Source-linked AI summary
Power-law models for infectious disease spread
Sebastian Meyer, Leonhard Held
TL;DR
The paper addresses limited spatial-interaction formulations in infectious-disease surveillance models. It embeds power-law decay into individual-level and aggregated-count frameworks using Euclidean distance or neighbourhood order, and reports improved performance in both applications.
Problem
Existing surveillance models represented spatial interaction crudely, motivating models that better reflect human travel behaviour and occasional long-range transmission.
Method
The paper extends endemic–epidemic point-process and multivariate time-series models with power-law spatial interaction estimated jointly by likelihood methods.
Results
Power-law formulations performed better than previous Gaussian or first-order interaction models in both applications and were in line with unrestricted piecewise-constant alternatives.
Takeaways & Limitations
Power-law travel behaviour translates into a useful modelling formulation for infectious-disease spread across individual and regional surveillance data.
Takeaways & Limitations
In the meningococcal application, the estimated short-range scale was not interpretable because it was smaller than the data’s spatial resolution.
Abstract
from arXiv · showhide
Short-time human travel behaviour can be described by a power law with respect to distance. We incorporate this information in space-time models for infectious disease surveillance data to better capture the dynamics of disease spread. Two previously established model classes are extended, which both decompose disease risk additively into endemic and epidemic components: a spatio-temporal point process model for individual-level data and a multivariate time-series model for aggregated count data. In both frameworks, a power-law decay of spatial interaction is embedded into the epidemic component and estimated jointly with all other unknown parameters using (penalised) likelihood inference. Whereas the power law can be based on Euclidean distance in the point process model, a novel formulation is proposed for count data where the power law depends on the order of the neighbourhood of discrete spatial units. The performance of the new approach is investigated by a reanalysis of individual cases of invasive meningococcal disease in Germany (2002-2008) and count data on influenza in 140 administrative districts of Southern Germany (2001-2008). In both applications, the power law substantially improves model fit and predictions, and is reasonably close to alternative qualitative formulations, where distance and order of neighbourhood, respectively, are treated as a factor. Implementation in the R package surveillance allows the approach to be applied in other settings.
1. Introduction.
The paper extends infectious-disease surveillance models by incorporating power-law spatial interaction, motivated by human travel behaviour and intended to capture occasional long-range transmission. It applies the extensions to individual cases and aggregated counts, comparing them with earlier and qualitative alternatives.
- The study extends both a spatio-temporal point-process model for individual data and a multivariate time-series model for aggregated count data.
- Earlier surveillance models represented spatial interaction crudely, using a Gaussian kernel for individual cases and first-order adjacency for aggregated counts.
- Short-time human travel behaviour is described by a decreasing power law of distance, f(x) ∝ x^-d, with positive decay parameter d.
- The power law has a heavy tail, allowing occasional long-range infectious transmissions alongside principal short-range infections.
- The extensions are evaluated by reanalysing invasive meningococcal disease and influenza surveillance data against previous models and alternative qualitative interaction formulations.
2. Individual-level model.
The individual-level framework models infection events with endemic and observation-driven epidemic components, replacing Gaussian spatial interaction with power-law-based alternatives. Kernel parameters are estimated by likelihood methods and evaluated alongside qualitative distance effects.
- The point-process model describes individual infection events and their potential to trigger secondary cases through a conditional intensity.
- Its endemic component uses population-related offsets and exogenous covariates, while the epidemic component superposes infection pressures from previously infected individuals.
- Spatial infection pressure depends on Euclidean distance under an isotropic assumption, with temporal decay described separately as a function of elapsed infectious time.
- The proposed power-law interaction permits occasional long-range transmission, unlike the previously used Gaussian kernel.
- The study considers basic, lagged, and Student power-law kernels, plus an unconstrained step function treating distance intervals as qualitative categories.
- Parameters are estimated by full likelihood maximization, with kernel parameters estimated on the log-scale and numerical integration required for most spatial kernels.
3. Count data model.
The count-data framework models regional disease counts with endemic, autoregressive, and spatio-temporal epidemic components. It replaces first-order adjacency weights with power-law decay over neighbourhood order and assesses predictions using proper scoring rules.
- The multivariate time-series model assumes spatially and temporally aggregated counts follow a negative binomial distribution with an additively decomposed mean.
- The mean combines an endemic component, temporal autoregression, and transmission from cases in other regions through a spatio-temporal epidemic component.
- Region-specific predictors contain fixed and random effects, exogenous covariates, and application-specific time effects such as harmonic waves.
