Spatial Regression of Crime
Also known as: Spatial Lag Model of Crime, Spatial Error Model of Crime, Geographically Weighted Regression of Crime, Spatial Econometric Crime Models
Spatial regression models explain crime rates across areal units — neighborhoods, census tracts, counties — while explicitly accounting for the fact that nearby places tend to have similar crime levels. Ordinary regression assumes each unit's residual is independent, an assumption crime data routinely violate, biasing standard errors and sometimes the coefficients themselves. Spatial econometric models, formalized in Luc Anselin's 1988 framework, introduce a spatial weights matrix and add a spatial lag of the outcome or a spatially correlated error so that the dependence between neighboring areas is modeled rather than ignored.
Key highlights
- Corrects the biased standard errors that ordinary regression produces when crime is spatially autocorrelated.
- Distinguishes substantive spillover (spatial lag) from spatially patterned omitted variables (spatial error).
- Provides formal diagnostics (Moran's I, Lagrange-multiplier tests) to detect and specify spatial dependence.
- Geographically weighted variants reveal how covariate effects vary across a city rather than assuming one global effect.
- Directly operationalizes neighborhood theories of crime such as social disorganization across connected areal units.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use spatial regression of crime when your outcome is a crime count or rate measured across areal units, you have explanatory covariates, and the residuals of an ordinary regression show spatial autocorrelation. It is the standard tool for testing neighborhood theories of crime — social disorganization, concentrated disadvantage — without the false precision that ignoring spatial dependence produces. A spatial lag model suits questions about diffusion and spillover; a spatial error model suits situations where unmeasured causes are spatially patterned; geographically weighted regression suits questions about how relationships vary across a city. It is less appropriate for individual-level data, for point events better handled by point-pattern or kernel methods, or when units are too few to estimate spatial parameters reliably.
Strengths & limitations
- Corrects the biased standard errors that ordinary regression produces when crime is spatially autocorrelated.
- Distinguishes substantive spillover (spatial lag) from spatially patterned omitted variables (spatial error).
- Provides formal diagnostics (Moran's I, Lagrange-multiplier tests) to detect and specify spatial dependence.
- Geographically weighted variants reveal how covariate effects vary across a city rather than assuming one global effect.
- Directly operationalizes neighborhood theories of crime such as social disorganization across connected areal units.
- Results depend heavily on the spatial weights matrix, an analyst choice with no single correct specification.
- Aggregating to areal units invokes the modifiable areal unit problem and the ecological fallacy when inferring about individuals.
- Distinguishing a spatial lag from a spatial error process is statistically difficult and can be theoretically ambiguous.
- Standard models assume Gaussian outcomes, so crime counts often need transformation or count-data spatial extensions.
- Spatial parameters can absorb genuine causal effects, making interpretation of spillover versus confounding delicate.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How do I decide between a spatial lag and a spatial error model?
Start by fitting ordinary regression and testing the residuals for spatial autocorrelation with Moran's I. If dependence is present, the robust Lagrange-multiplier tests on the OLS residuals indicate which specification is more appropriate: the lag model when the outcome itself is spatially dependent (diffusion), the error model when only the disturbances are. Theory matters too — choose the lag model when you have a substantive reason to expect neighbors' crime to influence a unit's crime, and the error model when you suspect spatially patterned omitted variables.
Why does the spatial weights matrix matter so much?
The weights matrix W defines which units are neighbors and how strongly they are linked, and every spatial parameter is estimated relative to that structure. A contiguity-based W, a distance-based W, and a k-nearest-neighbor W can yield different ρ, λ, and even different coefficient estimates. Because there is no single correct choice, good practice grounds W in the substantive process, often row-standardizes it, and reports how results change under alternative neighbor definitions.
Can I use spatial regression with crime counts rather than rates?
The classical spatial lag and error models assume an approximately Gaussian outcome, so analysts often model log rates or transformed counts. For small areas with low counts, where rates are unstable and the normality assumption fails, count-based spatial models — spatial Poisson or negative-binomial regression, or Bayesian conditional autoregressive (CAR) models — are more appropriate and increasingly standard in small-area crime analysis.
Sources
- 1.Anselin, L. (1988). Spatial Econometrics: Methods and Models. Kluwer Academic Publishers.ISBN 9789024737352
- 2.Anselin, L., Cohen, J., Cook, D., Gorr, W., & Tita, G. (2000). Spatial analyses of crime. Criminal Justice 2000, 4, 213–262.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 22). Spatial Regression of Crime. ScholarGate. https://scholargate.app/criminology/spatial-regression-crime