A spatial model of socioeconomic and environmental ...
Transcript of A spatial model of socioeconomic and environmental ...
A spatial model of socioeconomic and
environmental determinants of dengue fever in Cali,
Colombia
Abstract. Dengue fever has gradually re-emerged across the global South, particularly affecting
urban areas of the tropics and sub-tropics. The dynamics of dengue fever transmission are
sensitive to changes in environmental conditions, as well as local demographic and
socioeconomic factors. In 2010, the municipality of Cali, Colombia, experienced one of its worst
outbreaks, however the outbreak was not spatially homogeneous across the city. In this paper, we
evaluate the role of socioeconomic and environmental factors associated with this outbreak at the
neighborhood level, using a Geographically Weighted Regression model. Key socioeconomic
factors include population density and socioeconomic stratum, whereas environmental factors
are proximity to both tire shops and plant nurseries and the presence of a sewage system
(R2=0.64). The strength of the association between these factors and the incidence of dengue
fever is spatially heterogeneous at the neighborhood level. The findings provide evidence to
support public health strategies in allocating resources locally, which will enable a better
detection of high risk areas, a reduction of the risk of infection and to strengthen the resilience of
the population.
Keywords: Colombia; Dengue Fever; Environmental and Socioeconomic Determinants; GIS;
Geographically Weighted Regression (GWR).
1. Introduction
Dengue fever is a fast emerging mosquito-borne viral disease of high public health relevance
which is prevalent in many urban and semi-urban environments in the tropics and sub-tropics. It
is transmitted among humans by female mosquitos of the genus Aedes (i.e. Ae. aegypti and Ae.
albopictus) (Messina et al. 2015). The spatial distribution and impact of the disease is determined
by a complex interplay of environmental and socioeconomic factors. Changes in temperature,
precipitation patterns and humidity affect the distribution of these diseases through a number of
effects on mosquito habitats, the vector, vector activity and pathogen development within the
vector (Rogers and Randolph 2006; Semenza et al. 2012). Socioeconomic (e.g. education,
knowledge, behavior, occupation, income, livelihoods, etc.), biological (e.g. age, acquired
immunity, health status) and institutional factors (e.g. access to health care, quality of care,
control and prevention strategies, etc.) contribute to the vulnerability of the population and are
therefore fundamental drivers of disease risk (Bates et al. 2004a; Bates et al. 2004b; Hagenlocher
et al. 2013). Finally, the rapid, uncontrolled expansion of urban environments and increasing
international mobility/air travel has exacerbated the potential of its spread (Gardner and Sarkar
2013; Nunes et al. 2014; Alirol et al. 2011). To date, approximately 4 billion individuals are at
risk of contracting dengue fever, resulting in nearly 400 million worldwide infections per year
(Bhatt et al. 2013; Messina et al. 2015).
In the Americas, despite efforts to eradicate the vector during the 1950s and 1960s (OPS
1960), dengue fever has reemerged as a result of a reduction in surveillance and control
strategies (Dick et al. 2012; Wilson and Chen 2002; Torres and Castro 2007; Guzman and Kouri
2003; Istúriz, Gubler, and Castillo 2000; San Martín et al. 2010). Between the years 2001 and
2007 alone, nearly 4.5 million cases were reported in 30 countries. Examples reported in the
literature include Peru (Chowell et al. 2008), Colombia (Restrepo, Baker, and Clements 2014),
Brazil (Rodriguez-Barraquer et al. 2011; Teixeira et al. 2009; Nogueira, Araújo, and Schatzmayr
2007) and northern Argentina (Vezzani and Carbajo 2008).
Several studies have been conducted in urban environments across the Americas, where
population growth and unplanned urbanization have accelerated the prevalence of dengue fever.
Considerable research efforts have focused on (1) mapping spatial and spatiotemporal disease
patterns and associated risks (Castillo et al. 2011; Delmelle et al. 2013; Getis et al. 2003; Tran et
al. 2004; Rotela et al. 2007; Stoddard et al. 2013; Teixeira et al. 2002; Chiaravalloti-Neto et al.
2015; Siqueira-Junior et al. 2008; Vazquez-Prokopec et al. 2010) and –but to a lesser extent, on
(2) estimating vulnerability to the disease by evaluating the role of socioeconomic and
environmental determinants (Braga et al. 2010; de Mattos Almeida et al. 2007; Hagenlocher et
al. 2013; Kienberger et al. 2013; Honorio et al. 2009).
In Colombia, several measures were adopted to eradicate the presence of the mosquito in the
1950s and 1960s, by means of fumigation and the elimination of mosquito foci (Dick et al.
2012). Although these efforts were tangible for a while, a re-infestation of the Aedes Aegypti
occurred during the 1970s and beyond due to the reduction in surveillance and control program
support (Romero-Vivas, Leake, and Falconar 1998). Since the 1970s, DF has become endemic in
many areas of Colombia and adjacent countries (Messina et al. 2014). Dengue fever is now
considered endemic in certain areas of the country with periodic outbreaks including 1991, 1994,
1998, 2001, 2006, 2010, and 2013 (Restrepo, Baker, and Clements 2014; Ocampo et al. 2014;
Cali 2010; Villegas, Aristizabal, and Rojas 2010). The disease has exhibited an endemo-
epidemic pattern repeating every 2 to 3 years. While the temporal pattern is influenced by the El
Niño oscillation (Eastin et al. 2014), we hypothesize that the variation of rates across the city is
largely driven by environmental and socioeconomic factors.
