Our analysis captured strong spatial autocorrelation. The observed pattern could reflect residual variation from a potential inadequate covariate adjustment (including disease spread), spatial variation of pre-existing conditions, other unknown spatial confounders or a combination from all above. In a sensitivity analysis, we observed that the factors associated with disease transmission left the latent field unchanged (Supplemental Material Fig. S21), as did the inclusion of hypertension, diabetes and COPD (Supplemental Material Fig. S33). When we restricted the analysis to the pre-lockdown period, the latent field for both pollutants captured London and Birmingham, i.e. the cities with the first outbreaks. Considering the above, and the fact that COVID-19 is an infectious disease, we believe that large variation of Fig. 4 is likely due to disease spread, which is not adequately captured in the disease progression covariates.
Fig. 4 Median posterior spatial relative risk (exponential of the spatial autocorrelation term) and posterior probability that the spatial relative risk is larger than 1 for the models with NO2 and a spatial autocorrelation term and the fully adjusted NO2 model.