Variation in SARS-CoV-2 outbreaks across sub-Saharan Africa

Reported SARS-CoV-2 case counts, mortality and testing in SSA as of December 2020

Variables and data sources for reporting data

The numbers of reported cases, deaths and tests for the 48 SSA countries studied (Supplementary Table 1) were sourced from the Africa CDC dashboard on 20 December 2020 (and previously on 23 September and 30 June 2020). The Africa CDC obtains data from the official Africa CDC Regional Collaborating Centre and member state reports. Differences in the timing of reporting by member states results in some variation in the recency of data within the centralized Africa CDC repository, but data should broadly reflect the relative scale of testing and reporting efforts across countries. For Mauritius (https://covid19.mu/) and Rwanda (https://covid19.who.int/region/afro/country/rw), reporting to the Africa CDC was confirmed by comparison to country-specific dashboards.

The countries or member states within SSA in this study follow the United Nations and Africa CDC-listed regions of Southern, Western, Central and Eastern Africa (excluding Sudan). From the Northern Africa region, Mauritania is included in SSA.

For comparison to non-SSA countries, the number of reported cases in other geographical regions were obtained from the Johns Hopkins University Coronavirus Resource Center on 23 September 2020 (https://coronavirus.jhu.edu/map.html).

Case fatality ratios (CFRs) were calculated by dividing the number of reported deaths by the number of reported cases and expressed as a percentage. Positivity was calculated by dividing the number of reported cases by the number of reported tests. Testing and case rates were calculated per 100,000 population using population size estimates for 2020 from the United Nations Population Division (https://population.un.org/wpp/Download/Standard/Population/). Since reported confirmed cases are likely to be an underestimate of the true number of infections, CFRs may be a poor proxy for the IFR, defined as the proportion of infections that result in mortality⁴.

Variation in testing and mortality rates

Testing rates among SSA countries varied by multiple orders of magnitude as of 30 June and remain highly variable as of 23 September and 20 December 2020. The number of tests completed per 100,000 population ranged from 19.84 in Burundi to 13,508.13 in Mauritius in June 2020; from 65.98 in the Democratic Republic of the Congo to 18,321.83 in Mauritius in September 2020; and from 100.9 in the Democratic Republic of the Congo to 23695.0 in Mauritius in December 2020 (Extended Data Fig. 1a). Tanzania (6.50 tests per 100,000 population) has not reported new tests, cases or deaths to the Africa CDC since April 2020. The number of reported infections (that is, positive tests) was strongly correlated with the number of tests completed in June 2020 (Pearson’s correlation coefficient, r = 0.9667, P < 0.001), September 2020 (r = 0.9689, P < 0.001) and December 2020 (r = 0.9750, P < 0.001) (Extended Data Fig. 1b). As of June 2020, no deaths due to SARS-CoV-2 were reported to the Africa CDC for five SSA countries (Eritrea, Lesotho, Namibia, Seychelles, Uganda). As of December 2020, still no deaths due to SARS-CoV-2 were reported to the Africa CDC for two of those countries (Eritrea and Seychelles). Among countries with at least 1 reported death, the CFR varied from 0.22% in Rwanda to 8.54% in Chad in June 2020; from 0.21% in Burundi to 6.96% in Chad in September 2020; and from 0.26% in Burundi to 5.40% in Chad in December 2020 (Extended Data Fig. 1c). Limitations in the ascertainment of infection rates and the rarity of reported deaths (for example, the median number of reported deaths per SSA country was 25.5 as of June 2020, 71.0 as of September 2020 and 101.0 as of December 2020), indicate that the data are insufficient to determine country-specific IFRs and IFR by age profiles for most countries. As a result, global IFR by age estimates was used for the subsequent analyses in this study.

