Modelling the geographical distribution of soiltransmitted helminth infections in Bolivia
 Frédérique Chammartin^{1, 2},
 Ronaldo GC Scholte^{1, 2},
 John B Malone^{3},
 Mara E Bavia^{4},
 Prixia Nieto^{3},
 Jürg Utzinger^{1, 2} and
 Penelope Vounatsou^{1, 2}Email author
DOI: 10.1186/175633056152
© Chammartin et al.; licensee BioMed Central Ltd. 2013
Received: 10 November 2012
Accepted: 8 May 2013
Published: 25 May 2013
Abstract
Background
The prevalence of infection with the three common soiltransmitted helminths (i.e. Ascaris lumbricoides, Trichuris trichiura, and hookworm) in Bolivia is among the highest in Latin America. However, the spatial distribution and burden of soiltransmitted helminthiasis are poorly documented.
Methods
We analysed historical survey data using Bayesian geostatistical models to identify determinants of the distribution of soiltransmitted helminth infections, predict the geographical distribution of infection risk, and assess treatment needs and costs in the frame of preventive chemotherapy. Rigorous geostatistical variable selection identified the most important predictors of A. lumbricoides, T. trichiura, and hookworm transmission.
Results
Results show that precipitation during the wettest quarter above 400 mm favours the distribution of A. lumbricoides. Altitude has a negative effect on T. trichiura. Hookworm is sensitive to temperature during the coldest month. We estimate that 38.0%, 19.3%, and 11.4% of the Bolivian population is infected with A. lumbricoides, T. trichiura, and hookworm, respectively. Assuming independence of the three infections, 48.4% of the population is infected with any soiltransmitted helminth. Empiricalbased estimates, according to treatment recommendations by the World Health Organization, suggest a total of 2.9 million annualised treatments for the control of soiltransmitted helminthiasis in Bolivia.
Conclusions
We provide estimates of soiltransmitted helminth infections in Bolivia based on highresolution spatial prediction and an innovative variable selection approach. However, the scarcity of the data suggests that a national survey is required for more accurate mapping that will govern spatial targeting of soiltransmitted helminthiasis control.
Keywords
Bayesian modelling Bolivia Geostatistical variable selection Mapping Soiltransmitted helminthsBackground
Soiltransmitted helminth infections are mainly caused by the intestinal worms Ascaris lumbricoides, Trichuris trichiura, and the two hookworm species Ancylostoma duodenale and Necator americanus[1]. They are the most prevalent neglected tropical diseases, and they are widely distributed across Latin America [2, 3]. Soiltransmitted helminthiasis and other neglected tropical diseases primarily affect lowincome populations, causing chronic conditions, learning disabilities, and reduced productivity and income earning capacity in later life. Morbidity control and, where resources allow, local elimination are now recognised as a priority for achieving the millennium development goals [4]. In 2009, the Pan American Health Organization (PAHO) developed a plan to eliminate neglected and other povertyrelated diseases in Latin America and Caribbean countries. Soiltransmitted helminthiases were identified as target diseases to be controlled through preventive chemotherapy and by promoting access to clean water, improved sanitation, and better hygiene behaviour [5]. Control programmes require reliable baseline information of the geographical distribution of the number of infected people and disease burden estimates in order to enhance the spatial targeting and costeffectiveness of planned interventions [6, 7].
Bolivia is ranked last among the Western Hemisphere countries in terms of key health indicators. For example, child mortality rate is the worse in South America and, according to the 2001 census, 64% of the population did not have enough income to meet their basic needs [8]. The prevalence of soiltransmitted helminth infection is estimated at around 35% [9]. However, the geographical distribution and burden of soiltransmitted helminth infections is poorly documented.
In the past 20 years, progress in geographical information system (GIS) and remote sensing techniques, coupled with spatial modelling, enabled a better understanding of helminth ecology and mapping at high spatial resolution [6, 7, 10–13]. Ecological niche and biologydriven models have been used in assessing the distribution of helminth infections [14–16]. Bayesian geostatistical models offer a robust methodology for identifying determinants of the disease distribution and for predicting infection risk and burden at high spatial scales [17]. These models have been widely used in assessing the relationship between helminth infection with demographic, environmental, and socioeconomic predictors, at subnational [11, 18], national [19], or regional scales [13, 20, 21]. In the Americas, high resolution, geostatistical, modelbased risk estimates have been obtained for the whole continent [22] as well as for Brazil [23]. A key issue in geostatistical modelling is the selection of the predictors. Most of the variable selection methods in geostatistical applications rely on standard methods, such as stepwise regression or bivariate associations that are appropriate for nonspatial data [10, 11]. However, ignoring spatial correlation leads to incorrect estimates of the statistical significance of the predictors included in the model. Recently, Bayesian variable selection has been introduced in geostatistical disease mapping [21, 24].
The purpose of this paper was to map the geographical distribution of A. lumbricoides, T. trichiura, and hookworm in Bolivia, and to estimate the risk, number of infected schoolaged children, and the costs related to treatment interventions in the country. Survey data were extracted from published and unpublished sources. Bayesian geostatistical models were employed using rigorous variable selection procedures.
Methods
Disease data
Search strategy identification of for soiltransmitted helminth infection prevalence survey data in Bolivia
Keywords  Exclusion criteria  Quality control measures 

Bolivi* AND helminth* (OR ascari*, OR trichur*, OR hookworm, OR necator, OR ankylostom*,OR ancylostom*,OR strongy*, OR hymenolepis*, OR toxocara*, OR enterobius*, OR geohelminth*, OR nematode)  Hospitalbased study; casecontrol study (except control group); clinical trials (except baseline); drugefficacy study (except placebo group); displaced population (travellers, military, expatriates, nomads); population treated for the infection during the past year; unclear location of the survey; sample size <10  Double check of each entry; search and elimination of duplicates; recalculation of prevalence; verification in Google Maps that point level coordinates correspond to human settlement 
Environmental, socioeconomic, and population data
Data sources and properties of the predictors explored to model soiltransmitted helminth infection risk in Bolivia
Data type  Source  Date  Temporal resolution  Spatial resolution 

