hpd.summary: Summarize an emmGrid from a Bayesian model

Description

This function computes point estimates and HPD (or quantile) intervals for each factor combination in object@emmGrid. While this function may be called independently, it is called automatically by the S3 method summary.emmGrid when the object is based on a Bayesian model. (Note: the level argument, or its default, is passed as prob).

Usage

hpd.summary(object, prob, by, type, point.est = median,
  ci.method = get_emm_option("post.ci.method"), delta,
  bias.adjust = get_emm_option("back.bias.adj"), sigma, ...)

Value

an object of class summary_emm

Arguments

object: an emmGrid object having a non-missing post.beta slot
prob: numeric probability content for HPD intervals (note: when not specified, the current level option is used; see emm_options)
by: factors to use as by variables
type: prediction type as in summary.emmGrid
point.est: function to use to compute the point estimates from the posterior sample for each grid point
ci.method: character value matching "HPD" (default) or "quantile" (but actually not case-sensitive). The default is to use HPD intervals (which are the shortest possible intervals). Alternatively, choosing "quantile" uses the quantiles of the posterior having equal tail probabilities.
delta: Numeric equivalence threshold (on the linear predictor scale regardless of type). See the section below on equivalence testing.
bias.adjust: Logical value for whether to adjust for bias in back-transforming (type = "response"). This requires a value of sigma to exist in the object or be specified.
sigma: Error SD assumed for bias correction (when type = "response". If not specified, object@misc$sigma is used, and a warning if it is not found or invalid. Note: sigma may be a vector, as long as it conforms to the number of observations in the posterior sample.
...: required but not used

Equivalence testing note

If delta is positive, two columns labeled p.equiv and odds.eq are appended to the summary. p.equiv is the fraction of posterior estimates having absolute values less than delta. The odds.eq column is just p.equiv converted to an odds ratio; so it is the posterior odds of equivalence.

A high value of p.equiv is evidence in favor of equivalence. It can be used to obtain something equivalent (in spirit) to the frequentist Schuirmann (TOST) procedure, whereby we would conclude equivalence at significance level $\alpha$ if the $(1 - 2\alpha)$ confidence interval falls entirely in the interval $[-\delta, \delta]$. Similarly in the Bayesian context, an equally strong argument for equivalence is obtained if p.equiv exceeds $1 - 2\alpha$.

A closely related quantity is the ROPE (region of practical equivalence), obtainable via bayestestR::rope(object, range = c(-delta, delta)). Its value is approximately 100 * p.equiv / 0.95 if the default ci = 0.95 is used. See also bayestestR's issue #567.

Finally, a Bayes factor for equivalence is obtainable by dividing odds.eq by the prior odds of equivalence, assessed or elicited separately.

Examples

Run this code

if(require("coda")) 
    emm_example("hpd.summary-coda")
    # Use emm_example("hpd.summary-coda", list = TRUE) # to see just the code