📊 statistics

Bayesian disease mapping from incomplete hospital-based surveillance: accounting for catchment and missing case origins in dengue surveillance in Makassar, Indonesia

This paper presents a hierarchical Bayesian point-process framework that improves dengue risk mapping in Makassar, Indonesia, by jointly modeling hospital catchment behavior and unobserved case origins to effectively handle incomplete, facility-based surveillance data with varying levels of spatial resolution.

Original authors: Bobby Poerwanto, Ian Hall, Emma L Fairbanks, Thomas House, Olatunji Johnson

Published 2026-10-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bobby Poerwanto, Ian Hall, Emma L Fairbanks, Thomas House, Olatunji Johnson

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Dengue fever is a mosquito-borne illness that thrives in tropical cities, causing widespread illness and straining healthcare systems. To stop it, public health officials need to know exactly where the risk is highest. However, a major obstacle stands in the way of clear answers: the data we have is often incomplete. When a person gets sick, they go to a hospital, and that hospital records the case. But the hospital is not always the patient's home. Sometimes, a patient travels far from their neighborhood to get care, and sometimes the records are so vague that officials only know the patient came from a general area, or not even that. If researchers simply map where patients were treated, they end up mapping where hospitals are and how people travel, rather than where the mosquitoes are actually breeding. This confusion makes it difficult to target interventions effectively, leaving some dangerous areas overlooked while resources are wasted on others.

To solve this puzzle, a team of researchers from the University of Manchester and Universitas Negeri Makassar developed a new way to look at dengue data from the city of Makassar, Indonesia. Instead of treating the hospital records as a direct map of the disease, they built a statistical model that separates the disease itself from the behavior of the people seeking care. They started with the idea that dengue risk is not spread evenly across a city; it clusters in specific neighborhoods due to local conditions like vegetation and drainage. They also recognized that people do not always go to the nearest hospital; they choose facilities based on distance and other factors, a pattern known as a "catchment" area. The researchers created a framework that simultaneously estimates where the disease is most intense and how far people are willing to travel to get treated. This approach allowed them to work with messy data, including cases where the patient's home address was missing entirely, by treating the missing location as a mystery to be solved by the model rather than a dead end.

The study analyzed 1,842 dengue cases reported between 2021 and 2025 from eight hospitals in Makassar. The data was imperfect: some patients had their exact location recorded, some were known only to a subdistrict, and 220 cases had no residential information at all. The researchers tested their new method against older, simpler ways of handling this data. The old methods either assumed everyone went to the nearest hospital or guessed the missing locations based on population size. The new model, however, used a "gravity" concept, which assumes that the likelihood of a patient visiting a hospital drops as the distance increases, but not as sharply as older models assumed. This allowed the model to infer the most probable home locations for the 220 patients with missing data, effectively filling in the gaps in the map.

When the researchers tested their model by hiding the known locations of some hospitals and seeing if the model could guess them correctly, it performed significantly better than the traditional methods. The new approach was far more accurate at predicting where cases actually came from, proving that it could successfully untangle the difference between where a person lives and where they seek treatment. This success gave the researchers confidence to apply the model to the full dataset to find the true high-risk areas. The results revealed a surprising truth: the neighborhoods with the most reported cases were not necessarily the ones with the highest risk per person. For instance, the Panakukkang subdistrict had the highest number of recorded cases, but because it is also very populous, the risk to an individual living there was actually lower than average.

The model identified a different area as the true hotspot. The Manggala subdistrict emerged as having the highest estimated relative risk, with a risk level nearly three times the city average. This finding held up even after accounting for population size and the travel patterns of patients. The researchers also identified three other subdistricts—Tamalate, Rappocini, and Tallo—as high-risk areas. The study suggests that Manggala's high risk is linked to its specific environment: it is a low-lying area with poor drainage and a nearby landfill, creating stagnant pools of water that are perfect breeding grounds for the mosquitoes that carry dengue. While the study confirms that vegetation is generally associated with higher risk, it also showed that simply counting cases can be misleading if you do not adjust for how many people live there and where they go for help.

This research offers a powerful new tool for public health officials in cities like Makassar and beyond. By using a method that accounts for incomplete data and patient travel behavior, officials can move beyond simple case counts to see the true landscape of disease risk. The study does not claim to have solved the problem of dengue, but it provides a much clearer picture of where the danger lies. The findings suggest that resources should be directed toward areas like Manggala, where the risk to individuals is highest, rather than just the areas with the most hospital visits. Ultimately, the work demonstrates that even when data is messy and incomplete, it is possible to extract reliable, life-saving insights by understanding the human behavior behind the numbers.

Technical Summary: Bayesian Disease Mapping from Incomplete Hospital-Based Surveillance

Problem Statement
Dengue surveillance data in many settings, particularly in low- and middle-income countries, are collected at healthcare facilities rather than at patients' residences. Consequently, reported case locations reflect healthcare-seeking behavior and hospital catchment dynamics as much as the underlying disease process. A significant challenge arises from the heterogeneity of location data: some cases are geolocated to exact coordinates, some are recorded only by subdistrict, and others carry no residential information at all. Treating these incomplete data as fully observed or imputing missing origins via deterministic rules introduces spatial misclassification and biases risk estimates. Furthermore, standard spatial models often fail to explicitly distinguish between the latent disease intensity and the observation process (hospital attraction), leading to unreliable spatial inference.

