Fig. 2: Geographic regression maps showing interaction and individual effects for cognition and air pollution in Leeds with smoothing applied (95% of kernel density within a 5 km diameter).

Each row originates from a separate GeoSPM model comprising a shared set of covariates (not shown here) and a specific combination of cognition and pollution, while each column groups (from left to right) the effects for interaction, cognition and pollution. A colour scale from blue to red has been applied per column. Significant areas of a two-tailed t-test at p < 0.05 FEW (voxel-level family-wise correction) with at least ten participants for the interaction effect are marked by a white border and reproduced in the individual effects. The areas are numbered and referenced in the last column, in which the distribution of coefficients is summarised (min-max, IQR and median) per area, to compare their dominant sign and help in the interpretation of the interaction. Rising double bars indicate an interaction with positive increments for individual effects, falling double bars indicate an interaction with negative increments, crossed bars indicate an interaction with opposite increments, and missing bars indicate inconsistent signs between interaction and individual effects. UK Biobank provide locations at a resolution of 1 km² and each map shows an area of 26 by 26 km. The quintiles of the sample density are indicated by the size of a circle. A The cognition variable is reaction time and each row represents a different measure of pollution, from top to bottom: NO₂, NO_x and PM_2.5. B The cognition variable is pair matching completion time with the same measures of pollution as in (A).