In this paper, we evaluate the role of socioeconomic and environmental determinants
associated with rates of dengue fever during the 2010 outbreak in the city of Cali, Colombia. We
hypothesize that this association may not be spatially homogeneous (i.e. the dengue fever rate
estimated at one neighborhood is likely to exhibit similarities to rates at nearby neighborhoods)
and that a Geographically Weighted Regression (GWR), which explicitly incorporates the notion
of spatial autocorrelation, better captures this heterogeneity. Our paper is part of a larger,
collaborative effort to (1) map the dynamics of dengue fever in Cali (Kienberger et al. 2013;
Delmelle, Zhu, et al. 2014; Delmelle et al. 2013; Delmelle, Dony, et al. 2014), (2) estimating
vulnerabilities to dengue fever (Hagenlocher et al. 2013) and to (3) strengthen our understanding
of the role played by both environmental and socioeconomic factors (Eastin et al. 2014). The
findings from our work are critical for the planning of preventive measures and inform future
risk assessments.
2 Materials and Methods
2.1 Study area
The city of Cali in Colombia is the third largest metropolitan area in the country with an
estimated 2010 population of 2.3 million (average population density over 4,000/km2). The city
is administratively organized into 22 communes, which correspond to neighborhood groupings (a
total of 340 neighborhoods) with homogenous socioeconomic stratum. Neighborhoods are the
finest scale at which the census reports its socioeconomic and demographic information. Cali is
characterized by a tropical climate which favors annual transmission of dengue (Eastin et al.
2014). In recent decades, the city has experienced unorganized migrations of individuals
displaced by the Fuerzas Armadas Revolucionarias de Colombia (FARC). These populations
have settled in high-density, low income population neighborhoods at the periphery of the city.
These neighborhoods have suffered from unplanned urbanization, including squatter settlements
and contain limited sewer infrastructure with several open-air waste water channels.
2.2 Data pre-processing
We rely on three distinct sources of geospatial information, namely (1) individual dengue
fever cases, (2) socioeconomic and demographic datasets from the 2005 national Colombian
census (provided by the planning department of the city of Cali), and (3) 2010 environmental
variables from the city of Cali.
The epidemiological dataset used in this paper was provided by the Health Secretariat of
the Municipality of the city of Cali, based on the Colombian Public Health Surveillance System
(SIVIGILA). A total of 11,760 cases were reported in the year 2010 at local health facilities. A
case is defined as a patient identified by a doctor at a health facility with dengue fever symptoms
and confirmed via lab work. The doctor files a form which is logged into the country’s
epidemiological surveillance system (SIVIGILA). Using the residential address of each patient,
we geocoded 9,404 cases which were then aggregated at the neighborhood level due the
difference in scale between the case data and the socioeconomic data (data is reported at the
individual level while socioeconomic data are reported at neighborhood level). For each
neighborhood, rates of dengue fever were estimated by dividing the aggregated counts by the
total population (range: 0 to 4.86 per 100 individuals). Neighborhoods representing parks and
military installations were removed, resulting in a revised sample of 9,287 dengue cases across
323 neighborhoods.
Several authors have isolated the role played by socioeconomic and environmental
determinants on the variation of dengue fever in urban areas. Although conclusions may vary by
locations, there exists a certain level of consensus. Greater population density (Hsueh, Lee, and
Beltz 2012; Khormi and Kumar 2011; de Mattos Almeida et al. 2007), higher levels of
urbanization (Wu et al. 2009), larger proportion of elderly individuals (Wu et al. 2009; Siqueira-
Junior et al. 2008), lower level of educational attainment and income (Siqueira-Junior et al.
2008; Koyadun, Butraporn, and Kittayapong 2012; de Mattos Almeida et al. 2007), proximity of
canals and ditches (Chiu et al. 2014) and poor household conditions (Khormi and Kumar 2011)
have been associated with higher rates of dengue fever. In this paper, we use a set of 36
environmental, socioeconomic and demographic variables motivated by the aforementioned
studies and also based on discussion with local experts (Hagenlocher et al. 2013). These
variables, which are all standardized between zero and one, are organized according to a risk and
social vulnerability framework (Table 1), where “hazard” variables are those factors influencing
the spread and distribution of the pathogen and disease transmitting vector (here: the mosquito),
and “social vulnerability” refers to the pre-disposition of the population to be negatively affected
by the disease as a result of their susceptibility and their lack of resilience, which comprises the
lacking capacity to anticipate infection, to cope with infection or to recover from infection
(Hagenlocher et al. 2013; Hagenlocher and Castro 2015).
To estimate the potential exposure of individuals in each neighborhood to the presence of
mosquitos in parks and parkways, rivers, tire shops, plant nurseries, water pumps and cemeteries
we used a kernel density estimation (KDE) approach, reflecting the density (i.e. probability
surface) of those features across the study area. A total of 9 cemeteries, 49 plant nurseries, 23
water pumping stations and 499 tire shops were geocoded, while 4 rivers (Cali, Aguacatal,
Cauca and Jarillon) were hand digitized; all served as an input for the KDE technique. To
compute proximity to rivers and green areas such as parks and parkways (linear and polygon
features), we converted those features to gridded points and used them as an input for the KDE
procedure. Zonal statistics at the neighborhood level were then implemented in a GIS to
summarize the average KDE value of the “hazard” variable for each of these neighborhoods.
<Table 1 around here>
2.3 Spatial autocorrelation
We use the global Moran’s I statistic (Moran 1948) to test the hypothesis that dengue fever
rates are not stationary across the study area. Its local version - the Local Moran’s I statistic
(Anselin 1995), maps local clusters of unusually high rates. The statistic evaluates the difference
in rate from one neighborhood i to its surrounding neighborhoods j. Positive [negative] values of
the statistic are indicative of a homogeneous cluster of high [low] rates. Both global and local
Moran’s I statistic are estimated in Geoda (Anselin, Syabri, and Kho 2006).
2.4 Geographically Weighted Regression
Global regression models such as the Ordinary Least Squares (OLS) capture the average
strength and significance of the predictor variables, but makes the assumption that rates at a
neighborhood i are independent of rates at neighboring units j and that the residuals are normally
distributed. As such, OLS does not consider small-scale spatial variation and spatial
autocorrelation in the residuals are indicative that those assumptions are violated (Anselin 2002).
