A Bayesian Hierarchical Modeling Framework for Geospatial Analysis of Adverse Pregnancy Outcomes

Cecilia Balocchi, Ray Bai, Jessica Liu, Silvia P. Canelón, Edward I. George, Yong Chen, Mary Regina Boland

Abstract

Studying the determinants of adverse pregnancy outcomes like stillbirth and preterm birth is of considerable interest in epidemiology. Understanding the role of both individual and community risk factors for these outcomes is crucial for planning appropriate clinical and public health interventions. With this goal, we develop geospatial mixed effects logistic regression models for adverse pregnancy outcomes. Our models account for both spatial autocorrelation and heterogeneity between neighborhoods. To mitigate the low incidence of stillbirth and preterm births in our data, we explore using class rebalancing techniques to improve predictive power. To assess the informative value of the covariates in our models, we use posterior distributions of their coefficients to gauge how well they can be distinguished from zero. As a case study, we model stillbirth and preterm birth in the city of Philadelphia, incorporating both patient-level data from electronic health records (EHR) data and publicly available neighborhood data at the census tract level. We find that patient-level features like self-identified race and ethnicity were highly informative for both outcomes. Neighborhood-level factors were also informative, with poverty important for stillbirth and crime important for preterm birth. Finally, we identify the neighborhoods in Philadelphia at highest risk of stillbirth and preterm birth.

Introduction

Despite significant medical advancements, stillbirth rates in the U.S. have remained largely stagnant for decades, and preterm birth rates are on the rise. While clinical factors are well-studied, public health experts increasingly recognize that "place matters." Adverse pregnancy outcomes are geographically heterogeneous, yet standard epidemiological models often struggle with sparse data at the neighborhood level and fail to account for spatial autocorrelation—the tendency for neighboring areas to share similar risk profiles. This research bridges the gap between individual clinical data and neighborhood-level social determinants of health to provide a nuanced map of risk in Philadelphia.

Methodology

The study employs a Bayesian mixed-effects logistic regression model designed to handle spatial dependencies.

  • Spatial Autocorrelation: A Conditional Autoregressive (CAR) prior (Leroux formulation) allows the model to "borrow strength" from adjacent neighborhoods. This improves estimates in census tracts with few recorded births.
  • Computational Efficiency: To manage large-scale Bayesian inference, the researchers utilized a Pólya-Gamma latent variable strategy. This data augmentation technique transforms the logistic regression into a form that allows for efficient Gibbs sampling, significantly accelerating the Markov Chain Monte Carlo (MCMC) algorithm.
  • Risk Stratification: Following neighborhood probability estimation, k-means clustering categorized Philadelphia’s census tracts into "lower-risk," "moderate-risk," and "higher-risk" groups.
  • Marginal Analysis: A secondary regression analyzed how neighborhood characteristics correlated with these risk clusters to identify place-based drivers.

Data Integration and Feature Selection

The study integrated three primary data sources from 2010 to 2017: Penn Medicine EHR (45,919 deliveries), U.S. Census Bureau data, and Philadelphia Police Department crime statistics.

Table 1: Patient-level and neighborhood-level variables used in the geospatial analysis.

The inclusion of both individual and community-level variables (as shown in the table above) allowed the model to account for the "masking" effect common in purely patient-level analyses. For instance, while poverty and public assistance are often correlated at the individual level, neighborhood-level analysis clarifies the broader impact of community-wide economic disadvantage.

Key Results and Insights

The spatial models significantly outperformed independent models, confirming that unmeasured factors—such as local pollution or "food deserts"—are likely clustered geographically.

Risk Drivers

  • Patient Factors: Self-identifying as Black was associated with a significantly higher risk of stillbirth (OR=2.23) and preterm birth (OR=1.55). Multiple births were the strongest individual predictor for preterm birth (OR=10.56).
  • Neighborhood Stressors: High rates of poverty and public assistance were strongly linked to higher risk. Notably, violent crime emerged as a significant neighborhood stressor for both outcomes, suggesting that the chronic stress of living in an unsafe environment has physiological consequences for pregnancy.
  • Protective Factors: Higher rates of college-educated women or women in the labor force were associated with lower neighborhood risk.

Geographical Risk

The highest-risk census tracts exhibited 2.68 times the average risk of stillbirth and 2.01 times the average risk of preterm birth compared to lower-risk tracts. These high-risk clusters were primarily concentrated in North and West Philadelphia, while the lowest-risk areas were centered around Center City.

Conclusion and Future Directions

This Bayesian framework demonstrates that neighborhood-level stressors drive pregnancy risks independently of individual patient characteristics. However, the study has limitations: it used static residency at delivery, did not account for mothers with multiple pregnancies in the dataset, and used deterministic clustering. Future work aims to incorporate spatio-temporal interactions to track how neighborhood risks evolve over time and explore more robust density-based clustering methods like DBSCAN.

Summary updated

Abstract

Studying the determinants of adverse pregnancy outcomes like stillbirth and preterm birth is of considerable interest in epidemiology. Understanding the role of both individual and community risk factors for these outcomes is crucial for planning appropriate clinical and public health interventions. With this goal, we develop geospatial mixed effects logistic regression models for adverse pregnancy outcomes. Our models account for both spatial autocorrelation and heterogeneity between neighborhoods. To mitigate the low incidence of stillbirth and preterm births in our data, we explore using class rebalancing techniques to improve predictive power. To assess the informative value of the covariates in our models, we use posterior distributions of their coefficients to gauge how well they can be distinguished from zero. As a case study, we model stillbirth and preterm birth in the city of Philadelphia, incorporating both patient-level data from electronic health records (EHR) data and publicly available neighborhood data at the census tract level. We find that patient-level features like self-identified race and ethnicity were highly informative for both outcomes. Neighborhood-level factors were also informative, with poverty important for stillbirth and crime important for preterm birth. Finally, we identify the neighborhoods in Philadelphia at highest risk of stillbirth and preterm birth.
BibTeX
@misc{balocchi2021bayesian,
  author = {Cecilia Balocchi and Ray Bai and Jessica Liu and Silvia P. Canelón and Edward I. George and Yong Chen and Mary Regina Boland},
  title = {{A Bayesian Hierarchical Modeling Framework for Geospatial Analysis of Adverse Pregnancy Outcomes}},
  year = {2021},
  eprint = {2105.04981},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2105.04981},
}
Figure