Synthesizing factors that increase or decrease SARS-CoV-2 epidemic risk in SSA

Variable selection and data sources for variables associated with an increased probability of severe clinical outcomes for an infection

To characterize epidemic risk, defined as potential SARS-CoV-2 related morbidity and mortality, we first synthesized factors hypothesized to influence risk in SSA settings (Supplementary Table 2). Early during the pandemic, evidence suggested that age was an important risk factor associated with morbidity and mortality associated with SARS-CoV-2 infection⁴⁰, a pattern subsequently confirmed across settings^2,11,41. Associations between SARS-CoV-2 mortality and comorbidities including hypertension, diabetes and cardiovascular disease emerged early⁴⁰ and have been observed across settings, with further growing evidence for associations with obesity^11,42, severe asthma¹¹ and the respiratory effects of pollution⁴³. Specific to Africa, vulnerability scores based on these hypothesized associations or combinations of risks factors have been developed (for example, refs. ^44,45).

Many possible sources of bias complicate interpretation of these associations⁴⁶; while they provide a useful baseline, inference is also likely to change as the pandemic advances. To reflect this, our analysis combined a number of high-level variables likely to broadly encompass these putative risk factors (for example, noncommunicable disease-related mortality and healthy life expectancy) with more specific measures encompassed in evidence to date (for example, prevalence of diabetes, obesity and respiratory illness, such as chronic obstructive pulmonary disease). We also included measures relating to infectious diseases, undernourishment and anemia given their interaction and effects in determining health status in these settings⁴⁷. Although interactions with such infectious diseases have been suggested, evidence is limited to date, except for HIV, where effects have been suggested to be minor⁴⁸. We also note that the key concern raised around such infections to date is associated with disruption to routine screening (for example, for malaria⁴⁹), treatment⁵⁰ or prevention programs⁵¹.

Data on the identified indicators were sourced in May 2020 from the WHO Global Health Observatory database (https://www.who.int/data/gho), World Bank (https://data.worldbank.org/) and other sources detailed in Supplementary Table 3. National-level demographic data (population size and age structure) was sourced from the United Nations World Population Prospects and data on subnational variation in demography was sourced from WorldPop²⁷. Household size data was defined by the mean number of individuals in a household with at least 1 person aged >50 years, taken from the most recently available Demographic and Health Surveys data (https://dhsprogram.com). All country-level data for all indicators can be found online at the SSA-SARS-CoV-2 tool (https://labmetcalf.shinyapps.io/covid19-burden-africa/).

Comparisons of national-level estimates sourced from the WHO and other sources are affected by variation within countries and variation in the uncertainty around estimates from different geographical areas. To assess potential differences in data quality between geographical areas, we compared the year of the most recent data for the variables (Extended Data Fig. 2). The mean (range varied from 2014.624 to 2014.928 by region) and median year (2016 for all regions) of the most recent data varied little between regions. To account for the uncertainty associated with the estimates available for a single variable, we also included multiple variables per category (for example, demographic and socioeconomic factors, comorbidities, access to care) to avoid reliance on a single metric. This allowed exploring variation between countries across a broad suite of variables likely to be indicative of the different dimensions of risk.

Although including multiple variables that were likely to be correlated (see the principal component analysis (PCA) methods below for further discussion) would bias inference of cumulative risk in a statistical framework, we did not attempt to quantitatively combine risk across variables for a country, nor project risk based on the variables included in this study. Rather, we characterized the magnitude of variation among countries for these variables (see Fig. 3 for a subset of the variables and Fig. 4b for the bivariate risk maps following Chin et al.⁵²) and then explored the range of outcomes that would be expected under scenarios where the IFR increases with age at different rates (Fig. 4).

Variable selection and data sources for variables modulating the rate of viral spread

In addition to characterizing variation among factors likely to modulate burden, we also synthesized data sources relevant to the rate of viral spread, or pace, for the SARS-CoV-2 pandemic in SSA. Factors hypothesized to modulate viral transmission and geographical spread include climatic factors (for example, specific humidity), access to prevention measures (for example, handwashing) and human mobility (for example, international and domestic travel). Supplementary Table 2 outlines the dimensions of risk selected and references the previous studies relevant to the selection of these factors.

Climate data were sourced from the global, gridded ERA5 dataset⁵³ where model data were combined with global observation data (see Methods for climate-driven modeling of SARS-CoV-2 section for further details).

International flight data were obtained from a custom report from OAG Aviation Worldwide (UK) and included the departure location, arrival airport, date of travel and number of passenger seats for flights arriving to 113 international airports in SSA (see International air travel to SSA section).

