Spatial Entropy Balancing
Also known as: spatial EB, geographically-weighted entropy balancing, spatial reweighting
Spatial entropy balancing extends standard entropy balancing to observational settings where units are embedded in geographic space, incorporating spatial structure into the reweighting process so that balance is achieved while respecting spatial proximity, clustering, or spillover dependencies between units.
Key highlights
- Achieves exact covariate balance on specified moments without iterative caliper tuning or propensity model specification.
- Explicitly incorporates geographic proximity into counterfactual selection, reducing confounding from spatially varying unobservables.
- Preserves maximum entropy among weights, retaining more of the original data diversity than nearest-neighbor or caliper matching.
- Transparent and reproducible: the optimization problem has a unique solution and balance is verifiable by construction.
- Extends naturally to multiple spatial scales by choosing different neighborhood radii or spatial kernel bandwidths.
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 entropy balancing when you have observational data where treatment assignment and outcomes have a geographic dimension, and spatial proximity is a plausible source of confounding or spillover. It is appropriate when you need exact covariate balance but also want to restrict comparisons to geographically plausible counterfactuals — for example, comparing treated municipalities to nearby untreated ones. It is not appropriate when the data have no meaningful spatial structure, when the treated and control units are completely geographically segregated with no overlap, or when sample sizes are too small to support stable reweighting (fewer than roughly 30 control units).
Strengths & limitations
- Achieves exact covariate balance on specified moments without iterative caliper tuning or propensity model specification.
- Explicitly incorporates geographic proximity into counterfactual selection, reducing confounding from spatially varying unobservables.
- Preserves maximum entropy among weights, retaining more of the original data diversity than nearest-neighbor or caliper matching.
- Transparent and reproducible: the optimization problem has a unique solution and balance is verifiable by construction.
- Extends naturally to multiple spatial scales by choosing different neighborhood radii or spatial kernel bandwidths.
- Requires a defensible choice of spatial weights matrix or bandwidth; results can be sensitive to this choice and sensitivity should be reported.
- Exact balance is achieved only on the moments the researcher specifies; unmeasured spatially varying confounders remain unaddressed.
- With small control pools (after spatial restriction) the reweighting may concentrate mass on very few units, inflating variance.
- The spatial restriction may eliminate valid distant controls that happen to be good counterfactuals on observed covariates.
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 is spatial entropy balancing different from standard entropy balancing?
Standard entropy balancing reweights any control unit to match covariate moments, regardless of geographic distance. Spatial entropy balancing adds constraints or restrictions that limit or penalize the use of geographically distant controls, so that the balanced sample also respects spatial proximity.
How do I choose the spatial bandwidth or neighborhood radius?
The bandwidth should be chosen on substantive grounds — how far could spatial spillovers plausibly reach? — and then tested for sensitivity. Reporting results across a range of bandwidths is recommended to show robustness.
Do I need a spatial weights matrix?
Not necessarily. A simple radius or k-nearest-neighbor restriction on the eligible control pool is sufficient in many applications. A formal spatial weights matrix is more useful when incorporating spatial autocorrelation in the outcome model as well.
What if the optimization returns extreme weights on a few control units?
This usually signals that the spatial restriction has left too few control units or that the moment constraints are over-specified. Consider widening the spatial bandwidth, reducing the number of moment conditions, or checking for geographic overlap between treated and control pools.
How do I compute standard errors?
Bootstrap the entire procedure — spatial restriction, optimization, and ATT calculation — in each replicate. Do not use analytical standard errors that treat weights as fixed, as this ignores the estimation uncertainty from the reweighting step.
Sources
- 1.Hainmueller, J. (2012). Entropy Balancing for Causal Effects: A Multivariate Reweighting Method to Produce Balanced Samples in Observational Studies. Political Analysis, 20(1), 25-46.
- 2.Entropy balancing. Wikipedia.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Spatial Entropy Balancing. ScholarGate. https://scholargate.app/causal-inference/spatial-entropy-balancing