19 climatic variables related to temperature and precipitation  WorldClim^{1}  19502000    1 km 
Altitude  SRTM^{2}  2000    1 km 
Land cover  MODIS/Terra^{3}  20002011  Yearly  1 km 
EVI / NDVI  MODIS/Terra^{3}  20002011  16 days  1 km 
Soil acidity / soil moisture  ISRICWISE^{4}  19602000    10 km 
Unsatisfactory basic needs (UBN)  Census^{5}  2001  10 years  Municipality 
Infant mortality rate (IMR)  CIESIN^{6}  2005  Yearly  5 km 
Human influence index (HII)  LTW^{7}  2005    1 km 
Human development index (HDI)  PAHO^{8}  2005    Municipality 
Population density  WISE3^{4}  2010    10 km 
Schoolaged children proportion  IDB^{9}  2010    Country 
For prediction purposes, a 5 × 5 km spatial resolution grid was created. Environmental data available at 1 × 1 km spatial resolution, were averaged over their closest neighbours. Soil acidity, soil moisture, and infant mortality rate were linked to the prediction pixel with the closest distance. UBN and HDI were rescaled by assigning to each grid pixel the value of the administrative unit they belong to. Rescaling was performed in ArcMap version 10.0 (Environmental Systems Research Institute; Redlands, CA, USA).
Geostatistical model
Disease survey data are typically binomially distributed and modelled via a logistic regression. More precisely, let Y_{ i }, n_{ i }, and p_{i} be the number of infected individuals, the number of individuals screened, and the prevalence or risk of infection at location i, respectively, such as Y_{ i } ~ Bn ( n_{i,} p_{ i }). Spatial correlation is taken into account by introducing locationspecific parameters φ_{ i } that are considered as unobserved latent data from a stationary spatial Gaussian process. We modelled a temporal trend, the selected predictors (i.e. environmental and socioeconomic factors) X_{ i } and φ_{ i } on the logit scale: logit(p_{ i }) = X_{ i }^{ T }β + φ_{ i }. The temporal trend was modelled by a binary variable T_{ i } indicating whether a survey was carried out before or from 1995 onwards. We assumed that $\underset{\xaf}{\phi}~\mathrm{MVN}\phantom{\rule{0.25em}{0ex}}\left(\underset{\xaf}{0},\Sigma \right)$ with variancecovariance matrix Σ. Geographical correlation was modelled by an isotropic exponential correlation function of distance, i.e. ${\Sigma}_{\mathit{cd}}={\sigma}_{\mathit{sp}}^{2}\mathit{\text{exp}}\left(\rho {d}_{\mathit{cd}}\right)$, where d_{ cd } is the Euclidean distance between locations c and d, σ_{ sp }^{2} is the geographical variability known as the partial sill, and ρ is a smoothing parameter controlling the rate of correlation decay. The geographic dependency (range) was defined as the minimum distance at which spatial correlation between locations is less than 5% and is calculated by 3/ρ. To facilitate model fit, the model was formulated using a Bayesian framework of inference. Vague normal prior distributions $\underset{\xaf}{\beta}~\mathrm{N}\left(0,{\sigma}^{2}I\right)$ were adopted for the regression coefficients, an inverse gamma distribution ${\sigma}_{\mathit{sp}}^{2}~\mathit{IG}\left({a}_{{\sigma}_{\mathit{sp}}^{2}},{b}_{{\sigma}_{\mathit{sp}}^{2}}\right)$ was chosen for the variance σ_{ sp }^{2}, and a gamma distribution was assumed for the spatial decay ρ, ρ ~ G(a_{ ρ }, b_{ ρ }).
Geostatistical variable selection
Bayesian stochastic search variable selection [26] was performed to select the most important predictors among the 40 socioeconomic and environmental predictors, while taking into account the spatial correlation in the data. Predictors were either standardised or categorised if they presented a nonlinear bivariate association with the observed helminthiasis prevalence (on the logit scale). Furthermore, we considered a spike and slab prior distribution for the regression coefficients [27], which improves convergence properties of the Markov chain Monte Carlo (MCMC) simulation and allows selection of blocks of covariates such as categorical ones. In addition, we assessed correlation between the predictors and forced the model to choose only one (or none) predictor among those highly correlated (i.e. absolute value of Pearsons correlation coefficient larger than 0.9). The geostatistical variable selection explores all possible models and the final model is the one presenting the highest posterior probability.
The geostatistical variable selection specification is summarised in Figure 1. In particular, predictors were classified into 19 groups b, (b = 1, …, 19), depending on their mutual correlations. Thirteen predictors that were only moderately correlated with any other predictors were separated into single variable groups. Highly correlated predictors were divided into six groups, each containing 38 variables ${X}_{{j}_{b}},{j}_{b}=1,\dots ,{J}_{b}$. The regression coefficients are defined as the product of an overall contribution ${\alpha}_{{\mathrm{j}}_{\mathrm{b}}}$ of the predictor ${X}_{{j}_{b}}$ and the effect ${\xi}_{l{j}_{b}}$ of each of its elements (i.e. categories), ${X}_{l{j}_{b}},l=1,\dots ,L$ categories (excluding baseline) of the predictor ${X}_{{j}_{b}}$. We assigned a spike and slab prior [27, 28], which is a scaled normal mixture of inversegamma to ${\mathit{\alpha}}_{{\mathit{j}}_{\mathit{b}}}$, that is ${\alpha}_{{j}_{b}}~N\left(0,{\tau}_{{j}_{b}}^{2}\right)$, where ${\tau}_{{j}_{b}}^{2}\sim {\gamma}_{1b}\phantom{\rule{0.5em}{0ex}}{\gamma}_{2{j}_{b}}\phantom{\rule{0.5em}{0ex}}\mathit{IG}\left({a}_{\tau},{b}_{\tau}\right)+\left(1{\gamma}_{1b}\phantom{\rule{0.1em}{0ex}}{\gamma}_{2{j}_{b}}\right)\phantom{\rule{0.1em}{0ex}}{\upsilon}_{0}\mathit{IG}\left({a}_{\tau},{b}_{\tau}\right)$.