Thesis
On the simultaneous inference of susceptibility distributions and non-pharmaceutical interventions from epidemic trajectories
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2026
- Thesis identifier
- T18115
- Person Identifier (Local)
- 202279132
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- Individual variation in susceptibility can slow an epidemic through selective depletion of highly susceptible individuals, while non-pharmaceutical interventions (NPIs) reduce transmission through changes in contact behaviour. Because both mechanisms can produce similar epidemic slowdowns, distinguishing their effects from aggregate incidence data alone is difficult. This creates a practical identifiability problem for heterogeneous-susceptibility SEIR models and can lead to biased inference, misleading attribution of epidemic decline, and unreliable forecasts. This thesis studies inference in SEIR models with heterogeneous susceptibility and timevarying NPIs, with emphasis on identifiability, forecasting, and robustness to modelling assumptions. Using simulation experiments, it first compares heterogeneous and homogeneous susceptibility models under both homogeneous and heterogeneous data-generating settings. The results show that the heterogeneous model can reproduce homogeneous-like behaviour when the population is effectively homogeneous, while providing more reliable inference and forecasts when susceptibility variation is present. In contrast, a homogeneous model may fit the observed part of an epidemic while misattributing selective depletion to stronger interventions and misforecasting subsequent dynamics. The thesis then shows that single-epidemic inference is strongly affected by compensation between susceptibility heterogeneity and intervention strength. Across multiple inference methods and alternative NPI parameterisations, the coefficient of variation in susceptibility and the sustained contact level exhibit a pronounced trade-off, indicating that one epidemic trajectory often contains insufficient information to estimate both reliably. To address this, a joint-epidemic framework is developed in which related epidemics share common parameters but differ in their initial conditions. Joint fitting introduces additional contrasts in epidemic growth and susceptible depletion, thereby constraining the compensating directions present in single-epidemic analysis and substantially improving practical identifiability. To examine robustness to the assumed form of susceptibility variation, the thesis develops discretisation methods for continuous-trait models beyond the standard Gamma formulation and compares Gamma and Lognormal susceptibility families, with a homogeneous model included as a reference. The results show that distributional shape becomes important at moderate to high heterogeneity: although matched on mean and coefficient of variation, the Gamma and Lognormal models can produce different peak magnitudes, peak times, final sizes, parameter estimates, and forecasts. A misspecified heterogeneous model may still fit the observed epidemic well by shifting the estimated level of heterogeneity, and joint fitting can convert this hidden misspecification into precise but biased inference. Nevertheless, across the scenarios examined, using the wrong heterogeneous family is generally less damaging for forecasting than ignoring susceptibility heterogeneity altogether. The framework is finally applied to first-wave COVID-19 mortality data from England and Scotland using joint hierarchical models with different NPI specifications. The joint analysis produces informative estimates of susceptibility heterogeneity alongside country specific intervention effects, but the inferred magnitude of heterogeneity remains sensitive to the assumed NPI form, the infection-to-death delay, the prior structure, and the fact that only two epidemics are available. Overall, the thesis shows that separating susceptibility heterogeneity from NPI effects remains difficult from a single epidemic, but that joint inference from related epidemic time series, combined with explicit assessment of model misspecification, provides a more informative and transparent framework for inference and forecasting in heterogeneous epidemic models. Parts of the simulation studies reported in this thesis have been published in [1, 2]; related work underlying parts of the distributional misspecification analysis is archived as the preprint [3] and submitted for publication.
- Advisor / supervisor
- Robertson, Chris
- Gomes, Gabriela M.
- Resource Type
- DOI
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
PDF of Thesis T18115 | 2026-08-31 | Public | Download |