建立以氣候資訊為基礎的台灣登革流行預測模式：1998-2007年登革流行病學和氣象學因子之時序分析

(1)

i

國立臺灣大學公共衛生學院流行病學研究所博士論文

Graduate Institute of Epidemiology College of Public Health

National Taiwan University Doctoral Dissertation

建立以氣候資訊為基礎的台灣登革流行預測模式：

1998-2007 年登革流行病學和氣象學因子之時序分析 Establishment of a Better Prediction System for Dengue

Epidemics in Taiwan: Temporal Analyses of Epidemiological and Meteorological Factors of Dengue during 1998-2007

尚君璽

Chuin-Shee Shang

指導教授：金傳春博士、方啟泰博士 Advisors: Chwan-Chuen King, Dr. PH

Chi-Tai Fang, MD, PhD

中華民國 99 年 6 月 June, 2010

(2)

誌謝

感謝天父讓我有幸進入博士班就讀，並在八年之內陸續完成了結婚、生育

子女、畢業等大事。一路走來，酸甜苦辣兼而有之，感謝家人、朋友及教會的

支持，特別是外子詠銘的幫助、陪伴，以及父母的栽培與鼓勵；感謝已經在天

上的外婆、祖母的養育之恩，讓我可以也成為照顧別人的人。也感謝過去各個

求學階段的恩師益友，這些教導與砥礪為我日後的研究生涯奠定了重要的基礎。

謝謝金傳春老師將我帶進傳染病的領域，讓我一窺學術殿堂的奧秘；謝謝

方老師費盡心思且無私無我的指導，為人處事及鑽研學問的嚴謹態度都是我的

好榜樣，幫助我可以將畢業論文順利發表於口碑極佳的國際期刊；謝謝媲美智

多星的柳中明老師，常常迅速提供氣象學的豐富資訊及各種疑難排解，總是從

容、溫和又充滿智慧；謝謝張念台老師這幾年來在研究計畫上提供的寶貴指導

與協助，對於提攜後進也是不遺餘力；謝謝梁賡義老師在口試階段的拔刀相助，

百忙之中仍盡力給予詳盡的指導，大師風範令人如沐春風。

謝謝台大提供了絕佳的環境，讓一切的努力與創意都得到百分之百的支

持；在這裡當學生的時間已有十四年，我深深以此身分為榮。古云：登高必自

卑，行遠必自邇；正如今年畢業典禮的主題：千里之行，始於足下。回首過往、

展望未來，我僅以感恩的心情向前邁步，認真迎接另一個精采生涯的開始。

(3)

iii

摘要

受登革病毒感染的旅客經常成為散播病毒至其他地區的重要途徑，甚而引

發他國的流行；然而這些感染的旅客入境後和當地氣候、病媒以及本地疫情之

間的互動關係並不清楚。由於境外移入病例與本地登革疫情關係的議題長期受

到忽略，本研究即在探究臺灣地區的境外移入登革病例和氣候因子，對於本地

疫情發生之影響，進而可讓基層與中央衛生人員將危險層級資訊馬上應用於疾

病防控策略。

我們使用羅吉斯（logistic）和普瓦松（Poisson）迴歸模式分析 1998 至 2007

年間臺灣南部地區經實驗室診斷證實的登革確定病例，以區辨在氣候因子的作

用下，境外移入和本地登革病例的時序相關性。結果發現本地登革疫情的發生

與境外移入病例數（2 至 14 週前）、高溫（6 至 14 週前）及低濕度（6 至 20 週

前）之間，具有延遲的相關性。此外，境外移入病例數和本地登革病例數僅在

流行被引發的「初期」階段，才有明顯數量上的相關性；一旦流行持續發生，

此種關係即不復見。另外，根據單變項分析的結論，挑選出的重要氣候因子所

建立的羅吉斯迴歸模式，可以進一步建立登革危險性指標的預警數值。經初步

運算，此法優於僅以前期登革危險性指標(dengue risk index, DRI)值預測本期 DRI

值的效果，在加入氣候因子後，模式預測值與觀察值的相關係數(Spearman

(4)

correlation)可達 0.71(屏東地區)、0.84(高雄地區)以及 0.86(台南地區)；至於泰國

地區也有 0.66-0.77 的水準。由此可知此法不但適用於台灣登革疫情層級之預

警，也適用於泰國海岸五省的疫情預測。

這些發現顯示，惟有氣象條件適宜時，境外移入登革病例才有可能引發本

地的疫情。據此，經由境外移入病例的快速實驗診斷、早期發現以及管理，可

以遏止其後大規模登革/登革出血熱流行的發生。因此整合氣象資訊的早期警示

監測系統，將是登革疫情尚未成為地方性流行的地區用以成功防治疫情的無價

利器。

關鍵字：登革熱/登革出血熱、本土流行、境外移入病例、氣象、迴歸模式、登

革危險性指標

(5)

v

Abstract

Travelers who acquire dengue infection are often routes for virus transmission to

other regions. Nevertheless, the interplay between infected travelers, climate, vectors,

and indigenous dengue incidence remains unclear. The role of foreign-origin cases on

local dengue epidemics has thus been largely neglected by research. This study

investigated the effect of both imported dengue and local meteorological factors on

the occurrence of indigenous dengue in Taiwan.

Using logistic and Poisson regression models, we analyzed bi-weekly,

laboratory-confirmed dengue cases at their onset dates of illness from 1998 to 2007 to

identify correlations between indigenous dengue and imported dengue cases (in the

context of local meteorological factors) across different time lags. Our results

revealed that the occurrence of indigenous dengue was significantly correlated with

temporally-lagged cases of imported dengue (2-14 weeks), higher temperatures (6-14

weeks), and lower relative humidity (6-20 weeks). In addition, imported and

indigenous dengue cases had a significant quantitative relationship in the onset of

local epidemics. However, this relationship became less significant once indigenous

epidemics progressed past the initial stage. Polytomous Logistic regression model

(6)

with relevant meteorological variables stepwise selected were able to fitting the value

of Dengue Risk Index (DRI) in both Taiwan and Thailand. The Spearman correlation

between observed DRI and model-expected DRI ranged from 0.71 to 0.86 in Taiwan

and from 0.66 to 0.77 in Thailand, respectively.

These findings imply that imported dengue cases are able to initiate indigenous

epidemics when appropriate weather conditions are present. Early detection and case

management of imported cases through rapid diagnosis may avert large-scale

epidemics of dengue/dengue hemorrhagic fever. The potential application of DRI

with meteorological modeling in both Taiwan and Thailand demonstrated that its

feasibility to be extended to other countries for the current important issue on global

warming and dengue. The deployment of an early-warning surveillance system, with

the capacity to integrate meteorological data, will be an invaluable tool for successful

prevention and control of dengue, particularly in non-endemic countries.

Keywords: dengue, meteorology, climate, weather, regression model, Dengue Risk

Index

(7)

vii

List of Tables

Table 1. Abbreviations of variables

Table 2. The definitions of risk category for dengue epidemics

Table 3. The monthly frequency of dengue risk category (DRI) in Taiwan and

Thailand

Table 4. Area-specific polytomous logistic regression models based on local

meteorology for predicting dengue risk index (DRI) in Taiwan from 1998 to

2007

Table 5. The consistence of DRI between observed and expected values for dengue

epidemics in Taiwan from 1998 to 2008

Table 6. Area-specific polytomous logistic regression models based on local

meteorology for predicting dengue risk index (DRI) in Thailand from 1996

to 2004

Table 7. The consistence of DRI between observed and expected values for dengue

epidemics in Thailand from 1996 to 2005

(10)

List of Figures

Figure 1. The rationale of local climate and imported effects on indigenous dengue

epidemics

Figure 2. Numbers of laboratory-confirmed dengue (including dengue and dengue

hemorrhagic fever) cases in Taiwan, 1998-2007.