There exist several techniques to incorporate spatial effects, such as the spatial regression (see
for instance Anselin 2009 for a summary) or the Geographically Weighted Regression (GWR)
technique, developed by Fotheringham, Brunsdon, and Charlton (2003). Unlike OLS or spatial
regression that provide global estimates, GWR is a local regression technique that is used to
measure how the strength of the relationships among the dependent and explanatory variables
differ from location to location. In essence, for each location, GWR pools data from adjacent
neighborhoods to conduct a local regression, resulting in an estimate of local regression
coefficients for each of the predictor variables. Using dengue fever rate as the dependent
variable, the GWR model can be expressed as follows:
�̂�(𝑖) = b0(i) + ∑ bk(𝑖)Xk(𝑖)Nk=1 + 𝜀(𝑖) (1)
with �̂�(𝑖) the predicted dengue fever rate at neighborhood i, 𝑏0(𝑖) the local intercept, bk(𝑖) the
coefficient of the kth variable at location i (k=1…N) and 𝜀(𝑖) the error term. GWR creates an
equation for every neighborhood, and its coefficients are calibrated using information from
nearby neighborhoods j. The importance of neighborhood j is weighted by the distance
separating neighborhood i and j. The ideal number of neighbors is determined by minimizing the
Akaike Information Criterion (AIC). Because each neighborhood has its own regression
equation, coefficients are allowed to vary across the study area. GWR results in a map of
coefficients, allowing to asses where each factor is a strong predictor of dengue fever, which is
particularly informative for public health prevention. Positive [negative] variable coefficients
(𝑏𝑘 < 0) suggest that an increase [decrease] in the amount of variable k, increases [decreases]
rates of dengue fever. Local regression coefficients are estimated in GWR 4.0 and visualized in
ArcGIS (ESRI, Redlands, USA). The GWR also results in a local coefficient of determination
R2, mapping the local strength of the model.
We use an OLS forward stepwise approach to select the variables entering the regression
model. Following this approach, variables are added sequentially until the coefficient of
determination, R2 does not increase significantly anymore. Significant variables are used in the
GWR model. We compare regression results to the ones obtained using an OLS approach.
3. Results
3.1Dengue fever rates
Table 2 and Figure 1(b) illustrates the rates of dengue fever per 100 individuals in the city
of Cali, suggesting substantial variation and the presence of a strong spatial autocorrelation
(Moran’s I = 0.34). Lower rates are observed in the eastern and northeastern region of the city
(numbers 4 and 5), while maximum rates of 4.86 cases per 100 individuals are observed both in
the center of town and the city’ southernmost neighborhoods.
<Figure 1 around here>
Following a local Moran’s I approach, clusters of higher rates are identified including 19
neighborhoods, which are grouped into homogeneous clusters numbered 1 through 3. Located in
the old core of the city and characterized by its strong economic activity, cluster 1 exhibits rates
ranging from 1.2 to 4.2 cases per 100 individuals. Although not as densely populated, individuals
are more at risk due to the presence of transient individuals. Cluster 2 (min: 1.3, max: 2.5 cases
per 100 individuals) is located on a military base area (Ejercito Tercera División) containing
several military housing complexes. Control processes for spraying of the mosquito are not the
same as in the rest of the city. Cluster 3 (min: 0.7, max: 4.86 cases per 100 individuals), which is
located in the southern edge of the city, is experiencing less housing development, but there is a
risk that transient population may import the virus into the area.
A total of 48 neighborhoods are characterized by lower rates and those are highlighted in
red in Figure 1. The population in region 4 is characterized by a low income status while a lack
of supply of health facilities reduces proper and timely access to care. The combination of these
two factors may contribute to underreporting. Low rates in region 5 were also attributed to
intensified spraying campaigns conducted by the Health Municipality1. In comparison to region
4, region 5 is safer in regards to criminal activity and as a result, control strategies can be
conducted more frequently.
3.2 Geographically Weighted Regression
Following an OLS stepwise regression, we identified six contributing socioeconomic and
environmental predictors (or risk factors) based on the 36 geospatial variables that significantly
contributed to the 2010 dengue fever outbreak (R2=0.295). These are proximity to tire shops,
proximity to plant nurseries, proportion of households with sewage system, proportion of
indigenous population, socioeconomic stratum2 and population density. We analyzed
collinearity metrics among the six variables (all variance inflation factor values were less than
1.632), suggesting no collinearity among the selected variables. Among those 6 variables, the
strongest correlation was between relative population density and socioeconomic stratum (r=-
0.541, p<0.01) and between relative population density and relative proportion of households
with sewage (r=-.292, p<0.01); individuals residing in neighborhood of higher socio-economic
stratum were more likely to have a sewage system and those neighborhoods also had lower
population density. We modified our OLS model to include interaction terms (population
density* socio-economic stratum) and (population density * proportion of households with
sewage system), however this did not improve our OLS model fit (in terms of R2: 0.201 and
0.182, respectively) from 0.295 (with the 6 variables mentioned earlier).
The spatial distribution of the six final variables that were used in the GWR is mapped in
Figure 2, while Table 2 summarizes the range of the regression coefficients for each predictor
variables.
<Figure 2 and Table 2 around here>
Our GWR model explains 64% of the variance among dengue fever rates (29.5% in a
standard OLS). Figure 3 illustrates the local variation of coefficients (𝑏𝑘) used in the GWR
procedure, and can be interpreted as follows: for each variable, a red [blue] colored region
suggests that an increase in the amount of that variable will increase [decrease] rates of dengue
fever. We use contour maps of varying sizes to map the standard error of the estimates for each
predictor variable in Figure 3(a)-(f), suggesting substantial greater error in the peripheral
neighborhoods. On the same Figure, local R2 values suggest an uneven coefficient of
determination across the city.
<Figure 3 around here>
1 Personal communication with J.H. Rojas, epidemiologist in the Secretaria de Salud Municipal, Cali. 2 Neighborhoods are classified based on socioeconomic strata into six classes, six being the highest.
