LGMFormula.jl

LGMFormula.jl

Tier-2 formula sugar (@lgm) for LatentGaussianModels.jl.

This package is optional. Core LatentGaussianModels.jl ships a complete explicit-constructor API (Tier 1); LGMFormula.jl adds a macro that expands to those constructor calls (Tier 2). Run @macroexpand @lgm(...) and you'll see a LatentGaussianModel(...) constructor call you could have written by hand.

See plans/macro-policy.md and ADR-008 for the design constraints.

Status

v0.2.0. Phase N PRs 1–6 closed:

• PR-1: core parser + single-likelihood single-f expansion.
• PR-2: component coverage roundtrips (~20 components).
• PR-3: multi-f roundtrip coverage.
• PR-4: tuple-LHS multi-likelihood (wide-format).
• PR-5: replicate / group term routing.
• PR-6: migration guide + vignette parity.

PR-4b (Copy augmentation) and PR-7 (SPDE-friendly coordinate forms) ship in a follow-up.

Quick example

using GMRFs, LatentGaussianModels, LGMFormula

df = (y = y, x = x, region = collect(1:n))
model = @lgm y ~ 1 + x + f(region, IID(n)) data=df family=PoissonLikelihood()
res   = inla(model, df.y)

The macro builds the projector A from the data columns and threads the components tuple and likelihood through to the constructor. See the migration guide for the full R-INLA → @lgm correspondence on Scotland BYM2 and Tokyo rainfall.

Installation

Not yet on the personal Julia registry. From a fresh Julia REPL:

using Pkg
Pkg.develop(url = "https://github.com/HaavardHvarnes/INLA.jl",
            subdir = "packages/LGMFormula.jl")

Why a separate package?

• Julia 1.9+ extensions cannot export new symbols. @lgm needs to be an exported macro, so it cannot live in a weakdep of LatentGaussianModels.
• StatsModels and its dep tree (StatsBase / StatsFuns / DataAPI / ShiftedArrays) is non-trivial. Users who prefer the explicit Tier-1 API should not pay for it.

See also

• LatentGaussianModels.jl — the core LGM package. This sub-package is useless without it.
• Phase N plan — the PR sequence behind this package.