Skip to main content
NCBI home page
GeoHealth logoLink to GeoHealth
. 2023 Mar 21;7(3):e2022GH000727. doi: 10.1029/2022GH000727

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

Gaige Hunter Kerr 1,✉, Hamada S Badr 2,3, Alisson F Barbieri 4, Josh M Colston 5, Lauren M Gardner 2, Margaret N Kosek 5, Benjamin F Zaitchik 6
PMCID: PMC10030230  PMID: 36960326

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: COVID‐19, Brazil, meteorology, non‐pharmacological interventions, generalized additive model, pandemic

Key Points

  • Unmeasured sources of between‐state heterogeneity and recent waves of cases are the dominant drivers of SARS‐CoV‐2 transmission in Brazil

  • The impacts of policy, meteorology, and mobility on transmission vary in direction and magnitude within subperiods of our study

  • Relying on proven mitigation measures such as mass vaccinations should be the key priority in the continued fight against COVID‐19

1. Introduction

The COVID‐19 pandemic, caused by the severe acute respiratory syndrome coronavirus 2 (SARS‐CoV‐2), has ravaged Brazil. As of July 2022, the country had recorded the second‐highest number of cases and second‐highest number of deaths globally (Center for Systems Science and Engineering, 2022). Disinformation sowed by Brazilian politicians; the defense of ineffective treatment based on chloroquine; less restrictive social isolation measures in some states and municipalities, especially those aligned with the federal government; difficulty in controlling the virus in Brazil's favelas (informal settlements); and a strained healthcare system have been suggested as key reasons for Brazil's unfortunate ranking with respect to the pandemic (Ponce, 2020).

Seasonality and meteorology, particularly temperature and humidity, have been purported drivers of SARS‐CoV‐2 transmission based on their impact on aerosolization of virus droplets, virus survival on fomites, host susceptibility, and human behavior (Lowen & Steel, 2014; Tamerius et al., 2013; Yang et al., 2015). Yet, myriad early studies investigating the associations between meteorology and COVID‐19 have not always reached consistent findings regarding the role of meteorological factors (Colston et al., 2023; Ma et al., 2021; Sera et al., 2021), although these and other studies generally emphasize that while the associations between COVID‐19 and meteorological variables may be significant, they are small compared with disease control interventions and could not entirely explain excess disease burdens. Several factors have been suggested as reasons for these inconsistent findings: a short temporal data record; simplistic statistical frameworks such as correlation analyses that overlook confounding factors; and error‐prone variables such as case counts, which could be biased toward the null due to underreporting, testing delays, and the proliferation of at‐home testing (Kerr et al., 2021; Mecenas et al., 2020).

Brazil's diverse climate, spanning equatorial and tropical zones in the north to temperate zones in the south, provides a unique range of meteorological conditions over which to examine the role of meteorology on transmission. Equally diverse is the political spectrum in Brazil's federative system, which gives relative autonomy to states and municipalities. This autonomy resulted in an ensemble of uncoordinated approaches toward COVID‐19 including, for example, facilitating the propagation of the virus given both strict and relaxed measures at different times (Castro et al., 2021; Kortessis et al., 2020). Studies aiming to understand how environmental factors may impact COVID‐19 in Brazil have not always accounted for these aforementioned uncoordinated governmental responses to the virus (e.g., Prata et al., 2022; Rosario et al., 2020); draw conclusions from a small geographic area (e.g., Rosario et al., 2020); or conduct their analysis only using data from the first few months during the pandemic (e.g., Pequeno et al., 2020; D. N. Prata et al., 2020; Yin et al., 2022). With these major methodological differences, there is unsurprisingly little consistency in the role of environmental factors on COVID‐19 in Brazil reported by studies.

Here, we conduct a spatiotemporally disaggregated time series study examining the roles of mobility, policy, and meteorology on SARS‐CoV‐2 transmission in Brazil's 26 states and federal capital district (hereafter generically referred to as “states”). Our innovative space‐time disaggregation additionally allows us to document how the drivers of transmission varied throughout the pandemic in 2020–2021. By incorporating 15 months of data into a flexible and interpretable statistical framework, we advance our understanding of what drives SARS‐CoV‐2 transmission gleaned from earlier studies that only considered data from earlier periods of the pandemic.

2. Materials and Methods

2.1. Data Sources

Our data‐driven study synthesizes time series of meteorological variables, policy, and mobility for each of Brazil's states over the period 1 July 2020–31 August 2021. The start date of this study is a few months after the first COVID‐19 case was detected in Brazil (25 February) but represents a period in which surveillance capabilities in Brazil were likely more well‐developed. Our 15‐month study period allows us to understand the impacts of meteorology, policy, and behavior on COVID‐19 transmission dynamics over an entire annual seasonal cycle.

The Johns Hopkins unified environmental‐epidemiological data set synthesizes data on meteorology, demography, and COVID‐19 control policies (Badr et al., 2021). From this database we extracted state‐level time series of population‐weighted daily average temperature and specific humidity at 2 m, originally derived from the fifth generation European Centre for Medium‐Range Weather Forecasts (ECMWF) atmospheric reanalysis. This reanalysis assimilates surface‐level observations, satellite‐derived measurements, and model output, providing continuous maps of meteorological variables at a spatial resolution of 0.25° × 0.25° (Hersbach et al., 2020). Hourly 2‐m temperature and specific humidity from the reanalysis were transformed to daily means; weighted by the population derived from WorldPop (Tatem, 2017); and thereafter averaged over each Brazilian state using geographic state boundaries from the Instituto Brasileiro de Geografia e Estatística (Instituto Brasileiro de Geografia e Estatística, 2021) Population weighting temperature and humidity accounts for the uneven distribution of population in several states. For example, over 50% of the population of Brazil's largest state by area, Amazonas (1.6 million km2), live in the capital city of Manaus (Instituto Brasileiro de Geografia e Estatística, 2022). Failing to population weight prior to calculating state averages could inflate the importance of temperatures or humidities in sparsely populated regions of states. We also incorporated a state‐level policy index from the Oxford COVID‐19 Government Response Tracker (OxCGRT), which estimates the strictness of lockdown policies using information on containment and closure policies and public information campaigns (Hale et al., 2021).

