Bayesian analysis of verbal autopsy data using factor models with age- and sex-dependent associations between symptoms

Tsuyoshi Kunihama, Zehang Li, Samuel J. Clark, Tyler H. McCormick

In many parts of the world, people die without ever seeing a doctor or having their death officially registered. In these low-resource settings, public health officials rely on Verbal Autopsies (VAs)—interviews with grieving family members about the deceased's symptoms—to estimate what is killing the population.

A recent paper titled "Bayesian analysis of verbal autopsy data using factor models with age- and sex-dependent associations between symptoms" introduces a sophisticated new statistical framework to make these estimates more accurate by acknowledging a simple truth: symptoms don't mean the same thing for everyone.

TL;DR

Researchers have developed a new Bayesian factor model (BF-AS) that estimates cause-of-death distributions by accounting for how age and sex alter the relationships between symptoms. By moving beyond the "one-size-fits-all" assumption of symptom independence, this model achieves higher accuracy in identifying population-level mortality trends compared to industry-standard tools.

What Problem Does This Paper Address?

Estimating the Cause-Specific Mortality Fraction (CSMF)—the proportion of deaths due to specific causes like pneumonia or preterm delivery—is vital for global health policy. However, current algorithmic methods for VA data often rely on the "Naive Bayes" assumption: that symptoms (like coughing or fever) are independent of each other given a cause of death.

In reality, symptoms are deeply interconnected. More importantly, these connections change based on the individual. A fever in a two-day-old boy might imply a different risk profile than the same fever in a month-old girl. Most existing models treat age and sex as just two more symptoms on a checklist. This paper argues they should instead be treated as covariates that fundamentally shift the "structure" of how other symptoms relate to one another.

Methods

The authors propose a Bayesian Factor (BF) model with a twist. Standard factor models use latent (hidden) variables to capture the correlations between symptoms. The authors extend this using covariance regression.

Key technical highlights include:

  • Age- and Sex-Dependence: Instead of a static covariance matrix, the model allows the associations between symptoms to vary linearly based on the baby's age and sex.
  • Latent Factors: It uses two sets of latent factors—one for associations shared across all groups and another for associations that change with demographics.
  • Information Theory: To understand which questions in the VA survey actually matter, the authors used Mutual Information (MI) and Conditional Mutual Information (CMI). This helps distinguish between a symptom that is a "smoking gun" on its own versus one that only becomes useful when combined with other data.

Datasets

The model was evaluated using the PHMRC (Population Health Metrics Research Consortium) neonatal dataset.

  • Scope: 1,619 neonatal deaths across six geographic sites (e.g., India, Philippines, Tanzania, Mexico).
  • Data Points: 5 causes of death and 97 binarized predictors (symptoms and demographics).
  • Validation: The authors used a "leave-one-site-out" approach, training the model on data from five locations and testing it on the sixth to simulate a real-world scenario where local "gold standard" data is unavailable.

Key Insights and Ideas

  • The "Sex" Paradox: One of the most fascinating findings involves the "Sex" variable. In terms of raw Mutual Information, sex ranks near the bottom (96th out of 97). Knowing the baby's sex tells you almost nothing about the cause of death directly. However, sex had the highest Conditional Mutual Information. This means sex is a vital "key" that helps the model interpret the other symptoms correctly.
  • Symptom Clusters: The paper proves via Cramér’s V plots that symptom associations do vary significantly by age and sex. For example, the link between "late pregnancy complications" and "birth asphyxia" looks different for early-stage vs. late-stage neonates.
  • Top Predictors: Unsurprisingly, birth weight and size at delivery are the most informative predictors for neonatal death, followed by delivery timing (early/late) and physical abnormalities.

Key Results

The proposed BF-AS model was compared against several heavyweights in the field, including InterVA, Tariff, and InSilicoVA.

  • Accuracy: The BF-AS model achieved an average CSMF accuracy of 0.837, the highest among all tested methods.
  • Consistency: While no single method won every individual case, BF-AS was consistently in the top three across all six geographic test cases.
  • Uncertainty: Unlike "point-estimate" methods (like Tariff or NBC), the Bayesian approach provides a 95% credible interval, giving policymakers a sense of how much they can trust the numbers.

Limitations and Future Work

Despite the gains in accuracy, the authors note that the model's coverage rate is only 0.6 (meaning the true value fell within the 95% interval only 60% of the time). This suggests the model is still somewhat "overconfident" and needs better calibration.

Future research directions include:

  1. Location-Specific Modeling: Allowing symptom associations to vary by geographic "domain" or hospital type, not just age and sex.
  2. Broader Demographics: Incorporating socio-economic factors like mother's education or income levels into the factor structure.
  3. Sparsity: Using variable selection to prevent the model from overfitting when dealing with hundreds of potential survey questions.

Summary updated

Abstract

Verbal autopsies (VAs) are extensively used to investigate the population-level distributions of deaths by cause in low-resource settings without well-organized vital statistics systems. Computer-based methods are often adopted to assign causes of death to deceased individuals based on the interview responses of their family members or caregivers. In this article, we develop a new Bayesian approach that extracts information about cause-of-death distributions from VA data considering the age- and sex-related variation in the associations between symptoms. Its performance is compared with that of existing approaches using gold-standard data from the Population Health Metrics Research Consortium. In addition, we compute the relevance of predictors to causes of death based on information-theoretic measures.
BibTeX
@article{kunihama2025bayesian,
  author = {Tsuyoshi Kunihama and Zehang Li and Samuel J. Clark and Tyler H. McCormick},
  title = {{Bayesian analysis of verbal autopsy data using factor models with age- and sex-dependent associations between symptoms}},
  journal = {The Annals of Applied Statistics},
  year = {2025},
  doi = {10.1214/25-aoas2049},
}
Figure