ɑ_{τ} and b_{τ} are fixed parameters of noninformative inversegamma distribution, while υ_{ 0 } is a small constant shrinking ${\mathit{\alpha}}_{{\mathit{j}}_{\mathrm{b}}}$ to zero when the predictor is excluded. The presence or absence of the predictors is defined by the product of two indicators γ_{1b} and ${\underset{\xaf}{\gamma}}_{2\mathit{b}}={\left({\gamma}_{2b1},\dots ,{\gamma}_{2b{J}_{b}}\right)}^{T}$, where γ_{1b} determines the presence or absence of the group b in the model and ${\underset{\xaf}{\gamma}}_{2{b}_{j}},{j}_{b}=1,\dots ,{j}_{b}$ allows selection of a single predictor within the group. A Bernoulli and a multinomial prior distribution are assigned to y_{1b} and γ_{2b}, respectively, such as γ_{1b} ~ Bern(Ω_{1}) and ${\underset{\xaf}{\gamma}}_{2b}~\mathit{Multi}\left(1,{\Omega}_{2b1},\dots ,{\Omega}_{2b{J}_{b}}\right)$ with inclusion probabilities Ω_{1} and ${\underset{\xaf}{\Omega}}_{2\mathit{b}}$. To allow greater flexibility in estimating model size, these probabilities are considered as hyperparameters having noninformative beta and Dirichlet distributions. A mixture of two Gaussian distributions is assumed for ${\xi}_{{l}_{j}{}_{b}},{\xi}_{{l}_{j}{}_{b}}~N\left({m}_{{}_{{l}_{j}{}_{b}}},1\right),{m}_{{}_{{l}_{j}{}_{b}}}~1/2{\delta}_{1}\left({m}_{{}_{{l}_{j}{}_{b}}}\right)+1/2{\delta}_{1}\left({m}_{{}_{{l}_{j}{}_{b}}}\right)$, which shrinks ${\xi}_{{l}_{j}{}_{b}}$ towards 1 (multiplicative identity). For predictors moderately correlated, ${\gamma}_{2b{j}_{b}}$ is fixed to 1, while the effect of linear predictors is only defined by an overall contribution of α.
Implementation details
We considered the following values for the parameters of the prior distributions: σ^{2}=100, (ɑ_{ ρ, }b_{ ρ })=(0. 01,0.01), $\left({a}_{{\sigma}_{\mathit{sp}}^{2}},{b}_{{\sigma}_{\mathit{sp}}^{2}}\right)=\left(2.01,\phantom{\rule{0.22em}{0ex}}1.01\right)$,(ɑ_{ τ },b_{ τ })=(5,25), (ɑ_{Ω1}, b_{Ω1})=(1,1), $\left({\underset{\xaf}{a}}_{\Omega 2b}\right)=\left(1,\dots ,1\right)$ and υ_{0}=0.00025.
MCMC simulations were used to estimate model parameters. For variable selection, a burnin of 50,000 iterations was performed and another 50,000 iterations were run to identify the model with the highest posterior probability. For each infection, the best geostatistical model was fitted with one chain sampler and a burnin of 5,000 iterations. Convergence was assessed after an average of 50,000 iterations using the Raftery and Lewis [29] diagnostics. A posterior sample of 1,000 values was used for validation purposes and for prediction at unsampled locations. Prediction was carried out using Bayesian kriging [17] over a grid of 26,519 pixels of 5 × 5 km spatial resolution. The median and standard deviation of the predicted posterior distribution were plotted to produce smooth risk maps together with their uncertainty. Analyses were implemented in WinBUGS 14 (Imperial College and Medical Research Council; London, UK), while R version 2.7.2 (The R Foundation for Statistical Computing) was used for predictions. Nonspatial explorative statistical analyses were performed in Stata version 10.0 (Stata Corporation; College Station, USA).
Model validation
Models were fitted on a random training sample of 39 locations for A. lumbricoides and T. trichiura, and 37 locations for hookworm. Model validation was performed on the remaining 10 test locations (around 20% of the total locations). The predictive performance was calculated by the proportion of test locations being correctly predicted within the k^{th} Bayesian credible interval (BCI) of the posterior predictive distribution (limited by the lower and upper quantiles $\mathit{BC}{I}_{i\left(k\right)}^{l}$ and $\mathit{BC}{I}_{i\left(k\right)}^{u}$, respectively), where k indicates the probability coverage of the interval as: $\frac{1}{10}{\displaystyle \sum _{i=1}^{10}\mathit{min}\left(I\left(\mathit{BC}{I}_{i\left(k\right)}^{l}<{p}_{i}\right),I\left(\mathit{BC}{I}_{i\left(k\right)}^{u}>{p}_{i}\right)\right)}$ The higher the number of test locations within the narrowest and smallest coverage BCI, the better the model predictive ability.
Treatment needs and estimated costs
The number of infected schoolaged children was calculated for each pixel from the geostatistical modelbased estimated risk and the population density. According to guidelines put forward by the World Health Organization (WHO), all schoolaged children should be treated twice a year in highrisk communities (prevalence of any soiltransmitted helminth infection ≥50%) and once every year in lowrisk communities (prevalence of any soiltransmitted helminth infection between 20% and 50%). Largescale preventive chemotherapy is not recommended in areas where prevalence is less than 20%; indeed treatment should be delivered on a casebycase basis in such areas [30]. We estimated the number of albendazole or mebendazole treatments needed during one year in the schoolaged population, considering different units at which levels of risk were determined (i.e. pixel, municipality, province, and department). Hence, we followed the same methodology as for estimating annualised praziquantel needs against schistosomiasis [31]. To calculate the cost of a schoolbased deworming programme in Bolivia, the estimated number of treatments was multiplied by an average unit cost equivalent to US$ 0.25, which includes additional expenses for training, drug distribution, and administration [9, 32].
Results
Seven out of 59 identified peerreviewed publications reported soiltransmitted helminth infection prevalence data in Bolivia [33–39]. For the current investigation, additional data were obtained from a 2006 report of the Ministry of Health (MoH) in Bolivia [40].
Variables selected by the geostatistical variable selection approach
A. lumbricoides infection  T. trichiura infection  Hookworm infection  

