Spatial Doubly Robust Estimation
Also known as: Spatial DR, Spatial AIPW, Spatial augmented IPW, Doubly robust spatial causal estimation
Spatial doubly robust estimation is a semiparametric causal inference method that combines propensity score weighting with outcome regression modeling — providing protection against misspecification of either component — while explicitly accounting for spatial autocorrelation among units. It extends the classical augmented inverse probability weighting (AIPW) estimator to settings where treatment assignment and outcomes are geographically clustered or spatially dependent.
Key highlights
- Doubly robust protection: consistent if either the propensity score model or the outcome regression is correctly specified, reducing sensitivity to individual model misspecification.
- Valid inference under spatial dependence: uses variance estimators that account for autocorrelation, avoiding spuriously narrow confidence intervals common in spatially correlated data.
- Flexible nuisance modeling: accommodates machine-learning outcome and propensity models (via cross-fitting), making it robust to complex, high-dimensional confounding.
- Handles spatial confounding: explicitly models geographically varying covariates and spatial patterns in treatment assignment, capturing confounders that standard methods miss.
- Semiparametric efficiency: under correct specification of both models, achieves the semiparametric efficiency bound for the average treatment effect.
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 doubly robust estimation when you have geo-referenced observational data, a binary or continuous treatment, and reason to believe that (a) spatial autocorrelation in the outcome or treatment assignment may bias naive causal estimates, and (b) you cannot be confident that any single nuisance model is correctly specified. It is well-suited for environmental epidemiology (e.g., pollution exposure effects), urban economics, agricultural trials, and regional policy evaluations. Avoid it when the sample is too small to reliably estimate both a propensity model and a spatial outcome regression, or when units are genuinely independent — in that case a standard AIPW or IPW estimator suffices without added spatial complexity.
Strengths & limitations
- Doubly robust protection: consistent if either the propensity score model or the outcome regression is correctly specified, reducing sensitivity to individual model misspecification.
- Valid inference under spatial dependence: uses variance estimators that account for autocorrelation, avoiding spuriously narrow confidence intervals common in spatially correlated data.
- Flexible nuisance modeling: accommodates machine-learning outcome and propensity models (via cross-fitting), making it robust to complex, high-dimensional confounding.
- Handles spatial confounding: explicitly models geographically varying covariates and spatial patterns in treatment assignment, capturing confounders that standard methods miss.
- Semiparametric efficiency: under correct specification of both models, achieves the semiparametric efficiency bound for the average treatment effect.
- Requires specification and estimation of two separate models (propensity and outcome), both of which must incorporate spatial structure — substantially increasing modeling complexity.
- Spatially-aware variance estimation (e.g., choosing a bandwidth for a spatial HAC estimator) involves tuning decisions that can affect inference materially.
- Cross-fitting for machine-learning nuisances must respect spatial autocorrelation when partitioning data into folds; naive random splits violate independence assumptions.
- Assumes no unmeasured spatial confounders — if spatial confounding is driven by an unobserved variable, both nuisance models are jointly misspecified and the doubly-robust guarantee fails.
- Computationally intensive with large spatial datasets: spatially-varying coefficient or Gaussian process outcome models scale poorly without approximations.
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 does spatial doubly robust estimation differ from standard AIPW?
Standard AIPW assumes independent units and uses non-spatial propensity and outcome models with conventional variance estimators. Spatial DR incorporates geographic covariates or spatially-varying models into both nuisance components and uses variance estimators — such as spatial HAC or clustered bootstrap — that remain valid under autocorrelation.
What does 'doubly robust' mean in this context?
The estimator is consistent for the average treatment effect if either the propensity score model or the outcome regression model is correctly specified, but not necessarily both. This gives protection against misspecification of one component, though it does not protect against misspecification of both simultaneously.
How should I partition data for cross-fitting in a spatial setting?
Use spatial folds — contiguous geographic blocks rather than random splits — so that observations in the held-out fold are spatially separated from the training fold. Random splits allow spatial leakage across folds, biasing nuisance estimates and invalidating the cross-fitting guarantee.
What happens if there is spatial interference (spillovers)?
Standard SUTVA is violated and the estimator targets an undefined individual treatment effect. You should clearly define your estimand (e.g., direct vs. spillover effects) and use an estimation framework that explicitly models interference, such as network causal inference or partial population experiments.
How do I choose the bandwidth for a spatial HAC variance estimator?
Common choices are based on the empirical range of spatial autocorrelation in the residuals (e.g., from a variogram), or by data-driven bandwidth selection methods analogous to those used in time-series HAC estimation. Sensitivity analysis across plausible bandwidth values is strongly recommended.
Sources
- 1.Papadogeorgou, G., Mealli, F., & Zigler, C. M. (2019). Causal inference with interfering units for cluster and population level treatment allocation programs. Biometrics, 75(3), 778-787.
- 2.Kennedy, E. H. (2016). Semiparametric theory and empirical processes in causal inference. In H. He, P. Wu, & D.-G. Chen (Eds.), Statistical Causal Inferences and Their Applications in Public Health Research (pp. 141-167). Springer.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Spatial Doubly Robust Estimation. ScholarGate. https://scholargate.app/causal-inference/spatial-doubly-robust-estimation