Quantifying patient and neighborhood risks for stillbirth and preterm birth in Philadelphia with a Bayesian spatial model

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

TL;DR

Researchers developed a Bayesian spatial model to disentangle how individual health factors and neighborhood environments contribute to stillbirth and preterm birth in Philadelphia. The study reveals that high-risk neighborhoods face more than double the average risk of these outcomes, uncovering critical environmental stressors—like violent crime and poverty—that are often masked in traditional patient-level analyses.

What Problem Does This Paper Address?

Despite significant medical advancements, rates of stillbirth and preterm birth in the U.S. have remained stubbornly high or even increased. While doctors often focus on a patient’s immediate medical history or demographics, public health experts have long suspected that where a person lives plays a massive role in pregnancy outcomes.

The challenge is twofold:

  1. Data Sparsity: Stillbirth is a relatively rare event. In a single city, the numbers are often too small to draw statistically significant conclusions using standard methods.
  2. Multicollinearity: Neighborhood factors like poverty, education, and crime are often highly correlated. In a standard regression model, these factors "fight" each other, potentially masking the true impact of specific environmental stressors.

This paper seeks to bridge the gap between clinical data and urban sociology to identify geographic "hotspots" and the specific neighborhood characteristics that drive risk.

Methods

The authors employed a Bayesian spatial mixed-effects logistic regression model. The "secret sauce" of their approach is the use of a Conditional Autoregressive (CAR) prior (specifically the Leroux formulation).

Standard models treat every neighborhood as an island. The CAR prior, however, assumes that adjacent neighborhoods are likely to share similar risk profiles. This allows the model to "borrow strength" from neighboring census tracts, which is essential for stabilizing estimates in areas with sparse data.

To handle the computational complexity of logistic regression in a Bayesian framework, they used a Polya-Gamma latent variable strategy for their Markov Chain Monte Carlo (MCMC) algorithm. This is a sophisticated data-augmentation trick that makes the sampling process much more efficient and numerically stable. Finally, they used $k$-means clustering on the posterior risk probabilities to categorize Philadelphia’s census tracts into "lower," "moderate," and "higher" risk groups.

Datasets

The study linked two primary types of data from 2010 to 2017:

  • Electronic Health Records (EHR): De-identified data from Penn Medicine hospitals covering 45,919 deliveries. This provided patient-level variables like age, race/ethnicity, and whether the birth was a multiple (e.g., twins).
  • Neighborhood Data: Census-tract-level data from the U.S. Census Bureau (American Community Survey) and crime statistics from the Philadelphia Police Department. This included 14 variables ranging from poverty rates and education levels to log-transformed counts of violent and non-violent crimes.

Key Insights and Ideas

  • The "Masking" Effect: One of the most important takeaways is that neighborhood-level analysis caught associations that patient-level analysis missed. For example, while "public assistance" didn't look significant when looking at individual patients (because it was statistically "crowded out" by the poverty variable), the neighborhood-level marginal analysis showed it was a clear marker for high-risk areas.
  • Spatial Correlation Matters: The model confirmed that stillbirth risk is highly autocorrelated. If your neighbor is in a high-risk tract, you likely are too. This validates the "place-based" intervention theory—public health efforts shouldn't just target individuals; they should target blocks and zip codes.
  • Crime as a Health Factor: Even after controlling for poverty and race, violent crime remained a significant predictor of adverse birth outcomes. This suggests that the chronic stress of living in an unsafe environment has a physiological impact on pregnancy.

Key Results

  • Risk Multipliers: Census tracts categorized as "higher-risk" had 2.68 times the average risk of stillbirth and 2.01 times the average risk of preterm birth compared to lower-risk tracts.
  • Demographic Disparities: At the patient level, self-identified Black patients faced significantly higher risks (OR=2.23 for stillbirth) compared to White patients, even when accounting for neighborhood factors.
  • Protective Factors: Neighborhoods with higher rates of college-educated women or women in the labor force were significantly associated with lower risks, highlighting the protective nature of community-level socioeconomic stability.
  • Geography of Risk: The model identified clear clusters of high risk in North and West Philadelphia, providing a literal map for where the city might deploy mobile clinics or social support services.

Limitations and Future Work

The authors acknowledge several hurdles:

  • Residential History: The model only uses the address provided at the time of birth. It doesn't account for how long a patient lived there or if they moved during the pregnancy.
  • The "Snapshot" Problem: The study treats the eight-year period as a static block. Future work could incorporate spatio-temporal interactions to see if risks are changing as neighborhoods undergo gentrification or economic shifts.
  • Clustering Uncertainty: The use of deterministic $k$-means for risk stratification doesn't fully account for the uncertainty in the risk estimates themselves. The authors suggest exploring density-based clustering (like DBSCAN) in future iterations.
  • Scope: The findings are specific to Philadelphia. While the methodology is portable to any city, the specific risk drivers might differ in a more rural or differently structured urban environment.

Summary updated

Abstract

Stillbirth and preterm birth are major public health challenges. Using a Bayesian spatial model, we quantified patient-specific and neighborhood risks of stillbirth and preterm birth in the city of Philadelphia. We linked birth data from electronic health records at Penn Medicine hospitals from 2010 to 2017 with census-tract-level data from the United States Census Bureau. We found that both patient-level characteristics (e.g. self-identified race/ethnicity) and neighborhood-level characteristics (e.g. violent crime) were significantly associated with patients' risk of stillbirth or preterm birth. Our neighborhood analysis found that higher-risk census tracts had 2.68 times the average risk of stillbirth and 2.01 times the average risk of preterm birth compared to lower-risk census tracts. Higher neighborhood rates of women in poverty or on public assistance were significantly associated with greater neighborhood risk for these outcomes, whereas higher neighborhood rates of college-educated women or women in the labor force were significantly associated with lower risk. Several of these neighborhood associations were missed by the patient-level analysis. These results suggest that neighborhood-level analyses of adverse pregnancy outcomes can reveal nuanced relationships and, thus, should be considered by epidemiologists. Our findings can potentially guide place-based public health interventions to reduce stillbirth and preterm birth rates.
BibTeX
@misc{balocchi2021quantifying,
  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 = {{Quantifying patient and neighborhood risks for stillbirth and preterm birth in Philadelphia with a Bayesian spatial model}},
  year = {2021},
  eprint = {2105.04981},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2105.04981},
}
Figure