As an estimate of connectivity within subregions of countries, the population-weighted mean travel time to the nearest city with a population greater than 50,000 was determined; details are provided in the section on Subnational connectivity among countries in SSA. To obtain a set of measures that broadly represent connectivity within different countries in the region, friction surfaces from Weiss et al.²⁶ were used to obtain estimates of the connectivity between different administrative level 2 units within each country. Details of this, alongside the metapopulation model framework used to simulate viral spread with variation in connectivity are found in the Subnational connectivity section.

Figure 3 shows variation among SSA countries for four of the variables and Extended Data Fig. 3 links to visualizations of variation for all variables. Figure 4 shows variation for a subset of the comorbidity and access to care indicators as a heatmap and Extended Data Fig. 4 shows variation for all the variables (also available at https://labmetcalf.shinyapps.io/covid19-burden-africa/).

PCA of variables considered

Selection of data and variables

The 29 national-level variables from Supplementary Table 3 were selected for the PCA. We conducted further PCA on the subset of 8 indicators related to access to healthcare (category E) and the 14 national indicators variables related to comorbidities (category B).

We excluded disaggregated subnational spatial variation data (variables A2, C1, E2 and category F), disaggregated or redundant variables derived from variables already included (variables A4 and D2) and disaggregated age-specific disease data from the Institute for Health Metrics and Evaluation (IHME) global burden of disease study (variables B2, B4 and B13) from the PCA analysis. COVID-19 tests per 100,000 population (variable D4; Supplementary Table 1), per capita gross domestic product (GDP) (variable A8) and the Gini index of wealth inequality (variable A9) were used to visualize patterns among SSA countries.

In some cases, data were missing for a country for an indicator; in these cases, missing data were replaced with a zero value. This is a conservative approach since zero values (that is, outside the range of typical values seen in the data) inflate the total variance in the dataset and thus, if anything, deflate the percentage of the variance explained by the PCA. Therefore, this approach avoids mistakenly attributing predictive value to principal components due to incomplete data. See Supplementary Table 3 for data sources for each variable.

PCA

The PCA was conducted on each of the three subsets described above using the scikit-learn library⁵⁴. To avoid biasing the PCA due to large differences in magnitude and scale, each feature was centered around the mean and scaled to unit variance before the analysis. Briefly, PCA applies a linear transformation to a set of n features to output a set of n orthogonal principal components that are uncorrelated and each explain a percentage of the total variance in the dataset⁵⁵. A link to the code for this analysis is available at https://labmetcalf.shinyapps.io/covid19-burden-africa/.

The principal components were then analyzed for the percentage of variance explained and compared to: (1) the number of COVID-19 tests per 100,000 population as of the end of June 2020 (Supplementary Table 1); (2) the per capita GDP; and (3) the Gini index of wealth inequality. For the Gini index, estimates from 2008 to 2018 were available for 45 of the 48 countries (no Gini index data were available for Eritrea, Equatorial Guinea and Somalia).

The first 2 principal components from the analysis of 29 variables explain 32.6 and 13.1% the total variance, respectively, in the dataset. Countries with higher numbers of completed SARS-CoV-2 tests reported tended to associate with an increase in principal component 1 (r = 0.67, P = 1.1 × 10⁻⁷; Extended Data Fig. 5a). Similarly, countries with a high GDP seemed to associate with an increase in principal component 1 (r = 0.80, P = 6.02 × 10⁻¹²; Extended Data Fig. 5b). In contrast, countries with greater wealth inequality (as measured by the Gini index) were associated with a decrease in principal component 2 (r = −0.42, P = 0.0042; Extended Data Fig. 5c). Despite these correlations, a relatively low percentage of variance was explained by each principal component: for the 29 variables, 13 of the 29 principal components were required to explain 90% of the variance (Extended Data Fig. 5d). When only the access to care subset of variables is considered, the first 2 principal components explain 50.7 and 19.1% of the variance, respectively, and 5 of 8 principal components are required to explain 90% of the variance. When only the comorbidities subset is considered, the first two principal components explain 27.9 and 17.8% of the variance, respectively, and 9 of 14 principal components are required to explain 90% of the variance (Extended Data Fig. 5d).