Group 1  
Home with indoor plumbing^{1}  0  0  0  0  0  0  0  0  0 
People with drinking water at home^{1}  0  0  0  0  0  0  0  0  0 
Pipe network  0  0  0  0  0  0  0  0  0 
Population with high quality of life  0  0  0  0  0  0  0  0  0 
Population with UBN  0  0  0  0  0  0  0  0  0 
Population with sanitation at home  0  0  0  0  X  0  0  0  0 
Group 2  
Population with material UBN  0  0  0  0  0  0  0  0  0 
Population with low quality of life  0  0  0  0  0  0  0  0  0 
Group 3  
Minimum temperature coldest month^{1,2}  0  0  0  0  0  0  X  0  X 
Altitude  0  0  0  X  0  0  0  0  0 
Annual temperature  0  0  0  0  X  0  0  0  0 
Maximum temperature warmest month  0  0  0  0  0  0  0  0  0 
Temperature wettest quarter  0  0  0  0  0  0  0  X  0 
Temperature driest quarter  0  0  0  0  0  0  0  0  0 
Temperature warmest quarter  0  0  0  0  0  0  0  0  0 
Temperature coldest quarter  0  0  0  0  0  0  0  0  0 
Group 4  
Temperature annual range^{3}  0  0  0  0  0  0  0  0  0 
Temperature diurnal range  0  0  0  0  0  X  0  0  0 
Group 5  
Annual precipitation^{1,2,3}  0  0  0  0  0  0  0  0  0 
Precipitation wettest month^{1,2}  0  0  0  0  0  0  0  0  0 
Precipitation wettest quarter^{1,2}  X  0  0  0  0  0  0  0  0 
Precipitation driest month^{2,3}  0  0  0  0  0  0  0  0  0 
Precipitation driest quarter^{2}  0  0  0  0  0  0  0  0  0 
Precipitation warmest quarter^{3}  0  0  0  0  0  0  0  0  0 
Precipitation coldest quarter^{2}  0  0  0  0  0  0  0  0  0 
Group 6  
Enhanced vegetation index  0  0  0  0  0  0  0  0  0 
Normalized difference vegetation index  0  0  0  0  0  0  0  0  X 
Moderately correlated  
Soil acidity^{1,3}  0  0  X  0  0  0  0  0  0 
Precipitation seasonality^{1,3}  0  0  0  0  0  0  0  0  0 
Soil moisture^{2}  0  0  0  0  0  0  0  0  0 
Isothermality  0  0  0  0  0  0  0  0  0 
Temperature seasonality  0  0  0  0  0  0  0  0  0 
Human influence index  0  0  0  0  0  0  0  0  0 
Infant mortality rate  0  0  0  0  0  0  0  0  0 
Human development index  0  0  0  0  0  0  0  0  0 
Population with education UBN  0  0  0  0  0  0  0  0  0 
Population with overcrowding UBN  0  0  0  0  0  0  0  0  0 
Population with sanitation UBN  0  0  0  0  0  0  0  0  0 
Population with light at home  0  0  0  0  X  0  0  0  0 
Unemployment rate  0  0  0  0  0  0  0  0  0 
Posterior probability [%]  42.2  5.9  2.9  10.1  6.0  5.2  10.2  4.7  2.0 
Parameter estimates of nonspatial bivariate and Bayesian geostatistical logistic models with environmental and socioeconomic predictors
Bivariate nonspatial  Geostatistical model  