Methodology
The authors propose a unified hierarchical Bayesian framework that jointly models the latent disease process and the hospital observation process. The approach consists of two main components:

Latent Disease Intensity (IPP–BYM2): The underlying spatial intensity of dengue is modeled as an inhomogeneous Poisson process (IPP) with a population-adjusted log-intensity. The intensity structure follows a Besag–York–Mollié (BYM2) specification, which combines a spatially structured random effect (via an intrinsic conditional autoregressive prior) and an unstructured random effect. This allows for the estimation of per-person relative risk at the subdistrict level while accounting for spatial autocorrelation and unstructured heterogeneity.
Hospital Observation Process (Gravity Kernel): The link between a patient's residence and the reporting hospital is modeled using a distance-based gravity (power-law) kernel. The probability of a case from subdistrict $j$ being observed at hospital $h$ decreases with travel distance $d_{hj}$ according to $k(d) = (1 + \gamma_h d)^{-b}$ , where $\gamma_h$ is a hospital-specific scale parameter and $b$ is a shared shape exponent.
Handling Missing Data: The framework accommodates three levels of spatial resolution within a single likelihood function:
- Geolocated cases: Contribute exact travel distances.
- Subdistrict-level cases: Contribute the distance from the subdistrict centroid to the hospital.
- Unlocated cases: Treated as latent variables. Their origins are marginalized over all possible subdistricts, with the posterior distribution of their origins inferred jointly with the risk surface and catchment parameters. This propagates uncertainty rather than fixing origins via imputation.

Key Contributions

Methodological Innovation: The paper introduces a principled approach to disease mapping that explicitly separates the disease process from the observation mechanism. It demonstrates that treating unlocated case origins as latent variables within a single likelihood allows for the simultaneous estimation of risk, catchment behavior, and missing origins.
Validation of Catchment Kernels: Through out-of-sample cross-validation (leave-one-hospital-out), the study provides empirical evidence that a gravity (power-law) kernel substantially outperforms both the conventional exponential kernel and deterministic allocation baselines (uniform and population-proportional) in recovering the true residential origins of held-out cases.
Robust Risk Estimation: The framework successfully recovers high-risk areas that differ from rankings based on raw case counts, demonstrating the necessity of population adjustment and catchment modeling.

Results
Applied to 1,842 dengue cases from eight hospitals in Makassar, Indonesia (2021–2025), the model yielded the following findings:

Origin Recovery: The proposed model with the gravity kernel reduced the mean total variation distance (TVD) between inferred and observed origin distributions to 0.243, significantly outperforming the exponential kernel (TVD 0.370) and deterministic baselines.
Risk Identification: The model identified Manggala as the subdistrict with the highest estimated relative risk (RR = 2.73; 90% CrI: 2.41–3.06), with an exceedance probability (EP) of 1.00. Four subdistricts met the high-risk criterion (EP $\ge$ 0.80): Manggala, Tamalate, Rappocini, and Tallo.
Discrepancy with Raw Counts: Rankings by model-estimated relative risk differed markedly from rankings by recorded case counts. For instance, Panakukkang had the highest raw case count (259) but an estimated RR of 0.95, while Biringkanaya (the most populous subdistrict) had a low RR of 0.91 despite 94 recorded cases.
Covariate Effects: On a per-person scale, Normalised Difference Vegetation Index (NDVI) showed a positive association with dengue risk ( $\beta = 0.29$ ), suggesting vegetated environments may facilitate transmission. Health facility density showed a weak, non-significant negative association.
Catchment Behavior: Estimated mean catchment distances were consistent across hospitals (3–5 km), with 75–84% of inferred cases originating within 5 km of the facility.

Significance and Claims
The authors claim that this framework offers a "principled approach" to disease mapping from incomplete, facility-based surveillance data, a common scenario in epidemiology. By treating missing origins as latent variables and marginalizing over them, the method avoids the bias introduced by ad hoc imputation. The study emphasizes that the identification of high-risk areas (such as Manggala) is driven by the combined effect of population adjustment, catchment modeling, and the inference of missing origins, rather than population adjustment alone.

The authors note that the identification of Manggala as a hotspot is ecologically plausible, aligning with its low-lying basin geography, seasonal inundation, and proximity to a landfill, which create favorable micro-habitats for Aedes aegypti.

Limitations and Scope
The paper modestly acknowledges several limitations:

Ecological Fallacy: Associations are at the subdistrict level and cannot be interpreted as individual-level effects.
Cross-Sectional Nature: The analysis pools data from 2021–2025, failing to capture seasonal or interannual transmission dynamics.
Model Saturation: With only 14 subdistricts, the model is near-saturated, leading to weak identification of some parameters (e.g., spatial mixing $\rho$ ).
Data Scope: The data are limited to eight major hospitals and may miss cases from primary care or those not seeking formal care. The observation model treats hospital choice independently rather than as a competing-destinations problem.
Generalizability: Findings are specific to Makassar and may not transfer directly to settings with different populations or health systems.

The authors conclude that while the framework provides a robust method for current data, future work should explore spatio-temporal extensions, human mobility integration, and competing-destinations observation models.

Technical Summary: Bayesian Disease Mapping from Incomplete Hospital-Based Surveillance

More like this