Posterior Predictive P-value

Obtain the posterior predictive p-value (ppp) for a fitted blavaan object, computing it on demand if it was not already computed during model estimation.

Usage

ppp(object, ...)

Arguments

object

An object of class blavaan.

...

Other arguments passed to the underlying computation for single-level models with ordinal variables, or for two-level (multilevel) models (currently only thin and parallel have an effect for either case, plus em_control for two-level models; see Details).

Details

For most models, fitMeasures(object, "ppp") already returns the ppp computed during estimation, and ppp(object) simply returns that same value. The exceptions are single-level models (with a single group or multiple groups) that include ordinal variables, and two-level (multilevel) models, both fit with target = "stan" or "cmdstan": for these models, test defaults to "none" and ppp is not computed automatically, because the computation takes noticeably longer than for other models. Calling ppp(object) on such a fit computes it as a separate step. To have ppp computed automatically during estimation instead, pass test = "standard" (or, for two-level models, any value other than "none") to bcfa()/bsem()/bgrowth()/ blavaan().

For single-level ordinal/mixed models, the on-demand computation is a limited-information, pairwise comparison of observed vs model-implied associations, which better reflects fit for ordinal data than the previous Stan-computed default.

For two-level (multilevel) models, the on-demand computation is a saturated-model likelihood-ratio comparison (the same statistic previously computed inside the compiled Stan program), now computed in R by calling lavaan's own saturated-model EM algorithm once per retained posterior draw. Because this per-draw computation is comparatively slow, increasing thin and/or setting parallel = TRUE (together with future::plan("multicore") or future::plan("multisession")) is recommended for larger two-level fits; an em_control argument (a list with tol, max_iter, and acceleration elements) is also available to tune the per-draw EM refit's convergence settings. fixed.x variables are supported at the within level, the between level, or both (in different variables): they are held fixed at their observed values in every simulated replicate rather than being resimulated, matching how fixed.x is already handled for single-level on-demand computations. A variable that is fixed.x at both levels simultaneously is not supported (a pre-existing lavaan limitation, not specific to this computation) and will raise an error.

Value

A numeric ppp value (or vector, for multiple imputations/groups, matching whatever fitMeasures(object, "ppp") would otherwise return).

See also

ppmc, for more flexible posterior predictive model checks using arbitrary discrepancy functions.

Examples

if (FALSE) { # \dontrun{
data(HolzingerSwineford1939, package = "lavaan")

HS.model <- ' visual  =~ x1 + x2 + x3
              textual =~ x4 + x5 + x6
              speed   =~ x7 + x8 + x9 '

## make some indicators ordinal
HSord <- HolzingerSwineford1939
HSord$x1 <- as.integer(cut(HSord$x1, 3))
HSord$x2 <- as.integer(cut(HSord$x2, 3))

## ppp is not computed by default for this ordinal model
fit <- bcfa(HS.model, data = HSord, ordered = c("x1", "x2"))

## compute it on demand
ppp(fit)
} # }