NEWS.md
In stLMM: Bayesian Spatial and Space-Time Linear Mixed Models

stLMM 0.0.5

Use a size_t counter when iterating over CHOLMOD supernodes, removing a signed/unsigned comparison warning.
CRAN source builds exclude local article caches and the data downloads used only by website articles.
CAR terms now retain every node in a supplied graph even when some areas have no rows in the fitting data. Missing-response placeholder rows are no longer required for these areas; recovery and prediction retain the full graph support. This applies to proper and Leroux CAR terms, including spatially varying coefficients.
Fits now record monotonic elapsed seconds for compiled initialization, proposal-calibration warmup, and the post-warmup MCMC loop separately, in fit$timing$phases. Sampling time includes all work in the loop, including repeated factorizations and any requested in-chain latent-effect recovery.
The former timing$sampler field is renamed timing$backend to describe its existing scope accurately: the entire compiled call. Multi-chain fits use backend_by_chain/backend_total and add phases_by_chain/phases_total. The old timing names are not aliases. Saved fits from builds without phase timing cannot supply sampling-only timings. Config/stLMM/Timing: phases-v1 identifies this capability in DESCRIPTION; the package version remains 0.0.5 during development.
Updated printed timing labels and the runtime, blocking-diagnostics, and NIMBLE-comparison articles to distinguish sampling from broader elapsed times.
CAR graph connectivity checks now preallocate their traversal queue, avoiding repeated vector copying on large graphs while preserving component labels and graph validation. This also benefits DAGAR and space-time models that use the same graph constructor.

stLMM 0.0.4

Added family = "probit" for binary probit mixed models using Albert-Chib latent-normal augmentation.
Added fitted values, prediction, log-likelihood, recovery, tests, and documentation for probit models.
Added support for structured probit process models using saved in-chain latent process draws.

NNGP prediction with joint = TRUE now defaults to joint_method = "vecchia" instead of the dense full method. This avoids building a dense covariance over new prediction nodes for moderate and large prediction sets. Use joint_method = "full" to request the previous dense behavior for small prediction sets.
stats::binomial(link = "probit") now dispatches to the probit likelihood.
Unsupported stats::binomial() links such as "cloglog" and "cauchit" now error instead of silently using the logit/Polya-Gamma path.