Data collection
Weekly case counts of all diagnosed PTB cases who resided in the Brunei-Muara district, Brunei between January 2001 and December 2018 (18 years, 939 weeks) were compiled from the National TB Coordinating Centre (NTCC). Brunei-Muara district is the most populated district in the country where 69.7% of the population reside21, and where the capital city is located. NTCC was established as part of the National TB programme in Brunei, and has implemented TB surveillance, treatment and control programmes since 2000. All patients suspected to have any form of TB across the whole country are often referred to NTCC, or any respective district directly observed treatment, short course (DOTS) centre, for diagnosis, treatment and follow-up25. All modes of diagnosis for the PTB cases were included (such as smear-positive, smear-negative, and through chest X-ray and/or clinician’s decision). These case counts were summed up by epidemiological week and year, based on treatment start date. In cases where the treatment start date is missing, the NTCC registration date was used.
Daily data on climate variables for the same period were obtained from a local meteorological station, located at Brunei-Muara district. The variables provided includes total sunshine hours, total rainfall (in millimeters), average wind speed (in knots), relative humidity (RH) in percentage (minimum, mean and maximum), and temperature in degree Celsius (minimum, mean and maximum). These daily data were averaged by epidemiological week and year. Any missing daily values (n = 5) were replaced with the mean value for that particular month and year. Vapour pressure (a measure of absolute humidity) was calculated using the Clausius-Clapeyron equation26, by inputting the mean RH values and the standard temperature and pressure conditions.
Statistical analysis
Spearman’s rank correlation test was used to explore the correlation between each climate variable, and with PTB case counts. Stationarity of the time series for weekly PTB case counts and each climate variable were checked using the augmented Dickey-Fuller test.
We used distributed lag non-linear model (DLNM) framework to investigate the association between climate variables and PTB incidence. Under this model framework, negative binomial distribution was assumed to account for overdispersion, and crossbasis terms were constructed for each climate variables. These terms comprise of 2 dimensions: one specifying the conventional exposure–response relationship, and the other specifying the lag-response relationship27. Natural cubic splines with 7 degrees of freedom (df) per calendar year were used to account for long-term trends and seasonality. This adjustment was included based on previous similar studies for TB5,17, and the number of df was determined using the Akaike’s Information Criterion (AIC) value. Natural cubic splines with 3 df were used to describe both the lagged and non-linear effects of each climate variable.
The median incubation period for PTB ranges between few months to 2 years14, and there is often a delay in diagnosing TB, by about 5–6 months25,28. Considering these factors, we decided to specify lags of up to 12 months (52 weeks) to capture the delayed effects of climate variables. The rationale is to cover as much of the incubation period without sacrificing any loss of statistical accuracy and efficiency that could be caused by adding more lags4,17. The general model formula structure used is as follows:
$$log Eleft( {Y_{t} } right) = alpha + sum CB left( {M,lag} right) + nsleft( {t,df = 7/year times no. of years} right)$$
where E(Yt) is the expected number of PTB cases at week t, (alpha) is the intercept, CB is the cross-basis function used for each climate variable to be assessed (M), and ns is the natural cubic spline function applied to account for long-term trend and seasonality. The presence of any residual auto-correlation were assessed using partial autocorrelation function plots (PACF). Any remaining autocorrelation detected was accounted for by adding lags of the model’s deviance residuals into the final model.
Although not all variables give significant results during univariate analysis, we decided to include 5 crossbasis terms that represent different aspects of climate variables and that are also previously known to be associated with TB incidence. The rationale here is to include these variables to control for potential confounding. The AIC value was used to assess which variables to be included in final model. This resulted in the choice of the following 5 variables in the final model: average wind speed, total sunshine hours, total rainfall, mean RH and minimum temperature. To ensure minimal issues with multi-collinearity and/or correlation (due to the use of multiple crossbasis terms in a single model), consistency in results obtained between univariate and multivariate were checked using visual analysis and referring to the AIC value.
We reported the relative risk (RR) of weekly PTB cases at the 5th and 95th percentiles of each climate variable, compared to their median, with corresponding 95% confidence intervals (95% CI). For climate variables with significant results observed at either percentile, we further determined and reported the starting lag week at which this significant result can be found. Overall relationship patterns were also described using three-dimensional (3D) and contour plots. Lag plots were used to show trend differences at lags 0, 13, 26, 39 and 52 weeks (corresponding to 0, 3, 6, 9 and 12 months), with higher number of lags indicating longer lagged effect of the corresponding climate variable As an additional sub-analysis, we repeated the same analysis described above to report the RR of weekly smear-positive PTB cases. Sensitivity analyses were conducted by repeating the analysis using natural cubic splines of 5 and 9 df for long-term trend. All analyses were done and all figures were generated in R (ver. 4.1), using tseries, splines and dlnm packages29,30.
This study was approved by the Medical and Health Research and Ethics Committee (MHREC), Ministry of Health, Brunei (Ref: MHREC/UBD/2019/2). All methods were performed in accordance with the relevant guidelines and regulations. Informed consent was waived because all analyses were based on aggregated data which do not contain any identifying or sensitive information.
