GitHub

Geographically Weighted Regression

GWR

Estimates locally varying coefficients using a kernel-weighted regression at each observation:

m = fit_gwr(X, y, coords; bandwidth=AdaptiveBandwidth(40))

# Local coefficients: m.coef_local (N × K+1 matrix)
# Global R²: m.R2

Bandwidth Types

  • AdaptiveBandwidth(k) — k nearest neighbors (recommended)
  • FixedBandwidth(d) — fixed distance threshold

Kernel Functions

  • GaussianKernel() (default)
  • BisquareKernel()
  • ExponentialKernel()

MGWR (Multiscale GWR)

Each covariate gets its own optimized bandwidth:

m = fit_mgwr(X, y, coords)
# m.bandwidths — per-variable bandwidths
# m.coef_local — N × (K+1) local coefficients

MGWR is preferred when covariates operate at different spatial scales (e.g., temperature varies regionally but policy varies locally).