When Finer Maps Can Help Untangle Spatial Relationships

More detail, not more territory

Spatial datasets often become larger by dividing the same territory into smaller units: states become counties, and counties give way to census tracts. Christopher W. Engh’s preliminary preprint studies this setting, known as infill or fixed-domain asymptotics. Unlike approaches that add observations by expanding the study area, it asks what researchers can learn by looking more closely within an existing map.

The proposed source of information is local variation in treatment after accounting for controls. Broad geographic patterns may reflect omitted influences correlated with treatment. Nevertheless, finer contrasts can potentially reveal the treatment coefficient if the remaining confounding contributes sufficiently little to the estimate. Engh calls this combination of requirements “locally exogenous variation.” It does not require treatment to be uncorrelated with omitted influences in levels.

A model for regional averages

The paper starts with a spatial linear model relating an outcome to a treatment, controls and an error. Its target is one treatment coefficient, assumed constant across locations. Researchers observe regional averages rather than the underlying spatial functions.

Because the coefficients are constant, averaging preserves the model exactly. The theoretical experiment holds the underlying spatial functions fixed while observing increasingly fine regional partitions. It also keeps the number and identity of control functions fixed.

Estimation uses generalized least squares, or GLS. This method weights observations using a working covariance matrix describing assumed error variances and correlations. In the paper, researchers supply this matrix using geography and any conditioned-on regressor information; its selection cannot use the current outcome disturbances.

What makes finer observations informative?

A central quantity, called Q, measures treatment variation remaining after controls, weighted by the inverse working covariance. The consistency result requires Q to increase without bound as the regions become finer. Merely increasing the observation count is not enough.

Two further restrictions matter. First, the true variance of every weighted combination of regional errors must be bounded above by a common finite multiple of its working variance. Second, after subtracting a fixed combination of controls, the conditional error mean must be small relative to treatment information in the paper’s weighted measure.

Under these and the stated regularity conditions, Engh proves that the estimated treatment coefficient converges in probability to its target. This result permits persistent covariance misspecification and does not require Gaussian errors. It does not establish that arbitrary spatial confounding disappears when a map becomes more detailed.

Confidence intervals demand more

Reliable uncertainty estimates require stronger assumptions. The paper assumes a conditionally Gaussian centered error field and a working covariance that becomes uniformly accurate, up to an unknown positive scale, across all linear combinations of regional errors.

It also requires a bounded weighted measure of one fixed conditional error-mean function after controls. Together with the other assumptions, this makes slope bias negligible relative to the slope’s standard deviation, rather than merely making bias vanish.

The calculation uses one GLS regression, followed by a residual-based estimate of the covariance scale. The reported slope variance is that scale estimate divided by Q. Engh proves asymptotic validity for the resulting normal confidence intervals under the stronger conditions—not exact finite-sample coverage under confounding.

What the simulations show

The simulation uses all 51 U.S. state/DC regions and all 3,144 county-equivalents, with 500 simulated outcome surfaces per scenario and geographic level. Nominal 95% interval coverage was 94.2% for states and 96.2% for counties under the matched covariance.

A deliberately severe mismatch produced 85.0% coverage for states but 95.8% for counties. That mismatch simultaneously changed the covariance family, distance metric and range, so it cannot isolate any one change’s effect. The county result is compatible with nominal coverage in this particular design, not proof of general robustness.

The simulation also differs from the theoretical regional-average setting: it uses deterministic representative points and point-support covariances, not exact polygon-integral covariances. Census tracts were excluded because dense covariance calculations were outside the implementation’s computational scope.

Assumptions remain substantive

The paper explicitly notes that information divergence cannot be verified from a single finite dataset and that restrictions on unobserved confounding need substantive justification. Estimating covariance-shape parameters from the same outcome errors remains a separate problem. Its contribution is therefore a conditional framework, implemented through the accompanying infillGLS R package: finer geography can help, but additional detail alone guarantees neither identification nor trustworthy confidence intervals.