Climate factors influence seasonal influenza activity in Bangkok, Thailand

influenza viruses in two ways. A univariate time series used past data as dependent variables to predict the future, while multivariate time series used the combined data from dependent and independent variables for prediction. The forecast model used data from January 2011 to December 2017 to calibrate and estimate the parameters. The last 12 months of data were used to test the forecasting ability. Data from 2010 were excluded due to bias from the influenza pandemic of 2009. The Auto-Regressive Integrated Moving Average (ARIMA) (p,d,q) model was used in this study. The parameter p represents maximum lag in the auto-regression (AR). The non-sea- sonal differencing (d) is the order of integration (I) and represents the number of differences required to make the series stationary. The parameter q of the moving average (MA) repre- sents the linear regression of the output value on the time series when compared with the cur- rent and previous terms [34]. Additionally, seasonal autoregressive integrated moving average model SARIMA (p,d,q)(P,D,Q)s was also used. The parameters P, D, Q, and s represent the order of seasonal AR, the degree of seasonal differencing, the seasonal moving average, and the season’s length, respectively [35]. Forecasting incidence by using ARIMA model involved four steps. First, we used the aug- mented Dickey-Fuller (ADF) and the Kwiatkowski-Phillips Schmidt-Shin (KPSS) tests to assess whether the individual time series was stationary. The log-transformation and/or

differencing was applied to render the data stationary when necessary. Second, autocorrelation function (ACF) and partial autocorrelation function (PACF) were applied to estimate p and q based on the cutoff time lag. The P and Q order was obtained from ACF and PACF peak at time s lags. Third, residual diagnosis and Ljung-Box (modified Box-Pierce) test was used to identify the possibility of the model and to diagnose white-noise. Finally, the Ljung- Box portmanteau test was applied to determine whether there was a lack-of-fit of the residuals of the time series. The null hypothesis was that the model does not exhibit lack-of-fit. It was expected that the test failed to reject the null hypothesis and that the residuals series showed white noise (p-value >0.5) [36]. The model goodness-of-fit took into account the root mean square error (RMSE), the mean absolute percentage error (MAPE), and the corrected Akaike’s Information Criterion (AICc). The lowest parameter values suggest the optimal model [37]. For multivariate model, only cli- mate variables that provided significant lagged effect and presented the highest correlation coefficient among lag 1 up to lag 4 were considered [26, 38]. All statistical analysis was per- formed in SPSS version 23 (SPSS Inc., USA) and R version 3.6.0 (R Foundation, Vienna, Aus- tria) [39]. Alpha of 0.05 denotes statistical significance. Results Influenza surveillance In this study period, 6,678 (21.6%) of the ILI samples tested positive for influenza virus

(Table 1). Of these, 80.2% (5,357/6,678) were influenza A virus and 19.8% (1,321/6,678) were influenza B virus. Influenza A virus subtypes identified were 34.2% A(H1N1)pdm09 and 46.0% A(H3N2). From 2010 to 2018, overall influenza cases generally peaked in and around August to September, which coincided with the rainy season (Fig 1). Biannual frequency of predominantly influenza A(H1N1)pdm09 virus was more apparent than that of influenza B virus. Climate factors and influenza activity We hypothesized that certain climate factors were associated with increased influenza activity, therefore we examined monthly variations in the average temperature, relative humidity, and PLOS ONE Climate factors influence seasonal influenza activity in Bangkok, Thailand PLOS ONE | https://doi.org/10.1371/journal.pone.0239729 September 29, 2020 4 / 15