All JE cases in this study originated from central Sarawak an area with approximately 650,000 people (180,000 children ≤12 years). Sarawak has a generally stable equatorial climate that is hot and humid throughout the year, with two monsoon seasons: a northeast monsoon season with heavy rainfall occurring from November to March and a southeast monsoon season with relatively less rainfall occurring from May/June to September [17]. All cases of encephalitis in the region are referred to Sibu Hospital. With the introduction of the formalin-inactivated mouse derived JE vaccine (Biken, Japan), a significant reduction in cases has been observed [18], [19].

The JE vaccine is provided for free under the Malaysian National Immunisation Programme and the recommended JE immunization interval schedule in Sarawak is: the primary series at 9 and 10 months with subsequent booster at 18 months, 4.5, 8, 11 and 15 years; one injection of 0.5 mL is given at each interval [19]–[21]. However, in practice we are reasonably certain that the majority of children in our study received at least two doses of the vaccine between 9 and 12 months.

The original data collection of the study was approved by the Director of Health for Sarawak and the Ethics Committee of the Liverpool School of Tropical Medicine (Liverpool, UK). All data analyzed were anonymized by removing any potential unique identifiers before conducting the analysis. Informed consent was obtained from each child's parent or guardian before samples were taken [18].

The data on JE cases were obtained through a prospective study in Sibu Hospital and a JE surveillance programme set up by the Sarawak Health Department since 1997 [18], [19]. Briefly, patients with suspected JE were enrolled into the hospital-based surveillance study. Paired sera and paired CSF for each patient were considered to be the ideal specimen set. However, in reality the complete specimen set was not obtained from many cases, from whom we usually obtained a single serum and a CSF specimen. Specimens were tested for JEV specific immunoglobulin-type M by MAC-ELISA (Venture Technologies Sdn Bhd, Malaysia), which distinguishes IgM elicited by JEV from that elicited by dengue viruses [22], [23]. The sensitivity, specificity, positive predictive value and negative predictive value of the test used were 83%, 99%, 0.98 and 0.92, respectively for CSF and 91%, 95%, 0.92 and 0.94, respectively for serum [23]. All cases of encephalitis with specific IgM to JEV in serum and/or CSF were considered to have been infected recently with JEV and classified as confirmed JE cases. It is important to note that by the time a JE infection leads to encephalitis there is usually no virus in the periphery and thus a “confirmatory test” looking for evidence of virus or virus genome is not useful.

Monthly mean, maximum, minimum and range temperature (°C) and mean rainfall (cm/month) for central Sarawak was acquired from the KNMI Climate Explorer [24] based on the Climate Research Unit [25] dataset. The weather coverage window we used was latitude: 3.731–1.205° and longitude: 111.078–115.236° (see Supplemental Material, Figure S1). We used the updated CRU TS 3.1 dataset (1901–2009) at 0.5° (∼55.5 km) spatial resolution [26]. The ENSO index used here is the Southern Oscillation Index (SOI), air pressure difference calculated between Tahiti and Darwin, Australia [27]. Monthly SOI was acquired from the NOAA Climate Prediction Center (CPC) [28] via KNMI Climate Explorer.

We fitted a Poisson log-linear regression model to the time series of monthly incident counts using the R open source software environment (www.r-project.org). We conducted the analysis in 3 phases:

Since it is well established that for JEV transmission there is a lag effect of meteorological variables prior to initiation of pathogen transmission [6]–[9], the initial phase in our analysis was to correlate JE cases with climate variables and identify temporal lags. We explored our data using Pearson's correlation coefficient. Since JEV transmission is thought to occur in Malaysia year-round in epizootic cycles [5], we assessed lag times up to 12 months.

In the second phase of our study, the monthly number of serologically confirmed JE cases was modelled using a Poisson regression model. There was no over-dispersion in our analysis; therefore, we found no need to set the scale parameter to a Pearson χ² statistic. In our model, we adjusted for seasonality, trend effects associated with JE cases and lag effects of the meteorological variables, and used time series diagnostics (autocorrelation of standardised residuals) to check the assumption of uncorrelated residual variation.