Figure 3. Correlation between bi-weekly “occurrence” of indigenous dengue cases

and 1) the number of imported cases as well as 2)meteorological variables

across time lags from 1 to 12 bi-weeks.

Figure 4. Correlation between bi-weekly “increase” of indigenous dengue cases and

1) the number of imported cases as well as 2) meteorological variables

across time lags from 1 to 12 bi-weeks.

Figure 5. Correlation between bi-weekly number of indigenous dengue cases and 1)

the number of imported cases as well as 2) meteorological variables

across time lags from 1 to 12 bi-weeks.

Figure 6. Correlation between bi-weekly number of indigenous dengue cases and

the number of imported cases over time lags from 1 to 12 bi-weeks.

Figure 7. Fitting the number of dengue cases with meteorological regression models

Figure 8. Distribution of monthly Dengue Risk Index in Taiwan and Thailand,

(11)

xi

1998-2005

Figure 9. Fitting (1998-2007) and predicting (2008, after red dash line) DRI in

Taiwan with meteorological regression model

Figure 10. Fitting (1996-2004) and predicting (2005, after red dash line) DRI in

Thailand with meteorological regression model

(12)

Chapter 1 Introduction

1.1 The impact of climate on dengue epidemics

Dengue/dengue hemorrhagic fever is the world’s most widely spread mosquito-borne

arboviral disease and threatens more than two thirds of the world’s population. Cases are

mainly distributed in tropical and subtropical areas, between 23.5^oN and 23.5^oS, in accordance

with vector habitats for Aedes aegypti and Ae. albopictus. Climate is believed to have complex

and long-lasting effects on the epidemics of infectious diseases, especially vector-borne

diseases [1-4]. The meteorological factors interact with human and mosquito vector, and affect

the occurrence and distribution of diseases ecologically [5-8] through both infectious agents

and arthropod vectors biologically [9,10].

The possible role of international travel in cross-country and cross-continent transmission

dengue has been identified an increasing public concern in recent years [11-13]. More and

more infectious travellers carry viruses back to home countries, and disperse the virus within

the network established by indigenous populations of both mosquito vectors and human hosts.

Thousands of international travelers infected dengue viruses in endemic areas, where a disease

occurs continuously and with predictable regularity in a specific area or population

(http://www.cdc.gov/ncidod/dvbid/dengue/index.htm), have been speculated as an important

pathway in transmitting the disease into non-endemic areas [14-17]. Potasman et al. have

(13)

2

discovered that Israeli travellers were at higher risk for acquiring dengue infection when they

arrived in tropical countries in summer [18]. Seasonality of imported dengue is therefore

supposed to have potentially complex and long-lasting effects on the epidemics of dengue [13].

Nonetheless, some of the previous studies even underestimate or neglected the effects that

imported cases may have on the occurrence of epidemics [19,20]. The interaction between

seasonality and imported dengue is reasonably to play an important role in the occurrence of

indigenous cases, and thus worthy of detailed investigation.

1.2 Unsolved questions

Up to now, the effects of regional weather may have on vector-borne infectious diseases

are not well-understood. The most frequently studied meteorological factors have been

temperatures and rainfall [1,21-23]. However, many other meteorological factors have been

mostly neglected. The time lag between those meteorological factors and the incidence of a

studied disease has generally be overlooked or underestimated in investigating their real

correlations. How to use the best model to consider all important meteorological factors with

time lags and unique epidemiological characteristics of dengue in Taiwan becomes very

important.

(14)

Chapter 2 Literature review

2.1 Global status of dengue

Dengue fever is caused by dengue viruses (four serotypes), classified as flaviviridae.

Both the incidence and epidemic areas have been increasing in the past four decades. About

50% of population in the world is under the threat of dengue infection and 100 million cases

occur annually [24]. Its potential severer complications, including dengue hemorrhagic fever

(DHF) and dengue shock syndromes (DSS), result in the leading cause of death in children in

endemic areas [25,26]. Historically, after the first official case report of dengue in the 18^th

century, the major epidemics occurred mainly in Asia, especially in South East Asia, such as

Thailand, the Philippines, and Indonesia. Nevertheless, dengue epidemics have also been

frequent in the Pacific islands, South Asia, Australia, Africa, Latin America, etc. Even high

latitude areas are also involved through frequent and instant international transportation

[27,28].

2.2 Dengue in Taiwan

As early in 1870s, there was reported dengue-like disease. Sporadic cases occurred

mainly in Kaohsiung area every two or three years. The disease had not been reoported after

1940s. Dengue reappeared in Pingtung area (Liuchiu islets) in 1981 and Kaohsiung area in

(15)

4

1987 with thousands of cases. Hereafter, there are epidemics every year in southern Taiwan.

After 1990, Taichung (1995) and Taipei (1996, 2007) have been also involved in indigenous

dengue.

2.2.1 Surveillance of dengue in Taiwan

Surveillance of dengue in Taiwan is made up of three parts: passive reporting of

dengue, including dengue fever (DF) and dengue hemorrhagic fever (DHF) within 24 hour,

active and semi-active surveillance. In passive surveillance, dengue-like illness reports by

health care workers to local health authorities account for most confirmed dengue cases.

Active surveillance, including volunteer reporting and fever screenings at international

airports (identifying fever cases by infrared thermal scanner, which has been routinely

operated by the government since 2003) [29,30]. In semi-active surveillance, fever cases

are investigated in residential areas, schools, and work places with epidemiological linkage,

and specimens are taken once confirmed dengue cases are identified. These active and

semi-active components, serve to complement and reinforce in support of comprehensive

virus detection. Among active strategies, fever screening detects imported dengue cases

efficiently [30]. All febrile patients identified through fever screening are required to

submit blood samples for testing. In addition, public health professionals at local health

authorities monitor suspected cases for the development of dengue-like symptoms/signs

(16)

until dengue virus infection is excluded [31]. These strategies identify and manage

potential dengue cases before they enter into the community. In other words, Taiwan’s

dengue surveillance involves serologically detected mild and asymptomatic dengue cases

that are quite different from many dengue-endemic countries where most of severe DHF

cases are reported with less involving mild dengue cases. Under these circumstances, the

true relationship between meteorological factors and dengue can be appropriately under

investigation.

2.2.2 Case definition of dengue

The current definitions for dengue, including DF, DHF and dengue shock syndrome

(DSS) in Taiwan have been applied since the 1980s. Cases of “probable DF” are patients

with body temperatures >38^oC and two or more of the following clinical manifestations:

headache, retro-orbital pain, myalgia, arthralgia, rash, hemorrhagic manifestations and

leucopenia. Cases of “probable DHF” and DSS are further identified based on criteria

established by the World Health Organization [32]. Identified probable dengue cases must

provide blood specimens for laboratory confirmation tests. These laboratory tests include

molecular identification of dengue virus by reverse-transcriptase polymerase chain reaction

(RT-PCR) [33], single or paired serum samples testing for dengue-specific IgM

seropositives, 4-fold dengue-specific IgG serotiter rises (with the exclusion of Japanese

(17)

6

encephalitis virus infection) [32], or virus isolation [34,35].

