Adaptive Spatial Proximity

• Adaptive spatial proximity is the dynamic alteration of spatial relationships based on covariate-driven changes and varying environmental conditions.
• Tree-based models like decision trees, random forests, and BART recursively partition data to uncover nonlinear and context-sensitive proximity effects.
• This approach enhances variable selection and interaction detection, providing robust, interpretable insights for policy analysis versus conventional regression methods.

Adaptive spatial proximity refers to the flexible, context-sensitive modification of spatial relationships—distance, clustering, or distributions—between entities (individuals, objects, or events) within a shared environment, driven either by system dynamics or external covariates. In statistical and machine learning applications, adaptive spatial proximity captures how spatial effects or proximity-dependent influences are not constant but vary as a function of covariates, decision structures, or hierarchical organization. Tree-based methods, such as decision trees, ensembles (random forests), and Bayesian Additive Regression Trees (BART), provide a principled non-parametric approach to uncovering and quantifying such adaptive spatial proximity patterns by recursively partitioning data based on spatial distance and other predictors, enabling the model to “discover” where, how, and for whom proximity matters.

1. Adaptive Modeling of Spatial Proximity with Tree-Based Methods

Adaptive spatial proximity is operationalized as spatial dependence that varies systematically—and often nonlinearly—with respect to covariates or outcome-relevant attributes. Tree-based algorithms construct recursive data partitions, selecting splitting variables (such as spatial distance to events, landmarks, or boundaries) and associated cutpoints at each node to maximize within-partition homogeneity of the outcome.

At every split, the algorithm evaluates all candidate predictors—including spatial and temporal distances—and chooses the feature and threshold that provide the optimal partition. For example, a single regression tree might partition support for border policies at a distance threshold of 200 km from a major crossing, revealing that spatial proximity only influences a subset of the population, or that its effect is activated only for crossings above a certain size.

Ensemble methods, such as random forests or BART, expand this logic, stochastically aggregating results across many trees, reducing variance and yielding predictions and variable importance measures that reflect the aggregated (and thus more stable) adaptive spatial proximity structure discovered from the data.

2. Proximity Metrics, Predictors, and Algorithmic Features

Crucial to the adaptive analysis is the use of raw spatial distance measures—such as the Euclidean distance from a respondent to the nearest border crossing or mass shooting event—as predictors, often alongside event- or landmark-level features (e.g., crossing volume, number of fatalities).

Within the tree-building process, proximity measures can enter as continuous variables, and the resulting splits are not fixed a priori but adaptively set by the algorithm. For example, the split might be d < 150 km for one subgroup and d < 400 km for another, depending on the joint influence of other covariates (like party identification or region).

The identification of important proximity measures is quantified through permutation-based variable importance: for each feature p, importance is defined as

$$\text{importance}_p = \left(\frac{MSE_{(p)}}{MSE_0} - 1\right) \times 100,$$

where $MSE_{(p)}$ is the mean squared error after permuting $p$ and $MSE_0$ is the unpermuted baseline. High importance signifies that permuting the proximity metric greatly degrades predictive performance, indicating substantive spatial influence.

Both global (average over all data) and local (observation-level) importance can be derived, allowing the detection of heterogeneous spatial effects—an essential aspect of adaptivity.

3. Methodological Advantages Over Conventional Regression

Tree-based approaches to adaptive spatial proximity offer substantial methodological advantages:

• Non-parametric flexibility: No assumption of linearity, monotonicity, or specific cutoffs for distance; the model uncovers (possibly nonlinear or threshold) proximity effects from the data itself.
• Embedded variable selection: Trees and ensembles automatically select relevant spatial predictors, ignoring those that do not contribute to outcome variation, mitigating overfitting and multicollinearity concerns inherent in high-dimensional proximity settings.
• Interaction and discontinuity detection: The recursive partitioning allows for the automatic identification of higher-order interactions—e.g., whether spatial proximity influences attitudes only among certain political affiliations or demographic groups—without manual specification of interaction terms.
• Transparency in policy-relevant thresholds: Tree structures yield interpretable, scenario-specific threshold values (e.g., "proximity below 225 km matters for urban respondents"), facilitating substantive interpretation and application.

