Dynamics of influenza in tropical Africa: Temperature, humidity, and co‐circulating (sub)types

Background The association of influenza with meteorological variables in tropical climates remains controversial. Here, we investigate the impact of weather conditions on influenza in the tropics and factors that may contribute to this uncertainty. Methods We computed the monthly viral positive rate for each of the 3 circulating influenza (sub)types (ie, A/H1N1, A/H3N2, and B) among patients presenting with influenza‐like illness (ILI) or severe acute respiratory infections (SARI) in 2 Ugandan cities (Entebbe and Kampala). Using this measure as a proxy for influenza activity, we applied regression models to examine the impact of temperature, relative humidity, absolute humidity, and precipitation, as well as interactions among the 3 influenza viruses on the epidemic dynamics of each influenza (sub)type. A full analysis including all 4 weather variables was done for Entebbe during 2007‐2015, and a partial analysis including only temperature and precipitation was done for both cities during 2008‐2014. Results For Entebbe, the associations with weather variables differed by influenza (sub)type; with adjustment for viral interactions, the models showed that precipitation and temperature were negatively correlated with A/H1N1 activity, but not for A/H3N2 or B. A mutually negative association between A/H3N2 and B activity was identified in both Entebbe and Kampala. Conclusion Our findings suggest that key interactions exist among influenza (sub)types at the population level in the tropics and that such interactions can modify the association of influenza activity with weather variables. Studies of the relationship between influenza and weather conditions should therefore determine and account for co‐circulating influenza (sub)types.


| BACKG ROU N D
Influenza epidemics cause substantial morbidity and mortality worldwide. Weather conditions, such as humidity and temperature, have been identified as key factors shaping the dynamics of influenza transmission. In temperate regions, lower humidity and lower temperature have repeatedly been shown to be associated with the wintertime epidemics. [1][2][3][4][5] In tropical regions, however, few studies have been conducted and, from the limited number of studies, findings on the impact of weather conditions are inconsistent.
While some studies reported increased influenza circulation during the rainy seasons [5][6][7][8][9][10][11][12] and association with higher humidity and/or precipitation in the tropics, 5,7,13 others reported no or contradicting effects of these climate variables. [14][15][16][17] Here, we investigate the relationship between weather conditions and influenza transmission in Uganda, a tropical country, as well as factors that may contribute to the aforementioned inconclusiveness. We used regression models to test the impact of temperature, humidity, and precipitation on influenza transmission. The two studied cities are in close vicinity with similar weather conditions.
Comparing findings for these 2 cities thus allowed identification of inconsistencies and spurious associations. In addition, as the influenza samples were fully subtyped, we were able to address 3 questions that have been rarely, if ever, investigated: (i) How do effects of weather conditions vary by influenza (sub)type? (ii) How do interactions among influenza (sub)types manifest at the population level? And (iii) how do these viral interactions along with weather conditions shape the epidemic dynamics of each influenza virus?

| Ethics statement
In the context of routine public health surveillance, verbal consent was obtained from suspected cases ≥18 years of age and from parents or legal guardians for cases <18 years of age. The study was approved by the Research Ethics Committee at UVRI, the Uganda National Council for Science and Technology, and the Institutional Review Board at Columbia University Medical Center. The study's funders had no role in study design, data analysis and interpretation, manuscript preparation, or decision to publish.

| Data
At each sentinel site, clinicians identified patients with influenzalike illness (ILI) and severe acute respiratory infection (SARI) using an established protocol. 9,10 Per standard World Health Organization guidelines, patients met the case definition for ILI if they were ≥2 months of age, presenting with a fever (>38°C) and either cough or sore throat. SARI was defined as: (i) a child aged 2 months to <5 years requiring hospitalization, with recent onset of cough or difficulty breathing within 10 days of symptom onset and an additional indicator of respiratory distress; or (ii) a patient aged ≥5 years requiring hospitalization, with a history of fever presenting with cough, shortness of breath, or difficulty breathing within 10 days of symptom onset. 9,10 Naso-and/or oropharyngeal swab samples were collected at enrollment from ILI/ SARI patients, and all samples were tested for influenza viruses by (sub)type-specific RT-PCR using primers provided by the U.S.
Centers for Disease Control and Prevention. 10 18 These weather variables were also available for Entebbe, from the same UBoS reports. We therefore used the average of the maximum and minimum temperature for each month to represent the monthly mean temperature and used these values along with monthly precipitation in the partial statistical analysis for a comparison of the 2 cities (2.4).

