitted-object route control = list(adapt_delta = 0.99) was IGNORED: stan_args[[1]]$control came back NULL where 0.97 was asked for, and the mean accept_stat__ was 0.9266 against 0.9836 once the setting arrives. fit_control, start, data2, na.action and REML were silently ignored on that route too, and are refused by name. An abbreviation is refused as well: contro = fell into ..., reached rstan::sampling() and partial matched onto ITS control, bypassing every check.

• as_tmbstan() no longer returns an empty stanfit without saying so. That was worse than a missing refusal: it bricked the session, and every later model failed with empty_nested() must be true. Both doors share one check, and a run that produced nothing is worded apart from a model that cannot be sampled at all, because those want different fixes.

• frm_sample() reads the compatibility registry before taping and refuses a fit the sampler cannot run, in about half a second with rstan never loaded, against the roughly three seconds the failure used to take. Where the formula position settles it the match is made on position; where it cannot, the match is made on names, and ?frm_sample says that is the largest set justifiable that way rather than a sound one.

• frm_sample(control =) is the SAMPLER’s control list now, which is what brms means by the name. frm_sample(fit, control = list(adapt_delta = 0.99, max_treedepth = 12)) ports straight from a brm() call and reaches rstan through tmbstan. The fit-time options frm() takes are fit_control =, and they still apply only on the formula route. This is a breaking rename, and the old shape is refused rather than reinterpreted: a frmtmb_control() list arriving in control is caught by the field names it carries and the refusal names fit_control. The two vocabularies share no name, which is what makes the two lists decidable, and a test asserts they stay disjoint. An option rstan does not have is also refused by name, because rstan answers an unknown one by declining to sample and returning an empty fit, which used to be reported as a solver failure. So is an ABBREVIATION of the name: control follows ... in the formals, so only the exact spelling binds to it, and a shorter one would land in ... and then partial-match rstan’s own control, reaching the sampler unchecked. vignette("brms-posterior") carried a row saying adapt_delta had no route through this function; it now says the opposite.

• frm_sample() refuses an assembly argument on a fitted object instead of accepting and discarding it. fit_control, start, data2, na.action and REML build a model, and a fitted object is already built, so on that route they were read by nothing and said nothing: frm_sample(fit, REML = TRUE) sampled the model it had. data and family were already refused there; these five now are too, and the message points at control for the caller who meant the sampler. fit_control matters most, because it is the name the control refusal sends a confused user to.

• Breaking, and a return-contract change: as_tmbstan() never hands back an empty stanfit now. rstan answers some failures by declining to sample, printing its reason to the console and returning the shell of a stanfit with no error at all. frm_sample() has always refused that; as_tmbstan() passed it on, so a failed run looked like a successful one until an accessor came back empty. Both now raise the same message, naming the function that raised it. The wording distinguishes the two ways to get nothing out of this package, because they want different fixes: a model that cannot be sampled at all is refused by the compatibility pre-flight before any chain starts, and this one is a failure of the run in front of you. Nothing in the package or its vignettes relied on the empty object.

• Fixed: a sampler control list handed to a fitted object was discarded without a word. Up to and including 0.3.2, frm_sample(fit, control = list(adapt_delta = 0.99)) on the fit route ran the chains with rstan’s own defaults and said nothing: control was read only where the formula route assembles a model, and the fit route does not go there. A brms user porting a call that raised adapt_delta therefore got the untightened sampler. The argument now reaches rstan on both routes, and ds$stanfit@stan_args[[1]]$control records what was asked for.

• A pre-flight against the compatibility registry, on frm_sample() and on as_tmbstan(). Both read the rows that name frm_sample with the status "refused" before anything is taped, before the formula route calls frm(), and before the rstan and tmbstan namespaces are loaded, so a model the sampler cannot run costs the registry read and nothing else: 0.36 s to 0.42 s on the first call of a session against the 2.9 s failure it replaces, and that failure also left rstan unusable for the rest of the session. The refusal repeats the registered note and names the row. frmtmb.ode registers the first such row.

  What the check refuses on is deliberately narrow, because what identifies a feature in a model is WHERE its call sits in the formula. It decides three ways. A response FAMILY is matched by name, which is exact; it does not look inside mixture(). An ADDITION TERM is matched by POSITION, since it is written inside the bar on the left of the response formula and nowhere else, which separates y | se(v) ~ x from y ~ a * se(x) with a user’s own se(). Everything else the vocabulary writes as a call is matched by NAME, outside addition-term position, and only when the name is not also a base R function; that route is justifiable from names but not sound, so a refusal registered on mo(), ma(), cosy() or mm() would also fire on a user’s function of that name. Everything the check cannot decide warns and samples rather than refusing a model that may be correct. A registered refusal cannot be revoked by a later registration: display names are unique, core refuses a second registration of one, and a name the vocabulary writes without parentheses is declared not to be a call at all.

frmtmb.sample 0.3.2

Requires frmtmb 0.55.0, for the hazard-container lint that now runs in this package’s own check.

frmtmb.sample 0.3.1

• The sampling articles say why they are unevaluated, and the reason they used to give was false: four chains on the model in question take seconds, not the minutes claimed. They are unevaluated because the samplers are optional dependencies and because that page documents a Stan build which samples the wrong density silently. The two chunks that need no sampler now run, so the page shows output.

