neighbayes.dgp.simulate_sar_poisson_flow

neighbayes.dgp.simulate_sar_poisson_flow(n=None, rho_d=0.3, rho_o=0.2, rho_w=0.1, mean_count=8.0, signal=0.3, seed=42, *, T=None, pair_effects=None, time_effects=False, separable=False, pair_effect_sd=None, time_effect_sd=None, **kwargs)[source]

Generate synthetic O-D flow counts from a Poisson SAR flow DGP.

\[y_{ij} \sim \operatorname{Poisson}(\mu_{ij}), \qquad \log \boldsymbol{\mu} = A(\rho_d, \rho_o, \rho_w)^{-1} X\beta\]
Parameters:
n : int, optional

Approximate number of spatial units (N = n^2 flows).

rho_d : float

Destination, origin and cross spatial parameters.

rho_o : float

Destination, origin and cross spatial parameters.

rho_w : float

Destination, origin and cross spatial parameters.

mean_count : float, default 8.0

Target mean of the simulated counts.

signal : float, default 0.3

Magnitude of the regression coefficients, which sets the spread of eta. The default gives sd(eta) ~ 1 and roughly 5-10% zeros. Larger values give heavy-tailed designs with many zeros (signal=0.9: sd(eta) ~ 3, 40-50% zeros, counts in the hundreds), which the Poisson flow samplers recover. An earlier note here said they could not; that was the last-arrival augmentation bug fixed on 2026-10-03, which biased large counts. Sparsity itself is not the obstacle: what limits ρ is the total information, so a low mean_count on a small n gives a correct but wide posterior (Gibbs matches NUTS there).

seed : int, default 42

Random seed.

**kwargs

Forwarded to simulate_sar_negbin_flow().

Returns:

As simulate_sar_negbin_flow(), without alpha.

Return type:

dict