Generate predictions from a fitted longitudinal AME model. Returns a list of matrices (one per time point) on the requested scale.
# S3 method for lame
predict(
object,
newdata = NULL,
type = c("response", "link"),
h = 0L,
by_draw = FALSE,
interval = c("none", "credible"),
probs = c(0.025, 0.975),
newexposure = NULL,
n_draws = NULL,
seed = NULL,
...
)List of prediction matrices (one per time point).
Fitted LAME model object.
Optional list of T dyadic covariate arrays
([n_row, n_col, p] each, with the same actors as the fit) to
compute counterfactual predictions. When NULL, the training-data
predictions are returned.
Character; "response" (default) or "link".
Integer >= 0: forecast horizon. When h = 0 (default),
returns in-sample predictions as before. When h > 0, propagates
the AR(1) (or RW1) state-space model forward by h periods and
returns a list of h matrices (one per future period). Requires
at least one dynamic component on the fit. Warns when posterior
\(\rho_\beta\) is near 1.
When TRUE and h > 0, returns an
n x n x h x n_draws array of per-draw forecasts instead of
per-period means.
One of "none" (default) or "credible".
When h > 0 and "credible", the per-period output is
a list of length-3 lists with $lower, $median,
$upper matrices computed at the probs quantiles
across posterior draws. Ignored for in-sample (h = 0)
predictions.
Length-2 vector of lower / upper quantiles for the
credible interval when interval = "credible". Default
c(0.025, 0.975).
Optional length-h non-negative numeric
vector of future-period exposures (Poisson only). When omitted
and the fit has period_exposure stored, defaults to the
last observed exposure; when both are absent, defaults to 1.
Number of posterior draws to use when h > 0.
Default NULL uses all stored draws. Ignored for in-sample
(h = 0) predictions.
Optional RNG seed for the h > 0 forecast draws,
making forecasts reproducible. Ignored when h = 0.
Additional arguments (not used).