A Survival Framework for Estimating Child Mortality Rates using Multiple Data Types

Child mortality is an important population health indicator. However, many countries lack high-quality vital registration to measure child mortality rates precisely and reliably over time. Research endeavors such as those by the United Nations Inter-agency Group for Child Mortality Estimation (UN IGME) and the Global Burden of Disease (GBD) study leverage statistical models and available data to estimate child survival summaries including neonatal, infant, and under-five mortality rates. UN IGME fits separate models for each age group and the GBD uses a multi-step modeling process. We propose a Bayesian survival framework to estimate temporal trends in the probability of survival as a function of age, up to the fifth birthday, with a single model. Our framework integrates all data types that are used by UN IGME: household surveys, vital registration, and other pre-processed mortality rates. We demonstrate that our framework is applicable to any country using log-logistic and piecewise-exponential survival functions, and discuss findings for four example countries with diverse data profiles: Kenya, Brazil, Estonia, and Syrian Arab Republic. Our model produces estimates of the three survival summaries that are in broad agreement with both the data and the UN IGME estimates, but in addition gives the complete survival curve.

💡 Research Summary

The paper introduces a novel Bayesian survival‑analysis framework for estimating child mortality trends that unifies all data sources traditionally used by the United Nations Inter‑Agency Group for Child Mortality Estimation (UN IGME). Rather than fitting separate models for neonatal, infant, and under‑five mortality, the authors model the continuous survival function S(a | θₜ) – the probability of surviving to age a (in months) in year t – with a single parametric distribution. Two families of distributions are examined: a log‑logistic model with parameters μₜ and σₜ (constrained to σₜ > 1 to guarantee a non‑increasing hazard) and a piecewise‑exponential model with three constant hazards for the intervals 0‑1 month, 1‑12 months, and 12‑60 months. Both are re‑parameterised on the log or logit scale to enforce positivity and monotonicity, and to allow a simple normal likelihood for pre‑processed mortality rates (NMR, IMR, U5MR) expressed on the logit scale.

The hierarchical Bayesian model treats each country independently but shares a temporal structure across years. For a country with T estimation years, the T × K matrix of survival parameters (K = 2 for log‑logistic, K = 3 for piecewise‑exponential) follows a multivariate random‑walk prior, providing smoothness in time while still allowing abrupt changes when data demand it. Hyper‑parameters governing the random‑walk variance receive weakly informative priors, completing the Bayesian hierarchy.

Data integration is a central contribution. The authors incorporate three distinct data types:

1. Full Birth History (FBH) micro‑data from DHS and MICS surveys, which provide individual child ages at death or censoring. A pseudo‑likelihood is constructed by treating each month of exposure as a Bernoulli trial with success probability equal to the monthly death probability derived from the chosen survival distribution.
2. Vital Registration (VR) counts of births and deaths, modeled with a binomial likelihood directly linked to the survival function at the appropriate ages.
3. Pre‑processed summary mortality rates released by UN IGME, incorporated via a normal likelihood on the logit‑transformed probabilities of death at 1, 12, and 60 months.

Sample Vital Registration (SVR) systems (Bangladesh, China, India, Pakistan, South Africa) receive a special error term to reflect their intermediate quality. All likelihood components are combined multiplicatively, yielding a joint posterior from which posterior draws of θₜ are obtained via Hamiltonian Monte Carlo (Stan/NUTS). Posterior samples are transformed back to obtain annual estimates of NMR, IMR, and U5MR, together with full 95 % credible intervals and the continuous survival curve for any age between birth and five years.

The framework is illustrated for four countries with contrasting data environments:

• Kenya (rich FBH and VR data) shows close agreement with UN IGME B3 estimates while providing a smoother age‑specific hazard.
• Brazil (high‑quality recent VR plus historical survey data) demonstrates how the model blends disparate sources to produce consistent trends.
• Estonia (high‑quality VR only) validates that the method reproduces official vital statistics.
• Syrian Arab Republic (no VR, sparse survey data) highlights the model’s ability to generate plausible mortality trajectories where traditional methods yield wide uncertainty.

Across all examples, the Bayesian estimates align with UN IGME point estimates and fall within their reported uncertainty bounds, yet they additionally deliver the full survival function, enabling calculation of mortality for any arbitrary age interval (e.g., 6‑9 months) without re‑fitting models.

Key advantages identified by the authors include:

• Enforced monotonicity and non‑increasing hazard, reflecting demographic theory.
• Unified treatment of heterogeneous data, reducing information loss.
• Explicit quantification of uncertainty through posterior distributions.
• Flexibility to predict mortality for any age range, not just the three conventional summaries.
• Compatibility with standard Bayesian software, facilitating reproducibility and extensions.

Limitations are also acknowledged. The approach relies on the chosen parametric families; if real mortality patterns deviate substantially, bias may arise. Computational demands are non‑trivial, especially for a global implementation covering hundreds of countries and years. Prior specifications (e.g., random‑walk variance, σ > 1 constraint) can influence results in data‑scarce periods, and sensitivity analyses are limited. The model currently omits covariates (e.g., socioeconomic indicators) and spatial correlation, which could improve estimates for countries with sparse data.

Future directions proposed include:

• Exploring semi‑parametric or non‑parametric survival models (e.g., Bayesian splines, Gaussian processes) to relax distributional assumptions.
• Building a multi‑country hierarchical model that shares information across nations, potentially borrowing strength for data‑poor settings.
• Incorporating covariates and spatial random effects to capture systematic differences between regions.
• Extending the framework to sub‑national estimation and to other age groups (e.g., adolescent mortality).

In summary, the paper presents a coherent, statistically principled Bayesian survival framework that integrates all major data streams used for child mortality estimation, produces continuous age‑specific survival curves, and yields estimates consistent with established UN IGME outputs while offering richer analytical possibilities for researchers and policymakers.