In the hazard category, the association between dengue fever rates and the proximity to
tire shops is generally positive (𝑏𝑡𝑖𝑟𝑒 = 0.0023), but that trend declines eastward. Second, the
proximity to plant nurseries (𝑏𝑛𝑢𝑟𝑠 = 0.0071) exhibits a strong localized impact in the central
part of the city, coinciding with clusters 1 on Figure 1b (the southern part of the city also exhibits
a strong localized impact, but its standard error is very large). Third, a higher proportion of
households equipped with a sewage system is likely to reduce dengue fever rates (𝑏𝑠𝑒𝑤 =
−0.0073); however this trend is particularly true in areas south of the city (especially areas
around cluster 2 in Figure 1b). Neighborhoods in the southern part of the city are more modern
with a better sewage system and we hypothesize that individuals residing in this area may
maintain better sanitary practices of not keeping water containers than if they had lived in an
older neighborhood with a similar stratification. In the social vulnerability category, higher
proportions of indigenous population and higher socioeconomic stratum has a modest, yet
significant impact on the rate (𝑏𝑖𝑛𝑑 = 0.0122, 𝑏𝑠𝑡𝑟𝑎𝑡 = 0.0023), while higher population density
has a slight negative impact (𝑏𝑝𝑜𝑝 = −0.0067). The coefficient reflecting the proportion of
indigenous people in the local population demonstrates substantial variation from the northern
part of the city (slightly negative association) to the south (stronger positive association). Higher
socioeconomic strata in the center part of the city was associated with lower levels of dengue
fever rates (around cluster 1, Figure 1b). Interestingly, in the southern part of the city, houses are
typically bigger with relatively larger yards, which may provide a suitable habitat for the vector
to breed. Finally, we observe that population density has a slight negative impact on dengue
fever rates, but exhibits little spatial variation. Some of the densely populated neighborhoods in
the eastern part of Cali have a high concentration of Afro-Colombian population, which has been
suggested to be less susceptible to the virus (Villegas et al. 2010).
Overall, the local coefficient of determination R2(i) suggests a strong predictive power in
the central and southern part of the city (coinciding with cluster 1 and 3 in Figure 1b). Lower
R2(i) values in the southwestern part of the city (cluster 2 in Figure 1b) indicates a poorer
regression fit. We came to realize that this area coincides with military barracks around the
Ejército Tercera División. This area suffered from a large number of dengue fever cases in 2010,
possibly owing to the presence of military personnel living in close proximity from one another.
The area also houses important universities (Arquelogico Univalle, Universidad Santiago De
Cali and the Universidad Cooperativa de Colombia Sede Cali). Proximity to schools (including
universities) and military barracks were not included in our model.
Figures 4(a)-(b) summarize the spatial distribution of residuals and predicted dengue
fever rates3, respectively. Overall, residuals are confined in the range [-1.17 to 1.39] cases per
100 individuals with a mean nearly equal to 0. (compared to [-0.0145, 0.0336] following the
standard OLS). Although the spatial pattern in the predicted rates is higher than in the observed
rates (Moran’s I = 0.54) -likely due the GWR smoothing effects, there is no apparent clustering
3 For comparison purposes legend breaks and color schemes of Figure 4b were kept similar to the ones of
Figure 1b (observed rates).
in the residuals (Moran’s I = 0.005, compared to 0.133 for the standard OLS), indicative of a
random pattern (Table 2). Following a local Moran’s I on the predicted rates, higher rates are
identified in 28 neighborhoods (14 of those clusters matched with the 19 from the observed
rates) and lower rates in 62 neighborhoods (44 of those clusters match with the 48 from the
observed rates).
<Figure 4 around here>
4. Discussion Considering both environmental and socioeconomic factors as predictors of dengue
outbreaks was previously highlighted by other studies (Mondini and Chiaravalloti-Neto 2008;
Hsueh, Lee, and Beltz 2012; Koyadun, Butraporn, and Kittayapong 2012). While the community
seems to agree on the importance of environmental factors impacting the spread and distribution
of the vector and the pathogen, the choice and selection of socioeconomic variables reflecting the
vulnerability of the population, seems often to be context and scale specific, largely driven by
data availability. A generalization of the role played by socioeconomic determinants may be
challenging; however the findings of our work exhibit similarities with previous literature, such
as the role of population density (Hsueh, Lee, and Beltz 2012; Khormi and Kumar 2011; de
Mattos Almeida et al. 2007; Koyadun, Butraporn, and Kittayapong 2012) and socioeconomic
status (Siqueira-Junior et al. 2008; Koyadun, Butraporn, and Kittayapong 2012; de Mattos
Almeida et al. 2007). Finally, our study contributes to a growing body of recent literature
underlining the usefulness of regression techniques which incorporates spatial effects in the
prediction of dengue fever rates (Khormi and Kumar 2011; Lin and Wen 2011; Hsueh, Lee, and
Beltz 2012; Baharuddin, Suhariningsih, and Ulama 2014).
Our results underline the need of taking into account both environmental factors favoring
local transmission as well as those socioeconomic factors contributing to social vulnerability of
the population when assessing risks associated with vector-borne diseases. Interestingly, the six
factors identified as statistically significant predictors of dengue fever and their relationship with
dengue coincided with four factors proposed by local domain experts (n =4) during an earlier
study in the same area (Hagenlocher et al. 2013), suggesting that control strategies should
incorporate some of the dengue predictors explored in this paper. Rates of dengue fever were
shown to have a non-stationary relationship with the six covariates, further underlining the need
to incorporate spatial effects in predictive models, as these may reflect locally varying
determinants not captured in traditional regression models. GWR metrics (R2, range of residuals
and autocorrelation in the residuals) were better than when using the standard OLS approach.
