Mapping the Serological Prevalence Rate of West Nile Fever in Equids, Tunisia R

Transboundary and Emerging Diseases

ORIGINAL ARTICLE Mapping the Serological Prevalence Rate of West Nile fever in Equids, Tunisia R. Bargaoui1, S. Lecollinet2 and R. Lancelot3

1 Institut de la Recherche Vet erinaire de Tunisie (IRVT), Service de Virologie, Tunis, Tunisie 2 Agence Nationale de Securit e Sanitaire de l’Alimentation, de l’Environnement et du Travail (ANSES), UMR n°1161 Virologie ANSES, INRA, ENVA, Maisons-Alfort, France 3 CIRAD, UMR n°15 CMAEE (CIRAD, INRA), Montpellier, France

Keywords: Summary West Nile fever; equids; serological prevalence; environmental risk factors; risk West Nile fever (WNF) is a viral disease of wild birds transmitted by mosquitoes. map; Tunisia Humans and equids can also be affected and suffer from meningoencephalitis. In Tunisia, two outbreaks of WNF occurred in humans in 1997 and 2003; sporadic Correspondence: cases were reported on several other years. Small-scale serological surveys revealed R. Bargaoui. Institut de la Recherche the presence of antibodies against WN virus (WNV) in equid sera. However, clin- Vet erinaire de Tunisie (IRVT), Service de ical cases were never reported in equids, although their population is abundant in Virologie, Rue Djebel Lakhdhar – La Rabta 1006 Tunis, Tunisie. Tunisia. This study was achieved to characterize the nationwide serological status Tel.: +216 99 64 13 92; of WNV in Tunisian equids. In total, 1189 sera were collected in 2009 during a Fax: +216 71 56 96 92; cross-sectional survey. Sera were tested for IgG antibodies, using ELISA and mi- E-mail: [email protected] croneutralization tests. The estimated overall seroprevalence rate was 28%, 95% conﬁdence interval [22; 34]. The highest rates were observed (i) in the north-east- Received for publication September 28, 2012 ern governorates (Jendouba, 74%), (ii) on the eastern coast (Monastir, 64%) and (iii) in the lowlands of Chott El Jerid and Chott el Gharsa (Kebili, 58%; Tozeur, doi:10.1111/tbed.12077 52%). Environmental risk factors were assessed, including various indicators of wetlands, wild avifauna, night temperature and chlorophyllous activity (normalized difference vegetation index: NDVI). Multimodel inference showed that lower distance to ornithological sites and wetlands, lower night-time temperature, and higher NDVI in late spring and late fall were associated with higher serological prevalence rate. The model-predicted nationwide map of WNF seroprevalence rate in Tunisian equids highlighted different areas with high seroprevalence probability. These ﬁndings are discussed in the perspective of implementing a better WNF surveillance system in Tunisia. This system might rely on (i) a longitudinal survey of sentinel birds in high-risk areas and time periods for WNV transmission, (ii) investigations of bird die-offs and (iii) syndromic surveillance of equine meningoencephalitis.

mosquitoes (Hubalek, 2008). Many bird species have been Introduction found infected by WNV and migrating birds are involved West Nile fever (WNF) is a mosquito-borne viral infection in its long-distance transmission (Komar, 2003). transmitted in natural cycles between birds and mosqui- In humans, WNV infection is usually an asymptomatic toes, particularly Culex species (Kramer et al., 2008). Con- or mild febrile illness. However, meningoencephalitis cases sequently, WN virus (WNV) transmission is sensitive to are observed with some fatalities in older or immunocom- environmental conditions: it is strongly associated with the promised patients. WNV is also a cause of animal disease, presence of wetlands and the occurrence of rainfall and especially in equids (horses, donkeys and mules) in which ﬂooding, as well as the abundance of avifauna and possibly fatal meningoencephalitis cases are observed. Both

