📈 economics

Bandwidth Selection for Spatial HAC Standard Errors

This paper addresses the unresolved challenge of bandwidth selection for spatial HAC standard errors by documenting an inverse-U relationship between bandwidth and error magnitude, and subsequently proposing a novel, data-driven non-parametric selector that effectively controls false positive rates across diverse spatial correlation structures.

Original authors: Alexander Lehner

Published 2026-03-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alexander Lehner

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery using data. You have a map of the United States, and you want to see if two things are connected: for example, does the percentage of Black residents in a county predict the percentage of Hispanic residents?

In the world of statistics, you usually draw a line through your data points to find the answer. But there's a catch: things that are close together on a map tend to be similar. This is called "spatial autocorrelation." If a county in Texas has a high Hispanic population, the counties right next to it probably do too. They aren't independent; they are "friends" who hang out together.

The Problem: The "Too Quiet" Detective

When you ignore this friendship, your statistical tools get confused. They think they have more independent evidence than they actually do. It's like a detective who thinks they have 100 different witnesses, but in reality, all 100 witnesses are just the same person repeating the story.

Because the detective thinks they have so much evidence, they become overconfident. They say, "I'm 99% sure this connection is real!" when it might just be a coincidence. In statistics, this leads to false positives—finding patterns that aren't actually there.

To fix this, statisticians invented a tool called Conley Standard Errors. Think of this tool as a "correction lens" that tells the detective, "Hey, these neighbors are friends, so don't count them as 100 separate witnesses."

But here is the big problem: How wide should the lens be?

If the lens is too narrow, you miss the friends who are a little further away.
If the lens is too wide, you start counting strangers who have nothing to do with the story.

For years, researchers had no good rule for setting this lens. They just guessed. The common advice was: "Just make the lens huge to be safe." The thinking was, "If I include everyone, I can't possibly miss anything."

The Big Discovery: The "Goldilocks" Curve

Alexander Lehner, the author of this paper, discovered that the "bigger is better" advice is actually wrong.

He found that the relationship between the lens size and the accuracy of your results looks like a hill (an "inverse-U" shape):

Too Narrow (The Left side of the hill): You miss the friends. You are overconfident and find fake patterns.
Too Wide (The Right side of the hill): You start counting strangers. Surprisingly, this also makes you overconfident! It's like trying to hear a whisper in a crowded room by turning the volume up so high that you only hear static. The "correction" becomes so messy that it actually makes your standard errors smaller than they should be, leading you to trust your results too much.
Just Right (The Peak of the hill): You include exactly the right group of friends. This is where your results are most accurate.

The Analogy: Imagine you are trying to measure the temperature of a room.

If you only measure one spot (too narrow), you might miss a draft.
If you measure the whole house including the freezer and the oven (too wide), your average temperature reading becomes nonsense.
You need to measure just the living room where the people are sitting.

The Solution: The "Zero-Crossing" Map

Lehner proposes a simple, smart way to find that "Just Right" spot without guessing. He suggests looking at the residuals (the parts of the data your model couldn't explain).

Think of the residuals as the "leftover noise" after you've drawn your line.

He calculates how similar the noise is between two points as they get further apart.
He draws a graph of this similarity.
He looks for the exact distance where the similarity crosses zero.

The Metaphor: Imagine you are shouting across a canyon.

Close to you, your friend hears you clearly (high correlation).
As they walk away, your voice gets quieter.
Eventually, you reach a point where your friend can no longer hear you at all; the sound is just random wind noise (zero correlation).

Lehner's method says: "Stop counting neighbors once the sound of your data hits the wind." That distance is the perfect size for your lens.

What Happens When You Use This?

Lehner ran thousands of computer simulations (like playing the game "Statistical Detective" 5,000 times) to test his idea.

The Old Way (Guessing): Researchers who used a "super wide" lens thought they were being safe, but they were actually finding fake patterns about 30–60% of the time!
The New Way (The Zero-Crossing): By using his method, the researchers only found fake patterns about 5% of the time (which is the correct, honest rate).

He also tested different "shapes" for the lens (called kernels) and found that two specific shapes, the Bartlett and Epanechnikov kernels, worked the best.

Real World Example

He tested this on real US Census data (Black population vs. Hispanic population).

Using a tiny lens (25 km), the result looked very significant.
Using a massive, "safe" lens (2,500 km), the result also looked significant.
Using his smart lens (which calculated the distance where the correlation stopped, about 980 km), the result suddenly became insignificant.

This means that for this specific data, the "connection" people thought they saw was actually just a statistical illusion caused by using the wrong lens size.

The Takeaway

This paper gives researchers a simple, automatic rule to stop guessing. Instead of blindly making their safety net huge, they can now look at their data, find the distance where the "friendship" between data points ends, and set their lens exactly there.

It's like finally having a GPS for your statistical analysis, ensuring you don't get lost in a forest of false discoveries. The author even built a free tool (an R package called SpatialInference) so anyone can use this "smart lens" immediately.

1. Problem Statement

In applied econometrics, spatial autocorrelation in regression residuals leads to downward-biased standard errors when using conventional heteroskedasticity-robust (HC) estimators, resulting in inflated $t$ -statistics and incorrect inference (over-rejection of the null hypothesis). The standard correction is the Spatial Heteroskedasticity and Autocorrelation Consistent (HAC) estimator introduced by Conley (1999).