2.2.3 Definition of imported dengue cases

Epidemiological questions such as travel history, incubation period, and first day of

illness were evaluated to identify the possible origin of dengue infection. “Imported

dengue cases into Taiwan” were defined as laboratory-confirmed dengue cases with travel

history to endemic countries within 14 days before the date of onset of dengue (based on

Taiwan-CDC’s definition) [31].

2.2.4 Serotype of dengue viruses in major epidemics

According to the reports of Taiwan-CDC, the major epidemics areas located in

southern Taiwan: TN, KH and PT [31]. Information on the predominant serotype of

isolated dengue viruses (Figure 1C) showed that the dominant serotypes/genotypes of

epidemic DENV varied by year and area [30]. However, in 2002, a DENV-2 epidemic

attacked all these three areas. During our study period, KH had the most frequent

occurrence of dengue epidemics, with epidemics of DENV-2 in 1998 and 2001-2003;

DENV-1 in 2004; DENV-3 in 2006, and DENV-1 in 2007. TN had four major epidemics,

including DENV-3 in 1998 [36], DENV-4 in 2000, DENV-2 in 2002, and DENV-1 in

2007. PT had two major epidemics, including DENV-2 in 2001-2003 and

(18)

DENV-1/DENV-4 in 2004. Interestingly, we found that local dengue epidemics, with

geographical variations in these three areas, only had higher numbers of indigenous cases

during certain years.

2.3 The impact of imported cases

2.3.1 International transmission of dengue

Dengue outbreaks initiated by international tourists, immigrants, and foreign workers

have been reported in numerous developed areas and countries [11-13]. Not only

tropical/subtropical areas, but distant areas/countries are also under the threat of dengue

through international tourists, immigrants, foreign workers, and militaries [16,27].

Developed areas and countries have reported outbreaks in the recent decade, including

Hawaii [37] and Texas [38] in the United States, France[17], and Germany [15].

Nevertheless, the potential impact from the complicated interaction among infected

travellers, climate, vectors and incidence is unclear [13,15].

2.3.2 Correlation between season and imported dengue cases

Historically, the link between imported cases and indigenous cases has been

established through phylogenetical analysis and viral sequence comparisons [29,30].

(19)

8

However, these retrospective studies are not capable of providing timely, relevant

information about transmission dynamics, nor do they provide quantitative insight for

disease control strategies in a broader context. For example, epidemiological data has

indicated that imported dengue cases enter Taiwan almost every month from other

countries (Figure 1B) but have not always resulted in local dengue epidemics [18,30]. This

suggests that the timing of imported dengue’s entrance may have considerable effect on

domestic dengue epidemics [13,16]. However, the role of these foreign-origin cases in

local dengue epidemics has not yet been quantitatively assessed [20].

2.4 Meteorological effects

Climate has been identified to be correlated with disease incidence and thus may be

usable to predict the occurrence of infectious diseases, especially for those vector-borne ones

which are sensitive to seasonality [4,13,39,40]. Higher temperature is believed to increase the

occurrence of dengue because of it shortens the time of virus replication in mosquito vectors

(known as extrinsic incubation period, EIP) [10]. Warmer condition may also reduce the size of

adult mosquito with higher metabolism rates, and thus require more blood meal through higher

biting rate [41]. Rainfall related factors are also considered positively correlated with the

incidence of dengue by increasing the breeding sites of mosquito vectors [1,21,39]. Rainy days

(20)

have been reported significantly associated with DHF incidence in Thailand [42]. Extremely

heavy rains was thought to flush away mosquito eggs, larvae and pupae from the breeding sites

[42]. Nevertheless, a recent study with more precise experiments concluded that the larva and

pupa of Aedes aegypti populations, comparing with Culex pipiens, were slightly affected by

excessive rain [43]

.

They found that the fourth instars of Ae. aegypti were not affected by

flushing when exposed for longer rain intervals (30 versus 60 min) or at a colder water

temperature (24 versus 16°C). This difference was most dramatic during the pupal stage.

Higher relative humidity has also been reported correlated with higher mosquito density and

higher vector capacity [44,45] (e.g. “the daffy rate at which future inoculations arise from a

currently infective case, provided that all female adults (the original text: flies) biting that case

become infected” [46]), and thus used for the projection of global dengue [2].

2.5 Modeling for dengue incidence

In order to control the disease, recent studies attempt to predict dengue incidence with

regression models. ARIMA (Autoregressive Integrated Moving Average) has been popular in

time series fitted models designed to simulate or predict the number or incidence of dengue

with meteorological information [5,19,20,47,48]. ARIMA models are based on linear

regression, assuming that dengue incidence is under normal distribution. Although some

(21)

10

previous studies fit the incidence pretty well, the disease data is actually count in its nature and

thus need more reasonable assumptions.

2.6 Potential usage of dengue risk index (DRI)

2.6.1 A need for an understandable index

A globally understandable index for epidemics or risk of dengue is needed. Currently,

most countries in the world have individual or international surveillance systems for

dengue with a focus mainly on the total number of cases (e.g. DengueNet of World Health

Organization). If we are able to integrate their information about dengue epidemics with an

indexing system, each country or area may have its specific value within certain period.

This index will be able to alert people before they enter an endemic area of dengue, as well

as governments once they have travelers from high risky countries. One example of

international-accepted indexing system is the ultraviolet index (UVI), which transforms the

daily maximum UVB flux into a digital value, named as UVI [49], and then assigned it

into one of five interpretable ordinal categories: low, moderate, high, very high and

extreme. Another commonly used indexing system is the Pollutant standard index (PSI),

which helps one to be alert on the air pollution condition where he lives or he will visit.

Although the importance of imported cases to local dengue epidemics has been reported

(22)

worldwide [12,27], unfortunately, for the travelers, there has been no advice from a system

to quantify the risk of their destinations. Human movements are identified to have impact

on spreading of dengue rather than solely mosquito dispersing [35,50]. The relevance of

different spatial distances are from household, neighborhood, regional to international

spreading [51]. Most strategies for dengue control are limited within a country or an area,

such as insecticide spraying and case notification, however, the management and

prevention for global transmission is lacked.

2.6.2 Categorized number of dengue cases

Previous studies on prediction models for dengue mainly focused on calculating the

number of dengue cases with regression models mostly [5,20,47]. However, this method is

not appropriate especially for those non-endemic areas of dengue because of the

over-dispersion (super high variance) among the numbers of indigenous dengue in different

seasons. On the other hand, it is not necessary for the public to realize the exact case number

if there is an epidemic. Instead, categorization of the case number of dengue into an

accessible risk index may not only enhance the efficiency of disease prediction, but also

become more practical for disease control and prevention.

In this study, we propose a Dengue Risk Index (DRI) by establishing polytomous

(multi-level) Logistic regression models based on the information about local

(23)

12

epidemiological data. We applied this concept to calculate area-specific DRI with real data,

and compared the expected values with observed ones in Taiwan and Thailand, two

countries with different status of dengue epidemics. In addition, a preliminary forecasting

method was also developed to illustrate how to implement DRI on dengue predicting.

(24)

Chapter 3 Objectives and Specific Aims

3.1 Objectives and specific aims

The study used data of all imported and indigenous dengue cases nationwide that had

been confirmed by the Taiwan-CDC) [31,52] to investigate the relationship between imported

and indigenous dengue, and all concurrent meteorological characteristics with potential for

facilitating disease transmission.

The specific aims of this study were:

1) To investigate the relationship between imported and indigenous dengue, and all concurrent

meteorological characteristics with potential for facilitating disease transmission.

2) To clarify the relationship between imported dengue, local weather, and domestic

epidemics of dengue, and to further identify the role of imported cases (in different phases)

during a dengue epidemic in non-endemic areas such as Taiwan.