© 2013 Blackwell Verlag GmbH • Transboundary and Emerging Diseases. 1 West Nile Fever in Equids, Tunisia R. Bargaoui et al. humans and horses are dead-end hosts for WNV: viremia the winter counts of waterfowls (i.e. excluding passerines, is low and does not allow the infection of mosquitoes feed- raptors, etc.) recently conducted in Tunisian wetlands ing on these hosts (Dauphin and Zientara, 2007). revealed the presence of 106 species, and an annual average Sporadic cases and outbreaks of WNF in humans and eq- of about 430 000 birds (Hamdi and Chard-Cheikhrouha, uids have been reported after WNV discovery in 1937 in 2011). the West Nile Province of Uganda. In the late 1990s, out- An entomological survey conducted in Tunisia after the breaks were increasingly reported in Europe: Romania WNF epidemic of 1997 (Feki et al., 2005) showed the (1996), Russia (1999) and the Mediterranean basin (1994, abundance of Culex and Aedes species, potential vectors of 1997, 1998–2000), with hundreds of human cases (Murgue WNV in Tunisia (El Ghoul, 2009). Culex pipiens was abun- et al., 2001). The virus was also introduced in the USA in dant including in urban areas (Feki et al., 2005): it is an 1999 where it caused a major epidemic and epizootic in opportunistic mosquito, feeding on birds and mammals. birds, horses and humans, and progressively spread from Therefore, it might play the role of bridge species between Canada to Argentina and Brazil (Randolph and Rogers, birds and mammals. 2010). The Tunisian equid population reaches ca. 188 000 Recently, human outbreaks were reported from Central (123 000 donkeys, 40 000 mules and 25 000 horses), Europe and the Mediterranean basin: Albania, Hungary, unevenly distributed: 36% in the North, 46% in the Centre Israel, Italy, Macedonia, the Palestinian Territory, Roma- and 17% in the South. nia, the Russian Federation, Serbia, Spain, Ukraine, Tuni- No animal survey was carried out during the human out- sia, Turkey, Greece (2010–2012) (Calistri et al., 2010; Papa breaks of 1997 and 2003. Three small-scale cross-sectional et al., 2010; Sirbu et al., 2011; Garcıa-Bocanegra et al., surveys were implemented in 2005, 2007 and 2008, two of 2012). which involved the outbreak areas of 1997 and 2003 (El In Tunisia, two major outbreaks of WNF were reported Ghoul, 2009). They showed WNV had circulated in the in humans. At the fall of 1997, a total of 173 human cases coastal and continental areas of the country. Thus, the of meningoencephalitis were recorded, with 8 deaths (Triki national veterinary services (DGSV) decided to achieve a et al., 2001; Feki et al., 2005). In 2003, WNV infections national cross-sectional serological survey in equids to were confirmed in 21 patients with neurological signs, assess the geographical extent of WNV infection and pro- including 3 fatal cases (Hachfi et al., 2010). No case was vide information to implement better surveillance and con- reported in equids in 1997 and 2003. The WNV identified trol measures. We report here the results of this survey, in 1997 in Tunisia belonged to lineage 1a, and more pre- attempting to find environmental factors associated with cisely to its Israeli–American cluster. It clearly diverged the prevalence rate of WNV antibodies in equids and to from the Western Mediterranean strains (Sotelo et al., draw a national map of this epidemiological indicator. 2011). No WNV isolate was obtained during the 2003 epi- demics in Tunisia. Materials and Methods Tunisia shows a great diversity of eco-climatic patterns dominated by aridity. From 1999 to 2007, the northern Survey design region showed the highest rainfall with 577 mm year 1, In Tunisia, imadas are the smallest administrative units ranging from 324 to 843 mm year 1; the central region (n = 2073). In the absence of an exhaustive database listing was drier: annual rainfall of 280 mm year 1 (153–375); the national population of equids, a two-stage sampling (iii) the southern region had a Saharan climate with very frame was adopted, with imadas as the first-stage units. low rainfall: 106 mm year 1 (44–179) (Direction generale DGSV established a list of 700 imadas where equid popula- des ressources en eau, 2007). Wetlands are distributed tion was noticeable. We expected a high serological preva- throughout the country with a greater abundance in the lence rate, assuming that (i) WNV was regularly north-western and central regions, including marshes, bogs, transmitted in equids; (ii) most sampled equids would be lagoons, salt flats (sebkhats) and flood-inundation playas adults with a possibly long WNV exposition history. Con- (Chott) (Bryant and Rainey, 2002). Among these wetlands, sequently, we sampled 20 equids at 10% of imadas. With 20 sites covering >700 000 hectares were identified in the this sampling effort, the probability of detecting at least one frame of the Ramsar International Convention on Wet- positive equid was 95% if prevalence rate of WNV antibod- lands of International Importance. These areas are hot ies was at least 15%. spots of biodiversity and abundance of insects and avi- A list of 75 imadas was randomly drawn, and the list of fauna, and the resting place of many migratory and resident equid owners was established for each of them by the field birds, especially waterfowl (Hamdi and Chard-Che- veterinary officers. Then, a random sample of owners was ikhrouha, 2011). Indeed, Tunisia is on the flyway of wild drawn. Two animals belonging to the same owner were birds migrating between Africa and Europe. For instance, bled: 10-ml blood samples were taken from the jugular vein

