Evaluating the effect of neighbourhood weight matrices on smoothing properties of Conditional Autoregressive (CAR) models

Arul Earnest, Geoffrey Morgan, Kerrie Mengersen, Louise Ryan, Richard Summerhayes, John BeardInternational Journal of Health Geographics 2007 · DOI

TL;DR

This paper investigates how different ways of defining "neighboring" areas in spatial statistical models (specifically Conditional Autoregressive or CAR models) impact the smoothing of disease risk estimates. The authors found that the choice of neighborhood definition significantly affects model performance, suggesting that an exploratory analysis of these definitions is crucial before applying CAR models for risk mapping or ecological modeling.

What Problem Does This Paper Address?

In spatial epidemiology, researchers often use models like the Conditional Autoregressive (CAR) model to smooth out random variations in disease rates across geographic areas. This smoothing helps to produce more reliable estimates of risk, especially in areas with few cases. A key component of CAR models is the definition of "neighborhoods" – which areas are considered to influence each other's risk estimates. However, there's been little systematic evaluation of how different ways of defining these neighborhoods (e.g., based on shared borders, distance, or even shared characteristics) affect the smoothing process and the resulting risk estimates. This lack of understanding can lead to potentially misleading conclusions in disease mapping and related ecological studies.

Methods

The researchers employed a Conditional Autoregressive (CAR) modeling framework to analyze birth defect data in New South Wales, Australia. They specifically focused on how different neighborhood weight matrices influence the model's smoothing properties.

Here's a breakdown of their approach:

  • Neighborhood Matrix Creation: They generated a diverse set of neighborhood weight matrices:
  • Adjacency-based: Four matrices based on contiguity (sharing a boundary or vertex), including "Queen" (shares any point) and "Rook" (shares a boundary) methods, extended to second-order neighbors.
  • Distance-based: Seven matrices that incorporated distance between area centroids. These varied in how distance was used (e.g., inverse distance, inverse squared distance) and included "Gravity" (weighted by population and distance), "Entropy" (emphasizing immediate neighbors), and "Density" (weighted by population density and distance).
  • Covariate-based: One matrix ("Weight-7") where neighbors were defined based on the similarity of a key covariate – maternal age.
  • Model Fitting: They fitted CAR models using data from 1995-2000.
  • Model Evaluation: Model performance was assessed using:
  • Deviance Information Criterion (DIC): A measure of model fit and complexity.
  • Chi-squared Residual Sum of Squares (RSS): To assess how well the model's predicted counts matched observed counts.
  • Kappa Statistic: To quantify agreement between observed and predicted risk categories.
  • Sensitivity and Specificity: To evaluate the models' ability to correctly identify high-risk areas.
  • Receiver Operating Characteristic (ROC) analysis: To determine optimal thresholds for classifying risk.
  • External Validation: The models were evaluated on data from a separate time period (2001-2003) to assess their predictive performance.

Datasets

The study utilized two primary datasets from New South Wales (NSW), Australia:

  • NSW Midwives Data Collection (MDC): A population-based surveillance system for all births in NSW, providing information on live births and stillbirths.
  • NSW Birth Defects Register (BDR): A population-based surveillance system monitoring birth defects detected during pregnancy, at birth, or up to one year of age.

Data from 1995-2003 were used, with models built using data from 1995-2000 and evaluated on data from 2001-2003. The analysis was conducted at the Statistical Local Area (SLA) level, with 198 SLAs defined within the NSW study area.

Key Insights and Ideas

  • Neighborhood Definition Matters Dramatically: The choice of how to define "neighbors" is not a trivial detail; it fundamentally alters the smoothing properties of CAR models and, consequently, the estimated disease risks.
  • Distance-Based Models Often Outperform Adjacency: For this dataset, distance-based neighborhood definitions generally led to better model performance (e.g., higher Kappa, lower RSS) compared to simple adjacency-based definitions, likely due to the irregular shapes and sizes of the administrative areas (SLAs).
  • Covariate-Based Neighborhoods Can Be Powerful: The "Covariate model" (Weight-7), which used maternal age similarity to define neighbors, showed promising results, particularly in recovering underlying risk structures and achieving good sensitivity. This highlights the potential of incorporating domain knowledge into neighborhood definitions.
  • Trade-offs Exist: Different models excel at different tasks. For instance, distance-based models might be better for classifying areas into risk categories, while others might be preferred if the goal is to explain away spatial relationships.
  • Exploratory Analysis is Crucial: Researchers should not blindly pick a neighborhood definition. An exploratory analysis of different weight matrices is recommended to understand their impact on smoothing and model fit before committing to a final model.

Key Results

  • Smoothing Differences: Considerable differences in smoothing properties were observed across the various neighborhood weight matrices.
  • Performance Ranking:
    • Distance-based matrices generally performed better than adjacency-based ones in terms of agreement between observed and predicted relative risks.
  • The "Gravity" model (Weight-4) achieved the lowest DIC (2202) and a high Kappa (0.15), indicating good fit and substantial smoothing.
  • The "Covariate" model (Weight-7) showed strong performance in recovering underlying risk structures, correctly classifying 74% of high-risk areas with 47% specificity.
  • Adjacency-based models (e.g., Queen-1) showed lower fractions of spatial random effects, meaning less of the variation was attributed to spatial structure.
  • Risk Detection: Using posterior probabilities from distance-based models offered higher sensitivity for detecting elevated risk areas compared to using smoothed relative risks alone.
  • RSS Values: Distance-based models generally had lower RSS values, indicating less smoothing and a better ability to predict observed risk in an area.

Limitations and Future Work

  • Limitations:
    • The study focused on a specific type of CAR model (simple spatial formulation with a temporal term) and did not compare it with more complex CAR formulations (e.g., spatio-temporal interaction models, multivariate CAR models).
  • The Kappa values, while showing differences, were considered modest, and the models were evaluated on an "external" time-period, which might influence agreement metrics.
  • The "Gravity" model's strength in borrowing from sparsely populated neighbors was noted as a potential limitation, though the model's weighting scheme mitigates this.
  • Future Work:
    • Further research could examine if these findings replicate in more complex CAR model formulations.
  • The authors suggest that future work could incorporate landscape features and spatial smoothing based on additional covariates (e.g., multivariate similarity index) to improve CAR model performance in disease mapping.
  • The paper emphasizes the need for a sensitivity analysis on the choice of neighborhood weight matrix.
  • Future ecological modeling should consider the implications of neighborhood choice and potentially use decision-theoretic approaches to select the most appropriate weight matrix based on loss, cost, and inferential aims.

Summary updated

Abstract

We found considerable differences in the smoothing properties of the CAR model, depending on the type of neighbours specified. This in turn had an effect on the models' ability to recover the observed risk in an area. Prior to risk mapping or ecological modelling, an exploratory analysis of the neighbourhood weight matrix to guide the choice of a suitable weight matrix is recommended. Alternatively, the weight matrix can be chosen a priori based on decision-theoretic considerations including loss, cost and inferential aims.
BibTeX
@article{earnest2007evaluating,
  author = {Arul Earnest and Geoffrey Morgan and Kerrie Mengersen and Louise Ryan and Richard Summerhayes and John Beard},
  title = {{Evaluating the effect of neighbourhood weight matrices on smoothing properties of Conditional Autoregressive (CAR) models}},
  journal = {International Journal of Health Geographics},
  year = {2007},
  doi = {10.1186/1476-072x-6-54},
}

Loading related content...

Figure