• loo() on a group-unit matrix says which unit it left out. loo::loo.matrix() never sees the attribute the log-likelihood matrix carries, so the printed estimate was indistinguishable from a per-observation one.

frmtmb.sample 0.3.0

conditional_effects() on draws returns core’s frame, grid and display quantity; hypothesis() reports Savage-Dickey evidence ratios validated against brms; loo() says which unit it leaves out. Requires frmtmb 0.53.0 for the seam exports.

• conditional_effects() on a draws object now returns the frame core’s fit method returns, which closes the divergence 0.2.0 recorded. The columns and their order are core’s ce_frame(): the varied predictor(s), then every other model variable at the value it is held at, then cond__, cats__, effect1__ and effect2__, then the band. cond__ is always present as a factor over every condition level, so brms’s own plot() has something to facet on; a per-category ordinal display is keyed "x:cats__" rather than "x", so brms’s plot() reads the pair back out of the effects attribute; and int_conditions =, categorical = and seed = are formals here as they are on the fit method. A ported script that indexed the returned frame positionally, or that read ce$x on an ordinal fit, has to be read again.

• BUG FIX. conditional_effects() on draws drew the mu predictor where the fit method draws the EXPECTED RESPONSE, on every family whose mean is not the inverse link of mu and on every trunc() response. A zero-inflated fit was plotted 1 / (1 - zi) times too high, a mixture showed one component’s mean instead of the mixture’s, and a response truncated below at zero reported a NEGATIVE mean and drew it. The column names, the grid and the keys were all correct throughout, which is why the first round’s identity checks passed. The display quantity now comes from core’s own ce_pred_dpar(), so the two surfaces cannot make different choices again. Measured after the fix, largest deviation from the fit curve over a five-point grid: 1.4% zero-inflated, 1.0% truncated, 0.5% mixture, against 21.7% and a sign change before it. An explicitly named dpar = was always correct and is unchanged.

• conditional_effects() on draws accepts allow_new_levels and lme4’s allow.new.levels, which the fit method has always accepted and this method warned about as unknown. It reads them with core’s own ce_dots(), so the accepted set cannot drift again.

• hypothesis() refuses abs(x) = 0 instead of reporting a Bayes factor twice too large. The affine check probed only non-negative points, where abs() is the identity; it now probes both signs. The prior of |X| for X ~ N(0, 1) is a half-normal with twice the density at zero, so the denominator was half what it should have been, silently.

• The "evid_ratio_mcse" attribute is measured over at least four blocks of draws. With two chains it was the standard deviation of two numbers over sqrt(2) - one degree of freedom - and understated a measured 0.46 gap between two runs of one model as 0.18. Chains are split in half when there are fewer than four, which keeps each block consecutive and so keeps the autocorrelation the estimator exists to respect.

• A point hypothesis on a natural-scale summary says so. sigma = 0.9 used to refuse with “not a population-level coefficient”, which is true of the name and false of the model: sigma IS a coefficient here, reported on a scale its prior is not written on. That is now the reason given, for every such quantity.

• BEHAVIOR CHANGE. conditional_effects(re_formula = NULL) on draws conditions on a NEW group, as it does on a fit, instead of silently taking the first observed level. The grid’s grouping column reports NA and each POSTERIOR DRAW gets that group’s effects drawn from its own covariance parameters, which is brms’s construction; the fit method draws them per bootstrap replicate instead. The band is wider than the population band by the amount the new group’s variance adds (1.92 to 3.93 in mean width on the model in dev/sample-ce-findings.md). seed = makes the draw reproducible.

• conditional_effects() on draws reports an argument it cannot use instead of discarding it in silence.

• hypothesis() on a draws object reports evid_ratio and post_prob, brms’s Evid.Ratio and Post.Prob. A directional hypothesis gets the posterior odds of the claim, which needs no prior. A POINT hypothesis gets the Savage-Dickey density ratio: the posterior density of the tested quantity at the point, from a kernel density with stats::density()’s default nrd0 bandwidth over 4096 grid points read by spline, over the PRIOR density there, evaluated from the specification the model was sampled under rather than estimated from prior draws. brms estimates both halves by kernel density from sample_prior = "yes" draws; the two agree to the accuracy of that estimate, and the numbers are in dev/sample-ce-findings.md. The between-chain Monte Carlo error of each ratio rides on the returned object as the "evid_ratio_mcse" attribute, because a density ratio read without one is easy to over-believe.