We included two mobility indicators from Google's COVID‐19 Community Mobility Reports (Google, LLC, 2022) in our study. Specifically, we consider time series of Google's workplaces and residential indicators. These changes represent departures from a pre‐pandemic baseline period (3 January–6 February 2020) and account for day‐of‐week variations. While these two mobility measures are inversely correlated (Spearman's rank correlation coefficient = −0.72), they represent different population‐level behaviors with respect to trip purpose and are measured differently. The residential indicator measures daily changes in the time spent in places of residence, and the workplaces indicator measures daily changes in total visitors to places of work.

The time‐varying reproductive number of COVID‐19 (Rt, the mean number of secondary infections caused by an infected person) was generated with the widely used transmission model EpiNow2 (Abbott et al., 2020) for each Brazilian state and used as the response variable in our study. For a given day in each state, EpiNow2 estimates R t based on cases by date of infection (Abbott et al., 2020; Gostic et al., 2020). Estimated infections are mapped to observed data (e.g., cases by date of report) using estimates of (a) the serial interval characterized by a gamma distribution with a mean of 5.2 days and standard deviation of 2.8 days, (b) the incubation period characterized by a log‐normal distribution with a mean of 5.1 days and standard deviation of 1.65 days, and (c) reporting delays characterized by a truncated log‐normal distribution with a mean of 2.5 days, standard deviation of 1.5 days, and maximum of 21 days (Alene et al., 2021; Ganyani et al., 2020; Lauer et al., 2020; McAloon et al., 2020; Rai et al., 2021). While this transmission model accounts for delays between infection onset and case reporting using parameters from the literature, we do not explicitly account for under‐ or over‐reporting of cases, which could bias the results. This transmission model accounts for quantifiable sources of uncertainty and propagates these uncertainties from the inputs to the final R t estimates. Recent work suggests R t is likely centered around 1 in most of the world in contrast to previous studies that reported a substantially higher value (Abbott et al., 2020). We found a mean R t of 0.996 (95% CI 0.994–0.998) in Brazilian states during our study period. R t for the entire nation was also generated to contrast with state‐level estimates.

2.2. Statistical Analysis

The impact of the meteorological, mobility, and policy on R t was quantified using generalized additive models (GAMs), semiparametric models that estimate the response variable, in our case R t , as the sum of nonlinear variable combinations or “smooth functions” (Hastie & Tibshirani, 1990). Examining individual smooth functions allows us to see the impact of a single variable or interactions between variables on R t . GAMs have been extensively used to assess the environmental health outcomes and drivers of COVID‐19 variability (Colston et al., 2023; Dominici, 2002; Sera et al., 2021). Unlike other types of linear models that assume, for example, normality or no multicollinearity, GAMs are flexible and can handle non‐linear and non‐monotonic relationships between explanatory and response variables.

We specified six different GAMs to predict daily variations of R t within each Brazilian state. The first used data from the entire study period, and the others used data from five different 3‐month periods: June–August 2020 (JJA 2020), September–November 2020 (SON 2020), December 2020–February 2021 (DJF 2020–2021), March–May 2021 (MAM 2021), and June–August 2021 (JJA 2021). These different GAMs allow us to understand the role of the environment, mobility, and policy on SARS‐CoV‐2 transmission over an entire seasonal cycle but also disentangle seasonality from within‐season variability and investigate how factors affecting transmission could change with time.

Specifically, we fit our GAMs to daily, state‐level R t assuming a Gaussian distribution with a log link. Our choice of a Gaussian family reflects the quasi‐normal distribution of the outcome variable, R t (Figure S1 in Supporting Information S1) and is also consistent with recent work modeling R t with GAMs (Colston et al., 2023). For each time period of interest, our model has the form:

Rt,e∼gaussianμt,e (1)
logμt,e=stemperaturet,e+shumidityt,e+sGoogleresidentialt,e+sGoogleworkplacest,e+sOxCGRTpolicyt,e+temperaturet,e⊗humidityt,e+σ(temperature)e⊗humidityt,e+slaggedcasest,e+s(state), (2)

where t denotes time, which is day in our case; e is each Brazilian state (“estado” in Portuguese); laggedcases is the total number of confirmed COVID‐19 cases during the preceding 30 days, which we include to account for autocorrelation; and σ(temperature) represents the standard deviation of temperature, used as a proxy for daily temperature variability. Here, s(…) represents univariate smooths, and ⊗ denotes a tensor product interaction of two variables. We use thin plate regression splines as the smoothing basis for each smooth term in Equation 2 besides the final term, for which we use random effects as the basis. These random effects account for states with higher or lower transmission due to random conditions beyond the fixed effects captured by the covariates in the model. The basis dimension is the maximum possible degrees of freedom allowed for each smooth term in the model. For our model of daily, state‐level variations of R t , all smooth terms have a basis dimension of three as a larger basis could overfit the data, resulting in highly non‐linear smooths that cannot be justified by a priori knowledge on the relationship between transmission and model terms in Equation 2. We tested whether this basis dimension was adequate using the residual variance based on differencing residuals. These tests, detailed in Table S1 in Supporting Information S1, generally show that our choice for a basis dimension was not too low.