OR ^{ † }  95% CI ^{ † }  OR ^{ † }  95% BCI ^{ † }  
A. lumbricoides infection  
Survey period  
Before 1995  1.00  1.00  
1995 onwards  0.26  (0.24; 0.29)^{*}  0.94  (0.64; 1.42) 
Precipitation wettest quarter (mm)  
<350  1.00  1.00  
350400  1.42  (1.23; 1.66)^{*}  1.32  (0.56; 2.81) 
≥400  12.25  (10.95; 13.70)^{*}  12.52  (5.05; 25.56)^{*} 
Median  95% BCI ^{ † }  
σ ^{2} _{sp}  1.11  (0.72; 2.00)  
Range (km)  9.2  (1.3; 63.0)  
T. trichiura infection  
Survey period  
Before 1995  1.00  1.00  
1995 onwards  0.33  (0.29; 0.37)^{*}  0.85  (0.55; 1.30) 
Altitude  0.33  (0.31; 0.36)^{*}  0.37  (0.26; 0.56)^{*} 
Median  95% BCI ^{ † }  
σ ^{2} _{sp}  1.29  (0.77; 2.23)  
Range (km)  28.7  (3.2; 80.2)  
Hookworm infection  
Survey period  
Before 1995  1.00  1.00  
1995 onwards  0.45  (0.41; 0.50) ^{*}  0.72  (0.12; 4.19) 
Minimum temperature coldest month  6.25  (5.81; 6.72)^{*}  11.35  (5.00; 22.20) ^{*} 
Median  95% BCI ^{ † }  
σ ^{2} _{sp}  3.07  (1.50; 7.44)  
Range (km)  128.4  (39.8; 387.5) 
Yearly estimation of schoolaged children needing preventive chemotherapy against soiltransmitted helminthiasis in Bolivia
5 × 5 km  Municipality  Province  Department  

