📊 statistics

Spatially continuous modelling of aggregated outcome data

This paper proposes a spatially continuous block aggregation approach for modeling aggregated outcome data with fine-resolution covariates, demonstrating its ability to provide reliable inferences at any desired spatial resolution while outperforming traditional centroid-based and aggregated covariate methods.

Original authors: Stephen Jun Villejo, Peter Diggle, Finn Lindgren, Haavard Rue, Guangquan Li, Ella White, Matthew Wade, Marta Blangiardo

Published 2026-04-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Stephen Jun Villejo, Peter Diggle, Finn Lindgren, Haavard Rue, Guangquan Li, Ella White, Matthew Wade, Marta Blangiardo

Original paper licensed under CC BY 4.0 (http://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

The Big Problem: The "Mismatched Puzzle"

Imagine you are trying to figure out how healthy a city is. You have two types of information, but they don't fit together:

The Health Data (The Big Picture): You know how many people got sick in specific neighborhoods (like "Downtown" or "The West Side"). These are big, irregular chunks of land called blocks.
The Risk Factors (The Fine Detail): You have a high-resolution satellite map showing exactly where people live, how old they are, and how much money they make. This data is like a pixelated image, with millions of tiny squares.

The Dilemma: You want to know the health risk for every single pixel on your map, but your health data only exists for the big blocks.

The Old Way (The "Average" Trap):
Traditionally, scientists would take the big block, calculate the average of the tiny pixels inside it, and pretend the whole block is just one average point.

The Metaphor: Imagine trying to understand the flavor of a fruit salad by taking a spoonful of the whole bowl, mashing it into a paste, and saying, "This is what an apple tastes like." You lose the specific taste of the apple, the banana, and the grape. You also lose the ability to tell if the apple is in the corner or the center. This is called aggregation bias.

The New Solution: The "Continuous Canvas"

The authors (Stephen Villejo and colleagues) propose a new way to solve this. Instead of treating the world as a patchwork quilt of separate blocks, they treat it as a continuous canvas (like a smooth painting) that happens to be observed through a coarse grid.

Here is how their method works, step-by-step:

1. The Invisible "Heat Map" (The Latent Process)

Imagine there is an invisible, smooth "heat map" of risk covering the entire country. This map exists everywhere, even in places where we haven't measured anything yet.

In the old methods, this heat map was jagged and only existed at the center of the blocks.
In this new method, the heat map is smooth and continuous, like a flowing river. It flows under the blocks, connecting them seamlessly.

2. The "Bucket" Analogy (The Sampling Model)

The researchers imagine that the big blocks (where we have data) are actually buckets sitting on top of this smooth river.

The water level in the bucket (the number of sick people) isn't determined by the water level at just one point (the center).
Instead, the bucket collects water from every drop of the river that flows underneath it.
The model calculates the total "flow" of risk under the bucket to explain why that bucket is full or empty.

3. The Magic Trick: Reversing the Bucket

Once the model understands how the buckets fill up based on the smooth river underneath, it can do the reverse. It can look at the buckets and say, "Ah, if the bucket is full, the river underneath must be flowing fast here."

This allows them to disaggregate the data. They can take the big block data and "pour" it back out to predict the risk for every single tiny pixel on the map.

Why is this better? (The Simulation Results)

The authors ran computer simulations to test their method against the old ways.

When predicting big blocks: All methods (Old and New) were about the same. If you just want to know the average risk for "Downtown," the old way works fine.
When predicting tiny details: The new method shined.
- The Metaphor: If you want to know the risk for a specific street corner, the old method is like guessing the weather based on the average of the whole county. The new method is like having a weather station on that specific corner.
- The old methods often failed to capture sharp changes in risk (like a sudden spike in disease in a small area) because they smoothed everything out too much. The new method kept the sharp details.