• A point hypothesis whose parameter has no proper prior REFUSES BY NAME, with a warning that says which parameter and why, and leaves evid_ratio NA; the rest of the row is unaffected and the other hypotheses of the same call still report. frm_sample() leaves class "b" flat by default exactly as brms does, so this is the commonest answer until a prior is written. A hypothesis on a variance component, on a nonlinear function of the coefficients, or on the intercept (whose prior is about the intercept at the predictor means, not about the coefficient) refuses for its own stated reason.

• log_lik(), loo() and waic() accept a structured family that declares how its likelihood factorizes (frmtmb::frmtmb_structure(loglik_group = ) or (loglik_row = )), instead of refusing every family whose likelihood is not rowwise. The columns are the pieces the family declares, at the coarsest granularity it gives, because a family that groups its likelihood is saying its rows are not independently droppable. The matrix carries attr(x, "unit") naming what a column is when it is not an observation. A family that supplies one number for the whole response is still refused, in the words it already used.

frmtmb.sample 0.2.0

• BEHAVIOR CHANGE, following core. ranef() on a frmtmb_draws object is keyed by the GROUPING FACTOR rather than by the block, so ranef(draws)[["1 | g"]] returns NULL where it used to return an array and ranef(draws)$g returns it instead. The method delegates to core’s ranef() once per draw, so it follows core’s re-key exactly; the block label rides along in each array’s "term" attribute, as it does in core. A model with two blocks on ONE factor gives two entries under one name, and the method now assembles them BY POSITION: keyed by name, out[[tn]] <- st wrote the same name twice, so the SECOND block was dropped entirely and the list came back with one entry instead of two.

• conditional_effects() on a draws object inherits four grid changes from core, because it builds its grids with core’s ce_grids_build(): the two-variable grid now varies the FIRST effect slowest (brms’s row order), a numeric moderator is held at the exact mean +/- sd rather than at signif(mean +/- sd, 3), a mo() predictor gets one grid point per level instead of a 100-point continuous grid, and a trials() variable is held at 1 with a message that is new here. What it does NOT yet inherit is core’s frame shape: the returned columns, the always-present cond__, the "x:cats__" key for a per-category display, int_conditions =, categorical = and the new-group meaning of re_formula = NULL are core-only for now, so the two surfaces disagree in both directions until the draws method is migrated.

• frm_sample() refuses parallel chains on Windows when the model’s family comes from a namespace built by pkgload::load_all(), naming the package and the two remedies. A worker process cannot load such a namespace, and R silently substitutes the global environment for it, so the family failed at its first internal call and rstan reported only that the chains returned no draws.

• The defaults this package registers with frmtmb::get_prior() now answer route = "sample" only. Loading this package used to change the table get_prior() returned for every caller, an author asking about frm() included; it no longer does. Ask for get_prior(..., route = "sample") to see what frm_sample() applies, and with this package unloaded that route is refused rather than reported as flat. frm_sample() itself is unchanged: the same defaults under the same call-over-fit-over-defaults precedence.

• The frmtmb floor rises to 0.51.0. The registration in .onLoad() now passes frmtmb_register_compat(expects = ), which frmtmb gained in 0.51.0, and an older frmtmb stops the load outright with unused argument (expects = c("hmm", "lca")). The old floor of 0.46.0 permitted an installation that could not load, so this is a hard requirement rather than a preference.

• frmtmb.latent joins Suggests. hmm() and lca() left frmtmb for that package, so the tests here that sample one of them qualify the call and skip when it is absent. The two compatibility rules this package registers for those pairs are declared in frmtmb_register_compat(expects = ), which says they name a feature another package owns. Before frmtmb checked, those rules were dropped without a word in every session that did not also load frmtmb.latent; now frm_compat() reports them as unresolved until it is loaded.

frmtmb.sample 0.1.0

First release, extracted from frmtmb 0.46.0.

• NUTS sampling for frmtmb models through tmbstan: frm_sample() on a fit or a formula, with brms 2.23’s default priors on both routes, priors stacking most-explicit-first, and non-centered sampling of qualifying random-effect blocks. Draws carry frmtmb parameter names.
• The full posterior method surface on core’s generics and on the brms, loo, posterior, bayesplot, rstantools, coda and bridgesampling generics: log_lik(), loo(), waic(), posterior_epred(), posterior_predict(), hypothesis(), conditional_effects() on draws, bayes_R2(), mcmc_plot(), and the rest.
• check_laplace() measures the Laplace and Wald approximations against NUTS on the same objective, centered and with no added priors.
• At load, the package registers its sampling default priors with core’s get_prior() and its feature rows with the compatibility matrix, so both stay truthful whether or not this package is installed.
• Refuses a tmbstan built against StanHeaders 2.39 or later, whose code generator leaves one log-density overload unpatched so every chain silently samples a standard normal instead of the model.
• Parallel chains work on every platform, Windows included: tmbstan rebuilds the tape on each socket worker from the serialized objective closure, and frmtmb’s generated closures are self-contained.