TL;DR
This paper introduces a novel method, Restricted and Controlled Monte Carlo (RCMC), to map disease occurrences at a fine, near-individual level using only aggregated data. By disaggregating town-level birth defect data and incorporating population information, RCMC generates high-resolution maps that reveal spatial patterns and quantify uncertainty, overcoming limitations of traditional aggregate mapping.
What Problem Does This Paper Address?
Mapping disease patterns is crucial for understanding their links to environmental factors and public health. However, most existing disease maps rely on aggregate data from large, irregularly shaped administrative units (like towns or census tracts). This aggregation leads to several classic problems in spatial analysis:
- Modifiable Areal Unit Problem (MAUP): Results can change depending on how the boundaries of these units are defined.
- Small Number Problem: Data within a small area might be too sparse to be statistically reliable.
- Unrealistic Geographical Assumptions: Maps often assume uniform distribution or concentration at a single point (like a centroid), which isn't realistic.
- Unidentifiable Spatial Uncertainty: Aggregation obscures the true location of cases, and conventional maps offer no way to quantify this uncertainty.
These issues limit the precision of disease mapping and hinder the detection of associations with environmental factors. The paper aims to address this by developing a method to disaggregate aggregate data to a finer spatial resolution, enabling more accurate and informative disease mapping.
Methods
The core of this paper is the Restricted and Controlled Monte Carlo (RCMC) method, an extension of previous work. Here's how it works:
- Disaggregation:
- Instead of directly mapping aggregate data, the RCMC process first disaggregates it to an approximated individual level (represented by 100m x 100m pixels).
- Restricted Randomization: Each birth (both those with defects and the total "all births" cohort) is assigned a random location within its original town polygon. This randomization is restricted by the town boundaries.
- Controlled Randomization: The assignment probability for a pixel is controlled by detailed background population data (specifically, the density of women in relevant age groups). This ensures that births are more likely to be placed in areas with higher population density.
- Stochastic Nature: The process is repeated many times (e.g., 50 iterations). Each iteration produces a slightly different set of disaggregated locations, reflecting the inherent uncertainty from aggregation.
- Intensity Estimation (Kernel Ratio Estimation - KRE):
- Once disaggregated, the data is used to map disease intensity. The paper employs a case-side adaptive bandwidth Kernel Density Estimation (KDE), referred to as Kernel Ratio Estimation (KRE) in this context.
- This method calculates the local ratio between disease cases (birth defects) and the background population (all births). The bandwidth of the kernel adapts to the local density of births, providing a more robust estimate in areas with varying population densities.
- Statistical Significance and Uncertainty Quantification:
- Monte Carlo Simulation: To assess statistical significance, the RCMC process is combined with an unrestricted Monte Carlo simulation (UCMC). The UCMC simulates random distributions of birth defects to generate null hypothesis scenarios.
- P-value Calculation: By comparing the intensity from the actual RCMC disaggregation with the intensities from 99 UCMC simulations, a p-value is calculated for each pixel, indicating the statistical significance of the observed risk.
- Uncertainty Mapping: The variance across the multiple RCMC iterations (e.g., 50 iterations) is used to create a map of spatial uncertainty, directly quantifying the uncertainty introduced by the initial data aggregation.
- Hot Spot Identification: Hot spots are identified by combining the mean p-value and standard deviation across iterations, highlighting areas with statistically significant high intensity and low uncertainty.
Datasets
- New Hampshire Birth Defect Data (NHBCP): Data from 2003-2009, containing records of 2,289 birth defects. This data was processed to represent individual infants, excluding certain defects (e.g., fetal alcohol syndrome, chromosomal defects) and multiple births, resulting in a final dataset of 1,395 infants with defects.
- New Hampshire "All Births" Data: Birth certificates from the same period (N=101,435) from the NH Department of Human and Health Services. This dataset served as the background population, filtered to exclude non-NH addresses, multiple births, and mothers outside the 15-49 age range, resulting in 98,982 births.
- LandScan™ Data: Used for its detailed spatial population distribution, particularly in urban areas.
- U.S. Census Data: Used to provide demographic information (age and sex) that complements LandScan™ data.
The paper integrates LandScan™ and Census data to create detailed raster layers of women aged 15-39 and 40-49 years, with a 100m cell size, serving as the background population for disaggregation.
Key Insights and Ideas
- Disaggregation is Key: The fundamental insight is that disaggregating aggregate data to a finer spatial scale is essential for accurate disease mapping and for uncovering associations with environmental factors.
- RCMC as a General Approach: The RCMC method is presented not just for birth defects but as a general framework for incorporating auxiliary spatial information to improve disease mapping and quantify uncertainty, especially when dealing with entirely aggregate data.
- Stochastic vs. Deterministic: Unlike traditional dasymetric mapping, RCMC's stochastic nature (running multiple iterations) is crucial for explicitly modeling and quantifying the spatial uncertainty introduced by data aggregation.
- Three-Tiered Hierarchy: The study highlights a more complex case-cohort-population hierarchy compared to previous case-background models, necessitated by the specific nature of birth defects affecting a defined cohort (infants) rather than the general population.
- Uncertainty is Quantifiable: The paper demonstrates that spatial uncertainty stemming from aggregation is not an intractable problem but can be explicitly measured and visualized through repeated disaggregation iterations.
Key Results
- The RCMC method successfully disaggregated town-level birth defect data to a 100m pixel resolution.
- The study generated three types of maps:
- Mean p-value map: Shows the representative spatial distribution of birth defect risk, with smaller values indicating higher statistical significance.
- Standard deviation map: Visualizes the spatial distribution of uncertainty caused by data aggregation.
- Hot spots map: Identifies areas with significantly high intensity and low spatial uncertainty (marked if mean p-value + 2*std dev < 0.1).
- The hot spot map identified specific areas in New Hampshire with elevated birth defect risk, providing a more precise spatial focus for further investigation than traditional town-level maps.
- The method effectively maximizes the use of available spatial information by incorporating detailed population data into the disaggregation process.
Limitations and Future Work
- Number of Simulations: The paper acknowledges that 99 simulations for p-value calculation is a relatively small number, and more iterations would improve accuracy. This limitation is attributed to the computing capacity of the desktop PC used. The authors are working on migrating the program to a high-performance computing platform.
- Multiple Testing: The study recognizes the inherent problem of multiple testing in analyzing many grid cells independently. However, they argue that the kernel's overlap and the strict hot spot identification criteria (two standard deviations) mitigate this issue. A quantitative evaluation of these effects is suggested for future work.
- Interpretation of Hot Spots: The authors caution that large patches in the hot spot map do not necessarily mean the entire area is high-risk but rather that high-risk locations exist within those areas.
- Other Spatial Uncertainties: The paper notes that the quantified uncertainty primarily stems from spatial aggregation. Other sources of uncertainty, such as those from the Kernel Ratio Estimation (KRE) itself (e.g., bandwidth selection), are not fully addressed.
- Future Work: The authors suggest that the RCMC process can be further enhanced by incorporating additional auxiliary information (e.g., socioeconomic factors, environmental exposures) into the background data layer to improve spatial uncertainty and analysis quality. They also propose further research into quantitatively evaluating the effects of multiple testing mitigation.
Summary updated