Number of children targeted  1,481,605  1,749,136  1,907,658  2,180,101 
Number of treatment required  2,894,936  2,868,016  2,847,604  3,013,413 
Cost (US$)  723,734  717,003  711,901  753,353 
Discussion
We present spatially explicit estimates of the risk and number of schoolaged children infected with the three common soiltransmitted helminths in Bolivia using a rigorous geostatistical variable selection approach. Survey data were extracted from the literature, georeferenced, and made public via the openaccess GNTD database. Our study also identified important data needs and gaps. For example, most of the surveys were conducted along the subAndean region. On the other hand, only few survey locations were available in the less densely populated highlands and in the northern tropical areas. Rigorous geostatistical variable selection methods have been used to identify environmental and socioeconomic determinants that govern the distribution of soiltransmitted helminth infection in Bolivia. The country, nestled between the high Andean peaks (on the West) and the Amazon forest (on the East), presents specific ecological characteristics that shape helminth cycles in a complex way. High altitude and diverse topography, as well as the paucity of weather stations in remote areas can introduce interpolation bias in the climatic factors used in our analysis [42]. Bayesian variable selection helped in identifying the potential factors influencing the geographical distribution of the three common soiltransmitted helminth species. Our methodology enabled us to explore all possible models arising from 40 climatic and socioeconomic predictors, while accounting for spatial correlation in the data.
The parameterisation of the prior distribution of the regression coefficients as developed in this manuscript selects the best predictors among highly correlated ones, while addressing nonlinearity. The selected predictors are plausible in terms of helminth biology, ecology, and epidemiology. Indeed, the distribution of A. lumbricoides was positively associated with precipitation above 400 mm during the wettest month. High humidity is related with faster development of parasite eggs in the free environment. Low humidity, on the other hand, can cease embryonation of A. lumbricoides[43, 44]. The positive association between the minimum temperature of the coldest month and the prevalence of hookworm reflects inhibition of the development of the eggs by hostile cold temperatures [3, 45]. The preventive effect of high altitude on T. trichiura infection risk has already been highlighted and explained by subsequent unfavourable temperature, which limits the transmission [46]. The three soiltransmitted helminth infection risks did not decrease significantly over time and we are unsure whether Bolivia has implemented integrated control measures. In the absence of preventive chemotherapy and/or sanitation improvement, environmental contamination is considerable, which may explain our observations of fairly constant infection rates over time [47, 48].
The transmission of soiltransmitted helminthiasis occurs via contaminated food or fingers (A. lumbricoides and T. trichiura), or through the skin by walking on larvaeinfested soil (hookworm). People living in poor conditions are more exposed due to their living conditions, the lack of access to clean water, sanitation, and health facilities [49]. Hence, we would have expected soiltransmitted helminth infections to be associated with some of the socioeconomic factors investigated, such as the ones related to sanitation [50]. However, none of the socioeconomic variables were picked up by our geostatistical variable selection approach. This may indicate that our socioeconomic proxies were not able to capture the socioeconomic disparities across the country when aggregated at district or municipality scales. Historical data are aggregated over villages or larger areas and they are rarely available at household level. Often variation in socioeconomic status is larger within rather than between locations, and hence, it may be harder for socioeconomic data to explain geographical differences.
Bolivian soil also exhibits specific characteristics such as presence of salt and soil compactation arising from livestock farming, which may affect the transmission of soiltransmitted helminths. In our analysis, we explored different soil predictors, including land cover, the vegetation indices EVI and NDVI, soil acidity and soil moisture. However, these factors failed to explain the distribution of the infection risks.
The population of Bolivia is mainly concentrated in and around the three main cities La Paz, Santa Cruz, and Cochabamba, where large parts of the country are uninhabited. The absence of human hosts breaks parasite life cycles. Thus, although environmental conditions may be suitable for parasite survival, there is no risk of transmission. To avoid potential misinterpretation, we clearly delineate areas where no humans live.
The predicted risk maps for the three common soiltransmitted helminth species in Bolivia should be interpreted with caution, particularly for areas characterised by only sparse survey data or poor coverage. Sample design is not optimised regarding the surveyed population; 29% of the data did not report the survey type (schoolaged, communitybased) and might bias the raw prevalence, as it is widely acknowledged that schoolaged children are at higher risk of soiltransmitted helminths, particularly A. lumbricoides and T. trichiura, than their older counterparts [51]. Slightly less than half of the surveys stated the use of the WHOrecommended KatoKatz technique for soiltransmitted helminth diagnosis [41, 52]. Heterogeneity in the data regarding the sensitivities and specificities of the diagnostic methods might introduce measurement errors in the raw prevalence data. Furthermore, a zero hookworm prevalence was reported for 60% of the survey data. While these data suggest the nonendemicity of hookworm, the diagnostic approach might have underestimated the “true” prevalence due to diagnostic dilemmas [53, 54]. Indeed, single KatoKatz thick smears, low intensity infections, and delays in stool processing compromise sensitivity, particularly for hookworm diagnosis [55, 56]. Giardina et al.[24] developed a zeroinflated binomial geostatistical model to estimate malaria burden when data contain a high proportion of zeros. This model could be adopted for soiltransmitted helminth infection and implemented in Bolivia as soon as more survey data become available. In addition, data in the literature usually report on hookworm prevalence, without differentiation of the species (A. duodenale and N. americanus). It would be interesting to analyse the two species separately, as they may have different ecological preferences.
Our study indicates that in Bolivia almost half (48.4%) of the population is infected with at least one of the three common soiltransmitted helminths. Our empiricalbased estimates suggested that a total of 2,868,016 annualised treatments are required for preventive chemotherapy targeting schoolaged children at the level of the municipalities. This estimate is higher than the one previously reported in the country (4,774,672 treatments for a 5year campaign [9, 32]). Population dynamic models [57–59] could be used to predict the effect of preventive chemotherapy on the epidemiological pattern of the three common soiltransmitted helminths, to evaluate the community effectiveness of the programme and to plan the duration of control interventions.
Conclusions
In the framework of a preventive chemotherapy strategy, reliable maps of the distribution of infection risk and disease burden are needed to enhance costeffectiveness of the interventions. Our high resolution estimates are based on existing data and their scarcity may raise doubts on the value of modelling of the disease distribution. However, soiltransmitted helminth infections are driven by environmental factors and, in the absence of interventions, the existing data can establish the relation between the risk of infection and climate. Hence, the risk maps produced are able to identify areas of high infection. Validation indicated that the models had good predictive ability. We therefore believe that the estimated maps can provide important inputs in the sampling design of a national survey by indicating the areas requiring more surveys. Hence, a coherent and optimally designed national survey is warranted to more accurately estimate the distribution and the number of people at risk of infection, so that preventive chemotherapy and other control measures can be optimally targeted.
Abbreviations
 BCI:

Bayesian credible interval
 CI:

Confidence interval
 EVI:

Enhanced vegetation index
 GIS:

Geographical information system
 GNTD:

Global neglected tropical diseases (database)
 HDI:

Human development index
 HII:

Human influence index
 IMR:

Infant mortality rate
 MCMC:

Markov chain Monte Carlo
 MoH:

Ministry of Health
 NDVI:

Normalized difference vegetation index
 OR:

Odds ratio
 PAHO:

Pan American Health Organization
 UBN:

Unsatisfactory basic needs
 WHO:

World Health Organization.
Declarations
Acknowledgements
The authors are grateful for financial support of the Pan American Health Organization (PAHO) and the UBS Optimus Foundation. RGCS received further financial support from the Swiss Brazilian Joint Research Programme (BSJRP 011008). We thank the reviewers for providing valuable comments in an earlier version of the manuscript.
Authors’ Affiliations
References
 Bethony J, Brooker S, Albonico M, Geiger SM, Loukas A, Diemert D, Hotez PJ: Soiltransmitted helminth infections: ascariasis, trichuriasis, and hookworm. Lancet. 2006, 367: 15211532. 10.1016/S01406736(06)686534.View ArticlePubMedGoogle Scholar
 Schneider MC, Aguilera XP, da Silva Junior JB, Ault SK, Najera P, Martinez J, Requejo R, Nicholls RS, Yadon Z, Silva JC: Elimination of neglected diseases in Latin America and the Caribbean: a mapping of selected diseases. PLoS Negl Trop Dis. 2011, 5: e96410.1371/journal.pntd.0000964.PubMed CentralView ArticlePubMedGoogle Scholar
 Pullan RL, Brooker SJ: The global limits and population at risk of soiltransmitted helminth infections in 2010. Parasit Vectors. 2012, 5: 8110.1186/17563305581.PubMed CentralView ArticlePubMedGoogle Scholar
 Hotez PJ, Molyneux DH, Fenwick A, Savioli L, Takeuchi T: A global fund to fight neglected tropical diseases: is the G8 Hokkaido Toyako 2008 summit ready?. PLoS Negl Trop Dis. 2008, 2: e22010.1371/journal.pntd.0000220.PubMed CentralView ArticlePubMedGoogle Scholar
 PAHO: 61st session of the regional committee: Elimination of neglected diseases and other povertyrelated infections. Resolution CD49.R19. Fortyninth Directing Council. 2009, Washington DC: Pan American Health OrganizationGoogle Scholar
 Simoonga C, Utzinger J, Brooker S, Vounatsou P, Appleton CC, Stensgaard AS, Olsen A, Kristensen TK: Remote sensing, geographical information system and spatial analysis for schistosomiasis epidemiology and ecology in Africa. Parasitology. 2009, 136: 16831693. 10.1017/S0031182009006222.PubMed CentralView ArticlePubMedGoogle Scholar
 Soares Magalhães RJ, Clements ACA, Patil AP, Gething PW, Brooker S: The applications of modelbased geostatistics in helminth epidemiology and control. Adv Parasitol. 2011, 74: 267296.View ArticleGoogle Scholar
 WHO: Country cooperation strategy at a glance. 2007, Geneva: World Health OrganizationGoogle Scholar
 InterAmerican Development Bank, Pan American Health Organization PAHO/WHO, Sabin Vaccine Institute: A call to action: addressing soiltransmitted helminths in Latin America and the Caribbean. 2011, Available at: http://bit.ly/faHFO2 accessed: 3 May 2013Google Scholar
 Brooker S, Beasley M, Ndinaromtan M, Madjiouroum EM, Baboguel M, Djenguinabe E, Haye SI, Bundy DAP: Use of remote sensing and a geographical information system in a national helminth control programme in Chad. Bull World Health Organ. 2002, 80: 783789.PubMed CentralPubMedGoogle Scholar
 Raso G, Matthys B, N’Goran EK, Tanner M, Vounatsou P, Utzinger J: Spatial risk prediction and mapping of Schistosoma mansoni infections among schoolchildren living in western Côte d’Ivoire. Parasitology. 2005, 131: 97108. 10.1017/S0031182005007432.View ArticlePubMedGoogle Scholar
 Clements ACA, Firth S, Dembelé R, Garba A, Touré S, Sacko M, Landouré A, BosquéOliva E, Barnett AG, Brooker S: Use of Bayesian geostatistical prediction to estimate local variations in Schistosoma haematobium infection in western Africa. Bull World Health Organ. 2009, 87: 921929. 10.2471/BLT.08.058933.PubMed CentralView ArticlePubMedGoogle Scholar
 Schur N, Hürlimann E, Garba A, Traoré MS, Ndir O, Ratard RC, Tchuem Tchuenté LA, Kristensen TK, Utzinger J, Vounatsou P: Geostatistical modelbased estimates of schistosomiasis prevalence among individuals aged ≤ 20 years in West Africa. PLoS Negl Trop Dis. 2011, 5: e119410.1371/journal.pntd.0001194.PubMed CentralView ArticlePubMedGoogle Scholar
 Malone JB: Biologybased mapping of vectorborne parasites by geographic information systems and remote sensing. Parassitologia. 2005, 47: 2750.PubMedGoogle Scholar
 Mudenda NB, Malone JB, Kearney MT, Mischler PD, Nieto P, McCarroll JC, Vounatsou P: Modelling the ecological niche of hookworm in Brazil based on climate. Geospat Health. 2012, 6: S111S123.View ArticlePubMedGoogle Scholar
 Stensgaard AS, Utzinger J, Vounatsou P, Hürlimann E, Schur N, Saarnak CFL, Simoonga C, Mubita P, Kabatereine NB, Tchuem Tchuenté LA: Largescale determinants of intestinal schistosomiasis and intermediate host snail distribution across Africa: does climate matter?. Acta Trop. 2013, 10.1016/j.actatropica.2011.11.010. in pressGoogle Scholar
 Diggle PJ, Tawn JA, Moyeed RA: Modelbased geostatistics. J R Stat Soc Ser C Appl Stat. 1998, 47: 299326.View ArticleGoogle Scholar
 Raso G, Vounatsou P, Gosoniu L, Tanner M, N’Goran EK, Utzinger J: Risk factors and spatial patterns of hookworm infection among schoolchildren in a rural area of western Côte d’Ivoire. Int J Parasitol. 2006, 36: 201210. 10.1016/j.ijpara.2005.09.003.View ArticlePubMedGoogle Scholar
 Clements ACA, BosquéOliva E, Sacko M, Landouré A, Dembélé R, Traoré M, Coulibaly G, Gabrielli AF, Fenwick A, Brooker S: A comparative study of the spatial distribution of schistosomiasis in Mali in 1984–1989 and 2004–2006. PLoS Negl Trop Dis. 2009, 3: e43110.1371/journal.pntd.0000431.PubMed CentralView ArticlePubMedGoogle Scholar
 Clements ACA, Deville MA, Ndayishimiye O, Brooker S, Fenwick A: Spatial codistribution of neglected tropical diseases in the East African Great Lakes region: revisiting the justification for integrated control. Trop Med Int Health. 2010, 15: 198207. 10.1111/j.13653156.2009.02440.x.PubMed CentralView ArticlePubMedGoogle Scholar
 Schur N, Hürlimann E, Stensgaard AS, Chimfwembe K, Mushinge G, Simoonga C, Kabatereine NB, Kristensen TK, Utzinger J, Vounatsou P: Spatially explicit Schistosoma infection risk in eastern Africa using Bayesian geostatistical modelling. Acta Trop. 2013, 10.1016/j.actatropica.2011.10.006. in pressGoogle Scholar
 Chammartin F, Scholte RGC, Guimarães LH, Tanner M, Utzinger J, Vounatsou P: Soiltransmitted helminth infection in South America: a systematic review and geostatistical metaanalysis of survey data. Lancet Infect Dis. 2013, 13: 50518. .View ArticleGoogle Scholar
 Scholte RGC, Schur N, Bavia ME, Carvalho EM, Chammartin F, Utzinger J, Vounatsou P: Spatial analysis and risk mapping of soiltransmitted helminths in Brazil, using Bayesian geostatistical models. Geospat Health. 2013, in pressGoogle Scholar
 Giardina F, Gosoniu L, Konate L, Diouf MB, Perry R, Gaye O, Faye O, Vounatsou P: Estimating the burden of malaria in Senegal: Bayesian zeroinflated binomial geostatistical modeling of the MIS 2008 data. PLoS One. 2012, 7: e3262510.1371/journal.pone.0032625.PubMed CentralView ArticlePubMedGoogle Scholar
 Hürlimann E, Schur N, Boutsika K, Stensgaard AS, Laserna de Himpsl M, Ziegelbauer K, Laizer N, Camenzind L, Di Pasquale A, Ekpo UF: Toward an openaccess global database for mapping, control, and surveillance of neglected tropical diseases. PLoS Negl Trop Dis. 2011, 5: e140410.1371/journal.pntd.0001404.PubMed CentralView ArticlePubMedGoogle Scholar
