Bayesian spatio-temporal modelling of air pollution and disease risk with spatially misaligned data

Zhu, Qiangqiang (2026) Bayesian spatio-temporal modelling of air pollution and disease risk with spatially misaligned data. PhD thesis, University of Glasgow.

Abstract

The spatial support at which environmental and health data are observed often does not coincide with the support at which inference is required. In Scotland, air quality is measured at a sparse monitoring network, respiratory prescribing is recorded at general practitioner (GP) surgery level, and small-area population and socio-economic covariates are available on Data Zones (DZs). Assigning exposures to populations, realigning GP-level outcomes, and mapping small-area disease risk each require the analyst to bridge supports that do not align by design. This thesis develops statistical methodology for the combined predictions of air pollution exposures, estimation of the effects of air pollution on respiratory ill health treated in primary care, and the small-area risk estimation of respiratory ill health treated in primary care where the supports of the underlying data are partially unknown or misaligned.
The first methodological strand addresses exposure prediction. Statistical and machine learning candidate approaches are compared under spatial cross-validation, with random forests selected for NO2 and PM2.5 and a linear model for PM10. The resulting monthly 1 km2 prediction surfaces cover Scotland from 2016 to 2020 with associated 95% prediction intervals, and provide the exposure resource used in the subsequent chapters.
The second study concerns the estimation of the health effect that air pollution has at the GP surgery level. Two population informed neighbourhood matrices are introduced to model the spatial autocorrelation in the data, which define spatial closeness between GP patient catchments through the overlap of their registered patient populations rather than the geographic distance between surgery locations. Both substantially reduce residual spatial autocorrelation relative to a 5-nearest-neighbours baseline definition based on surgery locations. Air pollution exposures are also assigned to each GP surgery from the residential distribution of its registered patients rather than using concentrations at the surgery location. Under this specification, a consistent positive association is identified between PM10 and preventer medication prescribing, with an estimated 0.5% increase in prescribing rates per one standard deviation increase in PM10. This finding is robust to full propagation of the predictive uncertainty from the exposure model.
The third study reverses the direction of the support problem, estimating relative risks on Data Zones from counts observed on GP surgeries. A Bayesian spatio-temporal disease mapping framework links GP-level observations to a latent DZ-level risk surface through a population-weighted change-of-support mechanism, adjusting an existing multiple-membership formulation to the relative-risk scale. Two simulation studies show that this alignment model consistently lowers the root mean squared error of DZ-level risk estimates relative to naive interpolation across plausible data generating scenarios. Applied to Glasgow City over 2019–2020 with a socio-economic covariate, the framework recovers a persistent and monotonic deprivation gradient in respiratory prescribing risk, with time-averaged posterior median relative risks in the most-deprived SIMD 2020 decile approximately four times those in the least-deprived decile. The estimated surface also identifies a city-wide spike in March 2020 coinciding with the onset of the Covid-19 pandemic.
Taken together, the thesis shows that spatial support mismatch is a primary modelling concern that shapes inference at each stage from exposure reconstruction to small-area risk estimation. Modelling the supports explicitly yields credible exposure predictions, pollution–health effect estimates, and small-area risk maps of respiratory ill health treated in primary care resolved below the level at which prescribing counts are recorded. The resulting estimates support more targeted environmental and public health action on respiratory burden in Scotland at the small-area scale relevant to policy.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: Q Science > QA Mathematics; R Medicine > RA Public aspects of medicine > RA0421 Public health. Hygiene. Preventive Medicine
• Colleges/Schools: College of Science and Engineering > School of Mathematics and Statistics
• Supervisor's Name: Lee, Professor Duncan and Stoner, Dr. Oliver
• Date of Award: 2026
• Depositing User: Theses Team
• Unique ID: glathesis:2026-86158
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 03 Aug 2026 14:59
• Last Modified: 04 Aug 2026 08:27
• Thesis DOI: 10.5525/gla.thesis.86158
• URI: https://theses.gla.ac.uk/id/eprint/86158

Related URLs

• Article DOI
• Article DOI