Potential for using climate forecasts in spatio-temporal prediction of dengue fever incidence in Malaysia

Abstract

Dengue fever is a viral infection transmitted by the bite of female \textit{Aedes aegypti} mosquitoes. It is estimated that nearly 40\% of the world's population is now at risk from Dengue in over 100 endemic countries including Malaysia. Several studies in various countries in recent years have identified statistically significant links between Dengue incidence and climatic factors. There has been relatively little work on this issue in Malaysia, particularly on a national scale. This study attempts to fill that gap. The primary research question is `to what extent can climate variables be used to assist predictions of dengue fever incidence in Malaysia?'. The study proposes a potential framework of modelling spatio-temporal variation in dengue risk on a national scale in Malaysia using both climate and non-climate information.

Early chapters set the scene by discussing Malaysia and Climate in Malaysia and reviewing previous work on dengue fever and dengue fever in Malaysia. Subsequent chapters focus on the analysis and modelling of annual dengue incidence rate (DIR) for the twelve states of Peninsular Malaysia for the period 1991 to 2009 and monthly DIR for the same states in the period 2001 to 2009.

Exploratory analyses are presented which suggest possible relationships between annual and monthly DIR and climate and other factors. The variables that were considered included annual trend, in year seasonal effects, population, population density and lagged dengue incidence rate as well as climate factors such as average rainfall and temperature, number of rainy days, ENSO and lagged values of these climate variables. Findings include evidence of an increasing annual trend in DIR in all states of Malaysia and a strong in-year seasonal cycle in DIR with possible differences in this cycle in different geographical regions of Malaysia. High population density is found to be positively related to monthly DIR as is the DIR in the immediately preceding months. Relationships between monthly DIR and climate variables are generally quite weak, nevertheless some relationships may be able to be usefully incorporated into predictive models. These include average temperature and rainfall, number of rainy days and ENSO. However lagged values of these variables need to be considered for up to 6 months in the case of ENSO and from 1-3 months in the case of other variables.

These exploratory findings are then more formally investigated using a framework where dengue counts are modelled using a negative binomial generalised linear model (GLM) with a population offset. This is subsequently extended to a negative binomial generalised additive model (GAM) which is able to deal more flexibly with non-linear relationships between the response and certain of the explanatory variables. The model successfully accounts for the large amount of overdispersion found in the observed dengue counts. Results indicated that there are statistically significant relationships with both climate and non-climate covariates using this modelling framework. More specifically, smooth functions of year and month differentiated by geographical areas of the country are significant in the model to allow for seasonality and annual trend. Other significant covariates included were mean rainfall at lag zero month and lag 3 months, mean temperature at lag zero month and lag 1 month, number of rainy days at lag zero month and lag 3 months, sea surface temperature at lag 6 months, interaction between mean temperature at lag 1 month and sea surface temperature at lag 6 months, dengue incidence rate at lag 3 months and population density.

Three final competing models were selected as potential candidates upon which an early warning system for dengue in Malaysia might be able to be developed. The model fits for the whole data set were compared using simulation experiments to allow for both parameter and negative binomial model uncertainty and a single model preferred from the three models was identified. The `out of sample' predictive performance of this model was then compared and contrasted for different lead times by fitting the model to the first 7 years of the 9 years monthly data set covering 2001-2009 and then analysing predictions for the subsequent 2 years for lead time of 3, 6 12 and 24 months. Again simulation experiments were conducted to allow for both parameter and model uncertainty. Results were mixed. There does seem to be predictive potential for lead times of up to six months from the model in areas outside of the highly urbanised South Western states of Kuala Lumpur and Selangor and such a model may therefore possibly be useful as a basis for developing early warning systems for those areas. However, none of the models developed work well for Kuala Lumpur and Selangor where there are clearly more complex localised influences involved which need further study.

This study is one of the first to look at potential climatic influences on dengue incidence on a nationwide scale in Malaysia. It is also one of the few studies worldwide to explore the use of generalised additive models in the spatio-temporal modelling of dengue incidence. Although, the results of the study show a mixed picture, hopefully the framework developed will be able to be used as a starting point to investigate further if climate information can valuably be incorporated in an early warning system for dengue in Malaysia.