To explore the potential usage of categorized case number, named as Dengue Risk Index,

in describing and analyzing the epidemiology of indigenous dengue outbreaks, and to further

establish a predicting system with succinct meteorological regression model.

3.2 Hypotheses

3.2.1 Specific meteorological variables correlated with more indigenous dengue cases

(25)

14

3.2.2 Imported dengue cases positively correlated with more indigenous dengue cases

3.2.3 Dengue risk index is able to reflect the number of confirmed/reported dengue cases

3.2.4 Previous dengue risk index correlate with current dengue risk index, thus is able to

predict

3.2.5 Specific meteorological variables selected in a regression model are able to fit the

distribution of dengue risk index

(26)

Chapter 4 Materials and methods

4.1 Study areas

Confirmed indigenous dengue cases in three epidemic areas in Southern Taiwan [Tainan

(TN), Kaohsiung (KH), and Pingtung (PT)] were investigated. All three areas had identified

both Aedes aegypti and Ae. albopictus mosquitoes as vectors for transmitting dengue virus. KH,

including both metropolitan Kaohsiung and Kaohsiung County (total area is 2,946.27 km², and

total population was 2,770,050 in year 2008), had served as the location for the majority of

Taiwan’s dengue epidemics involving all four serotypes of dengue viruses. Smaller scale

epidemics of dengue also occurred in both TN and PT, located adjacent to Kaohsiung. For this

study, TN included Tainan City and County (total area is 2,191.65 km², and total population

was 1,873,678 in year 2008), while PT referred to Pingtung City and County (total area is

2,775.60 km², and total population was 884,067 in year 2008). The subtropical climate of

southern Taiwan presents an annual hot and rainy summer season lasting from June to August

and daily mean temperatures ranging from 18^o to 32^oC year round.

4.2 Epidemiological data

Information on these confirmed cases of dengue fever (DF) and dengue hemorrhagic

fever (DHF) were obtained from Taiwan-CDC from 1998 to 2007 through dengue surveillance

(27)

16

in Taiwan. Date of onset of dengue illness, age, gender, clinical manifestations, reporting

hospital, and laboratory results were all thoroughly documented for each dengue case.

4.3 Meteorological data

We systematically collected daily weather data for Taiwan that was publicly available

through the 26 branch stations of the Central Weather Bureau (http://www.cwb.gov.tw/). The

meteorological variables analyzed in this study were selected after comprehensive evaluation of

all available data with biological relevance to vectors or cases of dengue, including daily mean

temperature, daily maximum temperature, daily minimum temperature, daily mean relative

humidity, daily mean wind speed, daily accumulative rainfall, daily accumulative rainy hours,

daily sunshine accumulation hours, daily mean sunshine rate (from sunrise to sunset), and daily

sunshine total flux. Unlike weather stations in Tainan and Kaohsiung, Pingtung County’s

station is located a far distance from Pingtung City, where the majority of Pingtung’s dengue

cases occurred. We therefore used weather data collected by the Environment Protecting

Agency (EPA) at their station in Pingtung City. This EPA weather station was only able to

provide data regarding daily mean temperature, daily maximum temperature, daily minimum

temperature, daily mean wind speed, and daily accumulative rainfall. We then substituted

Kaohsiung’s data for Pingtung’s meteorological variables not provided by the EPA because of

(28)

Pingtung City’s close proximity to Kaohsiung City.

4.4 Thailand’s data

Thailand has been an important origin country of dengue viruses that result in local

epidemics [30]. Datasets were available for related analyses through international cooperation.

Considering the similar meteorological and demographical situation of dengue to Taiwan, five

coastal provinces in Thailand were selected to compare with the model fitting in Taiwan. The

five provinces are (a) Phetchaburi, (b) Prachuap Khirikhan, (c) Chumphon, (d) Surat Thani and

(e) Nakhon Sitammarat. Since there were only reported dengue cases available in Thailand’s

data, we replaced confirmed dengue by reported dengue in Thai’s models. Both meteorological

and epidemiological data were kindly offered by Dr. Mathuros Tipayamongkholgul in Mahido

University, Thailand.

4.5 Statistical analyses

All laboratory-confirmed daily dengue cases, according to the date of onset of dengue

illness, were summed into total case numbers in bi-weekly intervals for data analysis. The mean

value of each meteorological variable was also calculated for each biweekly interval.

Abbreviations of all variables analyzed are listed in the Table. As the effects of imported

(29)

18

dengue and meteorological factors on indigenous dengue logically had a time lag, we thus

tested different time lags for each variable from lag 1 up to lag 12 (lag 1 representing two

weeks, lag 2 representing four weeks, and so on)

4.5.1 Logistic and Poisson regression models

Logistic regression was used to analyze the correlation between the

occurrence/increase of indigenous dengue and the number of imported cases, as well as the

correlation between the occurrence/increase of indigenous dengue and each meteorological

variable across various time lags (from 2 weeks to 24 weeks). Poisson regression was used

to analyze the correlation between the number of indigenous dengue cases and the number

of imported cases, as well as the correlation between the number of indigenous dengue

cases and quantitative data of each meteorological variable across time lags from 2 weeks

to 24 weeks. Regression with the negative binomial model [53] was used for

over-dispersed data. All models were adjusted by area (two dummy variables, area_KH

and area_TN), popd (area-specific population density), and sine24 plus cosine24 (the

oscillatory sine and cosine functions were used to model seasonal variations of dengue

cases [54] ).

4.5.2 Three phases in epidemics outbreak

Because the quantitative relationship between indigenous and imported dengue cases

(30)

may exist only at the onset of local dengue epidemics, we further divided all bi-week

intervals into three categories for further analysis: 1) Period of “low intensity

transmission”: From March to May during our study period, 94.44% (170/180) of bi-week

intervals during these three months had no indigenous dengue cases in these studied areas.

2) Period of “early phase of outbreaks”: Those bi-week intervals presenting <10

indigenous dengue cases for months excluding March to May. 3) Period of “late phase of

outbreaks”: Those bi-week intervals presenting ≧10 indigenous dengue cases. Further

information on these regression models are listed in the Appendix 1.

4.5.3 The contents of regression models

1. Modeling the temporal correlation between “Occurrence” of indigenous dengue and

variables: Occurrence (y/n) ~ sin24 + cos24 + two dummy variables for areas+

population density + tested lagged variable, link = logistic, family = binomial

2. Modeling the temporal correlation between “Increase” of indigenous dengue and

variables: Increase (y/n) ~ sin24 + cos24 + two dummy variables for areas+

population density + tested lagged variable, link = logistic, family = binomial

3. Modeling the temporal correlation between the number of indigenous dengue and

variables: Case (count) ~ sin24 + cos24 + two dummy variables for areas+ population

density + tested lagged variable, link = log-linear, family = negative binomial

(31)

20

4. Modeling the temporal correlation between the number of indigenous and imported

dengue in Period of “low intensity transmission” (Those bi-week intervals were from

March to May):

Case (count) in Period of “low intensity transmission”~ sin24 + cos24 + two dummy

variables for areas+ population density + tested lagged variable, link = log-linear,

family = Poisson

dengue in Period of “early phase of known outbreaks” (Those bi-week intervals

presenting <10 indigenous dengue cases for months excluded March to May):

Case (count) in Period of “early phase of known outbreaks”~ sin24 + cos24 + two

dummy variables for areas+ population density + tested lagged variable, link =

log-linear, family = Poisson

dengue in Period of “late phase of known outbreaks” (Those bi-week intervals

presenting≧10 indigenous dengue cases):

Case (count) in Period of “late phase of known outbreaks”~ sin24 + cos24 + two

dummy variables for areas+ population density + tested lagged variable, link =

log-linear, family = negative binomial

(32)

4.5.4 The definition of Dengue Risk Index (DRI)

Five categories (Table 2) were distinguished and identified by a digit number of 0, 1,

2, 3 and 4 respectively, representing different levels of indigenous dengue condition based

on the number of indigenous cases confirmed during a time interval, e.g. one month or two

weeks. The cut-off points of categories are 0, 10, 100 and 1000, which are not only

mathematically meaningful, but also epidemiologically. For most non-endemic countries

of dengue, there is usually zero indigenous cases, thus “0” is supposed to be a major group

for case number (color green). When there are cases no more than 10, it may just because

of sporadic diseases initiated by imported cases (color yellow). Furthermore, if case

number gets to between 10 and 99, it may represent a limited, local epidemic is happening

(color orange). An epidemic is obviously spreading if cases are over 100 (color red). The

situation of dengue is out of control when the number indigenous cases is higher than

1,000 (color purple).