2 © 2013 Blackwell Verlag GmbH • Transboundary and Emerging Diseases. R. Bargaoui et al. West Nile Fever in Equids, Tunisia in a dry tube. Samples were sent to the Institute of Veteri- sera was infected and (v) the positive reference serum pro- nary Research of Tunis (IRVT) within 48 h. Sera were tected Vero cells from infection, and the average neutraliz- extracted and stored at 20°C until serological analyses. ing antibody titre for this positive reference serum was 90. A retrospective survey on equid mobility was imple- A serum was considered as negative if cells were found mented to assess whether the serological prevalence rate infected at any serum concentration. It was considered as observed in a given imada was an indicator of local WNV positive if cells were protected at the 1/10 serum dilution; transmission. A 2-stage sampling frame was adopted. The its titre was calculated as the inverse of the latest dilution at equid sample size (secondary units) was set to reach a 5% which cells were protected. precision under the assumption of a 10% frequency of equid entries or exits in the owner’s flock: 138 equids, Data set inflated to 200 to account for possibly large design effect (Bennett et al., 1991). A target size of 10 equids by imada For the i th imada (i ={1, … I}, with I the number of ima- was set, resulting in an imada sample size of 20, increased das in the first-stage sample), the response was the observed to 22 to account for field constraints. Imadas were ran- serological prevalence rate of antibodies against WNV, that domly selected among those involved in the serological sur- is, the proportion pi of yi positive tests over the number n i vey, split by geographical region. The number of imadas in of tested sera: pi = y i/n i. each region was proportional to the importance of the Rather than trying to identify new risk factors, we region in the serological survey: 10 (North), 10 (Centre), selected the environmental variables used to model pi and 2 (South). Each selected owner was visited and inter- according to established knowledge on WNV epidemiology viewed according to the 12Mo method (Lesnoff, 2008). (Eisen and Eisen, 2011). Surface water is necessary to the Firstly, present equids were listed and their origin (place, larval development of mosquitoes, as well as for the pres- date and occasion of entry) was recorded. Secondly, the ence and abundance of many wild-bird species – either resi- owner was asked to enumerate equid entries and exits dent or migratory. Therefore, we selected a vector layer (date, occasion) that occurred during the 12 previous (186 polygons) of inland water bodies, freely available in months. The survey was carried out by field veterinary offi- the DIVA-GIS repository http://www.diva-gis.org/. cers supervised by the first author. In the frame of Ramsar International Convention on Wetlands of International Importance, 20 wetlands of national importance for water birds were identified by Laboratory procedures Tunisia. We have georeferenced these wetlands. For each Sera were tested by a commercial ELISA detecting WNV serological prevalence data point, we have computed the antibodies recognizing structural pre-membrane (prM) shortest distances (i) to inland water bodies and (ii) to the and envelop (E) proteins (IDScreen WN competition Ramsar sites, and used them as two explanatory variables. ELISA kit, IDVET Innovative Diagnostics). Assays were We used a topographic wetness index (TWI) data set performed according to manufacturer’s instructions. provided by the Earth Resources Observation and Science A microneutralization test (MNT) was used to confirm Center (USA, http://eros.usgs.gov/) at a 1-km pixel resolu- the results of ELISA. Heat-inactivated sera, serially diluted tion. This index combines local upslope contributing area (1/5 to 1/3645) in Dulbecco’s modified Eagle’s medium and slope. It is related to soil moisture and vegetation (DMEM) were mixed with an equal volume (50 ll) of which are important for the larval development and adult DMEM containing 100 tissue culture infectious dose 50 resting sites of many insect species. It is used to quantify

