The regression parameters βzi(z = 0, 1, ...k) in (3.1) are realisations of continuous
functions β(u) at location point i(u) (with u a vector of geographical coordinates),
and ϵi is an independently normally distributed random error term ([9]). If β(u) is
constant across spatially independent sample observations, (3.1) can be estimated
with OLS. If β(u) varies following a decay function which reflects the geographical
principle of distance-decay, (3.1) can be estimated by GWR, which is a weighted least
squares estimator with weights conditioned on location u (eq. (3.2)). As an adaptive
decay weighting scheme, which varies according to the density in the neighbourhood
of each focal point (with less steep weight functions for e.g. regions with more sparse
districts), a bi-square (near-Gaussian) kernel is nested in (3.3) (for v = 2; wi(u) = 0
if di(u) ≥ δi). For samples with irregularly distributed locations and area sizes
such as the case of districts in Niger, adaptive kernels are more suitable than fixed
kernels ([25]). The bandwidth parameter δi delimits the maximum distance of local
spatial dependence around a location, and the related optimal number of kernel
points (Table 3: n) can be chosen based on minimum cross-validation regression
error sum-of-squares (omitting the ith observation) and/or lowest Akaike Information
Criterion (AICc, corrected for small sample bias and tendency to undersmooth in
kernel estimation; [14]: 8; [22]). In this analysis, both criteria have led to selection
of bi-square kernels with the same number of points.
The regression parameters βzi(z = 0, 1, ...k) in (3.1) are realisations of continuousfunctions β(u) at location point i(u) (with u a vector of geographical coordinates),and ϵi is an independently normally distributed random error term ([9]). If β(u) isconstant across spatially independent sample observations, (3.1) can be estimatedwith OLS. If β(u) varies following a decay function which reflects the geographicalprinciple of distance-decay, (3.1) can be estimated by GWR, which is a weighted leastsquares estimator with weights conditioned on location u (eq. (3.2)). As an adaptivedecay weighting scheme, which varies according to the density in the neighbourhoodof each focal point (with less steep weight functions for e.g. regions with more sparsedistricts), a bi-square (near-Gaussian) kernel is nested in (3.3) (for v = 2; wi(u) = 0if di(u) ≥ δi). For samples with irregularly distributed locations and area sizessuch as the case of districts in Niger, adaptive kernels are more suitable than fixedkernels ([25]). The bandwidth parameter δi delimits the maximum distance of localspatial dependence around a location, and the related optimal number of kernelpoints (Table 3: n) can be chosen based on minimum cross-validation regressionerror sum-of-squares (omitting the ith observation) and/or lowest Akaike InformationCriterion (AICc, corrected for small sample bias and tendency to undersmooth inkernel estimation; [14]: 8; [22]). In this analysis, both criteria have led to selectionof bi-square kernels with the same number of points.
การแปล กรุณารอสักครู่..