However, a critical unresolved issue in the literature is the selection of the kernel bandwidth ( $\zeta$ ), which defines the distance within which residuals are assumed to be correlated.

Current Practice: Researchers typically rely on ad-hoc choices, intuition, or domain knowledge, creating "researcher degrees of freedom."
Conventional Wisdom: There is a prevailing belief that choosing a wider bandwidth yields more conservative (larger) standard errors and safer inference.
The Gap: No formal, data-driven guidance exists for selecting this bandwidth in spatial settings, unlike the rich literature available for time-series HAC (e.g., Newey-West).

2. Methodology

A. The Inverse-U Relationship

The author first documents a fundamental empirical regularity regarding the relationship between bandwidth and the magnitude of spatial HAC standard errors:

Shape: The relationship is inverse-U shaped.
Mechanism:
- Too Narrow: Underestimates variance because it ignores valid spatial correlations.
- Optimal: Standard errors peak near the true correlation range.
- Too Wide: Standard errors decrease as the bandwidth expands beyond the true correlation range. This occurs because the estimator begins averaging in pairs of residuals that are uncorrelated (noise), which dilutes the positive covariance contributions from truly correlated pairs.
Implication: Contrary to conventional wisdom, using an excessively wide bandwidth can lead to underestimated standard errors and even more severe over-rejection than ignoring spatial correlation entirely.

B. The Proposed Solution: Covariogram Range Method

Lehner proposes a simple, non-parametric, data-driven bandwidth selector:

Estimate Residuals: Obtain regression residuals ( $\hat{\varepsilon}_i$ ).
Compute Empirical Covariogram: Calculate the average product of residuals ( $\hat{\varepsilon}_i \hat{\varepsilon}_j$ ) for pairs of observations binned by their distance ( $h$ ).
$\hat{C}(h) = \frac{1}{|N(h)|} \sum_{(i,j) \in N(h)} \hat{\varepsilon}_i \hat{\varepsilon}_j$
Select Bandwidth: The bandwidth ( $\hat{\zeta}$ ) is defined as the first distance at which the empirical covariogram crosses zero (or falls below a small tolerance $\eta$ ).
$\hat{\zeta} = \min \{ h_b : |\hat{C}(h_b)| \leq \eta \}$
Theoretical Basis: Under assumptions of stationarity, mixing, and compact support (or effective range), this estimator is consistent for the true correlation range. The method anchors the bandwidth to the physical scale of dependence rather than shrinking it relative to sample size (avoiding the undersmoothing issues of MSE-optimal rules).

C. Kernel Functions

The study evaluates six kernel functions to determine which provides the best size control:

Uniform, Bartlett, Epanechnikov, Parzen, Quartic Biweight, and Gaussian.

3. Key Contributions

Refutation of Conventional Wisdom: The paper formally demonstrates that "wider is better" is false for spatial HAC. Both narrow and excessively wide bandwidths lead to underestimation of standard errors.
Novel Estimator: Introduction of the Covariogram Range Method, a fully non-parametric selector that does not require parametric pilot estimators (unlike Kim and Sun, 2011) or complex spatial differencing.
Systematic Kernel Comparison: A comprehensive evaluation of six kernel functions in the spatial domain, identifying the Bartlett and Epanechnikov kernels as superior for size control.
Implementation: The method is implemented in the R package SpatialInference.

4. Results

Monte Carlo Simulations

The author conducted extensive simulations (5,000 iterations) calibrated to the geography of the contiguous United States using four sample configurations (county centroids and regular lattices of varying densities).

Size Control: The proposed covariogram-range method (using Bartlett or Epanechnikov kernels) maintains the false positive rate (FPR) at or near the nominal 5% level across a wide range of spatial correlation intensities.
Comparison to Alternatives:
- HC1: FPR rises sharply (up to 70% in extreme cases) as spatial correlation increases.
- Fixed Narrow (25 km): Fails to control size when the true correlation range exceeds the bandwidth.
- Fixed Wide (2,500 km): Contrary to intuition, this "conservative" choice often results in lower standard errors than HC1 and high FPR, confirming the inverse-U finding.
Kernel Performance: The Bartlett and Epanechnikov kernels consistently delivered the best size control. The Uniform kernel performed well at moderate levels but over-rejected at high levels.

Empirical Application

Using U.S. county-level data (Black population share vs. Hispanic population share):

The estimated covariogram range was approximately 980 km.
At this bandwidth, the standard error was maximized ($0.0884$), and the $t$ -statistic dropped to $-1.32$ (insignificant).
Using a "conservative" 2,500 km bandwidth reduced the standard error to $0.0561$, making the result significant ( $t = -2.08$ ). This illustrates how arbitrary wide-bandwidth choices can lead to spurious significance.

5. Significance

Practical Guidance: The paper provides the first clear, data-driven rule for selecting Conley standard error bandwidths, removing the reliance on ad-hoc researcher choices.
Inference Validity: By correctly identifying the correlation range, the method prevents both the underestimation of variance (from narrow bandwidths) and the "dilution" effect (from overly wide bandwidths), ensuring valid hypothesis testing in spatial econometrics.
Accessibility: The method is computationally efficient (taking a fraction of a second for typical datasets) and is readily available via the SpatialInference R package, making it accessible to applied researchers.
Theoretical Insight: It clarifies that valid inference in the Conley framework is achievable without abandoning the familiar regression structure, provided the bandwidth is chosen to match the physical scale of spatial dependence.