This study relies on several assumptions that may affect the validity of our results. First,
although the set of input variables selected in this study are those most often considered in
dengue risk assessments, several important variables are missing due to a lack of data
availability. Among others, these include intra-urban mobility and migration patterns, quality of
the health care system, treatment seeking behavior of different social groups as well as extent
and coverage of dengue control programs (Eisen and Lozano-Fuentes 2009). Lower local R2(i)
values in some neighborhoods of the city need to be further investigated, for instance by the
inclusion of additional variables. Second, we did not differentiate among the features that make
up the “hazard” variables. The size and cleanliness of a tire shop or a plant nursery may impact
the presence of the mosquito and increase the risk of dengue fever. Third, both “hazard” and
“vulnerability” variables used in this study were drawn from different years. Fourth, our
geocoding process used the place of residence of each patient diagnosed with dengue fever, and
this location may not be the place where the patient was infected (Vazquez-Prokopec et al.
2009).
Despite these limitations, our approach demonstrated several strengths. First, the
methodological framework is transferable to urban environments in other regions (e.g. Asia,
Africa, etc.) and at different spatial scales (depending on data availability), as well as to different
mosquito-borne diseases influenced by a similar complex interplay of environmental and
socioeconomic factors (e.g. malaria, chikungunya, Zika fever, etc.). A second major strength lies
in the comprehensive selection of potential risk factors/predictors (n = 36), representing those
factors that are most commonly used in dengue risk assessments. Finally, the regression
technique used in this paper incorporated spatial effects and reduced the spatial pattern in the
residuals.
5. Conclusions Using a Geographically Weighted Regression (GWR) model we demonstrated the role of
socioeconomic and environmental factors associated with a dengue outbreak that affected the
city of Cali, Colombia, in 2010. The results presented in this paper make two important
contributions. First, both environmental factors impacting the spread and distribution of the
vector and pathogen as well as socioeconomic factors shaping social vulnerability need to be
considered in disease risk assessments. Second, the contribution of these factors varies spatially,
thus demanding approaches that are able to capture this variability. Our findings may not have
been identified by traditional regression analysis. This further underlines the importance of
small-scale level studies.
Our findings also have deep ramifications for health policy and decision-making, since
interventions aiming at reducing 2016) risk need to be targeted locally to optimize health
outcomes and addressing the specific local conditions that ultimately result in high risk levels
(Ndeffo-Mbah, 2016) We suggest that efforts in control strategies and prevention be greater in
areas where both environmental and socioeconomic factors are known to pose a significant risk.
For instance, educating owners of tire shops and nurseries to monitor for the presence of larvae’s
could curb the spread of mosquitos, while increasing the number of households with sewage
system is also likely to reduce the potential for vector breeding.
We recognize that further research will be needed to evaluate whether the factors identified
in this paper as significant predictors of dengue fever at an urban scale (1) are also valid for other
urban environments with similar characteristics, (2) generalizable to coarser scales and (3)
whether our findings (i.e. relationships) would hold during other epidemic years. Additionally, it
would be worth investigating whether local strategies which were implemented to eradicate the
vector (and associated prevention campaigns) were beneficial in reducing the number of cases in
subsequential years. Further, models incorporating climate factors such as mean monthly
temperature, rainfall or humidity will likely provide better dengue fever estimates. This is
particularly important at national and regional scales, where spatial variations in climatic
conditions manifest. It would be interesting to compare the results of our approach with a mixed-
effect model where the study region is included as a random effect. The merits of other
spatially-explicit models are worth investigating, such as (1) the spatial regression (Anselin,
Syabri, and Kho 2006), (2) count regressions such as in Linard et al. (2007) who combine count
data with generalized linear models and (3) the geographically weighted Poisson regression
(Nakaya et al. 2005). These approaches are likely to further reduce the problem of spatial
dependence.
Acknowledgment The authors would like to thank Jorge H. Rojas, epidemiologist at the Secretaría de Salud
Pública Municipal in Cali, Colombia for his assistance with the manuscript.
References Alirol, E., L. Getaz, B. Stoll, F. Chappuis, and L. Loutan. 2011. Urbanisation and infectious diseases in a globalised
world. The Lancet Infectious Diseases 11 (2):131-141.
Anselin, L. Spatial regression. In: Fotheringham A. and P.R. Rogerson. 2009. The SAGE handbook of spatial
analysis, 1, 255-276.
Anselin, L. 1995. Local indicators of spatial association – LISA. Geographical Analysis 27(2):93-115.
Anselin, L. 2002. Under the hood issues in the specification and interpretation of spatial regression
models. Agricultural economics, 27(3), 247-267.
Anselin, L., I. Syabri, and Y. Kho. 2006. GeoDa: an introduction to spatial data analysis. Geographical Analysis 38
(1):5-22.
Baharuddin, B., S. Suhariningsih, and B. S. S. Ulama. 2014. Geographically Weighted Regression Modeling for
Analyzing Spatial Heterogeneity on Relationship between Dengue Hemorrhagic Fever Incidence and
Rainfall in Surabaya, Indonesia. Modern Applied Science 8 (3):p85.
Bates, I., C. Fenton, J. Gruber, D. Lalloo, A. Lara, S. Squire, S. Theobald, R. Thomson, and R. Tolhurst. 2004a.
Vulnerability to malaria, tuberculosis, and HIV/AIDS infection and disease. Part 1: determinants operating
at individual and household level. The Lancet Infectious Diseases 4(5):267 - 277.
Bates, I., C. Fenton, J. Gruber, D. Lalloo, A. M. Lara, S. B. Squire, S. Theobald, R. Thomson, and R. Tolhurst.
2004b. Vulnerability to malaria, tuberculosis, and HIV/AIDS infection and disease. Part II: Determinants
operating at environmental and institutional level. The Lancet Infectious Diseases 4(6):368-375.
Bhatt, S., P. W. Gething, O. J. Brady, J. P. Messina, A. W. Farlow, C. L. Moyes, J. M. Drake, J. S. Brownstein, A.
G. Hoen, and O. Sankoh. 2013. The global distribution and burden of dengue. Nature 496(7446), 504-507.
Braga, C., C. Luna, C. Martelli, W. de Souza, M. Cordeiro, N. Alexander, M. de Albuquerque, J. Junior, and E.
