MerQuR+Statistics Glossary

Statistics Glossary → Bandwidth (kernel smoothing)

Bandwidth (kernel smoothing)concept

In smoothing methods such as kernel density estimation and local regression, the parameter that sets how far away observations can be and still contribute to the estimate at a point, and hence how smooth the result is.

Kernel methods compute the estimate at each point as an average of the surrounding observations, with weights that decline with distance. The bandwidth is the size of this neighbourhood: in a Gaussian kernel it is the kernel's standard deviation, and in geographically weighted regression (GWR) it is either a fixed radius or the distance to the k nearest neighbours (an adaptive bandwidth). An adaptive bandwidth narrows where data are dense and widens where they are sparse.

Choosing the bandwidth is a bias-variance trade-off: a narrow bandwidth gives a rough, jumpy result that follows the noise, and a wide one gives an over-smoothed result that wipes out real detail. It can be chosen with rule-of-thumb formulas such as Silverman's and Scott's (which work well when the data are roughly normal), or with data-driven criteria such as cross-validation or AICc. If a theoretical scale is known (such as walking distance), choosing it by hand also makes sense. Because results can be sensitive to the bandwidth, the chosen value and method should be reported.

An everyday example

A density map of café locations is drawn for a city. With a 100-metre bandwidth, the map shows a separate little peak around each café and the overall pattern is lost; with 5 kilometres, the whole city becomes a single vague hill. A bandwidth of around 500 metres clearly reveals the few central districts where cafés cluster.

In MerQur

It appears as the 'Bandwidth' (Silverman, Scott, Manual, Scott normal, Likelihood CV, Berman-Diggle CV) and 'Bandwidth (km)' options in KDE Density Map, and as 'Bandwidth type' (adaptive k neighbours or fixed radius), 'Manual bandwidth' and 'Model' (GWR single bandwidth or MGWR covariate-specific bandwidths) in GWR. In Quantile Regression, the robust choice under 'SE / CI method' computes standard errors with the Hall-Sheather bandwidth and the Epanechnikov kernel.

Analyses where this concept appears

Quantile RegressionGWR (Geographically Weighted Regression)KDE Density Map

Related concepts

Kernel (weighting) function (Gaussian, bisquare, Epanechnikov)Kernel density estimation (KDE)Information criteria (AIC, AICc, BIC)