📊 statistics

Local Variable and Neighborhood Selection for Firearm Fatality in the Southeast USA

This paper proposes a two-step localized variable selection and inference framework using the SCAD penalty to identify county-specific socio-economic drivers of firearm fatalities in the Southeast USA while simultaneously determining the most appropriate spatial neighborhood structure for modeling spatially correlated, over-dispersed data.

Original authors: Debjoy Thakur, Lingyuan Zhao, Soutir Bandyopadhyay

Published 2026-01-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Debjoy Thakur, Lingyuan Zhao, Soutir Bandyopadhyay

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

The Big Picture: Why One Size Doesn't Fit All

Imagine the United States as a giant patchwork quilt. Each square on the quilt is a county. The paper focuses on the "South-East" squares. The researchers are trying to figure out why some squares have a lot of gun deaths (high fatality) and others have very few.

The problem with old methods is that they tried to use one single recipe for the whole quilt. They asked, "What causes gun deaths everywhere?" and gave one answer, like "It's mostly about poverty" or "It's mostly about age."

But the authors say: "That's not how the world works."

In one county, gun deaths might be driven by young men in a specific neighborhood.
In a neighboring county, it might be driven by older women in a completely different setting.
The "neighbors" that influence a county might be to the North in one place, but to the West in another.

This paper introduces a new tool that acts like a magnifying glass. Instead of looking at the whole quilt at once, it zooms in on one specific square (county) to see exactly what factors matter right there, and exactly which neighbors influence it.

The Three-Step Detective Process

The authors propose a two-step "detective" method to solve three specific mysteries:

Mystery 1: Which clues matter here? (Local Variable Selection)

The Analogy: Imagine you are a detective investigating a crime in a small town. You have a bag of 800 potential clues (variables like age, race, income, gun laws, etc.).

Old Way: You look at the whole country and say, "Clue #5 (Age) is important everywhere." So you ignore the other 799 clues.
New Way: You zoom in on this specific town. You realize that for this town, Clue #5 doesn't matter, but Clue #12 (a specific type of income) and Clue #45 (a specific age group) are the only ones that matter.
The Tool: They use a mathematical "filter" called SCAD. Think of this as a sieve that shakes out the 798 useless clues for that specific location, leaving only the 2 or 3 that actually explain the deaths in that county.

Mystery 2: Which way does the influence flow? (Directional Variation)

The Analogy: Imagine the county is a hill. The researchers want to know if the "gun death risk" flows down the hill like water.

Old Way: They assumed the risk spreads equally in all directions (like ripples in a pond).
New Way: They realized the "wind" might only blow from the West. Maybe the risk comes from a neighboring county to the West, but the county to the East has no influence.
The Result: Their method can say, "In this county, the risk flows strongly from West to East, but not North to South." This helps them understand the direction of the problem.

Mystery 3: Who are the real neighbors? (Local Neighborhood Selection)

The Analogy: In a neighborhood, who counts as a "neighbor"?

Old Way: Everyone assumed a neighbor is anyone touching your house (North, South, East, West).
New Way: The researchers found that sometimes, the "neighbor" is only the house to the South, or maybe only the houses to the South and West.
The Discovery: Instead of guessing the neighborhood shape, their method tests 15 different shapes (like a cross, a line, or a corner) and picks the one that fits the data best for that specific county.

How They Did It (The "Two-Step" Recipe)

Step 1: The Filter. They zoom in on a county and use the SCAD filter to pick the most important demographic clues (like "Median Age of Females" or "Number of Young Males"). They ignore the rest.
Step 2: The Map. Once they have the important clues, they ask: "How do these clues interact with the neighbors?" They test different neighbor shapes (North, South, East, West, or combinations) to see which one explains the data best.

What They Found (The Real-World Test)

They tested this on real data from the Southeast USA in 2024.

Floyd County, Kentucky: They found that the "neighbors" influencing gun deaths were mostly to the North and West. The key clues were the median age of males and females.
Franklin County, Florida: Here, the influence flowed mostly West. The key clue was the median age of females.
Dougherty County, Georgia: The influence flowed East-West. The key clues were the ages of white males and Native American populations.

The Takeaway: If you used the "Old Way" (global selection), you might have missed these specific patterns. You might have thought "Age" was the only factor, or that "North" was the only direction that mattered. But by zooming in, they saw that every county has its own unique recipe and its own unique wind direction.

Why This Matters

The paper argues that to fix public health issues like gun violence, we can't use a "one-size-fits-all" policy. We need to understand the specific local ingredients and the specific local connections in every single county. This new method gives us the map to see those local details clearly.

Technical Summary: Local Variable and Neighborhood Selection for Firearm Fatality in South-East USA

Problem Statement
Firearm-related deaths in the United States exhibit significant spatial heterogeneity driven by local variations in race, age, sex, and income distributions. While extensive literature exists on variable selection for independent data (e.g., LASSO, SCAD) and some on spatial data, a critical gap remains: the lack of methods for local variable selection and post-selection inference for spatially correlated, over-dispersed count data. Traditional approaches often assume global variable importance or impose ad hoc spatial neighborhood structures (e.g., fixed first-order or second-order neighbors) that fail to capture directional anisotropy or local variations in the latent spatial random effect structure. This paper addresses the challenge of identifying influential socio-economic factors and the appropriate spatial neighborhood structure specifically for locations where firearm fatality counts exceed a threshold.

