NOTE JURNAL TENTANG ANALISIS SPASIAL

An environmental assessment and risk map

of Ascaris lumbricoides and Necator americanus
distributions in Manufahi District, Timor-Leste
Rebecca Wardell1, Archie C. A. Clements1, Aparna Lal2, David Summers3,
Stacey Llewellyn4, Suzy J. Campbell1, James McCarthy4,5, Darren J. Gray1,5,6,
Susana V. Nery1*

To determine if the final models adequately accounted for spatial autocorrelation, semivariograms
were created based on the village random intercepts for each given model. Semivariograms
describe how data are related with respect to distance and direction, with a
semivariogram presenting the similarity between observations (semivariance) at different
separation distances. [39] Omni-directional semivariograms based on the village random
intercepts were developed in R using the geoR package.[40] Lack of distinctive spatial patterns
in these semivariograms (S1 Fig) supported that the final models adequately accounted for spatial
clustering at the village level.

Water, Sanitation and Hygiene (WASH) and

environmental risk factors for soil-transmitted
helminth intensity of infection in Timor-Leste,
using real time PCR
Suzy J. Campbell1*, Susana V. Nery1, Rebecca Wardell1, Catherine A. D'Este1, Darren
J. Gray1,2,3, James S. McCarthy3,4, Rebecca J. Traub5, Ross M. Andrews6,
Stacey Llewellyn4, Andrew J. Vallely7, Gail M. Williams3, Archie C. A. Clements1

Separate assessment of spatial autocorrelation was undertaken using semivariograms

of residuals from multivariable models of selected environmental variables, with
household and village random effects [19]; no additional autocorrelation was identified [19].

Bayesian Geostatistical Model-Based

Estimates of Soil-Transmitted Helminth
Infection in Nigeria, Including Annual
Deworming Requirements
Akinola S. Oluwole1, Uwem F. Ekpo1*, Dimitrios-Alexios Karagiannis-Voules2,3, Eniola
M. Abe1,4, Francisca O. Olamiju5, Sunday Isiyaku6, Chukwu Okoronkwo7, Yisa Saka7,
Obiageli J. Nebe7, Eka I. Braide4, Chiedu F. Mafiana8, Jürg Utzinger2,3,
Penelope Vounatsou2,3

(Cakupan satu negara)

We applied Bayesian binomial geostatistical models to relate STH infection risk with environmental
and socioeconomic predictors. We used integrated nested Laplace approximations
(INLA) [23] and a stochastic partial differential equations approach [24] for fast approximate
Bayesian inference. Analysis was carried out in R [25] and the INLA package (www.r-inla.org).
Details of how models were implemented are provided in Supporting Information (S1 Text)
[24,26,27].
We followed an approach detailed by Karagiannis-Voules et al. [28], which has also been
used for STH geostatistical modeling in Cambodia [29], to select the best predictive model. In
brief, we fitted Bayesian bivariate geostatistical models to select the functional form of the effect
of each predictor based on the cross-validated logarithmic score [30,31]. We considered linear
and categorical functional forms of effects. The categorical functional form of the covariates
was generated using 25th, 50th, and 75th percentile to group each covariate into specific categories.
Non-linearity was addressed through random walk processes of order 1 and 2 [32].
The
form of each predictor giving the lowest mean logarithmic score was chosen. To identify the set
of the most important predictors, we fitted geostatistical models with all possible combination of
covariates (i.e., 256 models for each STH species-specific infection) and selected the one, for
each of the three STH species, with the best logarithmic score. The final models were used to
predict infection risk at a grid of 3 x 3 km including areas where infection data were not available.
The form of the covariate that was included in the final model used in the prediction of
each species of STH is shown in Table 2. The posterior estimates and Bayesian credible intervals
for the effects of the predictors are presented in odds ratios. Additional details are provided
in Supporting Information (S1 Text).

Due to the large number of observed zero prevalence data, we additionally fitted zeroinflated
binomial models with invariant probability of zero-inflation. These models have
shown better predictive ability in geostatistical modeling of malaria. [33]. In the present study,
the zero-inflated models did not improve predictions (based on the cross-validated logarithmic
score). Hence, we report results from the binomial models.

Mapping the Risk of Soil-Transmitted

Helminthic Infections in the Philippines
Ricardo J. Soares Magalhães1,2,3☯*, Maria S. Salamat4☯, Lydia Leonardo4, Darren
J. Gray3,5, Hélène Carabin6, Kate Halton7, Donald P. McManus8, Gail M. Williams3,
Pilarita Rivera4, Ofelia Saniel9, Leda Hernandez10, Laith Yakob11, Stephen T. McGarvey12,
Archie C. A. Clements3,5

(Cakupan satu negara)

Georeferencing of barangays
The unit of analysis was the barangay, the smallest administrative unit in the Philippines. The
mean length of the longest axis of barangays was 11km (SD:10.3). Barangay centroids were estimated
using the geographical information system (GIS) software QuantumGIS (QGIS) version
1.7.3 (QGIS Development Team, 2011). This procedure was based on combined information
from shapefiles of the barangays of the Philippines, which were obtained from the geographic
data warehouses DIVA GIS (www.diva-gis.org/Data) and PhilGIS (www.philgis.org) for the
Philippines. A total of 214 barangays in Luzon, the Visayas and Mindanao were included in the
analysis (Fig 1).