| Full statistical analysis
As complete data for temperature, precipitation, relative and absolute humidity are available for Entebbe, we performed a statistical analysis of the relationship between influenza activity and these 4 weather variables for the city during January 2007-December 2015.
We refer to this analysis as the full statistical analysis hereafter. We first examined the relationship for each influenza (sub)type and all strains combined without adjustment for co-circulating influenza

| Influenza activity and weather conditions
We used logistic regression models 21 with 1 autoregressive term to examine the impact of weather variables on monthly influenza activity. The basic model for this analysis took the following form: where Logit (p) is the logit function, that is, ln[p/(1-p)]; Flu represents the monthly viral positive rate for either A/H1N1, A/ H3N2, B, or all influenza viruses combined (All); Flu at lag-1 is the lag-1 autoregressive term. Seasonality (t) is a function to account for seasonality in the data; here, we used harmonics of the form with t as the month index (eg, t = 1 for January and t = 2 for February). 22 (Note that as sin 2πk 12 t for k = 6 is always 0, there are 11 coefficients, ie, β 2-12 , in this formula.) Temp represents temperature. HUM represents 1 of the 3 humidity-related variables (ie, precipitation, RH, and AH). That is, only 1 humidity-related variable was included in the model in order to avoid over fitting (same for all other models in this study). For models that include both temperature and a humidity-related variable, we did not adjust for collinearity between the 2 terms, since this effect was not severe (the variance inflation factors 23 were <5 for all variable pairs).

| Influenza activity, weather conditions, and co-circulating (sub)types
To examine the potential confounding and modification effect of cocirculating (sub)types on the relationship between influenza activity and weather conditions, we also developed models that adjusted for activity of co-circulating (sub)types. The basic model for this analysis assumed the following structure: where co-Flu1 and co-Flu2 represent concurrent viral positive rates for the other 2 co-circulating (sub)types. For instance, for A/H1N1 (1) Logit (Flu) ∼ β 0 +β 1 (Flu at lag-1) F I G U R E 1 Monthly viral positive rates for different influenza viruses (y-axis on the left) and weather variables (y-axis on the right) in Entebbe and Kampala. Weather variables are standardized to have zero mean and unit variance. Humidity data (RH and AH) are not available for Kampala (ie, Flu=A/H1N1), co-Flu1 and co-Flu2 are A/H3N2 and B, respectively. Note that, the 3 Flu time series (ie, proportions of specimens testing positive for A/H1N1, A/H3N2, or B, respectively, among all ILI/SARI visits) are not dependent on one another due to sharing the same denominator, as the denominator is not the number of influenza positive visits and the 3 proportions at each time step do not sum to 100% (Figure 1). Other variable settings were the same as in Equation 1.

| Best-fit models
To thoroughly search for the best-fit model for each influenza time series, we tested all possible combinations of co-Flu1, co-Flu2, temperature, and a humidity-related variable. Note this test also included models without any co-Flu variable or weather variable.
We then pooled all models and selected the one with the lowest Bayesian information criterion (BIC) 24 as the best-fit model for each influenza time series.

| Leave-one-out cross-validation
We performed cross-validation of the 2 basic models (Equations 1 and 2), by running the models excluding data from one of the study years (ie, 2007-2015). For instance, in a leave-one-out crossvalidation for Year 2008, data for 2008 were excluded and the remainder used to fit the models. Model estimates for each covariate across years were then compared.

| Partial statistical analysis
Statistical correlations consistently identified in multiple locations would strengthen these inferred associations. Due to a lack of humidity data for Kampala, we were unable to perform a full analysis as performed for Entebbe (2.3) and compare results between the 2 locations. Nevertheless, as a partial validation, we fitted the 2.3.1 and 2.3.2 models for both Kampala and Entebbe using a subset of weather variables (ie, temperature and precipitation), for which data were available for both cities for at least 3 years. That is, only precipitation was used as an HUM variable in Equations 1 and 2. For this comparison, we restricted the study period to July 2008-December 2014 when data for influenza activity, temperature, and precipitation were available for both cities. In addition, to control for effects due to the 2009 pandemic, we excluded the 2009 pandemic (July-December 2009) as in the full analysis. As in 2.3, we pooled all models [ie, with or without adjustment for co-circulating influenza (sub) types] and selected the one with the lowest BIC as the best-fit model for each influenza time series. In addition, cross-validation was performed for both cities as described in 2.3.4.

| Software
All statistical analyses were performed in R language (R Foundation for Statistical Computing, Vienna, Austria). All logistic models were fitted using the "bayesglm" function in the "arm" package in R. 25 The "bayesglm" method, developed by Gelman et al, 26 takes a Bayesian approach to obtain stable logistic regression coefficients using weakly informative priors (the default prior is a Cauchy distribution).

| Summary statistics
The surveillance network detected 514 influenza cases among 4477 ILI/SARI visits (11.48%) Figure S1). In both cities, temperature is high year-round; mean temperature during the study period was above 20°C in both cities (Table 1 and Figure S1).
Typically, each year has 2 rainy seasons (March-May and September-November; Figure 1 and Figure S1), with much higher levels of precipitation during rainy seasons (Table 1). In Entebbe, where humidity data were available, ambient humidity is high year-round; monthly averages were always above 60% for RH and above 11 g/kg for AH during the study period and were slightly higher during rainy seasons (Table 1 and Figure S1).

