Bridging surveillance gaps in dengue: a hierarchical model integrating mixed data sources for transmission estimation and vaccine targeting

Problem Statement
Estimating the force of infection (FOI) for dengue is critical for understanding transmission dynamics and guiding intervention strategies, particularly for targeting vaccination programs. However, in endemic settings like Indonesia, routine surveillance data are often incomplete, inconsistent in format, and subject to significant reporting bias. While age-stratified case data provide rich information for FOI estimation, many districts report only aggregate annual totals. Conversely, seroprevalence surveys offer complementary data on cumulative exposure but are costly and spatially limited. Existing catalytic models typically require uniform age-stratified data across all locations or estimate parameters independently, failing to leverage the heterogeneity of available data types (age-stratified cases, aggregate cases, and seroprevalence) within a single framework. This creates a fundamental challenge in consistently estimating FOI and reporting rates across locations with varying data quality and formats.

Methodology
The authors developed a Bayesian hierarchical catalytic model designed to jointly fit age-stratified case data, aggregate case data, and seroprevalence surveys within a unified inference framework. The model operates across 128 districts in seven provinces of Java and Bali, Indonesia, covering the period 2016–2024.

Key methodological components include:

• Hierarchical Structure: A two-level hierarchy nests districts within provinces, allowing for partial pooling of information. This structure enables districts with different data completeness levels (e.g., some with 9 age groups, others with 4, and some with only aggregate totals) to contribute simultaneously.
• Epidemiological Assumptions: The model assumes four co-circulating dengue serotypes with equal per-serotype FOI. It distinguishes between primary and secondary infections, assuming secondary infections are more likely to be reported due to antibody-dependent enhancement (ADE).
• Covariate Integration: External covariates (e.g., Oceanic Niño Index, urbanization, elevation, rainfall, and surveillance capacity) are incorporated to inform heterogeneity in both FOI and reporting rates. This external structure aids in parameter identifiability, particularly resolving the trade-off between FOI and reporting rate that often plagues models using case data alone.
• Historical Reconstruction: To initialize population immunity at the start of the surveillance period, the model reconstructs historical FOI for a period preceding the data (10 years prior) and assumes a constant FOI for the distant past.
• Validation: The framework was validated using synthetic data under four scenarios of decreasing data completeness (from full age stratification to aggregate-only data), with and without covariates.

Key Results

• Synthetic Validation: The model successfully recovered true parameter values across all data completeness scenarios. Models incorporating covariates showed lower Root Mean Squared Error (RMSE) for FOI and seroprevalence, even in scenarios with the most limited age-stratified data (S4). However, the study found that covariates alone, while improving point estimates, resulted in poorly calibrated uncertainty (over-confident posteriors). Anchoring the model with at least one seroprevalence survey was necessary to bring credible interval coverage close to nominal levels.
• Real Data Application: Applied to Java and Bali, the model revealed substantial spatial heterogeneity in FOI and reporting rates.
  • Reporting Rates: Estimated reporting rates were generally low (median ~2–7.5% across provinces), with Bali showing the highest rates.
  • Transmission Drivers: Covariate analysis indicated that the Oceanic Niño Index (ONI) had the largest positive effect on FOI. Surprisingly, higher-than-average rainfall was associated with lower FOI, and higher elevation was associated with higher FOI, likely reflecting specific ecological and demographic contexts of Java (e.g., urban centers at higher elevations).
  • Temporal Trends: The model captured major inter-annual peaks (2016, 2019, 2022, 2024). The 2022 baseline FOI was notably high, suggesting drivers beyond El Niño variability.
• Vaccine Prioritization: Using WHO-suggested criteria (seroprevalence at age 9 > 60%), the model identified high-priority districts for vaccine introduction. There was a significant misalignment between model-based seroprevalence prioritization and incidence-based prioritization. Districts identified as high-priority by the model but missed by incidence data were concentrated in Central and East Java (areas with lower reporting rates), while districts flagged by incidence but not the model were in Bali and Yogyakarta (areas with higher surveillance capacity).

Significance and Claims
The paper claims that its primary contribution is the development of a flexible Bayesian hierarchical framework capable of integrating mixed data sources (age-stratified cases, aggregate cases, and seroprevalence) to estimate FOI and reporting rates consistently across heterogeneous surveillance settings.

The authors assert that:

1. Data Synthesis: This approach allows for the correct synthesis of disparate data types, overcoming the limitation of existing models that require uniform data formats.
2. Identifiability: The inclusion of covariates and seroprevalence anchors significantly improves parameter identifiability, particularly for the FOI-reporting rate trade-off, even if in-sample predictive performance (measured by LOO-CV) remains similar across scenarios.
3. Policy Relevance: Model-based seroprevalence estimation offers a more consistent basis for identifying high-priority districts for vaccine introduction compared to reported incidence. Relying solely on incidence data may systematically under-serve areas with weak surveillance and over-prioritize areas with better case detection.
4. Uncertainty Quantification: The study emphasizes the importance of propagating uncertainty into policy decisions, noting that many districts near the 60% seroprevalence threshold have substantial uncertainty, requiring conservative criteria (e.g., >75% posterior probability) for high-confidence prioritization.

The authors conclude that while the framework improves estimation in data-sparse settings, it requires investment in data infrastructure (seroprevalence surveys, covariate data) and analytical capacity. They acknowledge limitations, including the assumption of constant reporting rates over time and equal FOI across serotypes, suggesting future work should explore time-varying reporting and serotype-specific dynamics.