Source-linked AI summary
A space-time conditional intensity model for invasive meningococcal disease occurrence
Sebastian Meyer, Johannes Elias, Michael Höhle
TL;DR
The paper investigates whether the two common German meningococcal finetypes differ in spatio-temporal behaviour, addressing the relative scarcity of covariate-adjusted modelling in continuous space. It proposes a regression framework for spatio-temporal point processes and reports finetype-specific differences in spread tendencies, while noting limitations from observation-window edge effects.
Problem
The paper asks whether the two common German meningococcal finetypes differ in spatio-temporal behaviour, in a setting where explicit continuous-space covariate-adjusted epidemic modelling is rare.
Source: §1 · p. 3Method
The authors propose a spatio-temporal point-process regression framework with endemic and epidemic components for modelling, inference, and simulation.
Source: §1 · p. 3, §2.1 · p. 4Results
The analysis reports differences in IMD transmission dynamics by age and finetype, and no significant connection with concurrent influenza cases.
Source: §6 · p. 12Takeaways & Limitations
The analysis reports finetype-specific spread tendencies at a spatial scale smaller than district-level data allow it to examine.
Source: §6 · p. 12Because only Germany is observed, cross-border infections may make disease clusters wider than observed and potentially lead to underestimation of epidemic weight.
Source: §6 · p. 13
Abstract
from arXiv · showhide
A novel point process model continuous in space-time is proposed for quantifying the transmission dynamics of the two most common meningococcal antigenic sequence types observed in Germany 2002-2008. Modelling is based on the conditional intensity function (CIF) which is described by a superposition of additive and multiplicative components. As an epidemiological interesting finding, spread behaviour was shown to depend on type in addition to age: basic reproduction numbers were 0.25 (95% CI 0.19-0.34) and 0.11 (95% CI 0.07-0.17) for types B:P1.7-2,4:F1-5 and C:P1.5,2:F3-3, respectively. Altogether, the proposed methodology represents a comprehensive and universal regression framework for the modelling, simulation and inference of self-exciting spatio-temporal point processes based on the CIF. Usability of the modelling in biometric practice is promoted by an implementation in the R package surveillance.
1. Introduction
The paper develops a spatio-temporal point-process regression framework to investigate whether the two common German meningococcal finetypes differ in their dynamics. The motivating data show seasonal peaks and suggest spatial and temporal clustering, while case locations and infection timing are only observed through proxies.
1. Introduction
- The study asks whether the two most common German finetypes exhibit different spatio-temporal behaviour, using cases recorded from 2002–2008.A finetype is a unique combination of serogroup and sequence types of specified PorA and FetA protein regions.Source: §1 · p. 1
- The analysis uses postal-code patient residences as quasi-continuous spatial locations and specimen-sampling dates as the beginning of illness and infectivity.The spatial resolution available was the patient’s postal code; actual infection times were unknown.Source: §1 · p. 1
- A total of 636 infections were registered, comprising 336 finetype B and 300 finetype C; cases predominantly occurred during winter and early spring, with peaks of varying prominence.Monthly IMD case counts are shown separately for each finetype.Source: §1 · p. 2, Figure 1 · p. 2
- Finetype B cases had a maximum of 16 at one postcode, versus 4 for finetype C, while the maps also show district population density.The spatial patterns suggest endemic occurrence, and higher point intensity in metropolitan areas partly reflects population-density variation.Source: §1 · p. 2, Figure 2 · p. 3
- The authors aim to quantify IMD transmission dynamics and covariate dependence with spatio-temporal point processes, an approach described as rare in explicit continuous space.They note that balancing spatial resolution against case confidentiality is one explanation for the scarcity.Source: §1 · p. 3
- The proposed model is presented as a unifying regression framework for modelling, inference, and simulation of spatio-temporal point processes beyond epidemics.The article introduces the model before discussing inference and simulation, then applies it to IMD data.Source: §1 · p. 3
2. Spatio-Temporal Two-Component CIF Model
The model represents the conditional intensity as endemic and epidemic components, combining covariate-dependent imported cases with history-dependent person-to-person spread. It extends this framework to distinguish event types and their possible transmission pathways.
2. Spatio-Temporal Two-Component CIF Model
- The conditional intensity superimposes an endemic component and a history-dependent epidemic component for a process continuous in space and time.The endemic component represents otherwise imported cases; the epidemic component represents person-to-person spread.Source: §2 · p. 4
2.1 Specification of the Endemic Component h(t, s)
- On a space-time grid, the endemic rate is proportional to a known offset such as population density and varies with covariates through an exponential predictor.The model can include time-lagged influenza case counts as an endemic covariate.Source: §2.1 · p. 4
2.2 Specification of the Epidemic Component e∗(t, s)
- Each infectious individual contributes a triggering rate factored into mark, elapsed-time, and relative-location effects, with infectivity decaying over space and time.An event can contribute only within the specified maximum temporal and spatial interaction ranges.Source: §2.2 · pp. 4–5
2.3 Characteristics of the Model
- The model is called twinstim, and its individual-specific mean infection count provides a reproduction-number-like summary of infectivity.A unit increase in a continuous mark multiplies that mean by e^γ_l for its corresponding coefficient.Source: §2.3 · p. 5
2.4 Extension: Type-Specific twinstim
- The marked extension uses a transmission indicator matrix to specify which event types can trigger which others, including finetype-specific transmission for IMD.Endemic covariate effects are assumed homogeneous across types, while transmission regimes and interaction functions may depend on the infective event type.Source: §2.4 · pp. 5–6
3. Statistical Inference
Inference evaluates the marked point-pattern likelihood while treating unpredictable marks as given predictors. The additive model structure permits separate component integrations, followed by numerical parameter estimation and standard model assessment.
