Social Science & Medicine 55 (2002) 1015–1024

The spatial epidemiology of cholera in an endemic area ofBangladesh

Mohammad Alia,*, Michael Emchb, J.P. Donnayc, Mohammad Yunusd,R.B. Sacke

a International Vaccine Institute, Seoul, South KoreabDepartment of Geography, Portland State University, USA

cDepartment of Geomatics, University of Liege, Belgiumd ICDDR, B: Centre for Health and Population Research, Bangladesh

eJohns Hopkins University, USA

Abstract

This paper defines high-risk areas of cholera based on environmental risk factors of the disease in an endemic area of

Bangladesh. The risk factors include proximity to surface water, high population density, and low educational status,

which were identified in an earlier study by the authors. Cholera data were analyzed by spatially referenced extended

household units for two time periods, 1983–1987 and 1992–1996. These periods were chosen because they had different

dominant cholera agents. From 1983–1987 classical cholera was dominant and from 1992–1996 El Tor was dominant.

By defining high-risk areas based on risk factors, this study builds a spatial risk model for cholera. The model is then

evaluated based on the locations of observed cholera cases. The study also identifies the determinants of death due to

cholera for the two different time periods dominated by the different cholera agents. The modeled risk areas that were

based on the risk factors were found to correspond with actual distributions of cholera morbidity and mortality. The

high-risk areas of the dominant cholera agents are relatively stable over time. However, from 1983–1987 El Tor cholera,

which was not the dominant agent during that period, was not associated with high-risk areas, suggesting that the El

Tor habitat may have changed over time. The case fatality rate for cholera was related to proximity to a diarrhea

treatment hospital in the study area. r 2002 Published by Elsevier Science Ltd.

Keywords: Cholera; Disease risk; Spatial analysis; Bangladesh

Introduction and Background

An important part of health-needs assessment is the

identification of high-risk areas for a disease because

understanding the characteristics of high-risk areas may

indicate what is needed to improve health care provision

(Haining, 1996). However, most epidemiological studies

overlook the spatial components of disease and focus

solely on characteristics of the people who contract a

disease (Smoyer, 1998). While these studies are useful for

identifying biological factors of a disease, they usually

cannot establish accurate individual exposure levels for

the critical risk factors of the disease (Haining, 1998).

The incorporation of spatial components in health

studies facilitates the identification of high-risk disease

areas, the identification of sources of disease, the

definition of high-risk populations, and the optimal

allocation of health facilities (Jacquez, 2000). Identifying

high- and low-risk areas can help with the estimation of

resources needed for effective health planning and it may

help determine the underlying processes responsible for

the spatial patterns of disease.

Methods for identifying high-risk areas by disease

clustering techniques are well established (Myaux, Ali,

Chakraborty, & de Francisco, 1997a; Kulldorff &

Nagarwalla, 1995; Besag & Newell, 1991). However,

*Corresponding author. International Vaccine Institute,

Kwanak PO Box 14, Seoul, South Korea 1S1-600.

E-mail address: mali@ivi.int (M. Ali).

0277-9536/02/$ - see front matter r 2002 Published by Elsevier Science Ltd.

PII: S 0 2 7 7 - 9 5 3 6 ( 0 1 ) 0 0 2 3 0 - 1

defining high-risk areas based on case incidence may be

influenced by the size and age structure of the

population. Several methods, such as kernel estimation

and k-function have been proposed to remove such bias

in identifying clusters of disease (Gatrell, Bailey, Diggle,

& Rowlingson, 1996). Since socio-economic status varies

across households, defining risk areas based on house-

hold level case incidence may be influenced by the status

of the households. Another potential problem of using

disease incidence to define risk areas is the availability of

data sets. In developing countries, disease incidence

monitoring systems are not likely to be implemented on

a large scale because of resource constraints.