(TCID50) of WNV, Is98 strain (kindly provided by P. Des- topographic control on hydrological processes and proved pres, Institut Pasteur). Cell and virus (100 TCID50 of useful in ecological studies (Sørensen et al., 2006). WNV) controls were added onto each plate. Moreover, Temperature, relative humidity and vegetation are key 10 1,10 2,10 3 and 10 4 dilutions of the virus suspension variables for the development of mosquitoes. We selected were prepared for its back titration. After incubation of the two remotely sensed MODIS data sets provided by the plates at 37°C for 1.5 h, 2 9 104 Vero cells in 100 llof National Aeronautics and Space Administration (http:// DMEM were added to every well. Plates were incubated at modis.gsfc.nasa.gov/): 37°C for 3 days and cytopathogenic effects were observed 1 Night-time land surface temperature (NLST), because under a light microscope. Results were validated if (i) many mosquito species are mostly active from sunset to infected cells were absent in the cell controls, (ii) infected dawn, and temperature is a major factor for their activ- cells were present in the virus controls, (iii) virus titre was ity, as well as for the dynamics of the extrinsic virus cycle comprised between 75 and 125 TCID50 per well, (iv) no in WNV-infected mosquitoes, protective effect was seen with the negative reference 2 Normalized difference vegetation index (NDVI) serum, for example, every well with the negative reference is related to chlorophyllous activity. Vegetation is

important for adult mosquito resting sites and is a proxy defined as AIC = 2 9 log(L) + 2 9 k, where L is the for relative humidity which is a factor for mosquito sur- maximized likelihood, given the data and model parame- vival. ters, and k is the number of fitted parameters in the model. These MODIS data have a moderate spatial resolution The term 2 9 log(L) is the deviance, which decreased (pixel size of 1 km) and a high temporal resolution: each when more terms are added in the model. Because deviance pixel on the Earth surface is sensed twice a day. Because is penalized by 2 9 k, AIC is a trade-off between maximum serological data represented a cumulative exposition to fit and model parsimony. For a set of logistic regression WNV possibly spanning over several years (depending on models fitted on the same data set, the best model has the equids’ age), we needed environmental variables represen- smallest AIC. Because the number of sampled imadas (I) tative of a long time period. We selected NLST and NDVI was small compared with the number of k model coeffi- data recorded from 2001 to 2008 and processed by tempo- cients (I/k < 40), a small-sample correction was needed: ral Fourier analysis (TFA), which transforms a series of AICc = AIC + 2 9 k 9 (k + 1)/(n – k 1). Moreover, observations taken at intervals over a period of time into an the actual distribution of the response often has a variance arithmetic mean, and a set of (uncorrelated) sine curves, or greater than expected under the binomial assumption. If harmonics, of different frequencies, amplitudes and phases this is not accounted for, this can lead to the selection of a that collectively sum to the original time series. The most spurious number of variables. Lebreton et al. (1992) pro- important harmonics are often those for the annual, bian- posed to use the overdispersion parameter ^c (model devi- nual and tri-annual cycles of seasonal changes, because they ance divided by the number of residual degrees of freedom) often have a clear biological interpretation (Scharlemann to compute QAIC = AIC/^c and its small-sample version et al., 2008). Therefore, seven variables were used for each QAICc = QAIC + 2 9 k 9 (k + 1)/(n k 1). of NLST and NDVI: the arithmetic mean, the annual, bian- We fitted a set of R models (see below), which were nual, and tri-annual amplitudes, and the annual, biannual ranked according to increasing QAICc. The difference D j and tri-annual phases. These TFA data sets were built by was computed for model j: Dj = QAICj QAICmin, with Prof Rogers’ team, Oxford University, at a 1-km spatial res- QAIC min the smallest QAICc observed among the R mod- olution. els. The relative likelihood of model j given the data was

