An integrated NKDE-MGWR framework for smoothing-scale and network-coverage sensitivity in rural highway crash modeling.

Crash-model estimates can depend on how crash events are spatially smoothed, which roadway segments are retained, and whether global or spatially varying regression is used. This study evaluated these sensitivities on 1071 rural divided multilane highway segments (1064.098 km) in Hamadan Province, Iran, using 2017-2019 data. Network kernel density estimation (NKDE) produced continuous crash-density outcomes at 1000 and 4000 m. Ordinary least squares (OLS), geographically weighted regression (GWR), and multiscale geographically weighted regression (MGWR) were estimated under a full-network regime and a selected-segment regime formed by removing 365 zero-crash removable segments while preserving connectivity. Raw crash frequency was retained as an unsmoothed benchmark and evaluated separately with Poisson and negative-binomial regression. Spatial models achieved higher adjusted R2 than OLS in the NKDE-based cases. The full-network 4000-m MGWR specification yielded the highest within-sample explanatory performance (adjusted R2 = 0.801), although MGWR was not uniformly superior across all cases. Segment removal did not consistently improve model performance. The 1000-m and 4000-m surfaces emphasized localized and corridor-scale patterns, respectively, and should be treated as complementary analytical scales. For raw crash frequency, the negative-binomial model showed better distributional adequacy than Poisson (AICc 4309.51 vs. 6001.62; dispersion ratio 0.988 vs. 4.588), although residual spatial dependence remained. Within the focused nine-predictor diagnostic subset, global collinearity was unchanged across NKDE outcomes because the predictor matrix was fixed, whereas local conditioning was sensitive to the GWR calibration neighbourhood. The framework supports scale-aware network screening while requiring count-specific modeling and engineering validation.