Stack fitted models into a LOO-weighted ensemble

Combines two or more fitted models into one predictive distribution with leave-one-out predictive stacking. Each member contributes its Pareto-smoothed importance-sampling LOO predictive density; the weights maximize the stacked leave-one-out predictive density and so favour members that predict held-out observations best. Works across every family (single-season / community / dynamic / integrated occupancy, JSDM, and the cover hurdle), each scored by the marginal likelihood in .tobs_pointwise_loglik().

Usage

tobs_stack(..., method = c("stacking", "pseudobma"))

# S3 method for class 'tobs_stack'
print(x, ...)

# S3 method for class 'tobs_stack'
fitted(object, ...)

# S3 method for class 'tobs_stack'
predict(
  object,
  X.0 = NULL,
  quantiles = c(0.025, 0.5, 0.975),
  n.draws = 4000L,
  ...
)

Arguments

...

ignored.

method

weighting scheme passed to loo::loo_model_weights(): "stacking" (default) or "pseudobma".

x

a tobs_stack object.

object

a tobs_stack object.

X.0

optional occupancy design matrix for out-of-sample prediction. When supplied, all members must share the same occupancy design.

quantiles

credible-interval quantile levels.

n.draws

size of the pooled stacked-predictive sample (design-matrix mode); each member contributes a share proportional to its weight.

Value

An object of class tobs_stack: a list with weights (named numeric, summing to 1), fits (the members), loo (per-member loo::loo() objects), comparison (a data frame of elpd_loo, weight, and worst Pareto-k per member), and method. stats::predict() / stats::fitted() on the result return the weight-combined predictive.

Details

The members must be fit to the same observation set. They may otherwise differ freely – different covariates, priors, or inference methods. (An ensemble of pure seed-variants, e.g. from tobs(..., control = list(n.seeds = K)), is a valid but degenerate case: the members are statistically identical, so the weights come out roughly uniform.)

Examples

# \donttest{
sim <- simulate_occu(N = 100, J = 3, n_occ_covs = 2, n_det_covs = 1,
                     seed = 1)
ctrl <- list(verbose = FALSE, progress = FALSE)
f1 <- tobs(~ occ_cov1, data = sim$data, family = occu(),
           detection = ~ det_cov1, y = sim$y, method = "laplace",
           control = ctrl)
f2 <- tobs(~ occ_cov1 + occ_cov2, data = sim$data, family = occu(),
           detection = ~ det_cov1, y = sim$y, method = "laplace",
           control = ctrl)
ens <- tobs_stack(simple = f1, full = f2)
ens$weights
# weight-combined in-sample psi / p / z
str(predict(ens))
# }