Marques. 2010. Seroprevalence and risk factors for dengue infection in socioeconomically distinct areas of
Recife Brazil. Acta Tropica 113(3):234 - 240.
Cali, S. d. S. P. M. d. 2010. Historia del dengue en Cali. Endemia o una continua epidemia. Cali: Secretaria de
Salud Publica Municipal de Cali.
Castillo, K. C., B. Körbl, A. Stewart, J. F. Gonzalez, and F. Ponce. 2011. Application of spatial analysis to the
examination of dengue fever in Guayaquil, Ecuador. Procedia Environmental Sciences 7:188-193.
Chiaravalloti-Neto, F., M. Pereira, E. A. Fávaro, M. R. Dibo, A. Mondini, A. L. Rodrigues-Junior, A. P. Chierotti,
and M. L. Nogueira. 2015. Assessment of the relationship between entomologic indicators of Aedes
aegypti and the epidemic occurrence of dengue virus 3 in a susceptible population, São José do Rio Preto,
São Paulo, Brazil. Acta Tropica 142:167-177.
Chiu, C.-H., T.-H. Wen, L.-C. Chien, and H.-L. Yu. 2014. A Probabilistic Spatial Dengue Fever Risk Assessment by
a Threshold-Based-Quantile Regression Method. PloS one 9 (10):e106334.
Chowell, G., C. Torre, C. Munayco-Escate, L. Suarez-Ognio, R. Lopez-Cruz, J. Hyman, and C. Castillo-Chavez.
2008. Spatial and temporal dynamics of dengue fever in Peru: 1994–2006. Epidemiology and Infection 136
(12):1667-1677.
de Mattos Almeida, M. C., W. T. Caiaffa, R. M. Assunçao, and F. A. Proietti. 2007. Spatial vulnerability to dengue
in a Brazilian urban area during a 7-year surveillance. Journal of Urban Health 84 (3):334-345.
Delmelle, E., I. Casas, J. H. Rojas, and A. Varela. 2013. Spatio-temporal patterns of Dengue Fever in Cali,
Colombia. International Journal of Applied Geospatial Research 4 (4):58-75.
Delmelle, E., C. Dony, I. Casas, M. Jia, and W. Tang. 2014. Visualizing the impact of space-time uncertainties on
dengue fever patterns. International Journal of Geographical Information Science 28 (5):1107-1127.
Delmelle, E. M., H. Zhu, W. Tang, and I. Casas. 2014. A web-based geospatial toolkit for the monitoring of dengue
fever. Applied Geography 52:144-152.
Dick, O. B., J. L. San Martín, R. H. Montoya, J. del Diego, B. Zambrano, and G. H. Dayan. 2012. The history of
dengue outbreaks in the Americas. The American Journal of Tropical Medicine and Hygiene 87 (4):584-
593.
Eastin, M. D., E. Delmelle, I. Casas, J. Wexler, and C. Self. 2014. Intra-and Interseasonal Autoregressive Prediction
of Dengue Outbreaks Using Local Weather and Regional Climate for a Tropical Environment in Colombia.
The American Journal of Tropical Medicine and Hygiene 91 (3):598-610.
Eisen, L., and S. Lozano-Fuentes. 2009. Use of mapping and spatial and space-time modeling approaches in
operational control of aedes aegypti and dengue. PLoS Neglected Tropical Diseases 3 (4):e411.
Fotheringham, A. S., C. Brunsdon, and M. Charlton. 2003. Geographically weighted regression: the analysis of
spatially varying relationships: John Wiley & Sons.
Gardner, L., and S. Sarkar. 2013. A global airport-based risk model for the spread of dengue infection via the air
transport network. PloS one 8 (8):e72129.
Getis, A., A. Morrison, K. Gray, and T. Scott. 2003. Characteristics of the spatial pattern of the dengue vector,
Aedes aegypti, in Iquitos, Peru. American Journal of Tropical Medicine and Hygiene 69 (5):494 - 505.
Guzman, G., and G. Kouri. 2003. Dengue and dengue hemorrhagic fever in the Americas: lessons and challenges.
Journal of Clinical Virology 27 (1):1-13.
Hagenlocher, M., and M. C. Castro. 2015. Mapping malaria risk and vulnerability in the United Republic of
Tanzania: a spatial explicit model. Population Health Metrics 13 (1):2.
Hagenlocher, M., E. Delmelle, I. Casas, and S. Kienberger. 2013. Assessing socioeconomic vulnerability to dengue
fever in Cali, Colombia: statistical vs expert-based modeling. International Journal of Health Geographics
12 (1):36.
Honorio, N., R. Nogueira, C. Codeco, M. Carvalho, O. Cruz, M. Magalhaes, J. De Araujo, E. de Araujo, M. Gomes,
L. Pinheiro, C. da Silva Pinel, and R. Lourenco-de-Oliveira. 2009. Spatial evaluation and modeling of
Dengue seroprevalence and vector density in Rio de Janeiro Brazil. PLoS Neglected Tropical Diseases 3
(11):1 - 11.
Hsueh, Y.-H., J. Lee, and L. Beltz. 2012. Spatio-temporal patterns of dengue fever cases in Kaoshiung City, Taiwan,
2003–2008. Applied Geography 34:587-594.
Istúriz, R. E., D. J. Gubler, and J. B. d. Castillo. 2000. Dengue and dengue hemorrhagic fever in Latin America and
the Caribbean. Infectious Disease Clinics of North America 14 (1):121-140.
Khormi, H. M., and L. Kumar. 2011. Modeling dengue fever risk based on socioeconomic parameters, nationality
and age groups: GIS and remote sensing based case study. Science of the Total Environment 409 (22):4713-
4719.
Kienberger, S., M. Hagenlocher, E. Delmelle, and I. Casas. 2013. A WebGIS tool for visualizing and exploring
socioeconomic vulnerability to dengue fever in Cali, Colombia. Geospatial Health 8 (1):313-316.