In this study high-risk areas, presumed to be niches of

cholera, are defined based on environmental-risk factors

of the disease. The risk factors are proximity to surface

water, high-population density, and low-educational

status, which were identified in an earlier study by the

authors (Ali, Emch, Donnay, Yunus, & Sack, 2001). An

environmental niche can be defined as a region that is

characterized by a set of environmental variables

(Gatrell, 1983). The combination of factors that permits

a species to survive defines its niche. By defining high-risk

areas based on risk factors, this study builds a spatial-

risk model for cholera. The model is then evaluated

based on the locations of observed cholera cases.

Different cholera agents coexist in Bangladesh but only

one agent is dominant at any one time. This study

compares the spatial distributions of the different cholera

agents at different times with the modeled risk areas. This

study also identifies the determinants of death due to

cholera for the two different time periods dominated by

different cholera agents. The determinants that are

investigated include accessibility to treatment centers,

educational level, and whether or not a person lives in a

flood-controlled area. Accessibility to treatment centers

has been found to be an important determinant of

diarrhea-related mortality (Rahman, Aziz, Munshi,

Patwari, & Rahman, 1982). Myaux, Ali, Felsenstein,

Chakraborty and de Francisco (1997b) reported that

acute watery diarrheal mortality for children under five,

although not statistically significant, was higher inside a

flood-protected area than outside. Educational level is

hypothesized to be an indicator of mortality because it

indirectly determines healthcare seeking behavior.

The study area and data sources

The study area

This study was conducted in Matlab, a rural area of

Bangladesh with endemic cholera. It is 53 km southeast

of Dhaka, the capital of Bangladesh. The study area is in

the central plain of Bangladesh adjacent to the

confluence of the Meghna and the Ganges Rivers. The

Dhonagoda River bisects the study area into two

approximately equal parts. There are also numerous

canals that remain dry in the winter and fill with water

during the monsoon. In the late 1980s, a flood-control

embankment was built along the Dhonagoda and

Meghna Rivers. The embankment was built primarily

to protect the area from monsoon flooding so that crops

can be grown throughout the year. It protects 31% of

the study population from flooding. The people of the

study area live in groups of patrilineally related house-

holds called baris. An average of six households

constitute a bari.

The spatial database

A vector spatial database of the study area was created

in 1994 to facilitate spatial analysis in health and

population research. Features in the database that were

used in this study include baris, rivers, and the embank-

ment (Fig. 1). The spatial database was converted to the

raster format in which space is divided into discrete units

called pixels. The size of the pixels was set to 30 m, which

facilitates the representation of baris by single pixels. A

total of 7691 pixels represent the baris and each pixel has

a unique identification number so that health and

population data can be linked to bari locations.

Health and demographic databases

Individual health and demographic surveillance data

are regularly collected for all individuals living in the

Matlab study area. These individual level data

were aggregated by baris. Cholera cases were identified

from Matlab hospital surveillance records. The Matlab

hospital is the only diarrhea treatment center in this

rural area and it provides free treatment to all patients.

Therefore, it is unlikely that a diarrhea patient from

the area would miss the opportunity to receive

treatment. Stool samples are collected for all patients

who live in the study area, and the samples are screened

for enteric pathogens in the laboratory. Since the

cholera morbidity data were based on hospitalized

patients, the analysis was restricted to the baris located

within 9 km of the hospital because few patients

reported to the hospital from further away. Cholera

morbidity data were collected for two time-periods,

1983–1987 and 1992–1996, in order to understand the

spatial epidemiology of two different agents of cholera,

classical and El Tor. Classical cholera was the dominant

agent from 1983–1987 and El Tor cholera from 1992–

1996. From 1983–1987 there were 1236 classical cholera

cases and 719 El Tor cholera cases. A total of 1342 El

Tor cholera cases were identified from 1992–1996 and no

classical cholera cases were identified during that time

period.