 George EI, McCulloch RE: Variable selection via Gibbs sampling. J Am Stat Assoc. 1993, 88: 881889. 10.1080/01621459.1993.10476353.View ArticleGoogle Scholar
 Scheipl F, Fahrmeir L, Kneib T: Spikeandslab priors for function selection in structured additive regression models. J Am Stat Assoc. 2012, 107: 15181532. 10.1080/01621459.2012.737742.View ArticleGoogle Scholar
 Chammartin F, Hürlimann E, Raso G, N’Goran EK, Utzinger J, Vounatsou P: Statistical methodological issues in mapping historical schistosomiasis survey data. Acta Trop. 2013, 10.1016/j.actatropica.2013.04.012. in pressGoogle Scholar
 Raftery AE, Lewis S: How many iterations in the Gibbs sampler?. Bayesian Stat. 1992, 4: 763773.Google Scholar
 WHO: Preventive chemotherapy in human helminthiasis: coordinated use of anthelminthic drugs in control interventions: a manual for health professionals and programme managers. 2006, Geneva: World Health OrganizationGoogle Scholar
 Schur N, Vounatsou P, Utzinger J: Determining treatment needs at different spatial scales using geostatistical modelbased risk estimates of schistosomiasis. PLoS Negl Trop Dis. 2012, 6: e177310.1371/journal.pntd.0001773.PubMed CentralView ArticlePubMedGoogle Scholar
 Bitran R, Martorell B, Escobar L, Munoz R, Glassman A: Controlling and eliminating neglected diseases in Latin America and the Caribbean. Health Aff (Millwood). 2009, 28: 17071719. 10.1377/hlthaff.28.6.1707.View ArticleGoogle Scholar
 Basset D, Gaumerais H, BassetPougnet A: Intestinal parasitoses in children of an Indian community of Bolivian altiplano. Bull Soc Pathol Exot. 1986, 79: 237246.Google Scholar
 Benefice E, Monroy SL, Jiménez S, López R: Nutritional status of Amerindian children from the Beni River (lowland Bolivia) as related to environmental, maternal and dietary factors. Publ Health Nutr. 2006, 9: 327335. 10.1079/PHN2006852.View ArticleGoogle Scholar
 Cancrini G, Bartoloni A, Paradisi F, Nunez LE: Parasitological observations on three Bolivian localities including rural communities, cities and institutions. Ann Trop Med Parasitol. 1989, 83: 591594.PubMedGoogle Scholar
 Cancrini G, Bartoloni A, Zaffaroni E, Guglielmetti P, Gamboa N, Nicoletti A, Genchi C: Seroprevalence of Toxocara canisIgG antibodies in two rural Bolivian communities. Parassitologia. 1998, 40: 476475.Google Scholar
 Esteban JG, Flores A, Aguirre C, Strauss W, Angles R, MasComa S: Presence of very high prevalence and intensity of infection with Fasciola hepatica among Aymara children from the Northern Bolivian Altiplano. Acta Trop. 1997, 66: 114. 10.1016/S0001706X(97)006694.View ArticlePubMedGoogle Scholar
 Flores A, Esteban JG, Angles R, MasComa S: Soiltransmitted helminth infections at very high altitude in Bolivia. Trans R Soc Trop Med Hyg. 2001, 95: 272277. 10.1016/S00359203(01)902329.View ArticlePubMedGoogle Scholar
 Tanner S: A population in transition: health, culture change, and intestinal parasitism among the Tsimane of lowland Bolivia. 2005, University of Michigan, Anthropology department: PhD ThesisGoogle Scholar
 Mollinedo S, Prieto C: El enteroparasitismo en Bolivia. 2006, La Paz Elite: Ministerio de Salud y DeportesGoogle Scholar
 Montresor A, Crompton DWT, Hall A, Bundy DAP: Guidelines for the evaluation of soiltransmitted helminthiasis and schistosomiasis at community level. 1998, Geneva: World Health OrganizationGoogle Scholar
 Hijmans RJ, Cameron SE, Parra JL, Jones PG, Jarvis A: Very high resolution interpolated climate surfaces for global land areas. Int J Climatol. 2005, 25: 19651978. 10.1002/joc.1276.View ArticleGoogle Scholar
 Otto GF: A study of the moisture requirements of the eggs of the horse, the dog, human and pig ascarids. Am J Hyg. 1929, 10: 497520.Google Scholar
 Spindler LA: The relation of moisture to the distribution of human Trichuris and Ascaris. Am J Hyg. 1929, 10: 476496.Google Scholar
 Smith G, Schad GA, Warren KS: The ecology of the freeliving stages: a reappraisal. Hookworm diseasecurrent status and new directions. 1990, London: Taylor and Francis, 89104.Google Scholar
 Bundy DAP, Cooper ES: Trichuris and trichuriasis in humans. Adv Parasitol. 1989, 28: 107173.View ArticlePubMedGoogle Scholar
 Hotez J, Fenwick A, Savioli L, Molyneux DH: Rescuing the bottom billion through control of neglected tropical diseases. Lancet. 2009, 373: 15701575. 10.1016/S01406736(09)602336.View ArticlePubMedGoogle Scholar
 Smits HL: Prospects for the control of neglected tropical diseases by mass drug administration. Expert Rev Anti Infect Ther. 2009, 7: 3756. 10.1586/14787210.7.1.37.View ArticlePubMedGoogle Scholar
 de Silva NR, Brooker S, Hotez PJ, Montresor A, Engels D, Savioli L: Soiltransmitted helminth infections: updating the global picture. Trends Parasitol. 2003, 19: 547551. 10.1016/j.pt.2003.10.002.View ArticlePubMedGoogle Scholar
 Ziegelbauer K, Speich B, Mäusezahl D, Bos R, Keiser J, Utzinger J: Effect of sanitation on soiltransmitted helminth infection: systematic review and metaanalysis. PLoS Med. 2009, 9: e1001162View ArticleGoogle Scholar
 Albonico M, De Carneri I, Di Matteo L, Ghiglietti R, Toscano P, Uledi MK, Savioli L: Intestinal parasitic infections of urban and rural children on Pemba Island: implications for control. Ann Trop Med Parasitol. 1993, 87: 579583.PubMedGoogle Scholar
 Speich B, Knopp S, Mohammed KA, Khamis IS, Rinaldi L, Cringoli G, Rollinson D, Utzinger J: Comparative cost assessment of the KatoKatz and FLOTAC techniques for soiltransmitted helminth diagnosis in epidemiological surveys. Parasit Vectors. 2010, 3: 7110.1186/17563305371.PubMed CentralView ArticlePubMedGoogle Scholar
 Knopp S, Mgeni AF, Khamis IS, Steinmann P, Stothard JR, Rollinson D, Marti H, Utzinger J: Diagnosis of soiltransmitted helminths in the era of preventive chemotherapy: effect of multiple stool sampling and use of different diagnostic techniques. PLoS Negl Trop Dis. 2008, 2: e33110.1371/journal.pntd.0000331.PubMed CentralView ArticlePubMedGoogle Scholar
 Bergquist R, Johansen MV, Utzinger J: Diagnostic dilemmas in helminthology: what tools to use and when?. Trends Parasitol. 2009, 25: 151156. 10.1016/j.pt.2009.01.004.View ArticlePubMedGoogle Scholar
 Dacombe RJ, Crampin AC, Floyd S, Randall A, Ndhlovu R, Bickle Q, Fine PEM: Time delays between patient and laboratory selectively affect accuracy of helminth diagnosis. Trans R Soc Trop Med Hyg. 2007, 101: 140145. 10.1016/j.trstmh.2006.04.008.View ArticlePubMedGoogle Scholar
 Krauth SJ, Coulibaly JT, Knopp S, Traoré M, N’Goran EK, Utzinger J: An indepth analysis of a piece of shit: distribution of Schistosoma mansoni and hookworm eggs in human stool. PLoS Negl Trop Dis. 2012, 6: e196910.1371/journal.pntd.0001969.PubMed CentralView ArticlePubMedGoogle Scholar
 Anderson RM, May RM: Helminth infections of humans: mathematical models, population dynamics, and control. Adv Parasitol. 1985, 24: 1101.View ArticlePubMedGoogle Scholar
 Chan MS, Guyatt HL, Bundy DAP, Medley GF: The development and validation of an agestructured model for the evaluation of disease control strategies for intestinal helminths. Parasitology. 1994, 109: 389396. 10.1017/S0031182000078422.View ArticlePubMedGoogle Scholar
 Anderson R, Hollingsworth TD, Truscott J, Brooker S: Optimisation of mass chemotherapy to control soiltransmitted helminth infection. Lancet. 2012, 379: 289290. 10.1016/S01406736(12)601202.View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.