These data suggest that intercountry variation in this dataset is not easily explained by a small number of variables. Moreover, although correlations exist between principal components and high-level explanatory variables (testing capacity, wealth), their magnitude is modest. These results highlight that dimensionality reduction is unlikely to be an effective analysis strategy for the variables considered in this study. Despite this overall finding, the PCA on the access to care subset of variables highlights that the variance in these variables is more easily explained by a small number of principal components and hence may be more amenable to dimensionality reduction. This finding is unsurprising since, for example, the number of hospital beds per 100,000 population is likely to be directly related to the number of hospitals per 100,000 population (r = 0.60, P = 5.7 ×10 ⁻⁶ for SSA). In contrast, for comorbidities, the relationship between different variables is less clear. Given the low percentages of variation captured by each principal component, and the high variability between different types of variables, these results motivate a holistic approach to using these data for assessing relative SARS-CoV-2 risk across SSA.

Evaluating the burden emerging from the severity of infection outcome

Data sourcing: empirical estimates of IFR

Estimates of the IFR that account for asymptomatic cases, underreporting and delays in reporting are few; however, it is evident that the IFR increases substantially with age⁵⁶. We used age-stratified estimates of IFR from three studies (two published^2,4 and one preprint³) that accounted for these factors in their estimation (Supplementary Table 4).

To apply these estimates to other age-stratified data with different bin ranges and generate continuous predictions of IFR with age, we fitted the relationship between the midpoint of the age bracket and the IFR estimate using a generalized additive model using the mgcv package v1.8-33 (ref. ⁵⁷) in R v.4.0.2 (ref. ⁵⁸). We used a beta distribution as the link function for the IFR estimates (data distributed on [0, 1]). For the upper age bracket (80+ years), we took the upper range to be 100 years and the midpoint to be 90.

We assumed a given level of cumulative infection (20% in each age class, that is, a constant rate of infection among age classes) and then applied IFRs by age to the population structure of each country to generate estimates of burden. Age structure estimates were taken from the United Nations World Population Prospects (Supplementary Table 3) country-level estimates of population in 1-year age groups (0–100 years of age) to generate estimates of burden.

Comorbidities over age from the IHME

Applying these IFR estimates to the demographic structure of SSA countries provides a baseline expectation for mortality but depends on the assumption that mortality patterns in SSA are similar to those from where the IFR estimates were sourced (France, China and Italy). Comorbidities have been shown to be an important determinant of the severity of infection outcomes (that is, IFR). To assess the relative risk of comorbidities across age in SSA, estimates of comorbidity severity by age (in terms of annual deaths attributable) were obtained from the IHME Global Burden of Disease (GBD) study in 2017 (ref. ⁵⁹). Data were accessed through the GBD results tool for cardiovascular disease, chronic respiratory disease (not including asthma) and diabetes, reflecting three categories of comorbidity with demonstrated associations with risk (Supplementary Table 2). We assumed that higher mortality rates due to these noncommunicable diseases, especially among younger age groups, is indicative of increased severity and lesser access to sufficient care for these diseases, suggesting an elevated risk for their interaction with SARS-CoV-2 as comorbidities. While there are uncertainties in these data, they provide the best estimates of age-specific risks and have been used previously to estimate populations at risk²⁰.

The comorbidity by age curves for SSA countries were compared to those for the three countries from which SARS-CoV-2 IFR by age estimates were sourced. Attributable mortality due to all three noncommunicable disease categories was higher at age 50 in all 48 SSA countries when compared to estimates from France and Italy and for 42 of 48 SSA countries when compared to China (Extended Data Fig. 6).

Given the potential for populations in SSA to experience a differing burden of SARS-CoV-2 due to their increased severity of comorbidities in younger age groups, we explored the effects of shifting IFRs estimated by the generalized additive model of IFR estimates from France, Italy and China younger by 2, 5 and 10 years (Fig. 3).