M. Ali et al. / Social Science & Medicine 55 (2002) 1015–10241016

Medical assistants regularly report the cause of death

for all people living in the Matlab study area. Data on

deaths due to acute watery diarrhea (AWD), also

called cholera-like diarrhea, were collected from the

demographic database. Medical assistants assigned

causes of deaths by reviewing verbal autopsy forms

recorded using methods described elsewhere (Kielmann,

DeSweemer, Parker, & Taylor, 1983). The classification

of the causes of death was derived from the Interna-

tional Statistical Classification of Diseases, Injuries and

Causes of Death (WHO, 1977), and was adjusted

according to the Matlab surveillance reporting system.

The mortality data were aggregated by baris for the two

time-periods by summing total deaths for each time-

period.

Methods

Cholera morbidity and AWD mortality indices

This study defines cholera-risk areas based on three

risk factors. It also assesses how well the risk areas

correspond to the actual distributions of cholera

morbidity and AWD mortality. Cholera-morbidity and

AWD-mortality indices were calculated within the raster

Fig. 1. Vector spatial database of the study area.

M. Ali et al. / Social Science & Medicine 55 (2002) 1015–1024 1017

GIS (Ali et al., 2000). The rates were calculated by

using a spatial smoothing technique referred to as a

spatial moving average rate (Kafadar, 1996). A 7� 7-

pixel moving window was used to compute average

rates for observed bari points (Talbot, Kulldroff, Teven,

& Haley, 2000). The expression for computing the

indices is

jDi ¼

Pnj¼1 dj�kjPnj¼1 pj�kj

�1000;

where i refers to image pixel, and j refers to window

pixel. Therefore, ji=rates per 1000 per year in bari i;dj=number of cases in bari j; pj=total number of

people in bari j; kj=kernel values (unitary) of cell j of the

moving window.

Indices of cholera morbidity and AWD mortality

were computed for the two time periods. The ratios were

multiplied by 1000 to express the index (jDi) per 1000

people.

Defining high-risk areas

Since there were multiple risk factors and each

was responsible for different levels of risk, a multi-

criteria evaluation model (Voogd, 1983; Carver, 1991)

was built to create a single index of risk. The single

value of risk is the linear combination of the

factors multiplied by its weight. The index of risk is

expressed as

Ri ¼ Sxijwj ;

where, Ri=spatial risk for bari i; xij=standardized score

of factor j of bari i; wj=weight of factor j:Since the factors were measured in different units,

standardized scores of the factors were used. The

evaluation model requires that higher values be

positively correlated with the outcome. Therefore,

the standardized scores of the negatively correlated

factors (distance to surface water and educational

status) were reversed. The factors’ weights were

determined by the relative importance of the factors in

influencing disease incidence, which were derived

from the percentage change in an outcome variable

associated with change in a risk factor. The regression

model used to determine the risk factors from 1992–1996

was used in this process. The risk factors were the same

for both time periods, although the magnitude of

influence on the outcome variable was different. The

factors’ weights were calculated by taking the principal

eigen vector of the squared reciprocal matrix of pairwise

comparisons describing the relative importance of the

factors (Saaty, 1977). The spatial risk for the observed

locations was computed for each of the two time

periods.

Kriging was used to interpolate spatial risk at a

regularly spaced interval in space in order to map risk

areas. Kriging is one of the most flexible methods for

interpolating data at unsampled points (Collins, 1998).

Its estimates are unbiased and have a known minimum

variance (Oliver & Webster, 1990). The interpolated

values at unsampled points were computed using the

following formula:

Gi ¼Xn

i¼1

lijZj ;

where, Gi=interpolated data at point i; n=number of

data points used in interpolating the data, Zj ¼ Z value

at the jth data point, lij=weight associated with the jth

data for computing Gi

The weighting factor lij varies between 0 and 1 in a

continuous scale; the closer a data point is to a measured

point, the more weight it will carry. The sum of the

weighting factors used to calculate the value is 1. In

defining the grid-cells, spacing was set to 150 m in all

directions. Kriging was found to be so robust that even

when spacing was set to 300 m, little difference was seen

in the resulting maps. Contour mapping methods, which

used the krigged data, were used to define the surface in

low-, moderate-, and high-risk areas. The thresholds to

define the risk were based on the distribution of the

single value of risk. Values below the mean were

considered to be low risk; values between the mean

and the mean plus half of the standard deviation were

considered to be moderate; and those higher than that

were considered to be high risk.