We examined the relationships between the JE cases, vaccine, meteorological variables and a linear time trend. Fourier terms up to the second harmonic were included in the model to account for seasonality in JE cases. An assumed full model that adjusts for the effects of all the explanatory variables is:where the annual periodicity in JE cases, t is an index of time given as a sequence of numbers (i.e. 1,2,3…‥n), X_t is some covariate process (i.e. vaccine, temperature, rainfall and SOI) with associated temporal lags.

To select and retain variables that significantly contributed to the improvement of the model fit; we used a stepwise approach using Akaike's Information Criterion (AIC) as a measure of relative goodness of fit, where smaller values represents better fits. Risk ratios (RRs) derived from the estimated regression parameters from the final model illustrate the associations between the meteorological determinants and confirmed JE cases. We attempted to identify statistical interactions with all combination of seasonality, trend, vaccine coverage and meteorological variables, but none of the combinations were significant. We also determined the adjusted McFadden's pseudo R² given by:where is the sample size and is the number of parameters.

We used the monthly data from the 10-year surveillance period in Sarawak from April 1997 to December 2006 where suspected JE cases were serologically confirmed as JE positive or negative. This dataset was used to test the association between JE cases, vaccine and meteorological variables in Sarawak and develop our model. We lagged vaccine coverage by two months to account for the interval between the first and second vaccination. This made clinical and immunological sense since the protection the vaccine provides does not reach clinically acceptable protective efficacy until the first two doses (given 1-month apart) are provided (Centers for Disease Control and Prevention 1993). Though we had three temperature variables (mean, maximum and minimum), we only used minimum temperature in the final model since it had the strongest fit and including mean and maximum temperature would be redundant. Each lagged variable of SOI was individually and iteratively assessed with temperature, rainfall and vaccine and the final model was selected based on the models with lower AIC value. We omitted the trend (i.e. timeindex) from the model because the downward trend in JE cases over the 10-year study period is likely attributed to the vaccination program. Also, vaccine coverage year and trend (i.e. timeindex) were found to be moderately correlated with each other (Pearson correlation coefficient = 0.859, p<0.001); therefore, vaccine coverage year was retained in the model, while trend was removed and not considered further.

We developed two models: a vaccine only model (vaccine and seasonality) and an observed climate and vaccine model (vaccine, seasonality, minimum temperature, rainfall and SOI).

In the third and final phase, we used the observed climate and vaccine model developed in the second analysis phase to predict the number of JE cases that would have been seen if there was no vaccine coverage throughout the 10-year surveillance and if there was vaccine coverage throughout the 10-year surveillance period.

We also conducted sensitivity analysis using different vaccine effectiveness values from 0 to 100% to see the range in risk reduction for different models. We used models with the variables: i) vaccine only, and; ii) vaccine, minimum temperature, rainfall and SOI. A study of children in Thailand vaccinated with two doses of the formalin-inactivated mouse derived JEV vaccine (Biken, Japan) showed that the vaccine had a 91% vaccine efficacy [29], [30] ; therefore, we assumed that the vaccine strategy being used in Malaysia had at least a 91% effectiveness for the target population, Sarawak children ≤12 years of age.

One-hundred-and-thirty-three serologically confirmed JE cases were identified from April 1997 to December 2006 (10-year period), 84 before and 49 after the vaccine period of July 2001. There were relatively high levels of JEV transmission for four seasons prior to the incorporation of JE vaccine into the expanded programme of immunization (EPI) in the state of Sarawak in July 2001 (Figure 1A). After the vaccination program, JE cases dropped substantially (Figure 1A). The majority of JE cases generally occurred in the 4^th quarter of the year (October to December) (Figure 1B).

There was temporal variation in climate variables and JE cases (see Supplemental Material, Figure S2). The mean, maximum and minimum temperatures were correlated with JE cases at 5 to 7-months lag; minimum temperature at 6-month lag showed the strongest correlation with JE cases; and rainfall was only correlated with JE cases at 1-month lag (see Supplemental Material, Table S1).

