Skip to contents

Posterior checks with bayesplot

Source: R/bayesplot.R, R/scam-methods.R
pp_check.Rd

Compare observations with replicated responses using an optional installation of bayesplot. Call bayesplot::pp_check() after loading gratia.

Usage

pp_check.gam(
  object,
  ...,
  type = "dens_overlay",
  prefix = c("ppc", "ppd"),
  n_draws = NULL,
  seed = NULL,
  group = NULL,
  x = NULL,
  response = NULL,
  conditioning = NULL,
  sampling_args = list()
)

pp_check.bam(
  object,
  ...,
  type = "dens_overlay",
  prefix = c("ppc", "ppd"),
  n_draws = NULL,
  seed = NULL,
  group = NULL,
  x = NULL,
  response = NULL,
  conditioning = NULL,
  sampling_args = list()
)

pp_check.gamm(
  object,
  ...,
  type = "dens_overlay",
  prefix = c("ppc", "ppd"),
  n_draws = NULL,
  seed = NULL,
  group = NULL,
  x = NULL,
  response = NULL,
  conditioning = NULL,
  sampling_args = list()
)

pp_check.gamm4(
  object,
  ...,
  type = "dens_overlay",
  prefix = c("ppc", "ppd"),
  n_draws = NULL,
  seed = NULL,
  group = NULL,
  x = NULL,
  response = NULL,
  conditioning = NULL,
  sampling_args = list()
)

pp_check.scam(object, ...)

Arguments

object

a fitted GAM, BAM, gamm, or gamm4 model.

...

arguments passed only to the selected bayesplot plotting function.

type

character; the bayesplot plot name without its ppc_ or ppd_ prefix. Available names come from the installed version of bayesplot.

prefix

character; "ppc" compares responses with observations; "ppd" shows replicated responses alone.

n_draws

positive integer; total coefficient and response draws. Defaults to 50 for ordinary plots and 4000 for leave-one-out plots. With supplied coefficient draws, defaults to their number of rows.

seed

integer or NULL; seed for reproducible sampling.

group, x

a variable name in the fitting model frame, or a vector in fitting-row order. Missing fitting rows are removed with the response.

response

character or integer; response column or grouped-family component to check, for example "response2" or "component1". Required when the model contains several responses or grouped components.

conditioning

character or NULL; conditioning passed to the existing posterior predictive sampler. See posterior_predictive_plot().

sampling_args

named list; arguments for the posterior sampler, including method, draws, unconditional, and mvn_method. Prediction at new data is not supported. Fitted binomial trials must be retained. Use chains for multiple coefficient chains and diagnostics = TRUE for coefficient diagnostics. Leave-one-out plots use stored chain identities; for unlabelled supplied draws, include r_eff or chain_id here.

Value

The object returned by the selected bayesplot function, usually a ggplot object.

Details

These methods require bayesplot only when called. Leave-one-out plots also require loo and a likelihood supported by log_lik(). Their response and likelihood calculations use exactly the same coefficient draws. For a multivariate response, the weights leave out the entire response vector. Grouped-family checks use only the selected component's observations.

Some bayesplot plotting functions, including error_binned and LOO-PIT plots, also require bayesplot's optional rstantools package. Install it with install.packages("rstantools") to use those plots.

Binomial observations and replicated responses are shown as numbers of successes with fitted trial counts. Other gratia predictive plots use proportions. The error_binned plot uses expected responses rather than response simulations and requires an available finite response mean. For gamm and gamm4 it currently requires conditioning = "gam_component". Plotting functions requiring additional inputs, such as censoring indicators, receive those inputs through ....

Coefficient draws use a Gaussian approximation by default; additional family parameters remain at fitted values. See log_lik() for the assumptions of approximate leave-one-out calculations. Existing conditioning and response dependence restrictions of the diagnostic samplers also apply here.

See also

posterior_predictive_plot(), log_lik()

Examples

if (requireNamespace("bayesplot", quietly = TRUE)) {
  load_mgcv()
  dat <- data_sim("eg1", n = 100, seed = 2)
  fit <- gam(y ~ s(x0) + s(x1) + s(x2) + s(x3), data = dat)
  bayesplot::pp_check(fit, n_draws = 10, seed = 42)
}