MerQuR+Statistics Glossary

Statistics Glossary → Spatial lag and spatial Durbin models (SAR, SDM, SDEM)

Spatial lag and spatial Durbin models (SAR, SDM, SDEM)concept

Regression models that use a spatial weights matrix W to model how the outcome at one location is influenced by outcomes, or by explanatory variables, at neighbouring locations.

A spatial lag is the weighted average of a variable's values in neighbouring units (Wy or WX); W defines which units count as neighbours and what weight they receive. The SAR (spatial autoregressive lag) model is y = ρWy + Xβ + ε: neighbours' outcomes directly affect a unit's outcome, and this effect spreads through the whole system by feedback. The spatial Durbin model (SDM) adds the neighbours' explanatory variables (WXθ). The spatial Durbin error model (SDEM) contains WX and a spatially correlated error term instead of Wy; there is no feedback in this model.

Because of feedback, the β coefficients in SAR and SDM cannot be read directly as marginal effects; direct, indirect (spillover) and total effects (LeSage and Pace) must be computed. In SDEM, by contrast, the direct effect is β and the indirect effect θ. Since SDM nests simpler models, likelihood-ratio tests can check whether it reduces to SAR (θ = 0) or to the spatial error model (the common-factor restriction θ = -ρβ). Results are sensitive to the choice of W, and the models are estimated by maximum likelihood or instrumental-variable methods rather than ordinary least squares.

An everyday example

House prices in a district depend not only on that district's features but also on prices in neighbouring districts: when prices rise next door, buyers turn to this district and prices here rise too, which in turn feeds back on the neighbour. The SAR model captures this spread. A new park opening in a neighbouring district and affecting prices here is an effect of the neighbours' explanatory variable, modelled with SDM or SDEM.

In MerQur

It is chosen with the 'Model (SAR / Spatial Durbin SDM)' parameter in SAR and 'Model (SEM / SDEM)' in the Spatial Error Model; SDM-SAR and SDM-SEM likelihood-ratio tests and direct, indirect and total effects are reported.

Analyses where this concept appears

SAR (Spatial Autoregressive Lag)Spatial Error Model (SEM)

Related concepts

Spatial weights matrix (W)Spatial autocorrelation (Moran's I, Geary's C)