Statistics Glossary → Kernel (weighting) function (Gaussian, bisquare, Epanechnikov)
Kernel (weighting) function (Gaussian, bisquare, Epanechnikov)concept
A function that sets how the weight given to other observations decreases as their distance from a point grows; the basic building block of local estimation.
In kernel density estimation, a small 'bump' is placed around each observation and their sum gives a smooth density surface; in geographically weighted regression (GWR), a separate regression is fitted at each location, with nearby observations weighted more heavily. The Gaussian kernel never takes the weight all the way to zero; bounded kernels such as bisquare, Epanechnikov, quartic and tricube give zero weight to observations beyond the bandwidth; the boxcar gives equal weight to everyone within it.
In practice the results depend far more on the bandwidth than on the type of kernel: a narrow bandwidth gives a detailed but noisy surface, a wide one a smooth surface that erases detail. The Epanechnikov kernel is theoretically the most efficient by a certain criterion, but the differences from other kernels are usually small. For bounded kernels the bandwidth means a radius, while for the Gaussian kernel it acts like a standard deviation, so the same number gives smoothing on a different scale for the two.
An everyday example
A city has air-pollution monitoring stations, and pollution is estimated at the location of a school. A station 500 metres from the school counts strongly, one 3 km away only weakly; with a bisquare kernel, stations beyond 5 km are not counted at all.
In MerQur
'Kernel' (Bisquare, Gaussian, Exponential, Tricube, Boxcar) in GWR; 'Kernel type' (Gaussian, Quartic/biweight, Epanechnikov, Triweight) in the KDE Density Map. In the nonparametric 'Kernel density' method of Discriminant Analysis the kernel (Uniform, Normal, Epanechnikov, Biweight, Triweight) and the radius r are also chosen.
Analyses where this concept appears
Discriminant Analysis (LDA / QDA)GWR (Geographically Weighted Regression)KDE Density Map