Real-World Examples

The paper tested this on two real-life scenarios:

1. Wastewater Viruses (The Sewage Test)

The Setup: Scientists test sewage from specific treatment plants (irregular shapes). They want to know the virus levels in specific administrative towns (different shapes).
The Result: The new method successfully mapped the virus levels from the sewage plants onto the towns, even though the shapes didn't match up perfectly. It used population density (a fine-detail map) to help guess where the virus was likely hiding.

2. Heart Disease Hospitalizations

The Setup: They had hospital data for large districts but wanted to know the risk for tiny neighborhoods (LSOAs).
The Result: They used fine-detail data about poverty and age to predict exactly which tiny neighborhoods had the highest risk. The new method gave a much clearer, high-definition picture of where the danger was, whereas the old methods gave a blurry, low-resolution guess.

The Bottom Line

The Old Way: "Let's average everything out and hope for the best." (Good for big pictures, bad for details).
The New Way: "Let's assume the world is smooth and continuous, and our big data blocks are just buckets collecting from that smooth world." (Good for big pictures, excellent for details).

This approach allows scientists to take coarse, messy data and turn it into a high-definition map, helping public health officials target their resources exactly where they are needed, down to the neighborhood level.

1. Problem Statement

The paper addresses the issue of spatial misalignment, a common challenge in spatial epidemiology and environmental statistics where the response variable (outcome) and covariates (predictors) are observed at different spatial resolutions.

The Scenario: Response data ( $Y$ ) are aggregated over coarse administrative units (blocks, $B_i$ ), such as census tracts or sewage catchment areas. Conversely, covariate data ( $z$ ) are available at a much finer spatial resolution, typically as raster images or high-resolution grid cells ( $b_{ij}$ ).
Limitations of Current Methods:
- Centroid Approach: Treating block-level data as point data located at the block centroid and averaging covariates. This fails to capture spatial heterogeneity within blocks and suffers from aggregation bias in non-linear models.
- Markov Random Field (MRF): Aggregating covariates to the block level and modeling spatial dependence via a discrete graph. This limits inference to the specific block partition and cannot easily predict outcomes at finer resolutions (spatial disaggregation) or for arbitrary block configurations.
- Modifiable Areal Unit Problem (MAUP): Results can vary significantly depending on how blocks are defined, leading to unstable inferences.

2. Methodology

The authors propose a Block Aggregation Model that treats the underlying process as inherently continuous in space, regardless of the observation scale.

Core Model Structure

The model consists of a latent spatially continuous process $S(x)$ and a sampling model for the observed block-level response $Y_i$ .

Latent Process: At the fine resolution ( $b_{ij}$ ), the linear predictor is defined as:
$S(b_{ij}) = \beta^\top z(b_{ij}) + R(b_{ij})$
Where $\beta$ are fixed effects, $z$ are covariates, and $R$ is a latent spatially correlated Gaussian process (typically a Matérn field).
Sampling Model: The block-level response $Y_i$ $Y_{i}$ is determined by integrating the latent process over the block area $B_i$ $B_{i}$ via an inverse link function $f(\cdot)$ $f (\cdot)$ and a weighting function $w(x)$ $w (x)$ :
$\mu_i = E[Y_i] = \int_{B_i} w(x) f(S(x)) dx \approx \sum_{j=1}^{m_i} w(b_{ij}) f(S(b_{ij}))$
- Gaussian Case: Uses an identity link. The block mean is the weighted average of the fine-scale linear predictors.
- Poisson Case: Uses a log link. The block mean is the sum of exponentiated fine-scale predictors (analogous to a log-Gaussian Cox process).

