MerQuR+Statistics Glossary

Statistics Glossary → Inverse distance weighting (IDW)

Inverse distance weighting (IDW)concept

A method that estimates the value at an unknown location by weighting the values at surrounding known points by the inverse of a power of their distance.

IDW rests on the principle that 'near things are more alike than distant things'. Each measurement point at distance d from the target receives weight 1/d^p, and the estimate is the weighted average ẑ = Σ wᵢzᵢ / Σ wᵢ. The larger the power p, the more the nearest points dominate and the more 'local' the surface; p = 2 is the most common choice. If the target coincides with a measurement point, that point's value is used directly. The same idea is used in k-nearest neighbours to weight the neighbours' votes by distance.

IDW is simple and fast but deterministic: it does not model the spatial structure of the data (the variogram) and gives no uncertainty (standard error) for its estimates. Estimates always stay between the smallest and largest observed values, can form 'bull's-eye' rings around measurement points, and give too much weight to clustered points. Kriging, in contrast, models spatial dependence through the variogram and provides the best linear unbiased prediction together with its prediction variance; the two methods can be compared with their leave-one-out (LOO) cross-validation error.

An everyday example

You estimate the temperature of a neighbourhood using five weather stations, one 2 km away and the others 10 km away. With p = 2 the near station gets weight 1/4 and each distant station 1/100, so the near station has about six times more say than the other four combined. If the near station reads 18 °C and the others 22 °C, the estimate comes out at about 18.6 °C.

In MerQur

In Variogram + Kriging, the 'Compare with IDW' checkbox and the 'IDW power (p)' parameter compare it with kriging by LOO error; in k-Nearest Neighbors, the 'Weighting (Equal, By distance (1/d))' option weights neighbours by 1/distance. SAR and the Spatial Error Model have an 'Inverse distance 1/d' choice in the 'Neighbourhood method', and Getis-Ord Gi* has 'Inverse distance (within band)' in its 'Neighbourhood' option.

Analyses where this concept appears

SAR (Spatial Autoregressive Lag)Spatial Error Model (SEM)k-Nearest Neighbors (KNN)Getis-Ord Gi* HotspotVariogram + Kriging (Spatial Interpolation)

Related concepts

Distance measure (Euclidean, Manhattan, Gower, great-circle)Kriging (simple, ordinary, universal, block)