4.5.5 Meteorological regression model

Statistical significant variable in univariate analyses were recruited into regression

model selection. Relevant variables selected by stepwise procedure were used to fit

observed and expected DRI, but also conduct out-of-fit to verify the prediction with

different datasets. In Taiwan’s part, these models were applied to calculate the expected

(33)

22

DRI for each bi-week interval during 1998 to 2007. Furthermore, out-of-fit of fitted

models were performed with data of year 2008 to predict the values of DRI. Similarly,

Thailand’s data from 1996 to 2004 were used to build up regression models, and year 2005

to predict DRI.

4.5.6 Comparison of different DRI forecasting methods

In order to validate the meteorological regression models in the use of predicting

DRI, we performed two other strategies for comparison by applying both Taiwan’s and

Thailand’s data from 1998 to 2005. As for forecasting, two classical approaches are

usually concerned. First is “continuous method”, which means predicting according to

what just happened. It is to move DRI curve with a certain lag forward as the prediction.

The second one, “historic method”, is predicting with the mean value of the past years. For

example, if we want to predict monthly DRI in 2001, the mean DRI of each month from

1991 to 2000 may be conducted to predict.

In order to testify the consistence between observed and expected values, Gamma,

Kendall’s Tau-b, Stuart’s Tau-c and Spearman’s rank correlation on ordinal data were

conducted. All formulas are adapted from User’s Manual of SAS and listed in Appendix 2.

Two-tailed p < 0.05 was regarded as statistically significant. The statistical analysis was

(34)

conducted using S-PLUS Enterprise Developer Version 8.0.4 (TIBCO Software Inc., Palo Alto,

CA, USA) and SAS 9.1.3 Service Pack 4 (SAS Institute Inc., Cary, NC, USA).

(35)

24

Chapter 5 Results

5.1 Logistic regression models for the occurrence and increase of indigenous dengue

cases: Univariate analyses

Figure 3 displays estimates of regression coefficients of independent variables (Xs) in the

logistical regression model for the “occurrence” of indigenous dengue cases. We found that the

variables of the number of imported cases (imported, p = 0.0023~0.0315) and daily

maximum/mean/minimum temperatures (tmax/tmean/tmin, p = 0.0002~0.0495) were positively

correlated. On the contrary, relative humidity (rh) was negatively correlated with indigenous

case occurrences (p <0.0001 ~ p = 0.0433). These findings indicate that an increase in imported

cases, in conjunction with warmer and drier weather, is favorable for the occurrence of

indigenous dengue. Among other meteorological variables, one sunshine related variable and

wind speed did not exhibit consistently significant relationships with indigenous dengue (data

not shown). However, Figure 4 reveals that, the influence of both imported cases and weather

conditions on the “increase” of indigenous dengue was less significant. In addition, when

binary outcomes were replaced with indigenous case counts (Figure 5), the quantitative

relationships between imported and indigenous dengue cases became insignificant.

(36)

5.2 Impact of imported dengue on indigenous dengue at three epidemic phases

In Figure 6, we observed variation in the impact of imported dengue at different epidemic

phases (please see definitions in Methods). Using Poisson models, we found that the imported

dengue cases were significantly correlated with indigenous dengue with lag 4 (i.e. 8 weeks)

only in periods of “low intensity transmission” (Figure 6A). However, this relationship was

more statistically significant in the “early phase of outbreaks” (Figure 6B). Imported dengue

had their greatest impact on epidemics during this phase. When local epidemics entered a

period of “late phase of outbreaks”, these correlations disappeared (Figure 6C), suggesting

that imported cases were unlikely to have influence on indigenous cases during this period.

These findings may indicate that imported dengue cases initiate local dengue cases almost

exclusively during the onset of an epidemic.

5.3 Poisson regression model fitted with meteorological variables of dengue cases

We attempted to fit the numbers of dengue cases with Poisson regression models, but it

did not show satisfactory results because the large variances of case numbers led to

over-dispersion.

5.3.1 Square root of case number

As shown in Figure 7A, although the square roots of case number replaced the case

(37)

26

numbers to lower the variance, the fitted case numbers are still out of expectation. The

same situation was also observed in using cubic root of case numbers.

5.3.2 Principal component analysis

To reduce the number of independent variables, we then compiled different time

lags of the same variable, i.e. factor of relative humidity (RH) may represent original rh

lag 1~8 with different weighted or adjusted process. Nevertheless, the fitting result for

the numbers of dengue cases was still poor (Figure 7B). In addition, the different lag of a

same variable is not independent one another, so the application of PCA is inappropriate.

5.3.3 Zero-inflated Poisson

In order to overcome the problem that zero dengue cases occurred over half of

time intervals, we used zero-inflated Poisson regression models to fit the numbers of

dengue cases. Unfortunately, the variation among numbers of dengue cases were too large

to fit Poisson distribution and the model fitting thus failed.

5.4 The distribution of monthly DRI

5.4.1 DRI in Taiwan and Thailand.

Figure 8A and 8B demonstrates monthly DRI from January 1998 to December

(38)

2005 in Taiwan and Thailand, respectively. (Taking the year of 1998 as an example, the

Taiwan’s health authority could be aware of the earlier peaking of DRI in Thailand. On the

contrary, if tourists from Thailand planned to visit Taiwan in 2002, they would have better

to be notified about the trend in increasing DRI in those months before August.

The temporal distributions of DRI in Taiwan and Thailand were very different. As

shown in Table 3A, the percentages of bi-weekly DRI (levels of 0, 1, 2, 3) in Taiwan

during 1998-2008 were 69.19% (548/792), 16.92% (134/792), 10.98% (87/792) and 2.90%

(23/792), respectively. The majority was level 0 (no confirmed indigenous dengue), and

most (58.18%, 64/110) of higher DRI (including level 2 and level 3) appeared in autumn

(September to November). On the other hand, the distribution of monthly DRI in Thailand,

listed in Table 3B was 1.33% (8/600), 12.67% (76/600), 66.50% (399/600), 18.83%

(113/600) and 0.67% (4/600) from 1996 to 2005. Level 2 was the most frequent level of

DRI, and higher DRIs were usually observed in summer (June to August). More frequent,

earlier and higher DRI in Thailand than in Taiwan actually reflect different endemicity of

dengue in the two countries. In other words, Thailand’s DRIs were earlier and higher than

Taiwan’s.

