Both families model a pair of counts as y_1
= N_s + N_1 and y_2 = N_s + N_2,
where N_s is the shared component,
recorded by both sources, and N_1 and
N_2 are the source-exclusive
components. The three components are independent given their means \mu (mu), \lambda_1 (lambdaone), and \lambda_2 (lambdatwo). Six
functions return predictions from a fit, either as draws (the
posterior_predict*() functions) or as expectations (the
posterior_epred*() functions). In the table, \theta denotes the parameters under one
posterior draw.
| Function | Distribution or quantity | Returns |
|---|---|---|
posterior_predict() |
p(y_1 \mid \theta) | draws × observations |
posterior_epred() |
E[y_1] = \mu + \lambda_1 | draws × observations |
posterior_predict_y1_and_y2() |
p(y_1, y_2 \mid \theta) | draws × observations × 2 |
posterior_epred_y1_and_y2() |
E[y_1] = \mu + \lambda_1, E[y_2] = \mu + \lambda_2 | draws × observations × 2 |
posterior_predict_y1_given_y2() |
p(y_1 \mid y_2, \theta) | draws × observations |
posterior_epred_y1_given_y2() |
E[y_1 \mid y_2] = E[N_s \mid y_2] + \lambda_1 | draws × observations |
The third dimension of the array returned by the two
_y1_and_y2() functions is named y1 and
y2:
draws <- posterior_predict_y1_and_y2(fit)
draws[, , "y1"] # one row per posterior draw, one column per observation
draws[, , "y2"]The first four functions in the table depend on the parameters alone.
posterior_predict_y1_given_y2() and
posterior_epred_y1_given_y2() also condition on the
recorded y_2, and are used mainly to
impute y_1 on the unpaired rows of a
fit from bipois_partialobs() or
binegbin_partialobs(), where y_1 was not recorded. With
keep_observed = TRUE,
posterior_predict_y1_given_y2() and
posterior_epred_y1_given_y2() return the recorded y_1 on paired rows.
brms functions for post-processing
| Function | Quantity |
|---|---|
fitted(), predict()
|
Marginal mean and draws of y_1, summarised |
conditional_effects() |
Marginal mean of y_1 over a grid of predictor values |
pp_check() |
Recorded y_1 against marginal draws of y_1 |
bayes_R2(), loo_R2()
|
Variance of the marginal mean of y_1, divided by that variance plus the residual variance |
LOO-PIT, through
pp_check(type = "loo_pit_overlay")
|
Leave-one-out probability integral transform of the recorded y_1 |
loo(), waic(), and
kfold()
|
Joint probability of both counts on paired rows, and probability of y_2 on unpaired rows |
On an unpaired row of a _partialobs() fit,
y1 is not a recorded count. pp_check(),
bayes_R2(), loo_R2(), and LOO-PIT compare the
predictions with y_1. For such a fit,
these four functions are meaningful only with newdata
restricted to the paired rows, as in
bayes_R2(fit, newdata = subset(dat, y1_obs == 1)).
loo(), waic(), and kfold() need
no such restriction, because the likelihood term for an unpaired row
does not depend on y_1.
Conditional and joint draws
Given y_2, the shared component has the distribution
P(N_s = k \mid y_2) \propto p(k \mid \mu, \phi_s)\, p(y_2 - k \mid \lambda_2, \phi_{x2}), \qquad k = 0, \ldots, y_2,
where p(\cdot \mid m, \phi) is the
negative-binomial probability mass function with mean m and (inverse) dispersion parameter \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).
posterior_predict_y1_given_y2() adds an independent draw of
N_1 to a draw of N_s from this distribution.
posterior_predict_y1_and_y2() draws N_s once for each pair and adds the same draw
to both counts, so the drawn counts have \mathrm{Cov}(y_1, y_2) = \mathrm{Var}(N_s),
as in the model. The package has no separate function for y_2, because draws of y_2 made separately would not share their
draws of N_s with the draws of y_1.
Imputing missing values of y_1
To impute missing values of y_1 conditional on the recorded values of y_2:
- Code the observation indicator
y1_obsas1on paired rows and0on unpaired rows, and sety1to any integer on the unpaired rows. We suggest-1, because it cannot be confused with a recorded count.y1must not beNA. - Fit the model with
binegbin_partialobs()(orbipois_partialobs()), passing the indicator as the second argument ofvint(). - Draw the unrecorded counts with
posterior_predict_y1_given_y2(), which draws each y_1 given the y_2 recorded on its row.
# 500 units. y2 is recorded on every unit, and y1 on each unit with
# probability 0.2.
set.seed(20261115)
dpars <- mfd_to_dpars(
M = 5,
f = 0.7,
delta = atanh(0.2),
kappas = 0.7,
kappax = 0.5
)
dat <- do.call(rbicount, c(list(n = 500), dpars))
dat$y1_true <- dat$y1
dat$y1[rbinom(nrow(dat), 1, 0.2) == 0] <- NA
# Add the indicator y1_obs, 1 where y1 was recorded, and set the missing values
# of y1 to -1, so that brms keeps those rows
dat$y1_obs <- as.integer(!is.na(dat$y1))
dat$y1[is.na(dat$y1)] <- -1L
fit <- brm(
bf_mfd(
y1 | vint(y2, y1_obs),
logM ~ 1,
logitf ~ 1,
delta ~ 1,
kappas ~ 1,
kappaxone ~ 1,
kappaxtwo ~ 1
),
family = binegbin_partialobs(),
stanvars = binegbin_partialobs_stanvars(),
data = dat,
prior = c(
prior(normal(1.5, 1), class = "b", nlpar = "logM"),
prior(normal(0, 1.5), class = "b", nlpar = "logitf"),
prior(normal(0, 0.5), class = "b", nlpar = "delta"),
prior(normal(0, 1), class = "b", nlpar = "kappas", lb = 0),
prior(normal(0, 1), class = "b", nlpar = "kappaxone", lb = 0),
prior(normal(0, 1), class = "b", nlpar = "kappaxtwo", lb = 0)
)
)
# Draws of the unrecorded counts: one column per unpaired row
y1_imputed <- posterior_predict_y1_given_y2(fit)[, dat$y1_obs == 0]
# Draws of y1 on every row: recorded values on paired rows, imputed values
# on unpaired rows
y1_completed <- posterior_predict_y1_given_y2(fit, keep_observed = TRUE)
# Posterior of the total of y1 over all rows, against the simulated total
quantile(rowSums(y1_completed), c(0.025, 0.5, 0.975))
sum(dat$y1_true)Each row of y1_imputed is one posterior draw of the
unrecorded counts, so summaries of the draws include both the
uncertainty in the parameters and the variation of the counts around
their expected values. posterior_epred_y1_given_y2() gives
the conditional expectation E[y_1 \mid
y_2] instead, which omits the variation of the counts. In the
article Partially
paired designs, imputed values of y_1 are compared with values withheld from
the model.
Checking the model against both counts
# Variance of d = y1 - y2 on the paired rows, against marginal draws
paired <- subset(dat, y1_obs == 1)
yrep <- posterior_predict_y1_and_y2(fit, newdata = paired, ndraws = 500)
var_d <- apply(yrep[, , "y1"] - yrep[, , "y2"], 1, var)
mean(var_d >= var(paired$y1 - paired$y2))The last line is the posterior predictive p-value for the variance of d. A value near 0 or 1 suggests that the model misstates the spread of the difference.