We quantified the importance of model terms in Equation 2 by calculating the accumulated local effects (ALE) of each term on R t . The ALE are calculated as the change in modeled R t over a small range of a given model term using all data samples within that range and centered around 0 such that the value of the ALE curve can be interpreted as the difference to the mean prediction. For example, if ALE = 0.05 at a temperature of 20°C, it means that, at this temperature, R t is 0.05 higher than the average predicted value of R t . Other ways to quantify feature importance from GAMs (e.g., partial dependence, partial effects) can be biased by correlation among input variables and may result in unrealistic combinations of input variables. ALE, on the other hand, are unbiased in their estimated feature effect.

Our analysis was conducted using R (version 4.0.3) with packages mgcv (version 1.8‐38) (Wood, 2011), additive (version 0.0.3) (Badr, 2021), mgcViz (version 0.1.9) (Fasiolo et al., 2019), and vip (version 0.3.2) (Greenwell & Boehmke, 2020).

3. Results

By 31 August 2021, 20,785,196 COVID‐19 cases and 580,763 deaths had been reported in Brazil. São Paulo had the highest number of cases (4,262,684) and deaths (145,836). The number of cases per capita exhibited substantial spatial variability, but we note that states in Brazil's sparsely populated North and Central‐West regions had a higher number of cases per capita than the more densely populated coastal states (Figure 1a). The Northern state of Roraima, whose health care system has been strained by a recent influx of migrants and refugees from neighboring Venezuela (Doocy et al., 2019), had the highest case rate of all states: 20,468 per 100,000.

Figure 1.

Figure 1

Open in a new tab

(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.

National‐level R t and R t for individual states share some common features such as decreasing R t at the beginning of our study period following the first wave and an increase in austral summer (November 2020–February 2021; Figures 1b–1g, Figure S2 in Supporting Information S1). While the overall temporal variations in R t are, at times, qualitatively similar between the national‐ and state‐level time series, a closer inspection of R t across these spatial scales highlights numerous differences that support our analysis of drivers of SARS‐CoV‐2 transmission at the state level.

The predictive power of our GAMs was evaluated with several performance metrics (Figure S3 and Table S2 in Supporting Information S1). The temporal correlation between EpiNow2 and modeled R t for different 3‐month seasons was strong, generally ∼0.7, indicating that our modeled R t provides a satisfactory temporal fit to R t from EpiNow2. We do note that, despite this strong correlation, the GAMs slightly underpredict R t for all time periods. Additionally, the model performance is worse for DJF 2020–2021 and subsequent 3‐month periods compared with JJA and SON 2020, which may stem from increased underlying immunity with time or events that alter behavioral patterns and therefore COVID‐19 transmission (e.g., Carnaval, Natal/Christmas).

Temperature emerges as a significant (p < 0.05) predictor for all periods but JJA 2021 (Table S2 in Supporting Information S1). Within Brazilian states, increasing temperatures on daily timescales are associated with a decrease of R t relative to the mean for the first four 3‐month periods of our study; however, during JJA 2021 and for the full study period, JJA 2020–JJA 2021, we found essentially no change in R t with temperature (Figure 2a). The largest temperature effects on R t of ∼0.05 occurred in JJA 2020. The magnitude of these effects is roughly half the magnitude of the daily variability of R t (σ(Rt) = 0.10). We note, though, that the ∼9°C range of temperatures over which we observe the ∼0.05 change in the ALE is not commonly encountered in many Brazilian states besides a handful in Brazil's south (e.g., Mato Grosso do Sul, Paraná, Rio Grande do Sul; Figure S4a in Supporting Information S1).

Figure 2.

Figure 2

Open in a new tab

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).

Increased specific humidity was generally associated with an increase in R t except during MAM 2021, which exhibits an inverted U‐shaped relationship (Figure 2b). The effect of specific humidity on R t is significant for all study periods except MAM 2021 (Table S2 in Supporting Information S1) and roughly double in magnitude compared with temperature's effect; for example, the largest increase in R t relative to the mean associated with variations in specific humidity is ∼0.1, observed during JJA 2020–JJA 2021 (Figure 2b). As with temperature, most states have a tight range of specific humidity and do not experience daily variations in specific humidity equivalent to the range over which we observe this ∼0.1 effect (Figure S4b in Supporting Information S1).

The ALE of the OxCGRT policy and Google‐derived mobility variables are generally equivalent or slightly larger in magnitude than the effects of meteorological variables. However, in contrast to the generally consistent conclusions we draw regarding the sign of temperature and specific humidity effects on R t , the OxCGRT policy and Google‐derived mobility variables have inconsistent effects on R t across 3‐month study periods and the full study period. Specifically, the ALE of OxCGRT policy reverses direction between nearly every period (Figure 2c), and the direction of the residential and workplace mobility in SON 2020 and MAM 2021 differ from other periods in our study (Figures 2d and 2e).

The lagged cases term, which gauges the trajectory of the pandemic, has the largest effect on R t (Figure 2f) and is significant in every time period of our study (Table S2 in Supporting Information S1). Its ALE are several times larger than that of the meteorological or policy‐ and mobility‐related terms (note the scale in Figure 2f). The direction of this relationship, indicating that more cases in the previous 30 days are associated with lower transmission, likely reflects a reduction in the size of the susceptible pool following periods with a high number of cases. While our model accounts for the cumulative number of cases in the previous 30 days, we have also tested how examining the number of cases for longer periods (60 days) impacts results and found no substantive difference in key conclusions (not shown).

In our model we included a term to test R t differences in states with tropical (small daily variations in temperature) versus temperate (large variations) climates (σ(temperature); Equation 2). Since this term consists of one value per state per time period compared with other continuous variables shown in Figure 2 we present the ALE differently and show their distribution in Figure S5 in Supporting Information S1. These results reveal that the ALE of temperature variability on R t is larger in states with the largest daily variations in temperature, although the average impact of temperature variability on R t is not consistently positive or negative (Figure S5 in Supporting Information S1). This effect of temperature variability on transmission could explain some of the differences in COVID‐19 cases between Brazil's tropical states with fewer COVID‐19 cases (e.g., Pará, Maranhão, Figure 1a) and temperate states with more COVID‐19 cases (e.g., Rio Grande Do Sul, Santa Catarina).