Methodology
The authors propose a two-step localized variable selection and inference framework for spatially indexed negative binomial (NB) data.

Step 1: Local Variable Selection (LSPMLE):
- Localization: The method localizes the mean function around a reference location $s^{(0)}_i$ using a Taylor series expansion. This decomposes regression coefficients $\beta_k(s)$ into a local intercept ( $\gamma_k$ ), a horizontal slope ( $\delta_k$ ), and a vertical slope ( $\tau_k$ ), allowing for directional variation in predictor effects.
- Penalization: A Smoothly Clipped Absolute Deviation (SCAD) penalty is applied to the localized pseudo-log-likelihood to select relevant predictors. This is performed under the assumption of no random effects initially to identify the set of active covariates $\hat{A}^{(i)}_n$ for the specific location.
- Optimization: The optimization utilizes an Anderson-accelerated coordinate descent algorithm to handle high-dimensional local problems and collinearity efficiently. Tuning parameters (bandwidth $h$ and penalty $\lambda$ ) are selected via a Weighted Quasi-BIC (WQBIC) criterion.
Step 2: Local Neighborhood Inference:
- Refitting: Using the predictors selected in Step 1, the model is refitted without penalties to estimate regression coefficients and the spatial random effect.
- Neighborhood Selection: The method evaluates 15 distinct spatial neighborhood structures (including standard "rook" neighbors and directional anisotropic structures like North-only, East-West, or South-West combinations).
- Inference: The spatial weight matrix is localized, and the neighborhood structure that best fits the data (maximizing the likelihood contribution of the spatial random effect) is identified. This allows for the inference of directional variation in the spatial dependence structure.

Theoretical Properties
The paper establishes theoretical results under infill asymptotics, where the number of sampling points within a fixed spatial region increases.

Existence and Consistency: The authors prove the existence of a local maximizer for the penalized likelihood.
Convergence Rate: The estimator converges at a rate of $O_P(\alpha_n)$ , where $\alpha_n = \sqrt{K_n/N_n}$ ( $K_n$ is the number of predictors, $N_n$ is the sample size).
Sparsity Selection: Under specific regularity conditions (including assumptions on the Fisher information matrix and the behavior of the SCAD derivative), the method achieves consistent sparsity selection, meaning it correctly identifies zero coefficients with probability approaching one as the sample size grows.

Key Contributions

Two-Stage Localized Approach: The introduction of a framework that first selects predictors locally using SCAD (assuming no random effects) and then estimates regression coefficients and neighborhood topology using the selected predictors.
Directional Variation Characterization: The method explicitly models and identifies directional variation in the coefficient surface (e.g., a predictor may have a stronger effect in the East-West direction than North-South) and in the spatial neighborhood structure.
Data-Driven Neighborhood Selection: Instead of assuming a fixed spatial structure, the method identifies the most appropriate type of spatial neighborhood (e.g., first-order, directional anisotropy) for specific "hotspot" counties.
Application to Over-dispersed Areal Data: This is presented as the first work to perform local variable selection for over-dispersed areal data (modeled via Negative Binomial distribution).

Results

Simulation Study: The method (LSPMLE) was tested on simulated spatial count data with 800 predictors and varying sample sizes. Results demonstrated that LSPMLE outperformed global SCAD and global Adaptive LASSO (AL) in terms of True Local Variable Selection Probability (TLVSP) and True Global Variable Selection Probability (TGVSP), particularly as sample size increased. The method successfully recovered the true local variable sets and latent neighborhood structures.
Real-Data Analysis (Southeast USA, 2024): The framework was applied to firearm fatality data across counties in the Southeast US.
- County-Specific Findings: The analysis revealed that influential predictors vary by county. For example, in Floyd and Knott counties (Kentucky), median ages of male/female populations and Native Hawaiian/Pacific Islander female populations were selected, while in Dougherty County (Georgia), variables related to white male and American Indian populations were significant.
- Neighborhood Structure: The optimal spatial neighborhood structure varied by location. For instance, Franklin County (Florida) exhibited spatial dependence primarily along the West side, whereas Floyd County (Kentucky) was best modeled by a standard "rook" (first-order) structure.
- Comparison: Global methods (AL and SCAD) selected a uniform set of variables and a single global neighborhood structure (mostly North-directional), failing to capture the local heterogeneity identified by the proposed method.

Significance and Claims
The paper claims that its primary significance lies in addressing the "vacuum" in literature regarding local variable selection for spatially correlated, over-dispersed data. By moving away from global assumptions, the method provides:

Local Inference: The ability to select variables specific to a location's socio-demographic context (e.g., selecting young Black males as a risk factor in one county while older white females are the factor in another).
Directional Insight: Identification of how selected predictors exert influence directionally (e.g., stronger effects along specific geographic axes).
Structural Flexibility: The elimination of ad hoc assumptions about spatial neighborhood structures, allowing the data to dictate the appropriate spatial dependency type (e.g., anisotropic vs. isotropic) for each hotspot.

The authors conclude that this approach offers a more nuanced understanding of firearm fatality drivers, which is critical for targeted public health interventions. They note that future work could extend this to fully generalized linear models, unified penalized likelihood frameworks with spatial random effects, and spatio-temporal contexts.