Returns a bvarnet_priors object containing a bvarnet_prior
for every model parameter type. Any argument left as NULL uses the
package default.
Available prior distributions are:
normal(loc, scale)
student_t(loc, scale, df)
cauchy(loc, scale) For standart deviations and random effects, the prior is automatically converted to a half-prior (truncated at
loc) in the Stan code, so the printed format reflects this.
Usage
set_priors(
intercept = NULL,
beta = NULL,
phi = NULL,
sd_u = NULL,
kappa = NULL,
sigma = NULL
)Arguments
- intercept
Prior for the intercept. Only applies to gaussian and bernoulli models; for ordinal models the intercept is absorbed into the kappa (threshold parameter).
- beta
Prior for fixed-effect regression coefficients (slopes).
- phi
Prior for lag coefficients.
- sd_u
Prior for random-effect standard deviations (half-prior).
- kappa
Prior for the adjacent-category thresholds (ordinal models only).
- sigma
Prior for residual standard deviation (gaussian models only; half-prior).
Automatic scaling of Gaussian defaults
Gaussian outcomes are modelled on their raw scale, so bvar() widens the
default intercept, beta and sigma scales by the outcome
SD before passing them to Stan. A default beta ~ Normal(0, 1) on data
with sd(y) = 17.8 therefore becomes Normal(0, 17.8). This keeps
the unit-scale defaults weakly informative whatever the units of y,
but it means the default you see here is not the prior that was used.
A prior you pass explicitly is taken at face value and never rescaled. The
priors actually used are recorded on the fitted object as
priors_effective (see bvar), are reported by
print(), and are what bf_table divides by.