The significance of model terms and their effect on R t (Table S2 in Supporting Information S1, Figure 2) are somewhat different concepts than importance. We next explore how each term's relative influence on model prediction by adopting the root‐mean‐square error (RMSE) as an indicator of importance, with a higher RSME representing greater importance of a model term. Figure 3 shows that the random effects term, which accounts for unexplained state‐level heterogeneity, and the total number of confirmed COVID‐19 cases in the preceding 30 days are clearly the most important terms in our model.

Figure 3.

Figure 3

Open in a new tab

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.

Figure 3 also highlights the evolving role of model terms on R t . While specific humidity and OxCGRT policy are generally the most important terms in our model after the random effects and lagged cases terms, the precise order of importance changes for different time periods. During MAM 2021, temperature is the third most important term in our model (Figure 3), which is consistent with the large ALE of temperature during this time period (Figure 2a); however, for other periods (e.g., JJA 2021), temperature is the least important.

Given the importance of state‐level random effects in Figure 3, we also show how these random effects impact R t . The ALE of the random effects have spatial structure, and these effects are consistently positive in Brazil's South and Southeast Regions and negative in the North (Figure S6 in Supporting Information S1). The spatial structure of this map roughly resembles metrics related to healthcare and underlying health risks as well as forms of social vulnerability (Rocha et al., 2021), and the random effects could be accounting for these factors.

4. Discussion

As one of the countries hardest hit by the COVID‐19 pandemic, there is an acute need to characterize the drivers of SARS‐CoV‐2 transmission in Brazil to inform policy and other mitigative measures for future surges in cases. Earlier attempts to answer this question in the literature were often limited by short temporal record, methodological frameworks that were prone to the biases of input data and could not account for nonlinear relationships, and conclusions that raised questions the generalizability and robustness of their policy‐relevant conclusions to different time periods. Our study leverages 15 months of data within a flexible, nonparametric regression model, allowing us to understand drivers of transmission over an entire seasonal cycle, and investigates how the relationships between transmission and meteorology, policy, and human mobility change from season to season.

We generally found increasing SARS‐CoV‐2 transmission with decreasing temperature and increasing humidity (Figure 2). While these results may conflict with individual studies that have found relationships in different directions, our results can be reasonably expected given the mixed landscape of the literature. A review of approximately 160 studies that investigated how COVID‐19 is affected by meteorology found a fairly even split between studies that found a positive versus negative relationship between COVID‐19 spread and temperature (Tan & Schultz, 2022). A slight majority (∼60%) of correlation coefficients and regression analyses synthesized in this review found a negative relationship between COVID‐19 and humidity, yet ∼40% found a positive relationship, which is consistent with our results. Moreover, the World Meteorological Organization COVID‐19 Task Team responding to questions on meteorological and air quality factors affecting COVID‐19 found inconsistent results regarding the role of humidity on SARS‐CoV‐2 survival in a laboratory setting: some studies found a slight increase in virus survival at lower humidity, while others found greater survival at both medium and high humidity (World Meteorological Organization, 2021).

The changing sign and magnitude of drivers of SARS‐CoV‐2 transmission (e.g., Figures 2 and 3) and documented in Yin et al. (2022) could explain the inconsistencies between our work and other studies on COVID‐19 in Brazil and, more broadly, the variability in the published literature. Two studies focused on COVID‐19 in Brazil in early‐ to mid‐2020 (Pequeno et al., 2020; Rosario et al., 2020) found increased temperatures were associated with decreased severity, similar to our findings for JJA 2020 (Figure 2a). However, we have shown that this negative relationship between temperature and transmission does not persist later in 2020 or in 2021. This finding demonstrates the importance of analyzing each season separately. For example, considering the full study period might lead to the conclusion that temperature has a statistically significant but essentially null impact on transmission (Table S2 in Supporting Information S1, Figure 2a), while temperature bears a larger association with transmission in most of the 3‐month periods.

The relatively small impact of policy and mobility was surprising, given that at the early stage of the pandemic when transmissibility was high and immunity was low, disease control interventions are believed to have a stronger impact on transmission than any environmental driver (Carlson et al., 2020). There are at least two potential reasons for this finding. One explanation could lie in the evolving behavioral responses to the pandemic. Using the correlation between R t and the OxCGRT policy and Google workplaces and residential variables as a proxy for behavioral responses during periods of increased versus decreased transmission, we find considerable spatiotemporal variability between these variables and R t (Figure S7 in Supporting Information S1). Another explanation could be related to what these terms precisely measure and the data from which they are formed. The OxCGRT policy index measures how the government has implemented health and containment measures but does not show whether policy has been implemented effectively or measure compliance to policies. It is likely that the variability in restrictive measures over time by states and municipalities and the continued urban public transit, even during high periods of transition, favored population mobility and consequently the circulation of the virus (Castro et al., 2021; Kortessis et al., 2020). Brazil's federal government not only underestimated the impact of COVID‐19 but also did not coordinate efforts, at times even trying to influence state and municipal governments against measures of social distancing (Castro et al., 2021). Governmental responses are also reactive and might not have a substantial effect if enacted in response to a surge in cases. Additionally, the Google mobility terms derive from cell phone usage and internet access, which vary across Brazil (Instituto Brasileiro de Geografia e Estatística, 2018) and are unequally distributed among socioeconomic groups. We tested whether modeling the drivers of SARS‐CoV‐2 transmission only in states where the percentage of households with internet access exceeded the median value of all state‐level proportions of household internet access. We hypothesized that the inconsistent impacts of workplace and residential mobility might be more consistent with time in states with greater internet access and likely more‐accurate input data for the Google mobility variables. However, we found no major changes regarding the inconsistent directions of factors driving transmission, although the magnitude of the ALE slightly changed for some model terms and time periods (Figure S8 in Supporting Information S1).

