API reference

Latte fits latent Gaussian models (LGMs). The primary way to define one is the @latte macro: it reads your model body, classifies each ~ block, and returns a model object. That object is engine-agnostic — the same definition runs through inla, tmb, or hmc_laplace, and every engine returns results through one shared accessor API.

Latent Gaussian models

Every Latte engine targets the same model class:

$$p(\theta ,x,y)=\underset{\text{hyperprior}}{\underbrace{p(\theta )}}\underset{\text{Gaussian latent field (GMRF)}}{\underbrace{p(x\mid \theta )}}\underset{\text{cond.-independent observations}}{\underbrace{p(y\mid x)}}$$

A Gaussian latent field $x$ (a GMRF) governed by a few hyperparameters $\theta$, with observations conditionally independent given the latent field. GLMMs, spatial (Besag, SPDE) and temporal (AR, random-walk) models, splines, and disease mapping all fit this template. You specify the three pieces — the hyperparameter priors, the latent GMRF, and the observation model — once, and any engine consumes it.

What's here

• Defining models: the @latte macro — the primary entry point, how the macro builds the model, and what it does not yet handle.

• Working with results — the accessors that read latent and hyperparameter marginals off any engine's result.

• Lower-level construction — building a model directly from @hyperparams and LatentGaussianModel when you need full control.

Component pages cover the individual pieces:

• Observation models — the likelihood $p(y\mid x)$.

• Gaussian approximation — the inner Gaussian fit to $p(x\mid y,\theta )$.

• Marginalization — recovering latent-field marginals.

• Hyperparameter posterior — exploring and marginalizing over $\theta$.

For worked, end-to-end models see the tutorials; for the inference methods themselves and how to tune each, see the engine pages.