Variable selection and residual spatial variation

For the purpose of the analysis the presence of parasite eggs in stool identified by the Kato-
Katz method was used as the outcome variable and thus all individuals were categorized into
infected and not infected based on the presence of at least one egg. Initial variable selection
included age and sex because these two factors have been shown to be associated with STH
infections probably by influencing exposure and susceptibility to infection [1]. The environmental
variables (rainfall, DPWB, LST and NDVI) were also considered in the initial variable
screening stage. Correlations between environmental covariates were investigated using
Pearson’s correlation coefficients. Scatter plots were used to assess the relationship between the
barangay-level STH prevalence and the value of each of the environmental variables. Multivariable
logistic regression models for a Bernoulli-distributed outcome, with cluster correction by
barangay using robust standard errors, were built for each STH species for each region of the
Philippines, i.e., Luzon, the Visayas, and Mindanao using the statistical software Stata version
10.1 (Stata corporation, College Station, TX). Residuals of the final non-spatial models were
examined for spatial autocorrelation by generating a semivariogram using the geoR package of
R software v.2.15. One semivariogram was generated for each STH species for each region of
the Philippines to determine how much of the clustering of STH infections is explained by
location, and to establish the propensity and size of geographical clusters.

Spatial risk prediction and model validation

Bayesian logistic geostatistical models were built for each STH species for the regions of Luzon
and the Visayas combined and for the region of Mindanao separately using WinBUGS (MRC
Biostatistics Unit, Cambridge, and Imperial College London, UK). This decision was based on
the results of the semivariogram analysis (Fig 2), which indicated similar spatial dependence in STH
prevalence in the regions of Luzon and the Visayas.

The models included an intercept, the

individual level variables age (categorized into children aged <5 years and 5–19 years, and
adults aged >20 years) and sex, the environmental variables DPWB, LST, NDVI and rainfall,
and a geostatistical random effect (S1 Text). In addition, the model included adjustment for
diagnostic uncertainty by modeling sensitivity and specificity as random variables. The Kato-
Katz technique is widely used for detecting helminth eggs in stools. The sensitivity of the test is
influenced by changes in the number of eggs excreted in the feces from day to day. True prevalence
was modeled as a function of the observed prevalence and test sensitivity and specificity,
with the generalised linear model fit to the true prevalence parameter. Priors for the sensitivity
and specificity were specified as beta distributions; we used the alpha and beta parameters
reported in previous studies [20] (Table A in S1 Text).

The covariate effects were summarized

using the mean and 95% credible intervals (representing the range of values that contains the
true value with a probability of 95%); a significant result for a coefficient is indicated by where
the 95% credible interval does not cross zero. The geostatistical random effect modeled spatial
correlation as a function of the separating distance between pairs of barangays. Model predictions
for A. lumbricoides and T. trichiura were used to generate representative STH risk maps
for males aged 5–19 years (the subgroup with the highest prevalence for these STH) and model
predictions for hookworm were used to generate representative STH risk maps for males aged
>20 years (the subgroup with the highest risk for this STH) across the Philippines in ArcGIS
version 10.0. Note the overall mean predicted prevalence is specific to the age and sex group
(i.e. choice of a different age-sex group would result in a different spatial mean), with spatial
variation around the mean being influenced by the environmental variables and the spatial correlation
component of the model. This means that the relative differences between locations
are consistent and so the maps presented are representative of the spatial distribution of risk
for all age groups and both sexes. The priors used for the model parameters (spatial and nonspatial)
are given in the S1 Text.

To determine the discriminatory performance of the model

predictions relative to observed prevalence thresholds (20% and 50%; corresponding to prevalence
thresholds for WHO-recommended STH control strategies [21]), the area under the
curve (AUC) of the receiver operating characteristic was used (more detail in Table B in S1
Text). An AUC value of >0.7 was taken to indicate acceptable predictive performance [22].
We estimated the mean prediction error and the percentage of the overall observed mean
attributable to the error estimate.

Spatio-temporal distribution of soil-

transmitted
helminth infections in Brazil
Frédérique Chammartin1,2, Luiz H Guimarães3, Ronaldo GC Scholte4, Mara E Bavia5, Jürg Utzinger1,2
and Penelope Vounatsou1,2
(Cakupan satu negara)
Details on spatio-temporal model formulation and
variable selection are given in the Additional file 1.
In brief, risk factors included in the spatio-temporal
models were selected through a Bayesian stochastic
search variable selection approach [23]. We followed
our previous procedure, which consists of selecting
within a geostatistical framework the best predictors
among highly correlated ones, while addressing nonlinearity
of the predictors [16]. We further extended
this formulation in applications to large datasets, by
estimating the spatial process through a predictive
process approximation [15]. The inclusion of a variable
in the model was defined as the product of two
indicators: the first was assumed to be Bernoulli distributed
and suggests the inclusion of the group of
highly correlated variables, whilst the second followed a
categorical prior distribution for selecting a single predictor
within the group. In addition, regression coefficients
were a priori parameterised with parameter
expanded normal mixture of inverse-gamma (peNMIG)
distributions [24], which ensure a rigorous selection of
categorical variables. Models with the highest posterior
probability identified the predictors to include in the
final models.