We chose a relatively small number of model terms (Equation 2) compared with other studies which have included air pollutants (Wu et al., 2020); sociodemographic data and additional mobility indicators (Colston et al., 2023; Nottmeyer & Sera, 2021); and additional meteorological variables (Colston et al., 2023; Ma et al., 2021; Zhang et al., 2022). Our selected meteorological terms have an established precedent for shaping respiratory virus seasonality (Lowen & Steel, 2014) and were most‐commonly investigated in early studies on the meteorological drivers of SARS‐CoV‐2 transmission (Kerr et al., 2021). The policy and mobility terms represent plausible proxies for governmental responses and individual‐level behavior that likely affect transmission. Future modeling efforts might consider including additional terms to account for temporal trends, especially when modeling drivers of transmission over longer time periods than our 3‐month segments to account for changing dominant variants, vaccination rates, and the like.

We tested whether our main conclusions regarding the inconsistent and evolving roles of meteorology, mobility, and policy were robust to different formulations of our GAMs. Specifically in the main model described in Equations 1 and 2, we (a) modified certain terms, (b) removed terms, and (c) included additional terms related to demographics. Upon refitting the GAMs for our different periods with these different adjustments, we find no substantive changes to our conclusions. Regarding (a), to account for different temperature and specific humidity ranges across different climatic zones in Brazil, we standardized temperature and humidity within each state prior to fitting the GAMs. We find that the magnitude and direction of the ALE are robust to both non‐standardized or standardized meteorological variables, although the positive relationship between transmission and temperature in JJA is more pronounced when using standardized temperature (Figure S9 in Supporting Information S1). We also normalized the lagged cases variable by the state population but again found little impact on our overall conclusions (Figure S10 in Supporting Information S1). Regarding (b), the Google workplaces and residential represent different trip purposes but are correlated with each other. As such, these variables could introduce issues related to concavity within the GAMs. To test the impact that the inclusion of these two correlated variables have our on our results, we refit the our 3‐month and full study period GAMs without each of these Google mobility‐related variables and found the role of the other variables (e.g., temperature, policy, etc.) largely unchanged (Figure S11 in Supporting Information S1). Regarding (c), we included two additional, state‐level terms related to demographics: the gross regional product (GRP) per capita and the population density. As with the other variables, these terms were incorporated into our modeling framework as smooths with a basis dimension of three. Including these additional terms only marginally improved the performance of the model (e.g., the deviance explained in our main model for JJA 2020–JJA 2021 increased from 26.8% to 26.9% with these additional terms) and did not change the roles of the meteorological‐, policy‐, and mobility‐related terms (Figures S12a–S12f in Supporting Information S1). GRP per capita was a significant term during all periods. The ALE of GRP per capita indicated that increasing GRP per capita was associated with increased SARS‐CoV‐2 (Figure S12G in Supporting Information S1). Population density, on the other hand, was only significant in three of our six periods (SON 2020, DJF 2020–2021, and MAM 2021), but its ALE generally suggested increased transmission in more densely populated states (Figure S12h in Supporting Information S1). We acknowledge that including additional terms may further improve model performance or change the role purported impact of our chosen variables on R t . However, additional terms could also lead to overfitting or decreased interpretability if no clear mechanism to tie a particular term to transmission exists.

In addition to the ecological fallacy that challenges studies investigating drivers of COVID‐19 transmission, our study has several limitations. Our analysis was conducted at the state level rather than at the municipal level, the lowest level of political division, to provide the highest granularity possible without encountering missing data (e.g., the OxCGRT policy data does not have time series for all of Brazil's municipalities). This limitation is particularly relevant due to state and municipal alignments with the federal government which affected the intensity, duration, and timing of local responses against the disease (Castro et al., 2021). Brazil's mass vaccine campaign, which likely impacted underlying immunity and behavior, began in January 2021 and was not explicitly accounted for due to lack of information on vaccine rollout at the state level. Terms included in our model are not exhaustive, and other studies have highlighted additional drivers that bear a significant association with transmission. We also did not account for multiple SARS‐CoV‐2 variants and their different transmissibility; however, conducting our analysis in several 3‐month periods could partially mitigate this limitation. Since the end of our study period in August 2021, it appears that COVID‐19 has settled into indefinite endemic circulation in Brazil. With fewer non‐pharmaceutical interventions and less drastic changes in vaccination progress, a more pronounced, consistent role of meteorological‐ and mobility‐related factors could possibly be uncovered. We welcome this line of inquiry in future research and encourage the use of high‐resolution data sets and robust statistical frameworks, such as those discussed in our present study.

5. Conclusions

In summary, we found that meteorological variables play a statistically significant, but relatively small, role in explaining spatiotemporal variations in SARS‐CoV‐2 transmission in Brazil. Higher temperatures were generally associated with decreased R t , higher specific humidity with increased R t , and increased total visitors in workplaces with decreased R t (Figure 2), although these terms were not always significant in all time periods we examined (Table S2 in Supporting Information S1). On the other hand, the impact of governmental policies and time spent in places of residences were associated with both increases and decreases in R t , depending on the time period (Figure 2). Most variations in R t , though, were attributed to unexplained between‐state heterogeneity and the trajectory of the pandemic (i.e., the number of recent cases, Figures 2 and 3). While our GAMs obtained satisfactory predictions of R t using various performance metrics (Figure S3 and Table S2 in Supporting Information S1), the inconsistent roles of meteorology, policy, and human mobility with time and the aforementioned unexplained heterogeneity suggest that efforts to model SARS‐CoV‐2 transmission at an ecological level in Brazil could be hamstrung by the complex and evolving nature of transmission.