- Previous models restricted transmission during one period to first-order neighbours and normalised source-region weights so cases were distributed across adjacent regions.
- 3.2. Power-law extension.: Neighbourhood order is defined by the shortest route across distinct adjacent regions, yielding a symmetric discrete distance matrix.
- 3.2. Power-law extension.: The count-data extension applies power-law decay to neighbourhood order, with larger d assigning less importance to higher-order neighbours and d → ∞ recovering first-order dependence.
- Performance is compared through one-step-ahead forecasts using logarithmic and ranked probability scores, for which lower values indicate better predictions.
4. Applications.
Applications to meningococcal disease and influenza show that power-law spatial interaction can improve model fit and prediction while representing both local transmission and occasional longer-range spread. The analyses also identify data-resolution, sensitivity, and model-specification boundaries.
- IMD application: 635 IMD cases were analysed with power-law and alternative spatial-interaction models, accounting for tied observations caused by limited spatial and temporal resolution.The data recorded dates and residence postcodes, requiring interval-censoring treatment and random tie-breaking for the continuous-space, continuous-time point process.
- IMD application: The power-law kernel places more weight on localised IMD transmission while retaining a heavier tail that permits occasional long-range transmissions.Compared with the Gaussian kernel, it shows faster initial distance decay and supports broader geographical spread through its heavy tail.
- IMD application: The lagged power-law model estimated a short-range dispersal radius of ˆσ = 0.40 (95% CI: 0.18 to 0.86) kilometres, but this scale was not covered by the data resolution.Sensitivity replicates were more dispersed, with mean ∆AIC = −21.1, SD = 3.8 compared to the Gaussian kernel.
- IMD application: The step function confirms the power-law shape, but has a slightly better fit with mean ∆AIC = −6.9, SD = 4.0 compared to the power law.Its advantages are offset by dependence on knot choice, sensitivity to data artifacts, and loss of monotonicity.
- Influenza application: The influenza power-law formulation reduced the weight of the endemic component and increased the importance of the spatio-temporal component, while the combined epidemic measure remained broadly unchanged.Power-law weights allow information from other regions to contribute, so jumps to nonadjacent regions are no longer assigned only to the endemic component.
- Influenza application: For influenza, accounting for higher-order neighbours with power-law weights improved both logS and RPS relative to the previous first-order model.The estimated decay was ˆd = 1.80 (95% CI: 1.61 to 2.01), close to the exponent reported for short-time travel in the USA despite using neighbourhood order.
5. Discussion.
The discussion finds that power-law spatial interaction improves both disease-spread applications while clarifying practical benefits, limitations, and possible extensions.
- In both applications, power-law formulations outperformed previously used naive Gaussian or first-order interaction models.
- Alternative piecewise-constant interaction models agreed with the estimated power laws, supporting their qualitative adequacy.
- The heavy tail permits occasional long-range transmission in addition to predominantly short-range infections.
- Explicit immigration data could improve treatment of edge effects beyond the study’s distance-to-border proxy.
- Postcode-level censoring required randomly sampling case locations within postcode-centroid discs, although results were otherwise similar across tested radii.
- The 2008 influenza forecast improved overall predictions but failed to capture an onset two weeks earlier than in 2001–2007.
- Reported influenza counts may be affected by time-varying media attention, limiting interpretation of the observed case process.
- The approach is especially useful when movement-network data such as plane or train traffic are unavailable.
APPENDIX: SOFTWARE
The paper’s software appendix documents an R implementation of the model frameworks, power-law extensions, related functionality, and supporting computational packages.
- Both model frameworks and their power-law extensions are implemented in the R package surveillance.
- The package includes the analysed meningococcal-disease and influenza datasets.
- The implementations support alternative specifications of spatial interaction functions and neighbourhood weights.
- The package also includes the twinSIR() model for time-continuous individual surveillance data in closed populations with fixed locations.
- Spatial integration, mapping, and animation relied on the R packages polyCub, sp, and animation, respectively.
Supplement A: Animations of the IMD and influenza epidemics
Supplement A provides visual materials for comparing observed epidemics, simulated influenza counts, and predictive calibration summaries.
- The supplement includes observed evolutions of the invasive meningococcal disease and influenza epidemics.
- It includes simulated counts from various models for the 2008 influenza wave.
- It includes weekly mean PIT histograms for the influenza predictions.
- Supplement B contains inference details, polygonal integration methods, and additional power-law figures and tables.