Forms a cheap estimate of the latent log-intensity field as the per-site
centred (and optionally scaled) `log1p` of the observed counts. Missing cells
are mean-imputed (0 after centring) and therefore contribute nothing to the
marginal likelihood.
Usage
build_plugin_field(obs_data, n, nt, value = "y_obs", standardise = TRUE)
Arguments
- obs_data
Data frame with `id` (site), `t` (time) and the count column
named by `value`.
- n
Number of sites.
- nt
Number of time points.
- value
Name of the count column (default `"y_obs"`).
- standardise
Logical; scale each site to unit variance after centring
(default `TRUE`).
Value
A numeric vector of length `n * nt`, ordered sites x times (time
fastest).
Details
Per-site centring removes the site intercept `mu_s`; per-site scaling
homogenises per-site variances so a single global `sigma^2` and the
*correlation* kernels apply.