International air travel to SSA

The number of passenger seats on flights arriving to international airports were grouped by country and month for January to April 2020 (Supplementary Table 5), the months when the introduction of SARS-CoV-2 to SSA countries was likely to have first occurred. The first confirmed case reported from an SSA country, according to the Johns Hopkins Coronavirus Research Center was in Nigeria on 28 February 2020. By 31 March 2020, 43 of 48 SSA countries had reported SARS-CoV-2 infections and international travel was largely restricted by April. Lesotho was the last SSA country to report a confirmed SARS-CoV-2 infection (on 13 May 2020); however, given the difficulties in surveillance, the first reported detections were likely delayed relative to the first importations of the virus. The probability of importation of the virus is defined by the number of travelers from each source location, each date and the probability that a traveler from that source location on that date was infectious. Due to limitations in surveillance, especially early in the SARS-CoV-2 pandemic, empirical data on infection rates among travelers were largely lacking. To account for differences in the status of the SARS-CoV-2 pandemic across source locations and thus differences in the importation risk for travelers from those locations, we coarsely stratified travelers arriving each day into 4 categories based on the status of their source countries: (1) travelers from countries with zero reported cases (that is, although undetected transmission was possibly occurring, SARS-CoV-2 had not yet been confirmed in the source country by that date); (2) those traveling from countries with more than 1 reported case (that is, SARS-CoV-2 had been confirmed to be present in that source country by that date); (3) those traveling from countries with more than 100 reported cases (indicating community transmission was likely beginning); and (4) those traveling from countries with more than 1,000 reported cases (indicating widespread transmission).

To determine reported case counts at source locations for travelers, no cases were reported outside China until 13 January 2020 (the date of the first reported case in Thailand). Over 13 January to 21 January, cases were then reported in Japan, South Korea, Taiwan, Hong Kong and the United States (https://covid19.who.int/). Subsequently, counts per country were tabulated daily by the Johns Hopkins Coronavirus Resource Center⁶⁰ beginning on 22 January (https://coronavirus.jhu.edu/map.html); we used the data from 22 January onwards and the WHO reports before 22 January.

The number of travelers within each category arriving per month is shown in Supplementary Table 5. This approach makes the conservative assumption that the probability a traveler is infected reflects the general countrywide infection rate of the source country at the time of travel (that is, travelers are not more likely to be exposed than non-travelers in that source location) and does not account for complex travel itineraries (that is, a traveler from a high-risk source location transiting through a low-risk source location would be grouped with other travelers from the low-risk source location). Consequently, the risk for viral importation is likely systematically underestimated. However, since the relative risk for viral importation will still scale with the number of travelers, comparisons among SSA countries can be informative (for example, SSA countries with more travelers from countries with confirmed SARS-CoV-2 transmission are at higher risk for viral importation).

Subnational connectivity among countries in SSA

Indicators of subnational connectivity

To allow comparison of the relative connectivity across countries, we used the friction surface estimates provided by Weiss et al.²⁶ as a relative measure of the rate of human movement between subregions of a country. For connectivity within subregions of a country (for example, transport from a city to the rural periphery), we used as an indicator the population-weighted mean travel time to the nearest urban center (that is, population density >1,500 per square kilometer or a density of built-up areas >50% coincident with population >50,000) within administrative 2 units⁶¹. For some countries, estimates at administrative 2 units were unavailable (Comoros, Cape Verde, Lesotho, Mauritius, Mayotte and Seychelles); estimates at the administrative-1 unit level were used for these cases (these were all island nations, with the exception of Lesotho).

Metapopulation model methods

Once SARS-CoV-2 has been introduced into a country, the degree of spread of the infection within the country is governed by subnational mobility: the pathogen is more likely to be introduced into a location where individuals arrive more frequently than one where incoming travelers are less frequent. Large-scale consistent measures of mobility are rare. However, recently, estimates of accessibility have been produced at a global scale²⁶. Although this is unlikely to perfectly reflect mobility within countries, especially since interventions and travel restrictions are put in place, it provides a starting point for evaluating the role of human mobility in shaping the outbreak pace across SSA. We used the inverse of a measure of the cost of travel between the centroids of administrative level 2 spatial units to describe mobility between locations (estimated by applying the costDistance function in the gdistance package v1.3-6 in R to the friction surfaces supplied in Weiss et al.²⁶). With this, we developed a metapopulation model for each country to develop an overview of the possible range of trajectories of unchecked spread of SARS-CoV-2.