For each raster data set (TWI and the TFA-processed computed: Lj = exp(0.5 9 D j), as well as the probability NLST and NDVI variables), we computed the mean value of modelPj given the set of R models, and available data: ¼ = R within a 5-km radius buffer around each equid serum sam- wj Lj r¼1 Lr. These probabilities were used (i) to build pling point (imada). These values were used as the explana- a ‘confidence interval’ of S plausible models, with S < R tory variables in the modelling steps. (e.g. the set of S models achieving a cumulated probability of 90%), (ii) to compute weighted coefficients and predictions from the set of S models and (iii) to compute Statistical analysis weighted variance for model coefficients and predictions. Given the adopted sampling frame, a clustering (within- The explanatory variables were ranked according to their imada) effect was expected, resulting in a standard error s frequency in the R models, thus providing an index of their for the estimate of prevalence rate p larger than the stan- importance. dard error si for simple random sampling frame (Bennett Exploratory data analysis was used to reduce the number 0.5 et al., 1991). The design effect D was such as s = D si of explanatory variables, and define the set of R models. We 0.5 with si = [(p (1 p)/n] and n the overall sample size. checked their correlations with bivariate scatter plots and Following Bennett et al. (1991), we computed correlation matrix. When high correlation (r > 0.8) or D = 1 + (b 1) q with b the average within-imada sam- strong nonlinear relationship was noticed between two ple size, and q the intracluster correlation coefficient esti- variables, we only kept the one with the most straightfor- mated using a mixed-effect model (Goldstein et al., 2002). ward bioecological meaning. To assess the relationship

We modelled pi with environmental variables at a between dependent and independent variables, we com- national scale, (i) to ﬁnd environmental patterns associated puted univariate, generalized additive models of serological with WNV activity, and (ii) to estimate a WNV seropreva- prevalence rate against smoothing splines of each explana- lence probability map for equids at the national scale. For a tory variable (Hastie and Tibshirani, 1990). After visual given imada i, pi was modelled using binomial logistic inspection (shape and span on the logit scale), explanatory regression. Because we had many explanatory variables, variables unrelated with the dependent variables were dis- model selection was an issue. We used the framework of carded (P < 0.25); variables showing nonlinear relationship information-theoretic criteria and multimodel inference with the dependent variable were categorized. The reduced (Burnham and Anderson, 2002). This framework is based set of variables, and their two-way interactions, was used to on the Akaike information criterion (AIC), which is build the set of R models. The model selection procedure

4 © 2013 Blackwell Verlag GmbH • Transboundary and Emerging Diseases. R. Bargaoui et al. West Nile Fever in Equids, Tunisia described above was applied to build a 90% confidence The mobility survey involved 226 animals. Their median interval of S models, which were subsequently used for age was 6 years, ranging from 2 to 15. Most animals were multimodel inference and prediction. For this purpose, the born after the 1997 WN fever epidemic, and only 18% of Tunisian territory was divided into 5-km pixels in which them were born when the 2003 epidemic occurred. Seven each explanatory variable was averaged. Serological preva- entries (3%) were reported by the farmers; no exit was lence rate was predicted for this new data set, using the S mentioned during the previous 12 months. Therefore, models. observed serological results at the imada level were good For model validation, the receiver operating characteris- markers of local WNV transmission. tic (ROC) curve was plotted (sensitivity = f(1 specific- There were neither between-species, nor between-gender ity)), and the area under the ROC curve (AUC) was significant differences in serological prevalence rates. calculated. To assess the predicted power of the fitted Strong spatial heterogeneities were observed in prevalence model of seroprevalence rates, the data set was randomly rates (Fig. 1). The highest serological prevalence rate was split into a training data set (75% of observations) and a vali- observed in the north-eastern governorates of Tunisia (Jen- dation data set (25% of observations). The positivity thresh- douba, 74%). High serological prevalence rates were also old defined for the training data set was used to compute the observed on the eastern coast (Monastir, 64%), as well as in sensitivity and specificity on the validation data set. lowlands of Chott El Jerid and Chott el Gharsa (Kebili, For further validation, we retrieved the geographical 58%; Tozeur, 52%), two ephemeral lakes belonging to a location of human cases during the 1997 and 2003 WNF large zone of playas stretching from the lowlands of south- outbreaks from hospital records. We were able to find the ern Tunisia to the Atlas Mountains of northern Algeria. location for 43 human cases in 1997 (173 cases reported by Triki et al., 2001), and 31 in 2003 (21 cases with central nervous system infection reported by Hachfi et al., 2010). Firstly, we projected these locations on the map of predicted serological prevalence rate, together with a two-dimensional kernel estimate of WNF incidence density rate for visual inspection (Venables and Ripley, 2002). Sec- ondly, we compared the mean value of predicted serological prevalence rate in pixels with and without clinical cases. For this purpose, 10 000 samples of n pixels were drawn with replacement in the box bounding clinical cases of each WNF outbreak, with n the number of clinical cases in each year (43 in 1997, 31 in 2003). The mean value of predicted serological prevalence rate was computed for each bootstrap sample. The desired probability was the proportion of such mean values higher than the observed mean value in pixels with clinical cases. We used the R software (R Core Team, 2012) for data analysis, as well as additional R packages ‘MASS’ (Venables and Ripley, 2002), ‘raster’, ‘rgeos’, ‘rgdal’, and ‘sp’ (Bivand et al., 2008) for spatial data and plotting, and ‘MuMIn’ for model averaging and multimodel inference (Barton, 2013).