Koyadun, S., P. Butraporn, and P. Kittayapong. 2012. Ecologic and sociodemographic risk determinants for dengue
transmission in urban areas in Thailand. Interdisciplinary Perspectives on Infectious Diseases 2012.
Lin, C. H., and T. H. Wen. 2011. Using geographically weighted regression (GWR) to explore spatial varying
relationships of immature mosquitoes and human densities with the incidence of dengue. International
Journal of Environmental Research and Public Health 8 (7):2798-2815.
Linard, C., Lamarque, P., Heyman, P., Ducoffre, G., Luyasu, V., Tersago, K., Vanwambeke S. and Lambin, E. F..
2007. Determinants of the geographic distribution of Puumala virus and Lyme borreliosis infections in
Belgium. International Journal of Health Geographics, 6(1), 1.
Messina, Jane P., Oliver J. Brady, Thomas W. Scott, Chenting Zou, David M. Pigott, Kirsten A. Duda, Samir Bhatt
et al. 2014. "Global spread of dengue virus types: mapping the 70 year history." Trends in microbiology 22
(3): 138-146.
Messina, J. P., O. J. Brady, D. M. Pigott, N. Golding, M. U. Kraemer, T. W. Scott, G. W. Wint, D. L. Smith, and S.
I. Hay. 2015. The many projected futures of dengue. Nature Reviews Microbiology.
Mondini, A., and F. Chiaravalloti-Neto. 2008. Spatial correlation of incidence of dengue with socioeconomic,
demographic and environmental variables in a Brazilian city. Science of the Total Environment 393
(2):241-248.
Moran, P. A. 1948. The interpretation of statistical maps. Journal of the Royal Statistical Society. Series B
(Methodological) 10 (2):243-251.
Nogueira, R. M. R., J. M. G. d. Araújo, and H. G. Schatzmayr. 2007. Dengue viruses in Brazil, 1986-2006. Revista
Panamericana de Salud Publica 22 (5):358-363.
Nakaya, T., Fotheringham, A. S., Brunsdon, C., & Charlton, M. 2005. Geographically weighted Poisson regression
for disease association mapping. Statistics in medicine, 24(17), 2695-2717.
Ndeffo-Mbah, Martial L., David P. Durham, Laura A. Skrip, Elaine O. Nsoesie, John S. Brownstein, Durland Fish,
and Alison P. Galvani. 2016. Evaluating the effectiveness of localized control strategies to curtail
chikungunya. Scientific reports 6.
Nunes, M. R., G. Palacios, N. R. Faria, E. C. Sousa Jr, J. A. Pantoja, S. G. Rodrigues, V. L. Carvalho, D. B.
Medeiros, N. Savji, and G. Baele. 2014. Air travel is associated with intracontinental spread of dengue
virus serotypes 1–3 in Brazil. PLoS Neglected Tropical Diseases 8 (4):e2769.
Ocampo, C. B., N. J. Mina, M. Carabalí, N. Alexander, and L. Osorio. 2014. Reduction in dengue cases observed
during mass control of Aedes (Stegomyia) in street catch basins in an endemic urban area in Colombia.
Acta tropica 132:15-22.
OPS. 1960. Guia de los informes de la campana de erradicacion del Aedes aegypti en las Americas. In Pan
American Health Organization, 15. Washington.
Restrepo, A. C., P. Baker, and A. C. Clements. 2014. National spatial and temporal patterns of notified dengue
cases, Colombia 2007–2010. Tropical Medicine & International Health 19 (7):863-871.
Rodriguez-Barraquer, I., M. T. Cordeiro, C. Braga, W. V. de Souza, E. T. Marques, and D. A. Cummings. 2011.
From re-emergence to hyperendemicity: the natural history of the dengue epidemic in Brazil. PLoS
Neglected Tropical Diseases 5 (1):e935.
Rogers, D., and S. Randolph. 2006. Climate change and vector-borne diseases. Advances in Parasitology 62:345-
381.
Romero-Vivas, C. M., C. J. Leake, and a. K. Falconar. 1998. Determination of dengue virus serotypes in individual
Aedes aegypti mosquitoes in Colombia. Medical and Veterinary Entomology 12:284-288.
Rotela, C., F. Fouque, M. Lamfri, P. Sabatier, V. Introini, M. Zaidenberg, and C. Scavuzzo. 2007. Space–time
analysis of the dengue spreading dynamics in the 2004 Tartagal outbreak, Northern Argentina. Acta
Tropica 103 (1):1-13.
San Martín, J. L., O. Brathwaite, B. Zambrano, J. O. Solórzano, A. Bouckenooghe, G. H. Dayan, and M. G.
Guzmán. 2010. The epidemiology of dengue in the Americas over the last three decades: a worrisome
reality. The American Journal of Tropical Medicine and Hygiene 82 (1):128-135.
Semenza, J. C., J. E. Suk, V. Estevez, K. L. Ebi, and E. Lindgren. 2012. Mapping climate change vulnerabilities to
infectious diseases in Europe. Environmental Health Perspectives 120 (3):385.
Siqueira-Junior, J. B., I. J. Maciel, C. Barcellos, W. V. Souza, M. S. Carvalho, N. E. Nascimento, R. M. Oliveira, O.
Morais-Neto, and C. M. Martelli. 2008. Spatial point analysis based on dengue surveys at household level
in central Brazil. BMC Public Health 8 (1):361.
Stoddard, S. T., B. M. Forshey, A. C. Morrison, V. A. Paz-Soldan, G. M. Vazquez-Prokopec, H. Astete, R. C.
Reiner, S. Vilcarromero, J. P. Elder, and E. S. Halsey. 2013. House-to-house human movement drives
dengue virus transmission. Proceedings of the National Academy of Sciences 110 (3):994-999.
Teixeira, M. d. G., M. L. Barreto, M. d. C. N. Costa, L. D. A. Ferreira, P. F. Vasconcelos, and S. Cairncross. 2002.