We assumed that the pathogen first arrives in each country in the administrative 2 level unit with the largest population (for example, the largest city) and the population in each administrative 2 level (of size N_j) is entirely susceptible at the time of arrival. We then tracked the spread within and between each of the administrative 2 level units of each country. Within each administrative 2 level unit, dynamics are governed by a discrete time susceptible (S), infected (I) and recovered (R) model with a time step of approximately one week, which is broadly consistent with the serial interval of SARS-CoV-2. Within the spatial unit indexed j, with total size N_j, the number of infected individuals in the next time step is defined by:

where β captures the magnitude of transmission over the course of one discrete time step; since the discrete time step chosen is set to approximate the serial interval of the virus, this will reflect the R₀ of SARS-CoV-2, and is thus set to 2.5; the exponent α = 0.97 is used to capture the effects of discretization⁶² and I_j,t captures the introduction of new infections into site j at time t. Susceptible and recovered individuals are updated according to:

where b reflects the introduction of new susceptible individuals resulting from the birth rate, set to reflect the most recent estimates for that country from the World Bank Data (https://data.worldbank.org/indicator/SP.DYN.CBRT.IN), and w reflects the rate of waning of immunity. The population is initiated with S_j,1 = N_jR_j,1 = 0, and I_j,1 = 0 except for the spatial unit corresponding to the largest population size N_j for each country since this is assumed to be the location of introduction; for this spatial unit, we set I_j,1 = 1.

We made the simplifying assumption that mobility linking locations i and j, denoted as c_i,j, scales with the inverse of the cost of travel between sites i and j evaluated according to the friction surface provided in Weiss et al.²⁶. The introduction of an infected individual into location j is then defined by a draw from a Bernouilli distribution following:

where L is the total number of administrative 2 units in that country and the rate of introduction is the product of connectivity between the focal location and each other location multiplied by the proportion of population in each other location that is infected.

Some countries show rapid spread between administrative units within the country (for example, a country with parameters that broadly reflect those available for Malawi; Extended Data Fig. 7), while in others (for example, reflecting Madagascar), connectivity may be so low that the outbreak may be over in the administrative unit of the largest size (where it was introduced) before introductions successfully reach other poorly connected administrative units. Where duration of immunity is sufficiently long, the result may be a hump-shaped relationship between the proportion of the population that is infected after five years and the time to the first local extinction of the pathogen (Extended Data Fig. 7, top right). In countries with lower connectivity (for example, resembling Madagascar), local outbreaks can go extinct rapidly before traveling very far; in other countries (for example, resembling Gabon), the pathogen goes extinct rapidly because it travels rapidly and rapidly depletes susceptible individuals everywhere. The U-shaped pattern diminishes as the rate of waning of immunity increases and is replaced by a monotonic negative relationship. With sufficiently rapid waning of immunity, local extinction ceases to occur in the absence of control efforts.

The impact of the pattern of travel between centroids is echoed by the pattern of travel within administrative districts: countries where the pathogen does not reach a large fraction of the administrative 2 units within the country in five years are also those where within-administrative-unit travel is low (Extended Data Fig. 7, right).

These simulations provide a window into qualitative patterns expected for subnational spread of the pandemic virus but there is no clear way of calibrating the absolute rate of travel between regions of relevance for SARS-CoV-2; this is further complicated by the remaining uncertainties around rates of waning of immunity. Thus, the time scales of these simulations should be considered in relative, rather than absolute terms. Variation in lockdown effectiveness, or other changes in mobility for a given country, may also compromise relative comparisons as might large volumes of land border crossings in some settings, which we have not accounted for in this study. Variability in testing and case reporting complicates clarifying this (Extended Data Fig. 7, bottom left and bottom right, respectively) but we have highlighted countries with less connectivity (that is, less synchronous outbreaks expected) relative to the median among SSA countries and with older populations (that is, a greater proportion in higher-risk age groups) (Extended Data Fig. 8).