Brazil is a global leader at administering COVID‐19 vaccines. Its vaccination rate—an estimated 80% as of July 2022—exceeds that of the United States, Germany, the United Kingdom, and several other developed nations that had the earliest access to vaccines (Johns Hopkins Centers for Civic Impact, 2022). The rate of vaccinations in Brazil has been credited with preventing approximately 1,000,000 deaths from COVID‐19 (Watson et al., 2022). While meteorology might weakly modulate transmission, we found no indication that daily or seasonal weather conditions alone will curb the virus in Brazil. At this point in time, disease control interventions and vaccines appear to be the greatest weapons to fight the pandemic in Brazil and throughout the world.

Conflict of Interest

The authors declare no conflicts of interest relevant to this study.

Supporting information

Supporting Information S1

Click here for additional data file. (33.2MB, docx)

Acknowledgments

We are grateful to two anonymous reviewers for their thoughtful and thorough comments. This work was supported by the NASA Health and Air Quality Program (Award 80NSSC18K0327).

Kerr, G. H. , Badr, H. S. , Barbieri, A. F. , Colston, J. M. , Gardner, L. M. , Kosek, M. N. , & Zaitchik, B. F. (2023). Evolving drivers of Brazilian SARS‐CoV‐2 transmission: A spatiotemporally disaggregated time series analysis of meteorology, policy, and human mobility. GeoHealth, 7, e2022GH000727. 10.1029/2022GH000727

Data Availability Statement

All data and source code used in this manuscript are publicly available at https://doi.org/10.5281/zenodo.7708308 (Kerr et al., 2023). Google's COVID‐19 Community Mobility Reports are available at www.google.com/covid19/mobility/.

