Examples

Every example is a runnable script under examples/, each with a README stating what it shows and its expected output. Run any of them from the repo root:

PYTHONPATH=src python examples/<name>/run.py

Most are covered by the CI suite, so the numbers in their READMEs are checked rather than remembered.

Reproducing published results

Example	Shows
columbus_spatial_econometrics	Anselin (1988), Table 12.1 — OLS reproduces the published values exactly; SAR lands next to the published ML spatial-error estimates, halving the apparent income effect. Walkthrough: case study
state_income_dynamic_network	A network that changes every year — real US state income and contiguity, recovering knocked-out panel cells ~5× better than the baselines. Walkthrough: case study

Start here

Example	Shows
general_lgm	The same model through both the Python and the YAML frontend
count_glm	A Poisson count model on the Laplace engine
method_comparison	pyLGM against a GLM, XGBoost and a Metropolis sampler — see comparison
boosted_offset	Gradient boosting and pyLGM in one linear predictor via offset= — halves the error of either alone, and shows how an in-sample offset silently decalibrates the posterior. Guide: combining with gradient boosting

Spatial and network structure

Example	Shows
disease_mapping	Besag spatial smoothing on Scotland lip cancer; fit to observed counts goes 0.63 → 0.96 against a non-spatial GLM. Walkthrough: disease mapping
weighted_network	BYM2 on a weighted firm-exposure graph — the CAR family on an economic network, not a map
directed_network_sar	SAR on a directed interbank-exposure network, estimating contagion strength ρ
columbus_spatial_econometrics	SAR reproducing Anselin's published Columbus results on real contiguity data
state_income_dynamic_network	DynamicSpatialPanel with one network per year, on 48 US states

Effect modifiers (weights, copy, replicate, group)

Example	Shows
grouped_panel	Grouped vs Replicated — a spatial field whose yearly copies are tied by an AR1, cutting latent error ~25% against independent copies by borrowing strength across years

Time, frequency and forecasting

Example	Shows
midas_nowcast	MIDAS smooth-lag regression of a low-frequency target on high-frequency lags
hybrid_nowcast	MIDAS + BYM2 + AR1 composed into one latent field that fits and predicts
synthetic_panel	A panel with both unit and time structure
nic_backtest	A rolling-origin backtest configuration with covariate-availability rules

Counts, durations and model choice

Example	Shows
count_regression	Horseshoe-crab Poisson GLM matching statsmodels, plus estimated overdispersion. Walkthrough: count regression
survival_duration	WeibullSurv unemployment durations with right-censoring and an IID frailty
predictive_selection	Choosing between candidate models by out-of-sample score

Hyperparameters and posterior integration

Example	Shows
empirical_bayes	Type-II ML estimation of a precision, prior-free
map_ii	The same fit penalised by a PC prior (MAP-II)
inla	Grid integration over hyperparameters instead of a point estimate
inla_sla	Simplified-Laplace latent marginals, with skewness
inla_full_laplace	Full-Laplace latent marginals and their quantiles
inla_criteria	DIC / WAIC / CPO / PIT model assessment