Creates a validated prior specification that can be used for the
intercept, the regression coefficients ("slopes") or the auxiliary
parameter (e.g. the residual standard deviation) of a model fitted with
fit_model_with_prior(). Argument names follow rstanarm.
Usage
prior_spec(
type = c("normal", "student_t", "cauchy", "laplace", "exponential", "flat"),
location = 0,
scale = 2.5,
df = 3,
rate = 1,
autoscale = FALSE
)Arguments
- type
One of
"normal","student_t","cauchy","laplace"(coefficients only),"exponential"(auxiliary parameter only) or"flat"(an improper uniform prior).- location
Prior location. For slopes, either a single value or one value per coefficient. Ignored for
"exponential"and"flat".- scale
Prior scale (positive). For slopes, a single value or one per coefficient.
- df
Degrees of freedom for
"student_t": a single value or, for slopes, one per coefficient.- rate
Rate for
"exponential"; the prior mean is1 / rate.- autoscale
Logical. If
TRUE, rstanarm rescales the prior using the standard deviations of the outcome and predictors.
Choosing a scale
A prior scale is only meaningful relative to the units of the outcome and
the predictors. A fixed scale such as normal(0, 2.5) is weakly
informative when variables are standardised but can be overwhelmingly
informative when they are not (for example, an outcome measured in
dollars). Either standardise your variables, use autoscale = TRUE
(rstanarm then rescales by the standard deviations of the data), derive
scales from substantive knowledge, or use empirical_bayes_priors(). See
Gelman et al. (2008) and vignette("bmbeR").
References
Gelman, A., Jakulin, A., Pittau, M. G., & Su, Y.-S. (2008). A weakly informative default prior distribution for logistic and other regression models. The Annals of Applied Statistics, 2(4), 1360–1383. doi:10.1214/08-AOAS191
See also
prior_config() to combine priors for a model.