Results In total, 1189 sera were sampled in 74 imadas, that is 807 donkeys, 273 horses, 107 mules and 2 equids of unstated species. A subsample of 102 sera was tested using the MNT, showing a 100% concordance between ELISA and MNT. The estimated overall serological prevalence rate was = 28% (n 1189), with a 95% conﬁdence interval [22; 34]. Fig. 1. Observed prevalence rate of antibodies against West Nile virus q^ = The intracluster correlation was 0.29, and mean sample in equids aggregated at the governorate level, Tunisia, 2009 (1189 size at the imada level was 16, leading to an estimated equids sampled in 74 imadas). Survey locations are shown as points. ^ design effect D = 5.4. Governorates not included in the survey are labelled as ‘NA’.

Table 1. Ninety per cent conﬁdence set of 10 binomial logistic regres- Table 2. Relative importance of explanatory variables according to the sion models of West Nile serological prevalence rate in Tunisia (1189 90% conﬁdence set of 10 binomial logistic regression models of West sera, 74 imadas), selected according to QAICc Nile serological prevalence rate in Tunisia (1189 sera, 74 imadas), selected according QAICc Model QAICc D w Explanatory variable (categorized) Importance 1356 114.3 0.00 0.61 12356 117.7 3.38 0.11 Distance to the nearest Ramsar site (DRams) 0.97 135 118.4 4.11 0.08 Arithmetic mean of night-time land surface 0.96 1236 119.6 5.33 0.04 temperature (NLST) 1235 119.8 5.48 0.04 Biannual phase of normalized 0.91 125 120.4 6.11 0.03 difference vegetation index 123 120.4 6.12 0.03 Interaction between DRams and the 0.79 13456 121.1 6.79 0.02 arithmetic mean of NLST 23 121.2 6.91 0.02 Distance to the nearest humid zone (Dhz) 0.29 25 121.9 7.62 0.01 Annual amplitude of NLST 0.02

1 distance to the nearest Ramsar site (Drams); 2 distance to the nearest humid zone (Dhz); 3 arithmetic mean of night-time land surface tem- Table 3. Model coefﬁcients and 95% conﬁdence intervals (CI) aver- perature (NLST); 4 annual amplitude of NLST; 5 biannual phase of nor- aged over a set of 10 binomial logistic regression models (Table 2) of malized difference vegetation index; 6 interaction between DRams and West Nile serological prevalence rate in Tunisia (1189 sera, 74 imadas), arithmetic mean of NLST. selected according QAICc

