posterior_predict.brmsfit: Samples from the Posterior Predictive Distribution

Description

Compute posterior samples of the posterior predictive distribution. Can be performed for the data used to fit the model (posterior predictive checks) or for new data. By definition, these samples have higher variance than samples of the means of the posterior predictive distribution computed by posterior_epred.brmsfit. This is because the residual error is incorporated in posterior_predict. However, the estimated means of both methods averaged across samples should be very similar.

Usage

# S3 method for brmsfit
posterior_predict(
  object,
  newdata = NULL,
  re_formula = NULL,
  re.form = NULL,
  transform = NULL,
  resp = NULL,
  negative_rt = FALSE,
  nsamples = NULL,
  subset = NULL,
  sort = FALSE,
  ntrys = 5,
  cores = getOption("mc.cores", 1),
  ...
)

Arguments

object

An object of class brmsfit.

newdata

An optional data.frame for which to evaluate predictions. If NULL (default), the original data of the model is used. NA values within factors are interpreted as if all dummy variables of this factor are zero. This allows, for instance, to make predictions of the grand mean when using sum coding.

re_formula

formula containing group-level effects to be considered in the prediction. If NULL (default), include all group-level effects; if NA, include no group-level effects.

re.form

Alias of re_formula.

transform

(Deprecated) A function or a character string naming a function to be applied on the predicted responses before summary statistics are computed.

resp

Optional names of response variables. If specified, predictions are performed only for the specified response variables.

negative_rt

Only relevant for Wiener diffusion models. A flag indicating whether response times of responses on the lower boundary should be returned as negative values. This allows to distinguish responses on the upper and lower boundary. Defaults to FALSE.

nsamples

Positive integer indicating how many posterior samples should be used. If NULL (the default) all samples are used. Ignored if subset is not NULL.

subset

A numeric vector specifying the posterior samples to be used. If NULL (the default), all samples are used.

sort

Logical. Only relevant for time series models. Indicating whether to return predicted values in the original order (FALSE; default) or in the order of the time series (TRUE).

ntrys

Parameter used in rejection sampling for truncated discrete models only (defaults to 5). See Details for more information.

cores

Number of cores (defaults to 1). Can be set globally via the mc.cores option.

...

Further arguments passed to prepare_predictions that control several aspects of data validation and prediction.

Value

An array of predicted response values. In univariate models, the output is as an S x N matrix, where S is the number of posterior samples and N is the number of observations. In multivariate models, an additional dimension is added to the output which indexes along the different response variables.

Details

NA values within factors in newdata, are interpreted as if all dummy variables of this factor are zero. This allows, for instance, to make predictions of the grand mean when using sum coding.

In multilevel models, it is possible to allow new levels of grouping factors to be used in the predictions. This can be controlled via argument allow_new_levels. New levels can be sampled in multiple ways, which can be controlled via argument sample_new_levels. Both of these arguments are documented in prepare_predictions along with several other useful arguments to control specific aspects of the predictions.

For truncated discrete models only: In the absence of any general algorithm to sample from truncated discrete distributions, rejection sampling is applied in this special case. This means that values are sampled until a value lies within the defined truncation boundaries. In practice, this procedure may be rather slow (especially in R). Thus, we try to do approximate rejection sampling by sampling each value ntrys times and then select a valid value. If all values are invalid, the closest boundary is used, instead. If there are more than a few of these pathological cases, a warning will occur suggesting to increase argument ntrys.

Examples

Run this code

# NOT RUN {
## fit a model
fit <- brm(time | cens(censored) ~ age + sex + (1 + age || patient), 
           data = kidney, family = "exponential", inits = "0")

## predicted responses
pp <- posterior_predict(fit)
str(pp)

## predicted responses excluding the group-level effect of age
pp <- posterior_predict(fit, re_formula = ~ (1 | patient))
str(pp)

## predicted responses of patient 1 for new data
newdata <- data.frame(
  sex = factor(c("male", "female")),
  age = c(20, 50),
  patient = c(1, 1)
)
pp <- posterior_predict(fit, newdata = newdata)
str(pp)
# }
# NOT RUN {
# }

Run the code above in your browser using DataCamp Workspace