3. Statistical Inference
- Likelihood inference estimates the CIF parameters from the observed marked spatio-temporal point pattern, with unpredictable marks treated as given predictors rather than modeled.The parameter vector includes endemic, mark-effect, and spatial and temporal interaction parameters.Source: §3 · p. 6
- The additive superposition allows endemic and epidemic CIF integrals to be evaluated separately after identifying potential infection sources.The endemic integral is a sum over the space-time grid.Source: §3 · pp. 6–7
- The spatial interaction integral is evaluated numerically with a two-dimensional midpoint rule, while typical temporal interaction integrals are analytically accessible.The midpoint rule was found best suited among the compared cubature rules.Source: §3 · p. 7
- Numerical optimization is required; uncertainty is estimated using expected Fisher information, with Wald or likelihood-ratio tests and AIC for model assessment.The score function is available to support optimization.Source: §3 · p. 7
4. Simulation Algorithm
Simulation uses Ogata’s modified thinning algorithm with piecewise upper bounds for the ground-process intensity. It proposes event times, accounts for rate changes, and then samples event sources, locations, and types.
4. Simulation Algorithm
- Ogata’s modified thinning algorithm simulates realizations using piecewise upper bounds for the ground-process intensity.The CIF is piecewise constant in time and changes when endemic covariates or the set of currently infectious individuals changes.Source: §4 · pp. 7–8
- An exponentially distributed waiting-time proposal is accepted according to the CIF rate, while a rate change before the proposed time triggers a restart at the first changepoint.The process repeats proposals until an event time is accepted.Source: §4 · p. 8
- After accepting an event time, the algorithm samples its source and then its type and location, using different rules for endemic and triggered events.Triggered event types follow the transmission matrix, and their locations are drawn relative to the infecting event using the spatial interaction density.Source: §4 · p. 8
5. Application to the IMD Data
The IMD application uses joint AIC-based model selection and goodness-of-fit checks to assess type- and age-specific spread. The fitted model indicates higher epidemic potential for serogroup B than C, while sensitivity and simulation checks identify important boundaries to interpretation.
5. Application to the IMD Data
- AIC selection used joint data from 630 complete cases, screened 600 configurations, then compared Gaussian spatial interactions for the top 10 models; ε was fixed at 30 days.The lowest-AIC model had AIC=18968.Source: §5 · pp. 8–9, Table 1 · p. 9
- Adding an epidemic intercept improved fit by ∆AIC=202.84; type C's spatial interaction was 43% of type B's, and ages 3–18 had the highest infectivity potential.A downward time trend with one harmonic described endemic behavior; adding past influenza counts did not improve the model.Source: §5 · p. 9
- Basic reproduction numbers were 0.25 (95% CI 0.19-0.34) for serogroup B and 0.11 (95% CI 0.07-0.17) for serogroup C.The confidence intervals indicate higher epidemic potential for the serogroup B finetype.Source: §5 · pp. 9–10
- Random U(0, 1) tie-breaking improved the goodness-of-fit plot relative to subtracting ε=0.01 days, with only minor relative changes in parameter estimates.The CDF assessment includes 95% Kolmogorov-Smirnov error bounds for both tie-breaking schemes.Source: §5 · p. 10, Figure 4 · p. 11
- Across 413 districts, 17 observed incidences exceeded the simulated 95% ranges from 100 realizations, many near Aachen, consistent with possible cross-border edge effects.The comparison uses district-specific seven-year incidences per 100,000 inhabitants.Source: §5 · p. 11
- Sensitivity checks preserved the type difference across tested ε ranges, but age-group findings varied and the data provided too little information to estimate ε.The 3–18-year group remained highest; above 35 days, the oldest group ranked second, so ε was fixed at a biologically plausible value.Source: §5 · p. 11
6. Discussion
The analysis found age- and finetype-dependent transmission patterns, while the continuous-space framework accommodates general space-time interactions and prospective change-point monitoring.
6. Discussion
- Although manifested cases leave the transmission network, case marks can proxy transmission links when surveillance data lack information on asymptomatic carriage.Secondary cases may acquire the strain from a case during incubation or from nearby asymptomatic carriers.Source: §6 · pp. 11–12
- Transmission dynamics differed by age and finetype; the analysis found no significant association with concurrent influenza cases and localized spread below district-scale resolution.The authors describe their analysis as the first report of finetype-specific differences in spread tendencies.Source: §6 · p. 12
- Continuous-space epidemic modelling differs from discrete-population SIR settings because cases do not arise from a predefined population or fit model compartments.The setting also requires population density and separates event covariates from surrounding-environment covariates.Source: §6 · p. 12
- The proposed space-time interaction functions are general, avoiding constraints imposed by earlier linear basis decompositions.Source: §6 · p. 13
- Germany-only observation can miss infections outside the study window, potentially underestimating epidemic weight, while edge correction remains an open methodological issue.The simulation identified higher-than-expected incidence around Aachen, consistent with cross-border spread.Source: §6 · p. 13
- The parametric framework could support prospective change-point monitoring for when an epidemic component becomes necessary to describe observed data.The authors suggest likelihood-ratio detectors within stochastic process control.Source: §6 · p. 13
- Inference and simulation methods for twinstim models are available in the R package surveillance.Source: §6 · p. 13