We selected vaccine at 2-months lag, minimum temperature at 6-months lag, rainfall a 1-month lag to be used in further analysis since they provided a better fit than other variables and lags (data not shown). Due to the persistence of the ENSO, where its impact on climate [31], and subsequently disease risk [32], can be felt for a whole year in the tropics, we iteratively assessed up to 12 lags of SOI variables. SOI at 6-months lag were used in the final model.

Table 1 shows the association between vaccine at 2-months lag and JE cases after controlling for seasonality (vaccine only model); Table 2 shows the association between vaccine coverage at 2-months lag, minimum temperature at 6-months lag, rainfall at 1-month lag, and SOI at lag 6-month lag, and JE cases after controlling for seasonality (observed climate and vaccine model). All variables were significantly associated with JE cases in the different models. The observed climate and vaccine model had a higher adjusted pseudo-R² and lower Akaike information criterion (AIC = 300.196, adjusted pseudo-R² = 0.377) than the vaccine only model (Akaike information criterion = 317.398, adjusted pseudo-R² = 0.278) suggesting the former is a “better” model. There was no significant autocorrelation in JE cases. Trend was not significant when included in the model with vaccine.

No interaction effects amongst the meteorological variables or vaccine variable were observed in our model suggesting no differential response of JE cases to the aforementioned variables.

For the vaccine only model, there seems to be an approximately 61% decrease in JE cases after the implementation vaccination programme (post July 2001) relative to prior the implementation (pre July 2001). In contrast, for the observed climate and vaccine models, the estimated decrease is approximately 45%.

The observed climate and vaccine models also suggest that a 1-unit increase in minimum temperature at a 6-month lag, rainfall at 1-month lag, and SOI at 6-month lag may be associated with approximately 100% (2-fold), 3% and 40% increases in risk of JE in Sarawak respectively. To better understand the effect of the scale of variability in temperature, rainfall and SOI on JE cases we also calculated the relative risk using the difference between the lower and upper quartiles of the explanatory variable. There was on average a 41%, 53% and 66% increase in JE risk for the upper-lower quartile difference for temperature, rainfall and SOI, respectively (see Supplemental Material, Table S2).

The observed climate and vaccine model performed relatively well with a pseudo-R² value of 0.370 suggesting an improvement 37.0% in the fitted model from the null (intercept only) model (Figure 2). Autocorrelation tests undertaken on the residuals of the fitted model to determine the goodness-of-fit showed no significant autocorrelation at different lags (see Supplemental Material, Figure S3).

Table 3 shows the sum of the observed and predicted JE cases modelled with and without vaccination over the 10-year surveillance period. The observed and predicted JE case sums were similar with the model slightly over-predicting the pre-vaccine period and under predicting the post-vaccine period and the total. When the model assumed no vaccine coverage over the 10-year surveillance period, 175 JE cases were predicted; when the model assumed vaccine coverage throughout the 10-year surveillance period, only 90 cases were predicted (Table 3). Figure 3 shows the time series of observed and predicted JE cases modelled with and without vaccination over the 10-year surveillance period.

Figure 4 shows the JE risk reduction at different vaccine efficacies (VE) for the vaccine only model and observed climate and vaccine model. In the vaccine only model, the sensitivity analysis shows that ranges in the risk reduction of JE cases due to the vaccine programme is approximately between 45 and 65% for vaccine efficacies ranging between 60 and 100% under the current vaccine strategy in Sarawak. However, the observed climate and vaccine model shows a lower risk reduction of JE cases due to the vaccine programme. The observed climate and vaccine model with SOI shows a risk reduction of approximately 30 to 45% for vaccine efficacies ranging between 60 and 100%.

Minimum temperature, rainfall and SOI were the key meteorological determinants associated with JE cases in Sarawak. These variables may be useful for developing an early-warning system based on seasonal climatic forecasting to temporally target vaccination programs to reduce the associated cost of vaccine distribution as well as scaling-up the current vaccine strategy to include un-prioritized or underserved populations. However, the mechanisms by which these variables impact JEV transmission is not entirely clear and require further investigation for development of a potential forecast system. For diseases like JE, vaccination for mitigation of disease cases should always be a top priority. However, to increase estimation accuracy of the vaccine effect, climate should be considered in epidemiological vaccine evaluations to determine the un-confounded impact of vaccination programs.