In the previous study (Ali et al., 2000) that identified

the three risk factors used to map cholera risk, proximity

to surface water was calculated using two different

methods. One method used the vector GIS database

to identify the main rivers and canals and the other

used Landsat Thematic Mapper satellite imagery to

identify all surface water that can be represented in

a 30-m resolution satellite image. The two data sources

were used in the analytical stage of this project because

they represent slightly different variables. The resulting

map of the risk areas, defined by using the main rivers

and canals variable as well as population density

and socio-economic status, is shown in Fig. 2. This

group of three variables including the main rivers and

canals is hereafter referred to as Model 1. The Model 1

risk map resulted in 78.67 km2 in the low-risk category,

59.18 km2 in the moderate risk category, and 46.26 km2

in the high-risk category. Fig. 3 represents the risk areas

obtained by using surface water derived from satellite

data as well as population density and socio-economic

status. This group of three variables including the

surface water derived from satellite imagery is hereafter

referred to as Model 2. The Model 2 risk map resulted in

77.98 km2 in the low-risk category, 55.06 km2 in the

moderate risk category, and 51.08 km2 in the high-risk

category.

M. Ali et al. / Social Science & Medicine 55 (2002) 1015–10241018

Assessment of risk areas

Spatial agreement between risk models

This study presents two different models of cholera

risk. It is therefore useful to know how the models agree

with each other in defining the risk areas. A k-statistic

was calculated to measure the spatial agreement between

the two models (Cohen, 1960). There was moderate

agreement between the moderate risk areas (k=0.48) for

the two models and good agreement between the high-

(k=0.67) and low-risk areas (k=0.72) (1=perfect

agreement; >0.6=good agreement; 0=no agreement).

Spearman’s correlation coefficients were also calculated

to measure agreement between the risk categories for the

two models. The results show that there is strong

agreement between the models (rs ¼ 0:75) suggesting

that the two measures of proximity to surface water are

possibly interchangeable. These results are expected

because only one factor between the models differs.

Risk areas and cholera morbidity

Cholera morbidity prevalence rates were calculated

for each of the three risk categories. The analysis was

done for both time periods using both data models.

From 1992–1996 there were 6012 baris within 9 km of

the hospital and from 1983–1987 there were 4399 baris

within 9 km of the hospital. The distribution of the baris

by risk area for each of the time periods is presented in

Table 1. Linear and logistic regression models were built

to determine whether cholera prevalence rates were

significantly different for the areas of different risk

categories. For instance, did the high-risk category have

Fig. 2. Map of the risk areas defined by Model 1.

M. Ali et al. / Social Science & Medicine 55 (2002) 1015–1024 1019

significantly higher cholera morbidity than the moder-

ate- or low-risk areas? In the linear-regression model, the

outcome variable was the spatially smoothed cholera-

morbidity rate, and in the logistic-regression model, the

outcome was defined by presence or absence of cholera

in a bari. In both of the regression models, the risk

categories were incorporated as dummy variables. Since

there are three risk categories, two binary dummy

variables were created, one for moderate risk and the

other for high risk. Table 2 shows the results of the

regression models, which indicate that the high-risk

areas had significantly higher cholera morbidity in each

of the models. The analysis was restricted to within 9 km

from the hospital. The table shows that the El Tor

prevalence rates from 1992–1996 is 22% higher in

moderate-risk areas and 24% higher in high-risk areas

compared to the rate in the low-risk areas as defined by

Model 1. Model 2 shows that the El Tor prevalence rate

is only 7% higher (not statistically significant) in

moderate-risk areas and 20% higher in the high-risk

areas compared to the rate in the low-risk areas. On the

other hand, from 1983–1987 the classical cholera cases

are significantly higher in the high-risk areas compared

to the rates in the low-risk areas. Table 2 also shows that

the findings of the logistic-regression analyses are the

same as the findings of the linear-regression analyses.