References

  1. Abbott, S. , Hellewell, J. , Thompson, R. N. , Sherratt, K. , Gibbs, H. P. , Bosse, N. I. , et al. (2020). Estimating the time‐varying reproduction number of SARS‐CoV‐2 using national and subnational case counts. Wellcome Open Research, 5, 112. 10.12688/wellcomeopenres.16006.2 [DOI] [Google Scholar]
  2. Alene, M. , Yismaw, L. , Assemie, M. A. , Ketema, D. B. , Gietaneh, W. , & Birhan, T. Y. (2021). Serial interval and incubation period of COVID‐19: A systematic review and meta‐analysis. BMC Infectious Diseases, 21(1), 257. 10.1186/s12879-021-05950-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Badr, H. S. (2021). additive: Bindings for additive TidyModels. Comprehensive R Archive Network (CRAN). Retrieved from https://hsbadr.github.io/additive/ [Google Scholar]
  4. Badr, H. S. , Zaitchik, B. F. , Kerr, G. H. , Nguyen, N.‐L. H. , Chen, Y.‐T. , Hinson, P. , et al. (2021). Unified real‐time environmental‐epidemiological data for multiscale modeling of the COVID‐19 pandemic [Preprint]. Epidemiology. 10.1101/2021.05.05.21256712 [DOI] [PMC free article] [PubMed]
  5. Carlson, C. J. , Gomez, A. C. R. , Bansal, S. , & Ryan, S. J. (2020). Misconceptions about weather and seasonality must not misguide COVID‐19 response. Nature Communications, 11(1), 4312. 10.1038/s41467-020-18150-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Castro, M. C. , Kim, S. , Barberia, L. , Ribeiro, A. F. , Gurzenda, S. , Ribeiro, K. B. , et al. (2021). Spatiotemporal pattern of COVID‐19 spread in Brazil. Science, 372(6544), 821–826. 10.1126/science.abh1558 [DOI] [PubMed] [Google Scholar]
  7. Center for Systems Science and Engineering . (2022). Johns Hopkins Center for Systems Science and Engineering (CSSE) Coronavirus Resource Center. Retrieved from https://coronavirus.jhu.edu/map.html
  8. Colston, J. M. , Hinson, P. , Nguyen, N.‐L. H. , Chen, Y. T. , Badr, H. S. , Kerr, G. H. , et al. (2023). Effects of hydrometeorological and other factors on SARS‐CoV‐2 reproduction number in three contiguous countries of Tropical Andean South America: A spatiotemporally disaggregated time series analysis. IJID Regions, 6, 29–41. 10.1016/j.ijregi.2022.11.007 [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Dominici, F. (2002). On the use of generalized additive models in time‐series studies of air pollution and health. American Journal of Epidemiology, 156(3), 193–203. 10.1093/aje/kwf062 [DOI] [PubMed] [Google Scholar]
  10. Doocy, S. , Page, K. R. , de la Hoz, F. , Spiegel, P. , & Beyrer, C. (2019). Venezuelan migration and the border health crisis in Colombia and Brazil. Journal on Migration and Human Security, 7(3), 79–91. 10.1177/2331502419860138 [DOI] [Google Scholar]
  11. Fasiolo, M. , Nedellec, R. , Goude, Y. , & Wood, S. N. (2019). Scalable visualisation methods for modern generalized additive models. (arXiv:1809.10632). arXiv. Retrieved from http://arxiv.org/abs/1809.10632
  12. Ganyani, T. , Kremer, C. , Chen, D. , Torneri, A. , Faes, C. , Wallinga, J. , & Hens, N. (2020). Estimating the generation interval for coronavirus disease (COVID‐19) based on symptom onset data, March 2020. Euro Surveillance, 25(17), 2000257. 10.2807/1560-7917.ES.2020.25.17.2000257 [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Google, LLC . (2022). Google COVID‐19 community mobility reports. Retrieved from https://www.google.com/covid19/mobility/
  14. Gostic, K. M. , McGough, L. , Baskerville, E. B. , Abbott, S. , Joshi, K. , Tedijanto, C. , et al. (2020). Practical considerations for measuring the effective reproductive number, R t . PLoS Computational Biology, 16(12), e1008409. 10.1371/journal.pcbi.1008409 [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Greenwell, B. M. , & Boehmke, B. C. (2020). Variable importance plots—An introduction to the vip package. The R Journal, 12(1), 343–366. 10.32614/RJ-2020-013 [DOI] [Google Scholar]
  16. Hale, T. , Angrist, N. , Goldszmidt, R. , Kira, B. , Petherick, A. , Phillips, T. , et al. (2021). A global panel database of pandemic policies (Oxford COVID‐19 Government Response Tracker). Nature Human Behaviour, 5(4), 529–538. 10.1038/s41562-021-01079-8 [DOI] [PubMed] [Google Scholar]
  17. Hastie, T. J. , & Tibshirani, R. J. (1990). Generalized additive models (1st ed.). [DOI] [PubMed]
  18. Hersbach, H. , Bell, B. , Berrisford, P. , Hirahara, S. , Horányi, A. , Muñoz‐Sabater, J. , et al. (2020). The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730), 1999–2049. 10.1002/qj.3803 [DOI] [Google Scholar]
  19. Instituto Brasileiro de Geografia e Estatística . (2018). 2018 Access to Internet and TV and ownership of cell phone for personal use. Continuous National Household Sample Survey, PNAD‐Contínua. Retrieved from https://www.ibge.gov.br/en/statistics/multi-domain/science-technology-and-innovation/18083-annual-dissemination-pnadc3.html?edicao=27537&t=resultados
  20. Instituto Brasileiro de Geografia e Estatística . (2021). Municipal mesh. Retrieved from https://www.ibge.gov.br/en/geosciences/territorial-organization/territorial-organization/18890-municipal-mesh.html?=&t=sobre
  21. Instituto Brasileiro de Geografia e Estatística . (2022). Population census, preliminary population of municipalities based on 2022 population census data collected up to December 25, 2022. Retrieved from https://www.ibge.gov.br/en/statistics/social/population/22836-2020-census-censo4.html?=&t=resultados
  22. Johns Hopkins Centers for Civic Impact . (2022). International vaccine data. Retrieved from https://coronavirus.jhu.edu/region/brazil
  23. Kerr, G. H. , Badr, H. S. , Barbieri, A. F. , Colston, J. M. , Gardner, L. M. , Kosek, M. N. , & Zaitchik, B. F. (2023). Dataset for “evolving drivers of Brazilian SARS‐CoV‐2 transmission: A spatiotemporally disaggregated time series analysis of meteorology, policy, and human mobility” (ver 2023.03.08). Zenodo. 10.5281/zenodo.7708308 [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Kerr, G. H. , Badr, H. S. , Gardner, L. M. , Perez‐Saez, J. , & Zaitchik, B. F. (2021). Associations between meteorology and COVID‐19 in early studies: Inconsistencies, uncertainties, and recommendations. One Health, 12, 100225. 10.1016/j.onehlt.2021.100225 [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Kortessis, N. , Simon, M. W. , Barfield, M. , Glass, G. E. , Singer, B. H. , & Holt, R. D. (2020). The interplay of movement and spatiotemporal variation in transmission degrades pandemic control. Proceedings of the National Academy of Sciences of the United States of America, 117(48), 30104–30106. 10.1073/pnas.2018286117 [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Lauer, S. A. , Grantz, K. H. , Bi, Q. , Jones, F. K. , Zheng, Q. , Meredith, H. R. , et al. (2020). The incubation period of coronavirus disease 2019 (COVID‐19) from publicly reported confirmed cases: Estimation and application. Annals of Internal Medicine, 172(9), 577–582. 10.7326/M20-0504 [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Lowen, A. C. , & Steel, J. (2014). Roles of humidity and temperature in shaping influenza seasonality. Journal of Virology, 88(14), 7692–7695. 