Returns a version of bipois() for partially paired designs, in which source 1
has missing values. The generative model, family name, dpars, and
post-processing methods are the same as for bipois(). Each row also
supplies an observation indicator as the second argument of vint(), after y2:
brm(
bf(y1 | vint(y2, y1_obs) ~ ...),
family = bipois_partialobs(),
stanvars = bipois_partialobs_stanvars(),
data = dat
)This template is for the native parameterisation. To fit the model in the
congruence–bias parameterisation, which we recommend, write the formula
with bf_mfd().
y1_obs is 1 on rows where both counts were recorded and 0 on rows where
only y2 was recorded. An indicator other than 0 or 1 makes the
sampler fail at initialisation, with a message in the chain output that
names the cause. The check also catches the two vint() arguments written
in reverse order.
Do not use NA for missing y1 values. brms drops any row with a
missing response with only a warning. On a row with y1_obs == 0, the
likelihood term does not depend on y1, so y1 can be any integer. We
suggest -1, because it cannot be confused with a recorded count:
dat$y1[dat$y1_obs == 0] <- -1LValue
bipois_partialobs() returns a brms custom_family object.
bipois_partialobs_stanvars() returns a stanvars object holding the
Stan code for the corresponding _lpmf. The returned family has the same
name as bipois(), so the post-processing methods documented there
apply to a fit made with either constructor.
Details
On a paired row, the likelihood term is the joint probability of
(y1, y2) under bipois(). On an unpaired row, the likelihood term is the
marginal probability of y2, the joint probability summed over every
possible value of the unrecorded y1. That sum has the closed form
y2 ~ Poisson(mu + lambdatwo), because a sum of independent Poisson
variables is Poisson.
To impute y1 on the unpaired rows, use posterior_predict_y1_given_y2()
or posterior_epred_y1_given_y2(); see vignette("prediction").
An unpaired row informs mu and lambdatwo only through their sum,
mu + lambdatwo, which is the expected value of y2. Two quantities are
therefore estimated from the paired rows alone: (i) lambdaone, which does
not appear in the likelihood term for an unpaired row, and (ii) the division
of mu + lambdatwo between the shared and source-2-exclusive components.
The congruence \(f\) depends on both, so \(f\) is also estimated from the
paired rows alone. With few paired rows, the posteriors of these quantities
stay close to their priors, however many unpaired rows the design contains.
bipois_partialobs() fixes the variance of each latent component equal to
the mean of that component. Where the counts are overdispersed relative to
a Poisson distribution, binegbin_partialobs() is the correct model and
bipois_partialobs() will understate the marginal variances. Where the
counts are not overdispersed, binegbin_partialobs() fits the counts only
with shapes, shapexone, and shapextwo very large (each kappa near
0), and sampling is poor when the dispersions are that large.
bipois_partialobs() has no dispersions, so it avoids the problem. Compare
the fits from the two constructors with loo(), which scores y1 and y2 on
a paired row and y2 on an unpaired row.
See also
bipois() for the fully paired case; binegbin_partialobs() for
the overdispersed counterpart.