Risk areas and AWD mortality

The AWD mortality rates were assessed by risk

category and relationships between several predictor

variables and AWD mortality were measured. All active

baris were studied because distance from the hospital

does not affect the mortality data since it is community-

Fig. 3. Map of the risk areas defined by Model 2.

M. Ali et al. / Social Science & Medicine 55 (2002) 1015–10241020

level data. There were 7467 baris from 1992–1996 in the

study area and 5718 baris from 1983–1987. The

distribution of the baris by risk area for each time

period is presented in Table 1. From 1992 to 1996 there

were approximately 3 AWD deaths per 10,000 people

per year. From 1983 to 1987 there were approximately 6

AWD deaths per 10,000 people per year. As with the

morbidity data, the mortality rate is significantly higher

in the high-risk areas for both periods (Table 3).

AWD mortality and predictor variables

Accessibility to treatment centers is an important

determinant of diarrhea mortality (Rahman et al., 1982).

In this study the cost distance to the nearest treatment

center was used to model accessibility. Movement in

space incurs a cost (time or money), which is a function

of frictions and forces that impede or facilitate move-

ment. Since there are numerous water bodies in the

study area, the linear distance would give imprecise time

costs for accessing the treatment centers in the study

area. Therefore, the cost distance was computed instead.

When calculating the time it takes to travel from

locations in the study area and the treatment centers,

traveling over water (rivers and canals) was assumed to

take 5 times more than traveling over the ground.

Myaux et al. (1997b) reported that AWD mortality for

children under five, although not statistically significant,

was higher inside the flood protection embankment. The

average number of years of education by bari was also

included in the list of the independent variables because

it indirectly determines healthcare seeking behavior.

Distance to the nearest treatment center significantly

influenced the mortality rate for both time periods

(Table 4). However, using the logistic model, there was

not a relationship between mortality and distance to a

treatment center. This might have resulted from using

continuous data with a binary outcome variable. From

Table 1

Frequency distribution of the baris by risk category

Categories of risk areas 1992–1996 1983–1987

Within 9 km from

hospital n ¼ 6012 no. (%)

Total study area

n ¼ 7467 no. (%)

Within 9 km from

hospital n ¼ 4399 no. (%)

Total study area

n ¼ 5718 no. (%)

Model 1a

low 2592 (43.1) 3037 (40.7) 1734 (39.4) 2143 (37.5)

moderate 1898 (31.6) 2268 (30.4) 1421 (32.3) 1760 (30.8)

high 1522 (25.3) 2162 (28.9) 1244 (28.3) 1815 (31.7)

Model 2a

low 2440 (40.6) 2985 (40.0) 1730 (39.3) 2236 (39.1)

moderate 1854 (30.8) 2212 (29.6) 1306 (29.7) 1629 (28.5)

high 1718 (28.6) 2270 (30.4) 1363 (31.0) 1853 (32.4)

a Please see the text for the definition of models.

Table 2

Regression analysis for risk area assessment: cholera morbidity

Categories of

risk areas

El Tor cholera, 1992–1996 Classical cholera, 1983–1987

Linear regression Logistic regression Linear regression Logistic regression

Rate/1000 % higher Odds ratio 95% CI Rate/1000 % higher Odds ratio 95% CI

Model 1a

Low 2.676 1 1.00 2.678 1 1.00

Moderate 3.274 22.3b 1.44 1.24–1.67b 2.942 9.9c 1.38 1.17–1.63b

High 3.317 24.0b 1.84 1.58–2.14b 3.332 24.4b 1.65 1.39–1.95b

Model 2a

Low 2.805 1 1.00 2.651 1 1.00

Moderate 2.997 6.9 1.20 1.03–1.40b 2.796 5.5 1.33 1.12–1.57b

High 3.375 20.3b 1.69 1.46–1.96b 3.474 31.0b 1.82 1.54–2.14b

a Please see the text for the definition of models.bpo0:01:cpo0:05:

M. Ali et al. / Social Science & Medicine 55 (2002) 1015–1024 1021

1992–1996 the area outside the flood-control embank-

ment had an 18% higher mortality rate compared to the

rate inside the embankment. The odds of dying were

1.53 (95% CI=1.18–1.99) times higher outside the

embankment than inside. The mean number of years

of education for each bari was 0.81. Level of education

was negatively related to mortality using both the linear-

and logistic-regression models for both time periods.

Discussion and Implications

Factors that put the people living in an area at risk for

cholera and AWD mortality have remained the same for

decades. The methods used in this study to define high-

risk areas are more robust than disease-clustering

methods based on case incidence rates. Defining high-

risk areas based on reported health events such as with

disease-clustering methods cannot be extrapolated to

other areas. Defining high-risk areas using risk factors

can be used in other areas where risk factor data are

available. However, when defining risk areas, the

different health outcomes used in this study (morbidity

and mortality) provide somewhat different risk surfaces.

This should be considered when planning the allocation

of services. This study provides new insight into the

spatial epidemiology of cholera by investigating the

spatial variation of the disease over time for two

different dominant agents of the disease. The high-risk

areas have persisted for decades for different dominant

cholera agents. This suggests that the niche for cholera is

in specific places in this endemic area. Higher morbidity

in the high-risk areas in both periods suggests that the

habitats of the present form of cholera, El Tor, and the

previous form, classical, are the same. Conversely, El

Tor risk areas during the two time periods are not the

same, which suggests that the present form of El Tor

cholera may have changed its environment and its

habitat is now the same as classical cholera before that

biotype disappeared.

The results show that only the dominant cholera

agents are associated with AWD mortality for both

Table 3

Regression analysis for risk area assessment: AWD mortality

Categories of risk areas AWD mortality, 1992–1996 AWD mortality, 1983–1987

Linear regression Logistic regression Linear regression Logistic regression

Rate/10,000 % higher Odds ratio 95% CI Rate/10,000 % higher Odds ratio 95% CI

Model 1a

Low 2.91 1 1.00 4.97 1 1.00

Moderate 2.76 �5.2 1.09 0.80–1.46 6.48 30.4b 1.52 1.21–1.92b

High 3.53 21.3c 1.74 1.33–2.29b 6.74 35.6b 1.82 1.45–2.27b

Model 2a

Low 2.90 1 1.00 4.96 1 1.00

Moderate 2.93 1.0 0.92 0.68–1.24 6.51 31.3b 1.41 1.12–1.78b

High 3.32 14.5d 1.38 1.06–1.81c 6.81 37.3b 1.73 1.39–2.15b

a Please see the text for the definition of models.bpo0:01:cpo0:05:dpo0:10:

Table 4

Regression analyses for independent variables: AWD mortality

Variables AWD Mortality, 1992–1996 AWD Mortality, 1983–1987

Linear regression Logistic regression Linear regression Logistic regression

ba p-value ba p-value ba p-value ba p-value

Level of education �5.418E-02 0.007 �0.2477 0.0467 �0.166 0.000 �0.4412 0.0000

Outside embankment 4.909E-02 0.025 0.4282 0.0012 �6.664E-02 0.073 �0.0903 0.3622

Distance to the nearest TC 6.376E-04 0.000 0.0006 0.4200 5.990E-04 0.005 0.0002 0.7589

a Coefficient.