10.1128/JVI.03544-13 [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Ma, Y. , Pei, S. , Shaman, J. , Dubrow, R. , & Chen, K. (2021). Role of meteorological factors in the transmission of SARS‐CoV‐2 in the United States. Nature Communications, 12(1), 3602. 10.1038/s41467-021-23866-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. McAloon, C. , Collins, Á. , Hunt, K. , Barber, A. , Byrne, A. W. , Butler, F. , et al. (2020). Incubation period of COVID‐19: A rapid systematic review and meta‐analysis of observational research. BMJ Open, 10(8), e039652. 10.1136/bmjopen-2020-039652 [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Mecenas, P. , Bastos, R. T. , da, R. M. , Vallinoto, A. C. R. , & Normando, D. (2020). Effects of temperature and humidity on the spread of COVID‐19: A systematic review. PLoS One, 15(9), e0238339. 10.1371/journal.pone.0238339 [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Nottmeyer, L. N. , & Sera, F. (2021). Influence of temperature, and of relative and absolute humidity on COVID‐19 incidence in England—A multi‐city time‐series study. Environmental Research, 196, 110977. 10.1016/j.envres.2021.110977 [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Pequeno, P. , Mendel, B. , Rosa, C. , Bosholn, M. , Souza, J. L. , Baccaro, F. , et al. (2020). Air transportation, population density and temperature predict the spread of COVID‐19 in Brazil. PeerJ, 8, e9322. 10.7717/peerj.9322 [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Ponce, D. (2020). The impact of coronavirus in Brazil: Politics and the pandemic. Nature Reviews Nephrology, 16(9), 483. 10.1038/s41581-020-0327-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Prata, D. , Rodrigues, W. , de Queiroz Trevisan, D. M. , Camargo, W. , Frizzera, H. , Carvalho, R. , et al. (2022). Climatic factors associated with economic determinants significantly affect the spread of COVID‐19 in tropical Brazil. One Health, 14, 100375. 10.1016/j.onehlt.2022.100375 [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Prata, D. N. , Rodrigues, W. , & Bermejo, P. H. (2020). Temperature significantly changes COVID‐19 transmission in (sub)tropical cities of Brazil. Science of the Total Environment, 729, 138862. 10.1016/j.scitotenv.2020.138862 [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Rai, B. , Shukla, A. , & Dwivedi, L. K. (2021). Estimates of serial interval for COVID‐19: A systematic review and meta‐analysis. Clinical Epidemiology and Global Health, 9, 157–161. 10.1016/j.cegh.2020.08.007 [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Rocha, R. , Atun, R. , Massuda, A. , Rache, B. , Spinola, P. , Nunes, L. , et al. (2021). Effect of socioeconomic inequalities and vulnerabilities on health‐system preparedness and response to COVID‐19 in Brazil: A comprehensive analysis. Lancet Global Health, 9(6), e782–e792. 10.1016/S2214-109X(21)00081-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Rosario, D. K. A. , Mutz, Y. S. , Bernardes, P. C. , & Conte‐Junior, C. A. (2020). Relationship between COVID‐19 and weather: Case study in a tropical country. International Journal of Hygiene and Environmental Health, 229, 113587. 10.1016/j.ijheh.2020.113587 [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Sera, F. , Armstrong, B. , Abbott, S. , Meakin, S. , O’Reilly, K. , von Borries, R. , et al. (2021). A cross‐sectional analysis of meteorological factors and SARS‐CoV‐2 transmission in 409 cities across 26 countries. Nature Communications, 12(1), 5968. 10.1038/s41467-021-25914-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Tamerius, J. D. , Shaman, J. , Alonso, W. J. , Bloom‐Feshbach, K. , Uejio, C. K. , Comrie, A. , & Viboud, C. (2013). Environmental predictors of seasonal influenza epidemics across temperate and tropical climates. PLoS Pathogens, 9(3), e1003194. 10.1371/journal.ppat.1003194 [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Tan, L. , & Schultz, D. M. (2022). How is COVID‐19 affected by weather? Metaregression of 158 studies and recommendations for best practices in future research. Weather, Climate, and Society, 14(1), 237–255. 10.1175/WCAS-D-21-0132.1 [DOI] [Google Scholar]
  42. Tatem, A. J. (2017). WorldPop, open data for spatial demography. Scientific Data, 4(1), 170004. 10.1038/sdata.2017.4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Watson, O. J. , Barnsley, G. , Toor, J. , Hogan, A. B. , Winskill, P. , & Ghani, A. C. (2022). Global impact of the first year of COVID‐19 vaccination: A mathematical modelling study. The Lancet Infectious Diseases, 22(9), 1293–1302. 10.1016/S1473-3099(22)00320-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Wood, S. N. (2011). Fast stable restricted maximum likelihood and marginal likelihood estimation of semiparametric generalized linear models: Estimation of semiparametric generalized linear models. Journal of the Royal Statistical Society: Series B, 73(1), 3–36. 10.1111/j.1467-9868.2010.00749.x [DOI] [Google Scholar]
  45. World Meteorological Organization . (2021). First report of the WMO COVID‐19 task team affecting the COVID‐19 pandemic (p. 42). (WMO‐No. 1262). Retrieved from https://library.wmo.int/doc_num.php?explnum_id=10555
  46. Wu, X. , Nethery, R. C. , Sabath, M. B. , Braun, D. , & Dominici, F. (2020). Air pollution and COVID‐19 mortality in the United States: Strengths and limitations of an ecological regression analysis. Science Advances, 6(45), eabd4049. 10.1126/sciadv.abd4049 [DOI] [PMC free article] [PubMed] [Google Scholar]
  47. Yang, W. , Lipsitch, M. , & Shaman, J. (2015). Inference of seasonal and pandemic influenza transmission dynamics. Proceedings of the National Academy of Sciences of the United States of America, 112(9), 2723–2728. 10.1073/pnas.1415012112 [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Yin, C. , Zhao, W. , & Pereira, P. (2022). Meteorological factors' effects on COVID‐19 show seasonality and spatiality in Brazil. Environmental Research, 208, 112690. 10.1016/j.envres.2022.112690 [DOI] [PMC free article] [PubMed] [Google Scholar]
  49. Zhang, S. , Wang, B. , Yin, L. , Wang, S. , Hu, W. , Song, X. , & Feng, H. (2022). Novel evidence showing the possible effect of environmental variables on COVID‐19 spread. GeoHealth, 6(3), e2021GH000502. 10.1029/2021GH000502 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Citations

  1. Badr, H. S. , Zaitchik, B. F. , Kerr, G. H. , Nguyen, N.‐L. H. , Chen, Y.‐T. , Hinson, P. , et al. (2021). Unified real‐time environmental‐epidemiological data for multiscale modeling of the COVID‐19 pandemic [Preprint]. Epidemiology. 10.1101/2021.05.05.21256712 [DOI] [PMC free article] [PubMed]

Supplementary Materials

Supporting Information S1

Click here for additional data file. (33.2MB, docx)

Data Availability Statement

All data and source code used in this manuscript are publicly available at https://doi.org/10.5281/zenodo.7708308 (Kerr et al., 2023). Google's COVID‐19 Community Mobility Reports are available at www.google.com/covid19/mobility/.

Articles from GeoHealth are provided here courtesy of Wiley

RESOURCES