Dynamics of dengue virus circulation: a silent epidemic in a complex urban area. Tropical Medicine &
International Health 7 (9):757-762.
Teixeira, M. G., M. d. C. N. Costa, F. Barreto, and M. L. Barreto. 2009. Dengue: twenty-five years since
reemergence in Brazil. Cadernos de Saúde Pública 25:S7-S18.
Torres, J. R., and J. Castro. 2007. The health and economic impact of dengue in Latin America. Cadernos de Salude
Publica 23:S23-S31.
Tran, A., X. Deparis, P. Dussart, J. Morvan, P. Rabarison, F. Remy, L. Polidori, and J. Gardon. 2004. Dengue spatial
and temporal patterns, French Guiana, 2001. Emerging Infectious Diseases 10 (4):615-621.
Vazquez-Prokopec, G. M., U. Kitron, B. Montgomery, P. Horne, and S. A. Ritchie. 2010. Quantifying the spatial
dimension of dengue virus epidemic spread within a tropical urban environment. PLoS Neglected Tropical
Diseases 4 (12):e920.
Vazquez-Prokopec, G. M., S. T. Stoddard, V. Paz-Soldan, A. C. Morrison, J. P. Elder, T. J. Kochel, T. W. Scott, and
U. Kitron. 2009. Usefulness of commercially available GPS data-loggers for tracking human movement
and exposure to dengue virus. International Journal of Health Geographics 8 (1):68.
Vezzani, D., and A. E. Carbajo. 2008. Aedes aegypti, Aedes albopictus, and dengue in Argentina: current
knowledge and future directions. Memorias do Instituto Oswaldo Cruz 103 (1):66-74.
Villegas, A. V., E. G. Aristizabal, and J. H. Rojas. 2010. Analisis epidemiologico de dengue en Cali. Cali: Secretaria
de Salud Publica Municipal.
Wilson, M. E., and L. H. Chen. 2002. Dengue in the Americas. Dengue Bulletin 26:44-61.
Wu, P.-C., J.-G. Lay, H.-R. Guo, C.-Y. Lin, S.-C. Lung, and H.-J. Su. 2009. Higher temperature and urbanization
affect the spatial patterns of dengue fever transmission in subtropical Taiwan. Science of the Total
Environment 407 (7):2224-2233.
List of Figures
Figure 1: Aggregated counts of 2010 reported dengue fever cases for the city of Cali, Colombia
in (a) and associated rates per 100 individuals in (b). Clusters of high rates are identified using
the Local Moran’s I and are outlined in black.
Figure 2: Spatial distribution of the socioeconomic and environmental variables selected in the
GWR model (standardized from 0 to 1), arranged according to the hazard & vulnerability
framework.
Figure 3: Local coefficients for the six explanatory variables used in the GWR model (a-f) and
associated standard error (SE), where a thinner line indicates a smaller SE. The figure to the
upper right summarizes the local R2.
Figure 4: Residuals and predicted dengue fever rates, in (a) and (b), respectively. Higher values
are determined using a Local Moran’s I statistic, and highlighted in black. Use observed rates in
Figure 1b for comparison.
List of Tables
Table 1. List of socioeconomic, demographic and environmental variables used in this study. All
variables were standardized in the [0, 1] interval.
Table 2: Coefficients and model performance for the GWR model. The Moran’s I statistic
reflects the global and local autocorrelation in predicted rates and associated residuals.
Figure 1
Figure 2
Figure 3
Figure 4
Table 1
Year Source Year Source
HAZARD SOCIAL VULNERABILITY
Environmental factors Lack of resilience
Relative proximity to parks 2010 Cali Households with phone (%) 2005 Census
Relative proximity to rivers 2010 Cali Individuals who can read/write (%) 2005 Census
Relative proximity to tire shops 2010 Cali Individuals who cannot read/write (%) 2005 Census
Relative proximity to water pumps 2010 Cali Individuals with no education (%) 2005 Census
Relative proximity to cemeteries 2010 Cali Individuals with low education (%) 2005 Census
Relative proximity to plant nurseries 2010 Cali Individuals with medium education (%) 2005 Census
Individuals with high education (%) 2005 Census
SOCIAL VULNERABILITY Mean hospital density (km2) 2010 Cali
Susceptibility Travel time to hospital 2010 Cali
Population density (km2) 2005 Census Walking distance to hospital 2010 Cali
Density of occupied households (per km2) 2005 Census Employed individuals (%) 2005 Census
Individuals of age 0-4 (%) 2005 Census Unemployed individuals (%) 2005 Census
Individuals of age 5-14 (%) 2005 Census Retired individuals (%) 2005 Census
Individuals of age 15-29 (%) 2005 Census Doing housework (%) 2005 Census
Individuals of age >30 (%) 2005 Census
Indigeneous population (%) 2005 Census
White population (%) 2005 Census Sources
Black population (%) 2005 Census Cali: City of Cali
Individuals with disabilities (%) 2005 Census Census: 2005 national Colombian census
Households with sewer (%) 2005 Census
Households with water (%) 2005 Census
Stratum (estrata) 2005 Census
Table 2
COEFFICIENTS Beta(min) Beta(max) Beta(avg)
HAZARD/DISEASE
proximity to tire shops -0.0271 0.0179 0.0023
proximity to plant nurseries -0.0057 0.0276 0.0071
proportion of housing with sewage -0.0308 0.0003 -0.0073
SOCIAL VULNERABILITY
proportion of indegeneous people -0.0117 0.2870 0.0122
SES stratum (estrata) -0.0049 0.0179 0.0023
population density -0.0162 0.0071 -0.0067
intercept -0.0076 0.0098 0.0069
MODEL PERFORMANCE
R2
residuals (range)
residuals (mean)
Moran's I (residuals)
Moran's I (predicted)
AIC -2694.00
0.6446
(-0.01177; 0.013941)
-0.000026
0.005
0.541