Inference Framework

Bayesian Inference: The model is fitted using a hierarchical Bayesian framework.
Linearised INLA: Since the link function (especially for Poisson) makes the predictor non-linear with respect to the latent variables, standard Integrated Nested Laplace Approximation (INLA) cannot be applied directly. The authors employ a linearisation approach (Lindgren et al., 2024):
- A first-order Taylor expansion approximates the non-linear function around a current estimate $u_0$ .
- The model is fitted iteratively: update the linearisation point, fit the linearised Gaussian model using INLA, and repeat until convergence (when the posterior mode matches the linearisation point).
SPDE Approach: The continuous spatial field $R(\cdot)$ is approximated using the Stochastic Partial Differential Equation (SPDE) approach with a triangulated mesh, enabling efficient computation.

3. Key Contributions

Unified Framework for Misalignment: The method provides a principled way to model data with mismatched resolutions without pre-aggregating covariates, thereby preserving fine-scale information.
Spatial Disaggregation Capability: Unlike MRF or centroid models, this approach naturally supports downscaling (predicting outcomes at the fine covariate resolution) and prediction over arbitrary partitions that do not align with the original observation blocks.
Handling Non-Linearity: The iterative linearised INLA method allows for the application of complex non-linear link functions (like Poisson) to block-aggregated data while maintaining computational efficiency.
Avoidance of Aggregation Bias: By integrating the non-linear function over the block rather than applying the function to the average covariate, the model avoids the bias inherent in standard pre-processing approaches.

4. Results

Simulation Study

The authors compared their Block Aggregation approach against two standard comparators:

Comparator A (Centroids): Treats blocks as points at centroids with averaged covariates.
Comparator B (MRF): Uses a discrete spatial graph on blocks with aggregated covariates.

Findings:

Block-Level Prediction: All three methods performed similarly when predicting at the original block level, except when the spatial correlation range was smaller than the block size. In these cases, the MRF approach performed significantly worse due to its inability to resolve fine-scale variation.
Fine-Resolution Prediction (Disaggregation):
- Gaussian Case: The Block Aggregation and Centroid approaches were equivalent (due to linearity), while MRF could not perform disaggregation.
- Poisson Case: The Block Aggregation approach significantly outperformed both comparators in predicting fine-scale ( $\mu_{ij}$ ) outcomes. The MRF and Centroid approaches suffered from aggregation bias, producing biased estimates and poor coverage for fine-scale predictions.
Coverage: The Block Aggregation method maintained nominal 95% credible interval coverage for both block and fine-scale predictions, whereas MRF often yielded low coverage.

Real-World Applications

Wastewater Epidemiology (Gaussian):
- Data: SARS-CoV-2 concentrations in sewage catchment areas (irregular shapes) vs. population density (1km grid).
- Goal: Predict virus concentration at the Lower Tier Local Authority (LTLA) level.
- Result: The Block Aggregation and Centroid methods produced nearly identical predictions for LTLAs. The MRF approach produced different, more variable results due to the difficulty of mapping irregular catchments to a regular LTLA graph.
Cardiovascular Hospitalisations (Poisson):
- Data: Hospitalisation counts at LTLA level vs. socio-demographic covariates at the finer LSOA level.
- Goal: Disaggregate hospitalisation risk to the LSOA level.
- Result: The Block Aggregation model successfully generated high-resolution risk maps. It revealed strong correlations with the proportion of elderly populations and weaker correlations with deprivation indices at the fine scale, demonstrating the utility of disaggregation for public health planning.

5. Significance

Methodological Advancement: The paper bridges the gap between geostatistics (continuous space) and areal unit modeling (discrete space), offering a flexible solution for the "spatial misalignment" problem.
Public Health Impact: The ability to perform spatial disaggregation is critical for precision public health. It allows researchers to map disease risk at the resolution of available covariates (e.g., neighborhood deprivation) even when only coarse administrative data is available, enabling more targeted interventions.
Computational Efficiency: By utilizing linearised INLA, the method avoids the computational burden of MCMC sampling, making it suitable for near-real-time surveillance systems and large-scale national datasets.
Robustness: The approach is robust to the definition of administrative boundaries (MAUP) because the underlying process is continuous, allowing for predictions over any arbitrary spatial partition required by the application.