Predictive check against simulated responses

The frequentist analog of brms's pp_check(): responses are simulated from the fitted model (marginally over the random effects) and handed to the corresponding bayesplot ppc_* function (bayesplot must be installed, but not necessarily attached).

Usage

pp_check(object, ...)

# S3 method for class 'frmtmb_fit'
pp_check(
  object,
  type = "dens_overlay",
  ndraws = 10,
  prefix = c("ppc", "ppd"),
  group = NULL,
  x = NULL,
  newdata = NULL,
  resp = NULL,
  ...,
  re_formula = NA
)

Arguments

object: A frmtmb_fit for a univariate model.
...: Passed to the ppc_* function.
type: The bayesplot check, i.e. the part after ppc_ ("dens_overlay", "hist", "stat", "stat_grouped", "scatter_avg", ...). With prefix = "ppc" a name that bayesplot::available_ppc() does not list is refused, as brms refuses it.
ndraws: Number of simulated response vectors.
prefix: "ppc" (the default) plots the observed response against the simulated ones; "ppd" plots the simulated ones alone, through bayesplot's ppd_* function of the same name.
group: The name of a model variable to stratify by, for the *_grouped types, which need it. The name is looked up in the model frame, as brms looks it up in the model's data, so a column the model does not use is refused.
x: The name of a model variable for the types that take an x ("intervals", "ribbon", "error_scatter_avg_vs_x", ...), looked up as group is.
newdata, resp: brms's arguments. With newdata the check is on its rows: simulate() draws for them, the observed series is its response column (needed for prefix = "ppc" only), and group and x are read from it, as brms reads them. A row with a missing response is dropped with brms's warning. On a multivariate fit resp names the one response to check, as in brms, which refuses the call without it: that response's observed values against its own simulated replicates, on its own rows (a subset() response on the rows of its subset). On a model with one response resp is accepted and ignored, which is what brms does with it there.
re_formula: The random-effect switch, in brms's spelling (pp_check() is a brms function). On a fit it is passed to simulate() and defaults to NA, which redraws every group-level effect in each replicate and holds every SMOOTH at its estimate, a smooth indexed by a grouping factor included; ~0 and ~1 mean the same, and a one-sided formula keeps the terms it names and redraws the rest (see simulate.frmtmb_fit()). On draws it is passed to posterior_predict() and defaults to NULL, because a draw already carries its own. lme4's re.form is refused. brms honors it on pp_check() and warns that it ignored it, which is a leak through its dots rather than a decision to copy.

Value

A ggplot object, as returned by the bayesplot ppc_* function that type selects.

Types a fit cannot draw

The loo_* types weight posterior draws by Pareto-smoothed importance sampling, and a maximum-likelihood fit has no posterior draws, so they are refused on a fit. pp_check() on the draws of frmtmb.sample::frm_sample() is the route to them. "error_binned" is refused on an ordinal, categorical or multinomial fit, as brms refuses it.

Examples

if (requireNamespace("bayesplot", quietly = TRUE)) {
  set.seed(1)
  dd <- data.frame(x = rnorm(100), g = factor(rep(1:10, 10)))
  dd$y <- rpois(100, exp(0.3 + 0.4 * dd$x + rnorm(10, 0, 0.6)[dd$g]))
  fit <- frm(bf(y ~ x + (1 | g)) + poisson(), data = dd)

  # the observed density against draws from the fit
  pp_check(fit, ndraws = 20)

  # any bayesplot ppc_* check, named by its suffix. A statistic the
  # model was not fitted to is the informative one: here, the share
  # of zeros, which is how zero inflation shows up.
  pp_check(fit, type = "stat", stat = function(y) mean(y == 0),
           ndraws = 50)
}
#> `stat_bin()` using `bins = 30`. Pick better value `binwidth`.