Skip to contents

posterior_predict_y1_given_y2() draws y1 from its distribution conditional on the recorded y2 and the model parameters, \(p(y_1 \mid y_2, \theta)\). posterior_epred_y1_given_y2() returns the conditional expectation \(E[y_1 \mid y_2]\). The main use of both functions is to impute y1 on the rows where y1 was not recorded. On an unpaired row of a fit from bipois_partialobs() or binegbin_partialobs(), the draws are the imputed values of the unrecorded y1, and the expectation is the mean of those imputed values.

posterior_predict() and posterior_epred() do not condition on y2. For a fit from any of the four constructors, these two functions return draws from the marginal distribution of y1 and its marginal mean mu + lambdaone, as for any other brms family. posterior_predict_y1_and_y2() draws both counts.

Usage

posterior_predict_y1_given_y2(
  object,
  newdata = NULL,
  re_formula = NULL,
  ndraws = NULL,
  draw_ids = NULL,
  keep_observed = FALSE,
  ...
)

posterior_epred_y1_given_y2(
  object,
  newdata = NULL,
  re_formula = NULL,
  ndraws = NULL,
  draw_ids = NULL,
  keep_observed = FALSE,
  ...
)

Arguments

object

A brmsfit from bipois(), bipois_partialobs(), binegbin(), or binegbin_partialobs().

newdata

An optional data frame; if NULL, the data of the fit are used. See Details.

re_formula

A formula for the group-level effects to include, passed to brms::prepare_predictions(). NULL (the default) includes all group-level effects, and NA includes none.

ndraws, draw_ids

Passed to brms::prepare_predictions() to select posterior draws.

keep_observed

Logical. If FALSE (the default), every row returns a conditional draw or the conditional expectation. If TRUE, paired rows return the recorded y1.

...

Further arguments to brms::prepare_predictions(), such as allow_new_levels.

Value

A numeric matrix with posterior draws in rows and observations in columns. The values from posterior_predict_y1_given_y2() are integers.

Details

Both families construct the pair as y1 = N_s + N1 and y2 = N_s + N2, where the shared component N_s has mean mu and dispersion shapes, and the source-exclusive components N1 and N2 have means lambdaone and lambdatwo and dispersions shapexone and shapextwo. The three components are independent given the parameters, so N1 is independent of y2, and \(y_1 \mid y_2\) is the sum of \(N_s \mid y_2\) and an independent draw of N1. The conditional distribution of the shared component is

$$P(N_s = k \mid y_2) \propto p(k \mid \mu, \phi_s)\, p(y_2 - k \mid \lambda_2, \phi_{x2}), \quad k = 0, \ldots, y_2,$$

where \(p(\cdot \mid m, \phi)\) is the negative-binomial probability mass function with mean \(m\) and inverse dispersion \(\phi\) for binegbin, and the Poisson probability mass function with mean \(m\) for bipois. For bipois, this distribution is Binomial with size \(y_2\) and probability \(\mu / (\mu + \lambda_2)\). For binegbin, the distribution has no closed form and is evaluated over the \(y_2 + 1\) values of \(k\). In both families,

$$E[y_1 \mid y_2] = E[N_s \mid y_2] + \lambda_1.$$

The exclusive-component dispersion in the conditional is shapextwo, the dispersion of the second exclusive component, because the count conditioned on is y2.

Recorded values of y1. With keep_observed = FALSE, the default, every row receives a conditional draw or the conditional expectation, including the paired rows. On a paired row, the conditional draw is the quantity to compare with the recorded y1 in a conditional predictive check. With keep_observed = TRUE, every draw on a paired row is the recorded y1, and the draws on unpaired rows are unchanged. Summed over rows, the draws returned with keep_observed = TRUE are draws of the total of y1: recorded where y1 was recorded, and imputed where y1 was not recorded. To reproduce results from posterior_predict() and posterior_epred() in bicountbrms 0.10.1, use these two functions with keep_observed = FALSE; given the same seed, the draws are identical.

New data. newdata must contain the recorded y2 of each row, because the result depends on y2. For a fit from a _partialobs() constructor, brms also requires the observation indicator named in vint(). The column y1 may be omitted, except that with keep_observed = TRUE a row indicated as paired needs its recorded y1, and the function stops with an error if that y1 is missing.

Stored fits whose family object declares the pre-0.7.0 rate names (lambdaem, lambdalb) or a single exclusive-component dispersion (shapex) are read under those names, as by the brms methods.

See also

posterior_predict_y1_and_y2() for draws of both counts from the marginal model; vignette("prediction") for a summary of the prediction functions.

Examples

if (FALSE) { # \dontrun{
library(brms)

# 500 units. y2 is recorded on every unit, and y1 on about 20% of units.
set.seed(20261115)
dat <- rbicount(
  500,
  mu        = 4,
  lambdaone = 1,
  lambdatwo = 2,
  shapes    = 2,
  shapexone = 4,
  shapextwo = 4
)
dat$y1_obs  <- rbinom(nrow(dat), 1, 0.2)
dat$y1_true <- dat$y1
dat$y1[dat$y1_obs == 0] <- -1L

fit <- brm(
  bf_mfd(
    y1 | vint(y2, y1_obs),
    logM      ~ 1,
    logitf    ~ 1,
    delta     ~ 1,
    kappas    ~ 1,
    kappaxone ~ 1,
    kappaxtwo ~ 1
  ),
  family   = binegbin_partialobs(),
  stanvars = binegbin_partialobs_stanvars(),
  data     = dat,
  prior    = c(
    prior(normal(1.5, 1), class = "b", nlpar = "logM"),
    prior(normal(0, 1.5), class = "b", nlpar = "logitf"),
    prior(normal(0, 0.5), class = "b", nlpar = "delta"),
    prior(normal(0, 1),   class = "b", nlpar = "kappas",    lb = 0),
    prior(normal(0, 1),   class = "b", nlpar = "kappaxone", lb = 0),
    prior(normal(0, 1),   class = "b", nlpar = "kappaxtwo", lb = 0)
  )
)

# Draws of y1 on every unit: recorded where y1_obs is 1,
# imputed from the y2 of each unit where y1_obs is 0.
y1_draws <- posterior_predict_y1_given_y2(fit, keep_observed = TRUE)

# Posterior of the source-1 total, against the simulated total
quantile(rowSums(y1_draws), probs = c(0.025, 0.5, 0.975))
sum(dat$y1_true)

# Imputed draws and E[y1 | y2] on the units where source 1 was not
# recorded only
unrec <- dat$y1_obs == 0
y1_imp <- posterior_predict_y1_given_y2(fit)[, unrec]
ep_imp <- posterior_epred_y1_given_y2(fit)[, unrec]
} # }