(39)

28

5.5 Correlation between monthly reported and confirmed DRI in Taiwan

The Pearson correlation coefficient between the number of confirmed and reported

dengue cases was 0.9833 (p<0.0001). Meanwhile, this Pearson correlation coefficient between

these numbers of confirmed dengue cases and DRI was 0.5770 (p<0.0001). On the other hand,

the Spearman correlation coefficient between these numbers of confirmed dengue cases and

DRIs was 0.9925 (p<0.0001), and such a coefficient between current DRI and previous DRI

was 0.7196 (p<0.0001). This analysis means that DRI is able to represent the tendency of

dengue incidence, as well as to be predicted by the DRI value of the previous month.

5.6 Comparison with different time unit of DRI

Daily levels of DRIs and the numbers of confirmed dengue cases were significant

correlated (R²=0.6780, p<0.0001), as well as biweekly DRIs and biweekly case numbers

(R²=0.9958, p<0.0001), and monthly DRIs and monthly case numbers (R²=0.9925, p<0.0001).

5.7 Predicting model of DRI

5.7.1 Meteorological models

5.7.1.1 Model fitting (1998-2007) and prediction (2008) in three areas in Taiwan

Table 4 shows area-specific regression models for indigenous DRI established

(40)

with local meteorological data during 1998 to 2007 in Taiwan. Selected weather

variables are temperatures and relative humidity with different time lag in different

areas. The DRI of previous bi-week explained most of correlation in models.

We examined the consistency comparison of observed and expected values of

DRI by area in Table 5A, and observed high concordance. The correlation diagnoses

in Table 5B support this finding. However, the harmony relationship disappeared in

area TN in 2008, the year after the ever largest epidemic of dengue in 2007. Figure 8

compares the two values with plotting curves, and the values of year 2008 after dash

line are out-of-fit. Most of time two curves go together, including most level zero and

infrequent level 3 in KH (2002, 2006) and TN (2007).

5.7.1.2 Model fitting (1996-2004) and prediction (2005) in five coastal provinces in

Thailand

Similar results were found in Thailand. In Table 6, most models include

relative humidity and temperatures in different time lags. The simplest model is area b,

province Prachuap Khirikhan, contains only daily mean temperature at lag 1 (tmean1).

The concordance between observed and expected values of DRI in Table 7 is

satisfying through areas during 1996 to 2004, while the predictability for DRI in year

2005 (Table 7B) is not stable in area a (province Phetchaburi) and e (province Nakhon

(41)

30

Sitammarat). The consistency between observed and expected category values is

varied among different provinces. Figure 9 compares the two values with plotting

curves, and the values of year 2008 after the red dash line are out-of-fit. The two

curves match most of time (Spearman correlation=0.66~0.77 in Table 7).

5.7.2 Comparison with different methods

In Table 7B and Table 5B, all fitted and predicted DRI were compared with the

correlation between current DRI and previous DRI (e.g. fitting with only previous DRI

and without meteorological variables). The results show that although the predictability

of DRI1 is good, the values fitted with regression models with meteorological variables

are better than the values of only previous DRI.

(42)

Chapter 6 Discussion

This study examined all laboratory-confirmed dengue cases detected through a

combination of active, semi-active, and passive surveillance, and found that imported dengue

are able to serve as an initial facilitator, or spark, for domestic epidemics. Nevertheless,

imported dengue cases do not have a noteworthy effect from March to May, during the low

transmission period of dengue in Taiwan. When these sparks meet suitable weather conditions,

the tinder, local dengue epidemics result. Eventually, this relationship does not persist once

biweekly indigenous case numbers rise over ten, indicating that a local epidemic has occurred.

Our findings thus provide evidence that a significant quantitative relationship between

Taiwan’s imported and indigenous dengue case numbers exists solely at the onset of an

epidemic and in the context of appropriate meteorological conditions. In addition, we also used

dengue risk index to describe the different endemicity of dengue in Taiwan and Thailand.

Although the DRI values in previous time intervals are able to fit current DRIs, we found

logistic regression models with meteorological variables improve the correlation between

observed and expected DRIs. The following discussion will focus mostly on the two major

parts of the results – the role of imported cases and meteorological factors and the dengue risk

index modeling.

(43)

32

6.1 The uniqueness of dengue analyses in Taiwan

To the best of our knowledge, this is the first study to simultaneously identify the

relationship between indigenous and imported dengue cases in the context of meteorological

factors. Our findings provide a highly accurate epidemiological portrait of dengue in Taiwan

because of the following components of the research: First, a better surveillance system was

instituted to actively rather than passively detect dengue cases. This system was also

laboratory-based to minimize confounding infection and manifestations [29,33,55,56]. Second,

we avoided a potential bias as a result of delays in dengue notification by analyzing all

confirmed dengue cases in accordance to their onset dates of illness rather than their reporting

dates. Third, all the dengue cases we analyzed are laboratory-confirmed plus the fact that fever

surveillance at airport to pick up mild cases and semi-active serological surveillance involving

even asymptomatically infected cases. All together makes our data source is closer to wide

spectrum of dengue infection, quite different from only severe dengue hemorrhagic cases are

required to be reported in many South East Asian countries where dengue has been endemic

and hyper-endemic.

(44)

6.2 The combined effect of imported dengue and meteorological factors

Because the numbers of imported dengue cases that initiate indigenous cases have been

increasing in non-endemic areas such as Taiwan [15,16,29,30] (further supported by the high

nucleotide identities of dengue viruses isolated from travelers with travel history to endemic

countries [29,30,33,36,57]), this study ventures to provide epidemiological evidence of the

combined impact of both imported dengue and weather conditions on local outbreaks.

Climate factors have provided helpful clues for monitoring dengue’s transmission in

affected areas [1,4,10,58]. Higher temperature has the effect of shortening the time intervals of

extrinsic incubation in the mosquito life cycle [10,59] and is positively correlated with more

occurrences of indigenous dengue in this study. This is consistent with previous findings that

demonstrate the suitability of warm or hot weather for the survivorship of adult mosquitoes and,

thus, dengue transmission [45,58]. In addition, although increased rainfall has been shown to

increase the number and quality of mosquito breeding sites (as well as the density of resting

sites) [1], lower rainfall and relative humidity (RH) were significantly related to indigenous

dengue in this study.

The correlation between lower RH and indigenous dengue with time lags was also

observed in previous studies in Thailand [60,61]. We explain this phenomenon as follows.

Drier conditions may facilitate dengue transmission through the increase of water storage

(45)

34

behavior, which result in an increase of breeding sites for Aedes mosquitoes, particularly in

areas without reliable water supplies [62-64]. Although piped water supply is available in 90%

of Taiwan (http://www.water.gov.tw/eng/08statistics/ sta_a_main.asp?bull_ id=4341), water

storage for gardening or agricultural use is popular during water restriction period in the dry

season, October to April, in southern Taiwan. In fact, a previous field survey identified water

buckets as the most common breeding sites of Ae. aegypti in southern Taiwan [65].

Entomologically, lower RH (50% vs. 90%) aids higher flight speed of female adult Ae. aegypti

at temperatures higher than 21 degrees of Celsius [66] thus facilitating dengue transmission.

This may explain why both RH and rainfall showed a negative correlation with the number of

indigenous dengue (Figure 4) and, that while higher temperatures occurred during July to

September in the summer of Taiwan, the number of indigenous dengue cases usually peak in

October-November.

On the other hand, although the correlation between drier conditions and increased

transmission is unlikely to be caused by higher temperatures, we acknowledge that the effects

of meteorological factors have a complex relationship. Unlike the consistent negative

correlation across lags 3-8 (rain) and lags 4-10 (rainhr) in Figure 4, the positive correlation of

“rain” and “rainhr” in Figure 3 occurred only in lag 9, and was therefore most likely a random

statistical anomaly rather than a conclusive finding. We believe that weather-based mechanisms

(46)

that support the proliferation of indigenous dengue therefore need further region-specific

investigation and more international collaboration.

