Predictive quality of census-based socio-economic indicators on Covid-19 infection risk at a fine spatial scale in France

Abstract

The COVID-19 pandemic in France induced the development of a national, high spatiotemporal resolution confirmed infection cases database. We aimed to estimate the predictive ability of census-based indicators on the infection risk to assess their potential usefulness in future pandemic response. We collected and aggregated all counts of biologically confirmed cases of SARS-CoV-2 infection in the Auvergne-Rhône-Alpes region in France at small-area statistical units between May 2020 and February 2021 (second wave). Ten census-based ecological covariates were evaluated as predictors of case incidence using a Poisson regression with conditional autoregressive (CAR) spatial effects. Benefits of CAR effects and covariates on model predictive ability was assessed comparing posterior predictive distribution of case incidence with the observed value for each statistical unit. Among 7,917,997 inhabitants, 438,992 infection cases over 5410 neighbourhoods were analysed. Spatial correlation was high for the periods before and after the epidemic peak, and illustrated with cartography. The addition of covariates to the null model led to an increase in satisfying prediction of + 5% from 14%, with a maximum of 21% across all periods. The ecological covariates assessed were insufficient to provide a satisfying prediction of infection risk without explicitly accounting for the spatial organization of the epidemic.

Keywords: Emerging diseases; Epidemiology; Modelling; SARS-CoV-2; Spatial statistics.

Fig. 1

Estimated incidence rate ratios for each covariate in the full model (M4), and their credibility intervals at 95%. The credibility interval for the Low incidence period is not shown to maintain clarity. Interpretation: during the Low incidence period, in an IRIS with around 17% of immigrants, the Incidence Rate is about 1.25 times that in an IRIS at the mean proportion of immigrants (about 7%). During the Peak and decrease period, in an IRIS with a population a quarter that of the regional mean (1,562), the Incidence Rate is about 0.75 times that in an IRIS with the population at the regional mean.

Fig. 2

Predicted incidence rate ratio of SARS-CoV-2 infection for the IRIS (infra-municipal spatial unit) of the French region Auvergne-Rhône-Alpes during the second epidemic wave: model with covariates and spatial effect (M4). From A to D: Low incidence, Growth, Peak and decrease, and Stabilization periods. Insets for each of the metropolitan areas labelled on the map (Clermont-Ferrand, Lyon, Grenoble, Saint-Etienne) are presented in the Supplementary material (Fig. S3 to S6). Produced in R, with packages ggplot2 (3.5.1), sf (1.0.16), ggrepel (0.9.5), ggspatial (1.1.9), and patchwork (1.2.0) (https://cran.r-project.org/).