neighbayes¶

neighbayes is a Python package for Bayesian estimation of spatial econometric models. Every model is a class, and specified with a Wilkinson formula via formulaic and a PySAL Graph (or any scipy.sparse matrix) for the spatial weights \(W\); fit() returns an arviz.InferenceData, so the standard posterior tooling (az.plot_trace, az.compare, az.loo) works without translation (though exercise caution as WAIC/LOO diagnostics aren’t valid for spatial models).

import libpysal
from neighbayes.dgp import simulate_sar
from neighbayes.models import OLS, SAR

gdf = simulate_sar(n=400, beta=[1, 0.4, 2.5], rho=0.6, create_gdf=True)
G = libpysal.graph.Graph.build_contiguity(gdf).transform("r")
form = "y ~ -1 + X_0 + X_1 + X_2"

# start with a baseline OLS and test the residuals for spatial dependence
ols = OLS(formula=form, W=G, data=gdf)
ols.fit()
print(ols.spatial_diagnostics_decision(format="ascii"))

# the diagnostics point to SAR here, so fit it and decompose the spillovers
sar = SAR(formula=form, W=G, data=gdf)
sar.fit(draws=2000, chains=4, random_seed=42)
sar.summary()
sar.spatial_effects()

neighbayes provides roughly fifty model classes behind one formula interface, each with a matching simulator in neighbayes.dgp. Every fit comes with Bayesian LM diagnostics (with a decision tree for specification search) and posterior decompositions of direct, indirect, and total effects. Every model (almost) has a custom Gibbs sampler for better mixing (with both numby and jax implementations), but can fall back to PyMC if requested.

Installation¶

conda install -c conda-forge neighbayes   # or: pip install neighbayes

Requires Python 3.12+. The installation guide covers development installs and the optional JAX backend.

Documentation¶

Model-by-model user guides, the diagnostics battery, effects interpretation, and the API reference are at pysal.org/neighbayes.

Validation¶

Every model includes thorough parameter recovery tests, and samplers have been validated against maximum likelihood estimators as well as published results (see docs).