6.3 The effect of vector control on epidemics of dengue

We consider that vector control efforts on dengue cases do not affect outbreak

initiation, but rather the size and magnitude of an outbreak. A dengue notification delay of over

one month allows for two transmission cycles, and increases the potential for a large outbreak

[67]. Vector control operations in Taiwan are unlikely to influence imported cases to initiate

local dengue epidemics because they are implemented after case notification [31]. By the time

indigenous dengue cases increase, the relationship with imported cases disappear (Figure 5C).

Hence, the focus of this study was to verify the correlation between imported dengue and the

onset of local dengue epidemics under appropriate weather conditions.

6.4 The increasing severity of dengue/DHF epidemics in Taiwan

Under suitable weather conditions, dengue viruses introduced via travelers are likely to

result in further domestic spread and subsequent occurrence of epidemics. In addition, the

introduction of more virulent genotypes of dengue viruses has been documented as a potential

factor for driving new epidemics [68-70]. For example, Thai strains belonging to the

(47)

36

1980-1994 clade within the genotype I of dengue virus serotype 1 (DENV-1) were replaced by

a 1990-2002 clade [30]. Additionally, an old clade in genotype I of DENV-3 during 1976-1978

was also replaced by a new 1991-2002 clade in genotype II [29,30]. Furthermore, cosmopolitan

genotypes of DENV-2, the causing agent of Taiwan’s largest-scale epidemic of dengue/DHF in

last thirty years, had been gradually and effectively replacing Asian genotype 2 in the

Philippines since 1998 and entered Taiwan in 2001 [71]. This cosmopolitan genotype of

DENV-2 is different from the Asian 1 and Asian 2 genotypes of Taiwan’s DENV-2 isolates

from 1981 to 1998 and the American/Asian genotype of Taiwan’s isolates in 2005, when the

majority of dengue cases were dengue fever [72]. In other words, the more virulent

genotypes/strains of the same serotype that have emerged during later years have resulted in

more severe and/or larger-scale epidemics of dengue/DHF in many Asian countries [68,70].

Based on phylogenetic analyses of dengue viruses isolated from imported cases [30] at

the micro-level, we find that local dengue epidemics in Taiwan typically originate in South

East Asia. It is therefore imperative to establish a stable surveillance system to detect the spread

of different genotypes of DENV. Currently, Taiwan’s comprehensive dengue surveillance

system is evolving and, hopefully, it may continuously monitor the possible evolution of

DENV in SE Asian countries through international collaboration. We believe that global

warming may have further impact on the incidence of imported dengue cases and future

(48)

dengue/DHF epidemics [73]. Advanced research integrating virus displacement and

meteorology will be necessary to provide a fuller understanding of both the macro and micro

changes contributing to the increasing severity of dengue/DHF epidemics.

6.5 Better interpretation with logistic and Poisson regression models

In order to construct the best possible regression models to reflect meteorological

conditions, we built alternative statistical models to demonstrate the role of imported cases in

the onset of dengue epidemics. Previous modeling studies using ARIMA (Autoregressive

Integrated Moving Average) found that the number of imported dengue cases was not

associated with the incidence of local dengue [19,20]. ARIMA examined a linear relationship

between case numbers of imported dengue cases and incidence of indigenous dengue cases

over several time lags. However, the quantitative relationship between imported and indigenous

dengue was likely limited to the onset (i.e. early phase) of outbreaks, and was therefore not

subject to linear modeling. We believe these conclusions by logistic and Poisson regression

models are not only demonstrable in countries with distinct seasonality, but also applicable in

non-endemic areas of dengue. However, meteorological conditions may need to be modified

for countries in higher altitudes.

(49)

38 6.6 Application of DRI to dengue control

This study performed a new categorical system, Dengue Risk Index, to quantify the risk

of local epidemics of dengue with both Taiwan and Thailand’s data. The risk of indigenous

dengue is able to be forecasted with previous case number and weather, such as weather

forecasting, as an early warning signal. Using real-time meteorological and epidemiological

information, public health officials are able to apply this DRI system and then easily

communicate with its prediction results to the public and international travelers. The timeliness

is better than currently used surveillance systems of dengue focused on human cases reporting

or entomological survey for earlier prevention and control of dengue before an epidemic starts.

For practical concerns, it is supposed to be more convenient to apply DRI in routine

surveillance. Since the calculation of DRI is to sum up total case numbers, once case are

confirmed, DRI is able to be announced immediately. Every month or two weeks, each country

or region may have area-specific real-time monthly or bi-weekly DRI. Accompanying with

weather forecasting information, region-specific DRI can be projected. In addition to evaluate

traveling risk, the most important usage of predicting DRI is for governmental disease

prevention and control. Moreover, international collaboration on dengue control will be

feasible through sharing newest nationwide DRI data from open-channel internet system.

(50)

6.7 The predictability of newly developed DRI models

Comparing with case number prediction in a variety of mathematical models for dengue

epidemics, polytomous (ordinal) logistic category models have much less variance in

distributed numbers. Therefore, it is much less fluctuated but more informative. The

cumulative-odds model in polytomous logistic regression is often considered most appropriate

in dealing with categorized data [74]. It is thus straightforward to interpret the effect of

independent variables with odds ratio by regression models.

The correlations of local weather variables with DRI in Taiwan were similar to in

Thailand, although their statuses of dengue endemicity were different. Temperatures and

relative humidity are the two statistically significant weather variables in regression models

after stepwise variable selecting process. Higher temperatures, including daily

mean/minimum/maximum temperature, and lower relative humidity were observed related with

higher DRI. However, when more than one lags of the same weather variable were selected, the

coefficient became puzzling. These regression models are not able to afford comprehensive

details, but offer a simple and direct method to detect possible occurrence of local dengue

epidemics.

The non-endemicity of Taiwan’s dengue status was observed in DRI exploration. The

annual local epidemics of dengue, caused by different serotypes and/or genotypes of dengue

(51)

40

viruses, have been mostly initiated by imported cases in the past three decades [30]. Therefore,

mean DRI in most months within one year is zero. Even the epidemics start, high values of DRI

seldom appear and usually vanish in winter. On the contrary, mean DRI in Thailand is two and

number of cases peak before summer.

6.8 Limitations

This study had notable limitations. First of all, meteorology alone does not initiate an

epidemic. Herd immunity also plays a decisive role in the spread of disease. Once a new or

more virulent genotype/strain of dengue virus is introduced, public health officials should alert

the public and implement prevention efforts regardless of meteorological conditions. Second,

local entomological data from Taiwan’s entomology surveillance was not included.

Non-government scholars do not have access to such data prior to 2002. Furthermore, the data

was divided by village or “Li” – the basic administration unit in Taiwan, and was not

systematically collected with a standardized process. Therefore, we did not use entomological

data because of its lack of consistency and inability to adequately represent the locations

covered in our study. Lastly, although socioeconomic status may influence vector habitat [75],

it was assumed to be relatively stable during the studied ten years.

(52)

6.9 Future direction

As an increase in viremic international travelers has led to global increasing DHF case

numbers to surge in recent two to three decades [24], efficient measures have to be instituted to

prevent imported dengue cases from igniting local dengue/DHF epidemics. Additionally, it has

been previously found that DHF cases with higher viral load [76] appeared when the number of

dengue fever cases increased rapidly, particularly in areas with higher dengue clusters [77].