M. Ali et al. / Social Science & Medicine 55 (2002) 1015–10241022

periods and that the location of high-risk areas is the

same for both time periods. In contrast, from 1983 to

1987 there was not a relationship between El Tor and

mortality. This suggests that the ecological state of the

high-risk areas might be involved in El Tor becoming a

more virulent agent. Therefore, it is imperative to

understand the aquatic environments of the high-risk

areas where the ecological conditions favor long-term

survival of cholera. Investigating the aquatic environ-

ment could reveal what ecological factors make Matlab

an endemic area and what changes have caused the

disappearance of classical cholera and increased the

virulence of El Tor.

Proximity to the treatment centers is a determinant of

AWD mortality as was found in a previous study

(Rahman et al., 1982). This suggests that access to

hospitals needs to be improved in the study area. Since a

cholera victim becomes dehydrated very rapidly, fast

access to health care facilities needs to be ensured. There

was also a relationship between educational level and

AWD mortality from 1983 to 1987. People with less

education are more likely to seek care from indigenous

medical practitioners, thus reducing the chance of

survival (Ali, Emch, Tofail, & Baqui, 2001). From

1992 to 1996 educational status did not influence

mortality as much as it did from 1983 to 1987. The

AWD mortality rate has also decreased over time.

Areas inside the embankment had higher cholera

morbidity than outside the embankment. Since the

embankment protects the area from flooding, there is

no longer natural flushing within the flood-protected

zone. The absence of natural flushing by floodwater has

changed the hydrological dynamics within the embank-

ment resulting in increased salinity (Siddique et al.,

1991). This has created a suitable environment for

cholera and therefore an increase in the disease

incidence. However, the higher morbidity did not result

in higher mortality inside the embankment. In contrast,

the embanked area had lower mortality compared to the

area outside the embankment. This could be due to the

effect of the proximity of baris to the nearest treatment

center. The average distance of the baris inside the

embankment to the nearest treatment center was greater

than those outside the embankment. Also, children who

are affected by enteric diseases are on average better

nourished inside the embankment (Myaux et al., 1997b)

and this may result in lower AWD mortality rates in the

area.

With advances in medical sciences, cholera is no

longer the deadly disease it used to be, but it is likely to

continue to cause concern and despair in Bangladesh

and other developing countries. This study shows that

geographical investigations are important in public

health for initiating effective control programs to reduce

health problems from environmental diseases. The water

and sewage infrastructure available in industrialized

countries will not be realized for some time in

developing countries. An intensive health intervention

program directed at high-risk areas in endemic areas will

assure that services are provided to the areas that need

them the most.

Acknowledgements

This research was funded by Belgian Administration

for Development Cooperation and ICDDR,B: Centre

for Health and Population Research. The ICDDR,B is

supported by countries and agencies which share its

concern for the health problems of developing countries.

Current donors providing unrestricted support include:

the aid agencies of the Governments of Australia,

Bangladesh, Belgium, Canada, Japan, the Netherlands,

Sweden, Sri Lanka, Switzerland, the United Kingdom

and the United States of America; international

organizations include United Nations Children’s Fund.

References

Ali, M., Emch, M., Donnay, J. P., Yunus, M., & Sack, R.B.

(2001). Identifying environmental risk factors of endemic

cholera: a raster GIS approach, Health and Place, revised

for publication.

Ali, M., Emch, M., Tofail, F., & Baqui, A. H. (2001).

Implications of health care provision on acute lower

respiratory infection mortality in Bangladeshi children.

Social Science and Medicine, 52(2), 267–277.

Besag, J., & Newell, J. (1991). The detection of clusters in rare

disease. Journal of the Royal Statistical Society A, 154, 143–

155.

Carver, S. J. (1991). Integrating multi-criteria evaluation with

geographical information systems. International Journal of

Geographical Information Systems, 5(3), 321–329.

Cohen, J. (1960). A coefficient of agreement for nominal scales.

Educational and Psychological Measurement, 20, 37–46.

Collins, S. (1998). Modelling spatial variations in air quality

using GIS, In GIS and Health (GISDATA Series 6), Ed.

