Julia posteriordb implementations

StanBlocks keeps its PosteriorDB implementations in the optional PosteriorDBExt extension. Loading PosteriorDB activates that extension and adds slic_implementation(::Val{:model_name}; data...) methods for the supported catalogue entries.

Why this page links to the extension

The legacy Quarto page used an include directive to paste the entire extension source into the rendered notebook. That produced a second, enormous copy of a maintained source file without turning it into a worked example. In the main docs, the extension itself is the canonical, line-addressable implementation; copying it here would drift immediately.

The implementations demonstrate a consistent set of StanBlocks mechanisms:

• per-model Val dispatch selects an implementation without a stringly typed runtime switch;

• @deffun supplies catalogue-specific recurrences and likelihood helpers;

• @lhs/@lpxf register custom distribution families for ordinary ~ statements and generated likelihood/prediction companions;

• @slic attaches PosteriorDB data and builds each model from the same public authoring surface used elsewhere in these examples.

Maintained model inventory

The extension currently defines 77 named implementations. Related names are variants of one statistical example rather than unrelated snippets:

• Generic GLM, rate, and introductory models: GLM_Binomial_model, GLM_Poisson_model, GLMM_Poisson_model, Rate_1_model, Rate_2_model, Rate_3_model, Rate_4_model, Rate_5_model, blr, logistic_regression_rhs, pilots, surgical_model, diamonds, dugongs_model, kilpisjarvi, and accel_splines. These cover binomial and Poisson regression, random effects, nonlinear growth, and spline bases.

• Canonical hierarchical and applied examples: eight_schools_centered, eight_schools_noncentered, seeds_model, seeds_centered_model, seeds_stanified_model, sesame_one_pred_a, election88_full, nes, nes_logit_model, and bym2_offset_only. The paired names expose centered/non-centred parameterisations or alternative spellings of the same likelihood.

• Child test-score regressions: kidscore_momhs, kidscore_momiq, kidscore_momhsiq, kidscore_mom_work, kidscore_interaction, kidscore_interaction_c, kidscore_interaction_c2, and kidscore_interaction_z. The suffixes add predictors, interactions, centering, or standardisation while preserving the same outcome.

• Earnings and height regressions: earn_height, log10earn_height, log10earn_height_male, logearn_height, logearn_height_male, logearn_interaction, logearn_interaction_z, and logearn_logheight_male. These make the response/predictor transforms and sex interaction explicit in the implementation name.

• Mesquite regressions: mesquite, logmesquite, logmesquite_logvolume, logmesquite_logva, logmesquite_logvas, and logmesquite_logvash. Successive variants change the response scale and add transformed canopy, shrub, and height predictors.

• Well-switching regressions: wells_dist, wells_dist100_model, wells_dist100ars_model, wells_dae_model, wells_dae_c_model, wells_dae_inter_model, wells_daae_c_model, wells_interaction_model, and wells_interaction_c_model. The family varies distance scaling, arsenic, education, interactions, and centered predictors.

• Radon multilevel family: radon_pooled, radon_county, radon_county_intercept, radon_partially_pooled_centered, radon_partially_pooled_noncentered, radon_hierarchical_intercept_centered, radon_hierarchical_intercept_noncentered, radon_variable_intercept_centered, radon_variable_intercept_noncentered, radon_variable_slope_centered, radon_variable_slope_noncentered, radon_variable_intercept_slope_centered, and radon_variable_intercept_slope_noncentered. Together they cover the pooling ladder and the centered/non-centred hierarchy choices.

• Time-series, Gaussian-process, mixture, and spatial specialists: arma11, garch11, gp_regr, gp_pois_regr, normal_mixture, low_dim_gauss_mix, and low_dim_gauss_mix_collapse. These are where the extension-local recurrence, custom-density, GP, and mixture helpers are most visible.

This inventory is intentionally explanatory rather than a pasted source dump. For example, every _centered/_noncentered pair is a parameterisation comparison, while the radon, wells, kid-score, earnings, and mesquite suffixes record the covariate/model-building progression. The full method bodies at the source link remain the executable authority.

For small, side-by-side executable examples of those mechanisms, use the feature atlas. For the complete catalogue mapping, read the extension source above; it is what the package loads and tests.