Advanced investigation and integration of immunological, virological, meteorological, and

entomological findings with prevention/control strategies will support a more comprehensive

understanding of the mechanisms that initiate dengue epidemics, and will help guide realistic

public health interventions in the era of rapid globalization and climate change [78].

The established DRI method from this dissertation research using ordinal logistic

regression incorporating with Taiwan’s data first and then extended to Thailand’s data imply

that DRI is a feasible approach from the initial concept of prediction with different warning

color flags to implementation for travelers crossing different countries. For public health

practical application, it is necessary to investigate the feasibility by using more data obtained

from different regions, such as other countries in the South East Asia, Caribbean and South

America, and to testify the reliability of meteorology-based DRI on epidemics prediction across

different geographical areas and climate zones. Using data of case number on WHO’s website

(53)

42

and the same log scale of dengue cases for different countries, the overall picture of where is

global endemicity of dengue with higher risk levels, better than case number alone, is easily

understandable to general public and local/regional public health officials. Future efforts in

combining with local weather information into regression models, the possible impact of

climate change on increasing DRI is able to be quantified and supported with probability

calculation. In other words, this study and further investigation efforts will enlighten decision

makers by providing an evidence-based projection of how many people will be under the risk

of dengue in the future.

建立以氣候資訊為基礎的台灣登革流行預測模式：1998-2007年登革流行病學和氣象學因子之時序分析

Graduate Institute of Epidemiology College of Public Health

National Taiwan University Doctoral Dissertation

感謝天父讓我有幸進入博士班就讀，並在八年之內陸續完成了結婚、生育

子女、畢業等大事。一路走來，酸甜苦辣兼而有之，感謝家人、朋友及教會的

支持，特別是外子詠銘的幫助、陪伴，以及父母的栽培與鼓勵；感謝已經在天

上的外婆、祖母的養育之恩，讓我可以也成為照顧別人的人。也感謝過去各個

求學階段的恩師益友，這些教導與砥礪為我日後的研究生涯奠定了重要的基礎。

謝謝金傳春老師將我帶進傳染病的領域，讓我一窺學術殿堂的奧秘；謝謝

方老師費盡心思且無私無我的指導，為人處事及鑽研學問的嚴謹態度都是我的

好榜樣，幫助我可以將畢業論文順利發表於口碑極佳的國際期刊；謝謝媲美智

多星的柳中明老師，常常迅速提供氣象學的豐富資訊及各種疑難排解，總是從

容、溫和又充滿智慧；謝謝張念台老師這幾年來在研究計畫上提供的寶貴指導

與協助，對於提攜後進也是不遺餘力；謝謝梁賡義老師在口試階段的拔刀相助，

百忙之中仍盡力給予詳盡的指導，大師風範令人如沐春風。

謝謝台大提供了絕佳的環境，讓一切的努力與創意都得到百分之百的支

持；在這裡當學生的時間已有十四年，我深深以此身分為榮。古云：登高必自

卑，行遠必自邇；正如今年畢業典禮的主題：千里之行，始於足下。回首過往、

展望未來，我僅以感恩的心情向前邁步，認真迎接另一個精采生涯的開始。

受登革病毒感染的旅客經常成為散播病毒至其他地區的重要途徑，甚而引

發他國的流行；然而這些感染的旅客入境後和當地氣候、病媒以及本地疫情之

間的互動關係並不清楚。由於境外移入病例與本地登革疫情關係的議題長期受

到忽略，本研究即在探究臺灣地區的境外移入登革病例和氣候因子，對於本地

疫情發生之影響，進而可讓基層與中央衛生人員將危險層級資訊馬上應用於疾

病防控策略。

我們使用羅吉斯（logistic）和普瓦松（Poisson）迴歸模式分析 1998 至 2007

年間臺灣南部地區經實驗室診斷證實的登革確定病例，以區辨在氣候因子的作

用下，境外移入和本地登革病例的時序相關性。結果發現本地登革疫情的發生

與境外移入病例數（2 至 14 週前）、高溫（6 至 14 週前）及低濕度（6 至 20 週

前）之間，具有延遲的相關性。此外，境外移入病例數和本地登革病例數僅在

流行被引發的「初期」階段，才有明顯數量上的相關性；一旦流行持續發生，

此種關係即不復見。另外，根據單變項分析的結論，挑選出的重要氣候因子所

建立的羅吉斯迴歸模式，可以進一步建立登革危險性指標的預警數值。經初步

運算，此法優於僅以前期登革危險性指標(dengue risk index, DRI)值預測本期 DRI

值的效果，在加入氣候因子後，模式預測值與觀察值的相關係數(Spearman

correlation)可達 0.71(屏東地區)、0.84(高雄地區)以及 0.86(台南地區)；至於泰國

地區也有 0.66-0.77 的水準。由此可知此法不但適用於台灣登革疫情層級之預

警，也適用於泰國海岸五省的疫情預測。

這些發現顯示，惟有氣象條件適宜時，境外移入登革病例才有可能引發本

地的疫情。據此，經由境外移入病例的快速實驗診斷、早期發現以及管理，可

以遏止其後大規模登革/登革出血熱流行的發生。因此整合氣象資訊的早期警示

監測系統，將是登革疫情尚未成為地方性流行的地區用以成功防治疫情的無價

利器。

關鍵字：登革熱/登革出血熱、本土流行、境外移入病例、氣象、迴歸模式、登

革危險性指標

Abstract

Travelers who acquire dengue infection are often routes for virus transmission to

other regions. Nevertheless, the interplay between infected travelers, climate, vectors,

and indigenous dengue incidence remains unclear. The role of foreign-origin cases on

local dengue epidemics has thus been largely neglected by research. This study

investigated the effect of both imported dengue and local meteorological factors on

the occurrence of indigenous dengue in Taiwan.

Using logistic and Poisson regression models, we analyzed bi-weekly,

laboratory-confirmed dengue cases at their onset dates of illness from 1998 to 2007 to

identify correlations between indigenous dengue and imported dengue cases (in the

context of local meteorological factors) across different time lags. Our results

revealed that the occurrence of indigenous dengue was significantly correlated with

temporally-lagged cases of imported dengue (2-14 weeks), higher temperatures (6-14

weeks), and lower relative humidity (6-20 weeks). In addition, imported and

indigenous dengue cases had a significant quantitative relationship in the onset of

local epidemics. However, this relationship became less significant once indigenous

epidemics progressed past the initial stage. Polytomous Logistic regression model

with relevant meteorological variables stepwise selected were able to fitting the value

of Dengue Risk Index (DRI) in both Taiwan and Thailand. The Spearman correlation

between observed DRI and model-expected DRI ranged from 0.71 to 0.86 in Taiwan

and from 0.66 to 0.77 in Thailand, respectively.

These findings imply that imported dengue cases are able to initiate indigenous

epidemics when appropriate weather conditions are present. Early detection and case

management of imported cases through rapid diagnosis may avert large-scale

epidemics of dengue/dengue hemorrhagic fever. The potential application of DRI

with meteorological modeling in both Taiwan and Thailand demonstrated that its

feasibility to be extended to other countries for the current important issue on global

warming and dengue. The deployment of an early-warning surveillance system, with

the capacity to integrate meteorological data, will be an invaluable tool for successful

prevention and control of dengue, particularly in non-endemic countries.

Keywords: dengue, meteorology, climate, weather, regression model, Dengue Risk

Index

Contents

Signature Page（口試委員會審定書） ... i

Acknowledgements（誌謝） ... ii