Anthony, C., Gatrell, Markku L .oyt .onen, London: Taylor &

Francis.

Gatrell, A. C. (1983). Distance and space: A geographical

perspective. Oxford: Clarendon Press.

Gatrell, A. C., Bailey, T. C., Diggle, P. J., & Rowlingson, B. S.

(1996). Spatial point pattern analysis and its application in

geographical epidemiology. Transactions of the Institute of

British Geography, NS21, 256–274.

Haining, R. (1996). Designing a health needs GIS with spatial

analysis capability. In M. Fischer, H. J. Scholten, D.

Unwin, (Eds.), Spatial Analytical perspectives on GIS,

GISDATA Series 4, London: Taylor & Francis.

Haining, R. (1998). Spatial statistics and the analysis of health

data. In A.C. Gatrell, M. L .oyt .onen (Eds.), GIS and health,

GISDATA Series 6, London: Taylor & Francis.

M. Ali et al. / Social Science & Medicine 55 (2002) 1015–1024 1023

Jacquez, G. M. (2000). Spatial analysis in epidemiology:

Nascent science or a failure of GIS. Journal of Geographical

Systems, 2, 91–97.

Kafadar, K. (1996). Smoothing geographical data, particularly

rates of disease. Satistics in Medicine, 15, 2539–2560.

Kielmann, A. A., DeSweemer, C., Parker, R. L., & Taylor, C.

E. (1983). Analysis of morbidity and mortality. Child and

Maternal Health Services in Rural India, The Narangwal

Experiment. Vol. 1: Integrated Nutrition and Health Care.

Baltimore, Johns Hopkins University Press, p.172–214.

Kulldorff, M., & Nagarwalla, N. (1995). Spatial disease

clusters: Detection and inference. Statistics and Medicine,

14, 799–810.

Myaux, J. A., Ali, M., Chakraborty, J., & de Francisco, A.

(1997a). Flood control embankments contribute to the

improvement of the health status of children in rural

Bangladesh. Bulletin of the World Health Organization,

75(6), 533–539.

Myaux, J. A., Ali, M., Felsenstein, A., Chakraborty, J., & de

Francisco, A. (1997b). Spatial distribution of watery

diarrhoea in children: Identification of ‘‘risk areas’’ in a

rural community in Bangladesh. Health & Place, 3(3), 181–

186.

Oliver, M. A., & Webster, R. (1990). Kriging: A method of

interpolation for geographical information systems. Inter-

national Journal of Geographical Information Systems, 4,

313–332.

Rahman, M. M., Aziz, K. M. S., Munshi, M. H., Patwari, Y., &

Rahman, M. (1982). A diarrhea clinic in rural Bangladesh:

Influence of distance, age, and sex on attendance and

diarrheal mortality. American Journal of Public Health,

72(10), 1124–1128.

Saaty, T. L. (1977). A scaling method for priorities in

hierarchical structures. Journal of Mathematical Psychology,

15, 234–281.

Siddique, A. K., Baqui, A. H., Eusof, A., Haider, K., Hossain,

M. A., Bashir, I., & Zaman, K. (1991). Survival of classical

cholera in Bangladesh. Lancet, 337, 1125–1127.

Smoyer, K. E. (1998). Putting risk in its place: Methodological

considerations for investigating extreme event health risk.

Social Science & Medicine, 47(11), 1809–1824.

Talbot, T. O., Kulldroff, M., Teven, P. F., & Haley, V. B.

(2000). Evaluation of spatial filters to create smoothed maps

of health data. Statistics in Medicine, 18, 2399–2408.

Voogd, H. (1983). Multicriteria evaluation for urban and

regional planning. London: Pion, Ltd..

WHO. (1977). Manual of the international statistical classifica-

tion of diseases, injuries and causes of death based on the

recommendations of the ninth revision conference, 1975.

Geneva: World Health Organization.

M. Ali et al. / Social Science & Medicine 55 (2002) 1015–10241024