This function creates default priors for brms-regression models, based on the same automatic prior-scale adjustment as in rstanarm.
auto_prior(formula, data, gaussian, locations = NULL)A formula describing the model, which just needs to contain
the model terms, but no notation of interaction, splines etc. Usually,
you want only those predictors in the formula, for which automatic
priors should be generated. Add informative priors afterwards to the
returned brmsprior-object.
The data that will be used to fit the model.
Logical, if the outcome is gaussian or not.
A numeric vector with location values for the priors. If
locations = NULL, 0 is used as location parameter.
A brmsprior-object.
auto_prior() is a small, convenient function to create
some default priors for brms-models with automatically adjusted prior
scales, in a similar way like rstanarm does. The default scale for
the intercept is 10, for coefficients 2.5. If the outcome is gaussian,
both scales are multiplied with sd(y). Then, for categorical
variables, nothing more is changed. For numeric variables, the scales
are divided by the standard deviation of the related variable.
All prior distributions are normal distributions. auto_prior()
is intended to quickly create default priors with feasible scales. If
more precise definitions of priors is necessary, this needs to be done
directly with brms-functions like set_prior().
As auto_prior() also sets priors on the intercept, the model
formula used in brms::brm() must be rewritten to something like
y ~ 0 + intercept ..., see set_prior.
data(efc)
efc$c172code <- as.factor(efc$c172code)
efc$c161sex <- as.factor(efc$c161sex)
mf <- formula(neg_c_7 ~ c161sex + c160age + c172code)
auto_prior(mf, efc, TRUE)
#> prior class coef group resp dpar nlpar lb ub source
#> normal(0, 38.96) Intercept <NA> <NA> user
#> normal(0, 9.74) b c161sex2 <NA> <NA> user
#> normal(0, 0.73) b c160age <NA> <NA> user
#> normal(0, 9.74) b c172code2 <NA> <NA> user
#> normal(0, 9.74) b c172code3 <NA> <NA> user
## compare to
# m <- rstanarm::stan_glm(mf, data = efc, chains = 2, iter = 200)
# ps <- prior_summary(m)
# ps$prior_intercept$adjusted_scale
# ps$prior$adjusted_scale
## usage
# ap <- auto_prior(mf, efc, TRUE)
# brm(mf, data = efc, prior = ap)
# add informative priors
mf <- formula(neg_c_7 ~ c161sex + c172code)
auto_prior(mf, efc, TRUE) +
brms::prior(normal(.1554, 40), class = "b", coef = "c160age")
#> prior class coef group resp dpar nlpar lb ub source
#> normal(0, 38.95) Intercept <NA> <NA> user
#> normal(0, 9.74) b c161sex2 <NA> <NA> user
#> normal(0, 9.74) b c172code2 <NA> <NA> user
#> normal(0, 9.74) b c172code3 <NA> <NA> user
#> normal(0.1554, 40) b c160age <NA> <NA> user
# example with binary response
efc$neg_c_7d <- ifelse(efc$neg_c_7 < median(efc$neg_c_7, na.rm = TRUE), 0, 1)
mf <- formula(neg_c_7d ~ c161sex + c160age + c172code + e17age)
auto_prior(mf, efc, FALSE)
#> prior class coef group resp dpar nlpar lb ub source
#> normal(0, 10) Intercept <NA> <NA> user
#> normal(0, 2.5) b c161sex2 <NA> <NA> user
#> normal(0, 0.19) b c160age <NA> <NA> user
#> normal(0, 2.5) b c172code2 <NA> <NA> user
#> normal(0, 2.5) b c172code3 <NA> <NA> user
#> normal(0, 0.31) b e17age <NA> <NA> user