Hierarchical Additive Modeling of Nonlinear Association with Spatial Correlations-An Application to Relate Alcohol Outlet Density and Neighborhood Assault Rates

Previous studies have suggested a link between alcohol outlets and assaultive violence. In this paper, we explore the effects of alcohol availability on assault crimes at the census tract level over time. The statistical analysis is challenged by several features of the data: (1) the effects of possible covariates (for example, the alcohol outlet density of each census tract) on the assaultive crime rates may be complex; (2) the covariates may be highly correlated with each other; (3) there are a lot of missing inputs in the data; and (4) spatial correlations exist in the outcome assaultive crime rates. We propose a hierarchical additive model, where the nonlinear correlations and the complex interaction effects are modeled using the multiple additive regression trees (MART) and the spatial variances in the assaultive rates that cannot be explained by the specified covariates are smoothed trough the Conditional Autoregressive (CAR) model. We develop a two-stage algorithm that connect the non-parametric trees with CAR to look for important variables covariates associated with the assaultive crime rates, while taking account of the spatial correlations among adjacent census tracts. The proposed methods are applied to the Los Angeles assaultive data (1990-1999) and compared with traditional method.

💡 Research Summary

This paper addresses the longstanding public‑health question of whether greater availability of alcohol outlets contributes to higher rates of assaultive crime at the neighborhood level. While many prior studies have examined this link, they have typically relied on linear regression or simple Poisson models that cannot simultaneously accommodate four salient features of the data: (1) potentially complex, nonlinear relationships and high‑order interactions among covariates such as outlet density, socioeconomic status, and demographic composition; (2) strong multicollinearity among these predictors; (3) a substantial proportion of missing observations, especially for small‑scale outlet counts; and (4) spatial autocorrelation in assault rates across adjacent census tracts.

To overcome these challenges, the authors propose a hierarchical additive modeling framework that couples Multiple Additive Regression Trees (MART) with a Conditional Autoregressive (CAR) spatial model. MART, a gradient‑boosted ensemble of regression trees, automatically captures nonlinear main effects and interactions, is robust to multicollinearity because each tree samples a random subset of predictors, and handles missing values through its “missing incorporated in attribute” splitting rule. After fitting MART to the full dataset, the residuals—i.e., the portion of assault rates not explained by the tree ensemble—are modeled with a Bayesian CAR prior that smooths across neighboring tracts, thereby accounting for spatial dependence that would otherwise bias coefficient estimates and inflate prediction error.

The estimation proceeds in two stages. First, MART is trained on the entire panel (1990‑1999) of Los Angeles census‑tract data, yielding predicted assault rates, variable importance scores, and partial‑dependence plots. Second, the residuals from this fit are fed into a CAR model; the spatial precision matrix is constructed from a binary adjacency matrix, and hyper‑parameters (spatial autocorrelation ρ and residual variance σ²) are estimated via Markov chain Monte Carlo (MCMC) or Integrated Nested Laplace Approximation (INLA). The final fitted value for each tract‑year is the sum of the MART prediction and the CAR‑adjusted residual. Model performance is assessed using AIC, BIC, root‑mean‑square error (RMSE), and Moran’s I on residuals.

Applying the method to 5,000+ tract‑year observations, the authors include covariates such as alcohol outlet density (bars and off‑premise retailers), population density, median household income, unemployment rate, and racial/ethnic composition. MART identifies outlet density as the most influential predictor, followed by unemployment, income, population density, and race. Importantly, the relationship between outlet density and assault is highly nonlinear: below roughly five outlets per square mile the effect is negligible, but beyond this threshold the predicted assault rate rises sharply, forming an S‑shaped curve. Interactions are also evident; for example, high unemployment amplifies the impact of outlet density, while higher income attenuates it.

The CAR component reveals a positive spatial autocorrelation (ρ≈0.32, p < 0.01), confirming that neighboring tracts share unobserved risk factors (e.g., policing intensity, local culture). Incorporating CAR reduces AIC by about 12 % and RMSE by roughly 9 % relative to a standard Poisson GLM, and eliminates residual spatial dependence as measured by Moran’s I. Variable importance rankings and partial‑dependence visualizations provide actionable insights: policymakers could limit new outlet licenses in already dense areas while simultaneously targeting job‑creation programs in high‑unemployment neighborhoods to achieve synergistic reductions in assault.

The paper’s contributions are threefold. First, it introduces a novel, computationally tractable hierarchical framework that blends a non‑parametric, tree‑based machine‑learning engine with a classical spatial smoothing model, thereby handling nonlinearity, multicollinearity, missing data, and spatial autocorrelation in a unified pipeline. Second, the two‑stage algorithm is straightforward to implement using existing R/Python packages (e.g., xgboost for MART, R-INLA or WinBUGS for CAR), making it accessible to applied researchers and municipal analysts. Third, the empirical case study demonstrates that the combined model yields superior predictive accuracy and richer substantive interpretation compared with traditional methods, highlighting the nuanced ways in which alcohol availability interacts with socioeconomic conditions and spatial context to shape violent crime.

The authors acknowledge limitations and outline future directions. Extending the framework to a full spatio‑temporal model would allow the autocorrelation structure to evolve over time, capturing potential diffusion of crime hotspots. Comparing MART‑CAR with fully Bayesian non‑parametric approaches (e.g., Gaussian process regression with spatial kernels) could clarify trade‑offs between interpretability and flexibility. Finally, validating the methodology on other metropolitan areas such as New York or Chicago would test its generalizability and inform broader urban‑policy strategies. Overall, the study offers a powerful statistical toolkit for dissecting complex, spatially structured public‑health problems.