Usable parameter names — variables • frmtmb

brms's names for the parameters of a fit, which are the names hypothesis() expressions accept: fixed-effect coefficients with brms's b_ prefix (b_Intercept, b_x, and b_sigma_Intercept for a coefficient of a sigma formula), natural-scale random-effect summaries (sd_<group>__<coef>, cor_<group>__<coef1>__<coef2>, with a distributional or nonlinear parameter's name in the coefficient part, sd_g__sigma_Intercept), and a distributional parameter nobody wrote a formula for, on its natural scale (sigma, shape, sigma_ya). Every name is spelled through brms's renaming: b_IxE2 for I(x^2), sd_g:h__Intercept for (1 | g:h).

Usage

variables(x, ...)

# S3 method for class 'frmtmb_fit'
variables(x, ...)

Arguments

x: A frmtmb_fit or frmtmb_draws.
...: Refused: an argument the method does not have is an error naming it, rather than silently changing nothing.

Value

A character vector.

Details

For sampled fits, variables() on the frmtmb.sample::frm_sample() result lists the draw columns. The coefficients, the natural-scale distributional parameters, an ordinal fit's thresholds (b_Intercept[1]) and the cs() coefficients (bcs_<column>[1]) have the names and the values this page gives them. So do the covariance summaries, sd_ and cor_ among them, which the draws carry as columns computed from each draw's covariance parameters; those are stored as the sampler sampled them, on the unconstrained scale and under their internal names (theta_1), beside them.

brms's variables() also lists what a fit has no counterpart of: group-level coefficients r_<group>[<level>,<coef>], the centered Intercept, lprior and lp__. A maximum-likelihood fit has no draws of those, and ranef() reports the conditional modes.

An ordinal fit's thresholds are b_Intercept[k], and a cs() category-specific term contributes bcs_<column>[k], one name per DESIGN COLUMN per category boundary. The column is the one stats::model.matrix() builds, as it is in brms, so cs(x) on a numeric predictor gives bcs_x[1], bcs_x[2], while cs(f) on a factor with levels a, b and c gives the treatment-contrast pairs bcs_fb[k] and bcs_fc[k]. See the "Category-specific effects, cs()" section of frm().

A residual correlation term (frmtmb-autocor) contributes its natural-scale parameters under brms's names: ar[1], ar[2], ma[1], cosy, cortime__<t1>__<t2>.

gr(cov = ), gr(prec = ) and equalto() blocks contribute sd_/cor_ names for their within-level covariance. A gp() term contributes brms's sdgp_<term> and lscale_<term>. Smooths, car() and spde() blocks contribute none: their parameters are not standard deviations. See the "Which random-effect blocks contribute names" section of hypothesis().

Examples

dd <- data.frame(x = rnorm(60), g = factor(rep(1:6, 10)))
dd$y <- rnorm(60, 1 + 0.5 * dd$x + rnorm(6, 0, 0.5)[dd$g], 1)
fit <- frm(bf(y ~ x + (1 | g)) + gaussian(), data = dd)
variables(fit)
#> [1] "b_Intercept"     "b_x"             "sd_g__Intercept" "sigma"