The University of Oxford’s Blavatnik School of Government generated composite scores of government response, interventions for containment and economic support provided, with each scored from 0 to 100 (Coronavirus Government Response Tracker; https://www.bsg.ox.ac.uk/research/research-projects/coronavirus-government-response-tracker). These data were compared with the day on which ten cases were exceeded in a country according to the Johns Hopkins dashboard data (Johns Hopkins Coronavirus Resource Center; https://coronavirus.jhu.edu/map.html).

While faster waning of immunity will act to increase the rate of spread of the infection, resulting in a higher proportion infected after one year, control efforts will generally act to slow the rate of spread of the infection (Extended Data Fig. 9). Since different countries are likely to have differently effective control efforts (Extended Data Fig. 9), this precludes making country-specific predictions as to the relative impact of control efforts on delay.

Modeling epidemic trajectories in scenarios where transmission rate depends on climate

Climate data sourcing: variation in humidity in SSA

Specific humidity data for selected urban centers comes from the ERA5 using an average climatology (1981–2017)⁵³; we did not consider year-to-year climate variations. Selected cities (n = 56) were chosen to represent the major urban areas in SSA. The largest city in each SSA country was included as well as any additional cities that were among the 25 largest cities or busiest airports in SSA.

Methods for climate-driven modeling of SARS-CoV-2

We used a climate-driven susceptible-infected-recovered-susceptible model to estimate epidemic trajectories (that is, the time of peak incidence) in different cities in 2020, assuming no control measures were in place or a 10 or 20% reduction in R₀ beginning 2 weeks after the total reported cases for a country exceeded 10 cases^25,63. The model is given by:

where S is the susceptible population, I is the infected population and N is the total population. D is the mean infectious period, set at 5 d following ref. ²⁵.

To investigate the effects on epidemic trajectories of a climate dependency of SARS-CoV-2 on cities with the climate patterns of the selected cities in SSA, we used parameters from the most climate-dependent scenario in ref. ²⁵, based on the endemic betacoronavirus HKU1 in the United States. In this scenario L, the duration of immunity, was 66.25 weeks (that is, >1 year and such that waning immunity did not affect the timing of the epidemic peak). We initially selected a range where R₀ declined from R_0max = 2.5 to R_0min = 1.5 (that is, transmission declined 40% at high humidity) since this exceeds the range observed for influenza and other coronaviruses for which data are available (from the United States). R_0max = 2.5 was chosen because 2.5 is often cited as the approximate R₀ for SARS-CoV-2. Thus, we initially assumed that the climate dependence of SARS-CoV-2 in SSA would not greatly exceed that of other known coronaviruses from the US context. Then, we explored the effects of different degrees of climate dependency (that is, wider ranges between R_0max = 2.5 to R_0min = 1.5 and scenarios where R_0min approached 1) (Extended Data Fig. 10).

Transmission is governed by β(t), which is related to the basic reproduction number R₀ by R₀(t) = β(t)D. The basic reproduction number varies based on climate and is related to specific humidity according to the equation:

where q(t) is specific humidity⁵³ and a is set at −227.5 based on estimated HKU1 parameters²⁵. We assumed the time of introduction for cities to be the date at which the total reported cases for a country exceeded 10 cases.

Sensitivity analysis

Selecting an R_0min value of 1, such that epidemic growth stops at high humidities, is likely implausible since simulations indicated no outbreaks would occur in cities such as Antananarivo (countered by the observation that SARS-CoV-2 outbreaks did in fact occur) (Extended Data Fig. 10b; see Supplementary Table 1 for the reported case counts at the country level). Expanding the range between R_0min and R_0max by increasing R_0max resulted in epidemic peaks being reached earlier after outbreak onset but did not increase the difference in timing between cities with different climates (Extended Data Fig. 10c; for example, the difference in timing between peaks in Windhoek and Lomé is similar in 10a and 10c). Finally, we explored scenarios where the R_0min was between 1.0 and 1.5. When R_0min > 1.1, epidemic peaks were seen in each SSA city with the difference in timing of the peak growing larger when smaller values of R_0min were selected (Extended Data Fig. 10d). However, the difference in timing, even when small values of R_0min were selected, was a maximum of 25 weeks and rapidly reduced to only a few weeks when R_0min approached 1.5.

Reporting Summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Source: Ecology - nature.com