In contrast, conventional regressions require a priori specification of the functional form—e.g., linear, quadratic, or dichotomized proximity effects—and, therefore, may yield highly sensitive or unstable results depending on how proximity is encoded.

4. Empirical Case Studies Illustrating Adaptive Spatial Proximity

Distance to Border Crossings and Immigration Reform

In a study of California voters, tree-based models were applied to assess whether support for pro-immigrant policies is modulated by spatial proximity to U.S.–Mexico border crossings. The analysis included both proximity measures and landmark attributes (e.g., crossing traffic volume).

Key findings:

• Conventional regressions produced results highly sensitive to proximity specification (choice of distance cutoff or use of raw distances).
• The tree-based approach identified that, for some subgroups (e.g., Hispanic Democrats), proximity to large crossings mattered, but for others (e.g., Republicans), proximity was irrelevant compared to party identification.
• The effect of an increase in 100 km distance to the border was estimated as a function of border crossing size using a conventional regression, but the tree method flexibly discovered where the effect changed or was activated.

Distance to Mass Shootings and Gun Control Support

The second case study examined the impact of spatial and temporal proximity to mass shootings on attitudes toward gun control. Tree-based models (including single trees, random forests, and BART) were compared to linear regressions.

Findings included:

• Conventional regression yielded a simple negative effect of distance (e.g., 100 km further from a shooting reduced gun control support by 0.05 units).
• Tree-based methods allowed for multiple proximity measures (e.g., to the most recent event, to the nearest school shooting, or average distance to several shootings) and event attributes.
• Ensemble models provided both improved predictive accuracy (lower MSE) and allowed for simulation-based, locally adaptive marginal effects, revealing that proximity effects are not uniform but vary over the spatial and demographic landscape.

5. Variable Importance and Marginal Effects

Ensemble tree methods offer robust assessment of which proximity metrics exert substantial influence on outcomes. By repeatedly permuting predictors and measuring error changes, global and local importance can be computed.

The use of multiple independent permutations (typically $K=25$) provides stability. Local variable importance measures allow for the examination of subpopulations or regions where adaptive spatial proximity effects are especially salient. Simulation-based marginal effect analysis further enables quantification of how changes in proximity would alter expected outcomes, both globally and within specific partitions.

6. Contributions and Implications for Spatial Analysis

Tree-based methods, as detailed in this framework, provide a principled strategy for assessing adaptive spatial proximity in substantive domains where the relationship between distance and outcomes is known to be fragmented, nonlinear, or contingent.

Key contributions documented in the paper include:

• Demonstration of the limitations and instability of proximity effect estimation using traditional regression, due to model form sensitivity.
• Robust methodology for identifying both threshold effects and complex interactions between spatial proximity and other covariates without manual model specification.
• Case-based validation in two policy-relevant domains—immigration reform and gun control—showing that the approach yields insights into contextually emergent proximity effects.
• Framework for automatic variable selection and the extraction of interpretable, policy-relevant effects, even in high-dimensional spatial predictor environments.

These features make tree-based modeling a valuable tool for the discovery and quantification of adaptive spatial proximity in social, policy, and event-based spatial data.

7. Summary Table: Comparison of Key Features

Feature	Conventional Regression	Tree-Based Methods
Model dependence	Requires prespecified form	Data-driven, non-parametric
Variable (feature) selection	Manual inclusion, risk of bias	Automated, embedded in algorithm
Nonlinear thresholds	Must specify a priori	Discovered adaptively
Interaction effects	Must specify	Automatically detected
Handling many proximity metrics	Sensitive to overfitting	Robust via importance/permutation
Interpretability	Global coefficients, fixed cuts	Recursive, local rules and effects

The full methodological development and empirical analysis supporting this entry appear in (Levin, 2024).