A flexible Bayesian framework to estimate age- and cause-specific child mortality over time from sample registration data

Austin E Schumacher, Tyler H. McCormick, Jon Wakefield, Yue Chu, Jamie Perin, Francisco Villavicencio, Noah Simon, Li Liu

This paper introduces a unified Bayesian framework for estimating child mortality by age and cause, moving away from "disconnected" multistage models toward a more statistically rigorous approach that better quantifies uncertainty and handles sparse data.

TL;DR

Researchers have developed a flexible Bayesian framework that simultaneously estimates all-cause and cause-specific child mortality. By moving away from traditional multistage "plug-in" methods, this approach provides more accurate uncertainty intervals and uncovers hidden trends in granular age groups that were previously masked by broader data aggregations.

What Problem Does This Paper Address?

In many low- and middle-income countries (LMICs), high-quality vital registration systems (which record every birth and death) are non-existent. Instead, health officials rely on Sample Registration Systems (SRS)—data collected from specific representative sites.

The challenge is that current methods for analyzing SRS data are often "multistage." They estimate the total number of deaths first, then separately estimate the fraction of deaths caused by specific diseases (like pneumonia or diarrhea), and finally multiply them together. This "disconnected" approach has a major flaw: it doesn't allow information or uncertainty to flow between the stages. It often underestimates how much we don't know, leading to overconfident (and potentially misleading) policy recommendations.

Methods

The authors propose a unified Bayesian framework based on a competing risks failure process. Key technical highlights include:

  • The "Poisson Trick": They demonstrate that the complex likelihood of individuals dying from different causes can be simplified into a product of Poisson distributions for tabulated death counts and person-years. This makes the model computationally efficient.
  • Log-linear Hazards: They model mortality rates on a log scale, allowing for flexible functional forms that include fixed effects (for regions and causes) and random effects.
  • Temporal Smoothing: To account for trends over time, they use second-order random walks (RW2). This ensures that mortality estimates don't jump erratically from year to year but instead follow a smooth, realistic trajectory.
  • Integrated Nested Laplace Approximation (INLA): Rather than using slow MCMC simulations, they use the INLA package in R, which allows for much faster estimation of complex Bayesian models.

Datasets

The primary real-world application uses data from the Maternal and Child Health Surveillance System (MCHSS) in China (1996–2015).

  • Scope: 20 years of data across six region-residency strata (e.g., "east rural," "west urban").
  • Granularity: Six age groups (ranging from 0–6 days to 24–59 months) and eight aggregated cause-of-death groups.
  • Simulations: The authors also generated synthetic datasets to test how their model handles "extra-Poisson variability" and correlated causes of death compared to traditional methods.

Key Insights and Ideas

  • Likelihood Misspecification Matters: The paper proves that multistage models suffer from "likelihood misspecification." Because they treat the log of the sum of mortality rates as normally distributed (rather than the log of the rates themselves), they consistently produce uncertainty intervals that are too narrow.
  • Granularity Reveals Hidden Risks: By breaking down the "0–1 month" neonatal period into "0–6 days" and "7–27 days," the model revealed that certain causes, like acute respiratory infections, remain a significant threat in the later neonatal stage even when they appear to be declining in the aggregate.
  • Borrowing Strength: The Bayesian approach allows the model to "borrow strength" across different strata. If data is sparse for a specific cause in one region, the model can use global trends and neighboring data to produce a more stable estimate.

Key Results

  • Better Coverage: In simulations, the unified model achieved near-nominal (95%) coverage for uncertainty intervals, whereas the multistage approach dropped as low as 80%, proving it was "overconfident."
  • China Case Study: The model successfully smoothed the noisy empirical data from China’s MCHSS. For example, in the "west rural" region, it accurately captured the high burden of prematurity-related deaths while providing a clearer picture of the decline in communicable diseases.
  • Policy Impact: The results showed that while some causes (like diarrhea) have plummeted, others (like injuries in older children) remain stubbornly high, suggesting that intervention funding needs to be more targeted by both age and region.

Limitations and Future Work

  • Data Sparsity for Correlations: While the framework can model correlations between different causes of death (e.g., if a measles outbreak also increases pneumonia deaths), the authors found that the China dataset was too small to estimate these correlation parameters reliably.
  • Exposure Time Estimation: The model requires "person-years at risk," which isn't always directly recorded in SRS data and must be estimated using demographic methods, introducing another potential source of error.
  • Future Directions: The authors suggest extending the framework to include more covariates (like environmental factors) and adapting it to handle "Verbal Autopsy" data, where the cause of death is assigned probabilistically by an algorithm rather than a physician.

Summary updated

Abstract

In order to implement disease-specific interventions in young age groups, policy makers in low- and middle-income countries require timely and accurate estimates of age- and cause-specific child mortality. High-quality data is not available in settings where these interventions are most needed, but there is a push to create sample registration systems that collect detailed mortality information. current methods that estimate mortality from this data employ multistage frameworks without rigorous statistical justification that separately estimate all-cause and cause-specific mortality and are not sufficiently adaptable to capture important features of the data. We propose a flexible Bayesian modeling framework to estimate age- and cause-specific child mortality from sample registration data. We provide a theoretical justification for the framework, explore its properties via simulation, and use it to estimate mortality trends using data from the Maternal and Child Health Surveillance System in China.
BibTeX
@article{schumacher2022flexible,
  author = {Austin E Schumacher and Tyler H. McCormick and Jon Wakefield and Yue Chu and Jamie Perin and Francisco Villavicencio and Noah Simon and Li Liu},
  title = {{A flexible Bayesian framework to estimate age- and cause-specific child mortality over time from sample registration data}},
  journal = {The Annals of Applied Statistics},
  year = {2022},
  doi = {10.1214/21-aoas1489},
}

Loading related content...

Figure