| Influenza activity and weather conditions, without adjusting for co-circulating (sub)types
We first modeled the relationship between influenza activity and all 4 weather variables (ie, temperature, precipitation, relative, and  Figure 2H).
Overall, although not statistically significant, A/H3N2 activity appeared to be negatively correlated with temperature, with a mean OR, averaged over 30 models, of 0.82 ( Figure 2H). Conversely, a positive correlation with temperature, although not statistically significant, was consistently estimated for influenza B; the mean OR was 1.19 across 30 models ( Figure 2L).
For all influenza viruses combined, the models did not identify any significant correlates ( Figure 2M-P)

| Influenza activity, weather conditions, and co-circulating (sub)types
In our second analysis, we further examined potential interactions among co-circulating influenza (sub)types and the resulting impact on the relationship of epidemic activity with weather variables. Consistent with our first analysis, in models including . The year left out in the cross-validation is shown on the y-axis (eg, "-2007" indicates data for Year 2007 were excluded, and "-none" indicates data for the entire study period were included). The associations with precipitation (Precip), relative humidity (RH), absolution humidity (AH), and temperature (Temp) are shown in columns 1 to 4, and (sub)type interactions are shown in columns 5 and 6; the vertical segments show the 95% confidence intervals for each variable; "x"s denote the mean, and "*"s indicate variables significant at the 5% level. For temperature (4th column), 3 estimates are shown for each dataset, corresponding to the 3 models using 1 of the 3 humidity variables (ie, precipitation, RH, and AH) shown in Figure S3. As shown in Figure 4, none of the 4 weather variables were included in the best-fit models. However, a negative association between influenza A/H3N2 and B activity was identified in the best-fit models for both viruses ( Figure 4B  , and the lower panel (D-F) shows estimates by the best-fit models (ie, the ones with the lowest BIC) among all tested models. The predictors are labeled along the y-axis; the vertical segments show the 95% confidence intervals for each predictor for Entebbe (in blue) and Kampala (in orange); "x"s denote the mean and "*"s indicate predictors significant at the 5% level

| Partial statistical analysis for Entebbe and Kampala
To verify the relationships between influenza activity with weather conditions and co-circulating (sub)types identified for Entebbe, we performed a similar statistical analysis for Kampala. However, due to a lack of data for humidity in Kampala, in this comparison, for both cities the models only included temperature and precipitation.
Model fits for both cities are shown in Figure S4. Figure  for Entebbe and 0.085 for Kampala; Figure 5B); however, this positive correlation was not mutually identified in the A/H1N1 models (P = .15 for Entebbe and 0.29 for Kampala; Figure 5A). For the 2 weather variables examined, lower temperature was associated with higher A/H3N2 for both cities (P = .054 for Entebbe and P = .00042 for Kampala; Figure 5B). These estimated associations were in general in agreement across years as shown in the leave-one-out crossvalidation ( Figure S5). This finding provides support for the impact of cross-immunity at the population level. Further, as influenza vaccination coverage in the 2 studied Ugandan cities was near zero, 9 the inter-(sub)type interactions reported here likely reflect the impact of cross-immunity conferred via natural infections, as opposed to vaccination.

| D ISCUSS I ON
Due to sparse observations of influenza infection and a lack of long-term disease surveillance, we used monthly data for the analyses. This coarse time resolution likely limited our ability to identify the impact of weather variables, which tend to act acutely (eg, lower AH ~2 weeks prior to the epidemic onset 4 ), and in part explains the null findings for relative and absolute humidity. This limitation stresses the need for enhanced surveillance and research on influenza in tropical regions, 39 in particular, in Africa. In addition, due to a lack of publically available weather data, we were only able to conduct a partial comparison between the 2 Ugandan cities. Findings from this partial comparison were not entirely consistent between the 2 cities (except for the negative association between A/H3N2 and B). This inconsistency could stem from a lack of statistical power due to low signal/noise ratios, or location-specific factors (eg, population immunity) that were not accounted for in the models. Further investigations into these issues are warranted.
Despite the limited disease data, we have shown differing correlations with weather variables for different influenza (sub)types as well as key viral interactions among co-circulating (sub)types at the population level in 2 tropical cities. Future work using data with better temporal resolutions and from more tropical locations hopefully will reveal a more comprehensive picture of the dynamics of influenza epidemics in tropical climates.

ACK N OWLED G EM ENTS
We thank Sasikiran Kandula for assistance in extracting the weather data. This study was supported by US NIH grants GM100467,