Evolving Drivers of Brazilian SARS-CoV-2 Transmission: A Spatiotemporally Disaggregated Time Series Analysis of Meteorology, Policy, and Human Mobility

Abstract

Brazil has been severely affected by the COVID-19 pandemic. Temperature and humidity have been purported as drivers of SARS-CoV-2 transmission, but no consensus has been reached in the literature regarding the relative roles of meteorology, governmental policy, and mobility on transmission in Brazil. We compiled data on meteorology, governmental policy, and mobility in Brazil's 26 states and one federal district from June 2020 to August 2021. Associations between these variables and the time-varying reproductive number (R _t ) of SARS-CoV-2 were examined using generalized additive models fit to data from the entire 15-month period and several shorter, 3-month periods. Accumulated local effects and variable importance metrics were calculated to analyze the relationship between input variables and R _t . We found that transmission is strongly influenced by unmeasured sources of between-state heterogeneity and the near-recent trajectory of the pandemic. Increased temperature generally was associated with decreased transmission and increased specific humidity with increased transmission. However, the impacts of meteorology, policy, and mobility on R _t varied in direction, magnitude, and significance across our study period. This time variance could explain inconsistencies in the published literature to date. While meteorology weakly modulates SARS-CoV-2 transmission, daily or seasonal weather variations alone will not stave off future surges in COVID-19 cases in Brazil. Investigating how the roles of environmental factors and disease control interventions may vary with time should be a deliberate consideration of future research on the drivers of SARS-CoV-2 transmission.

Keywords: Brazil; COVID‐19; generalized additive model; meteorology; non‐pharmacological interventions; pandemic.

Figure 1

(a) Cumulative cases of COVID‐19 per 100,000 population as of 31 August 2021 in selected Brazilian states and (b–g) time series of state‐level R _t generated by EpiNow2. Selected states represent five most populous states in 2021, and time series of additional states are shown in Figure S2 in Supporting Information S1. For contrast, the colorbar in (a) saturates at 4,000 and 16,000.

Figure 2

Accumulated local effects of (a) temperature, (b) specific humidity, (c) the Oxford COVID‐19 Government Response Tracker policy index, (d) Google workplaces mobility, (e) Google residential mobility, and (f) the cumulative number of cases in the preceding 30 days. Effects of model terms are shown for values between each term's 10th and 90th percentiles. Shaded bands for each curve denote the 95% confidence interval. Note the different scale of the vertical axis in panel (f).

Figure 3

Permutation‐based variable importance plot for generalized additive model (GAM) model terms using the root mean square error as the loss function. Larger values for a particular term indicate that removal of that variable causes the GAM to lose accuracy in its prediction. The zoomed‐in version of the gray region in the left panel is shown on the right.