Coefﬁcient Estimate Lower CI Upper CI

After the step of exploratory data analysis, 5 explanatory Intercept 0.76 1.77 0.25 variables were retained: cdrams2 1.83 3.21 0.44 1 NLSTa0: arithmetic mean of night-time land surface cdrams3 2.70 4.35 1.04 cLSTNa02 0.37 1.28 0.54 temperature, cLSTNa03 0.51 2.11 1.09 2 NLSTa1: annual amplitude of night-time land surface cNDVIp22 0.45 0.09 0.99 temperature, cNDVIp23 1.47 0.82 2.12 3 NDVIp2: biannual phase of normalized difference vege- cdrams2:cLSTNa02 1.44 0.31 2.57 tation index, cdrams3:cLSTNa02 2.37 1.06 3.67 4 DRams: distance to the nearest Ramsar site, cdrams2:cLSTNa03 3.04 1.80 4.29 5 Dhz: distance to the nearest wetland. cdrams3:cLSTNa03 3.30 2.11 4.49 cdzh2 0.40 0.90 0.09 Generalized additive models highlighted nonlinear rela- cdzh3 1.20 1.95 0.44 tionship between the seroprevalence rate and the explana- cLSTNa12 0.18 0.30 0.66 tory variables NLSTa1, NDVIp2 and DRams. Each of them cLSTNa13 0.07 0.44 0.57 was split into 3 categories according to the 33.3% quantiles of their empirical distribution. Each explanatory variable was split into 3 equal-size categories defined The 90% confidence set of models according to the by 33.3% quantiles numbered 1, 2 and 3. DRams distance to the nearest Ramsar site; Dhz distance to the nearest humid zone; NLSTa0 arith- cumulated probabilities wj (Table 1) included 10 models metic mean of night-time land surface temperature (NLST); NLSTa1 with a combination of 6 explanatory variables (main effects annual amplitude of NLST; NDVIp2 biannual phase of normalized differ- and interactions). For these 10 models, the ratio n/k ranged ence vegetation index; DRams/NLSTa0 interaction between Drams and from 4.3 to 11.2, thus confirming the need to use the small- NLSTa0. sample correction for the information criteria. The relative importance of the 6 variables is displayed in Table 2, and The distance to the nearest Ramsar site, the arithmetic multimodel averaged coefficients are presented in Table 3. mean of NLST and the biannual phase of NDVI had high The overdispersion parameter was high (^c = 3.0), confirm- and similar importance (0.97, 0.96 and 0.91). Coefficient ing the need to use QAICc. The AUC computed from the interpretation of the two former variables was difficult 10 models was 0.76. The positivity threshold minimizing given the existence of an interaction between them (fourth the distance between the ROC curve and the upper left cor- variable in importance: 0.79). Smaller distance to the Ram- ner of the ROC plot was 0.33, leading to a sensitivity of sar site, and lower mean NLST were associated with higher 0.65 and a specificity of 0.76. This model and set of aver- serological prevalence rate; however, the interaction coeffi- aged coefficients applied to the validation data set, with the cients were positive and strong, that is, higher values of same positivity threshold, provided a sensitivity of 0.76 and NLST combined with longer distance from the Ramsar sites a specificity of 0.70. were associated with higher serological prevalence rate.

(a) (b)

Fig. 2. Prediction of West Nile serological prevalence rate in equids, 2009, Tunisia (a) predicted probability, (b) coefﬁcient of variation of predicted probability. Estimates were obtained with multi-model inference on a set of 10 binomial logististic regression models of West Nile serological prevalence rate in 1189 Tunisian equids sampled in 74 imadas. Models were selected according to QAICc.

Higher values of the biannual phase of NDVI were asso- probability was estimated with a good precision in areas ciated with higher values of the serological prevalence rate: where it was high (Fig. 2b). The variation coefficient was humid late spring and fall were more favourable to WNV high in places where predicted probability was low, mostly transmission than drier occurrences of these seasons. in far-south Tunisia, an area with no wetlands and obvi- The two other variables had minor importance: distance ously low WNV transmission risk. to the nearest humid zone (0.29) and annual amplitude of The observed location of human cases during the 1997 NLST (0.02). For the former, sign of the associated coeffi- and 2003 WNF outbreaks corresponded to areas with high cients was as expected: the serological prevalence rate was predicted WNV transmission probability (Fig. 3). The higher for shorter distance to the humid zone. Regarding P-value associated with the bootstrap test was 3 9 10 4 in the latter, lower values of annual amplitude for NLST were 1997 and 0.020 in 2003 (B = 10 000 bootstrap samples in associated with lower serological prevalence rate. each case). The nationwide risk map for WNV antibody serological prevalence rate in Tunisian equids (Fig. 2a) highlighted dif- Discussion ferent areas with high serological prevalence probability from northern to southern Tunisia: Diagnostic methods and serological prevalence rate 1 Northern parts of Jendouba (74%), Beja (65%) and Biz- ELISA tests are commonly used to assess antibody preva- erte (30%) governorates, and coastal regions of Sousse lence in WNV transmission surveys, but they suffer from a (40%) and Monastir (64%) governorates; lack of specificity due to cross-reactions between antibodies 2 The southern part of Sfax (5%) and Sidi Bou Zid (32%); directed against WNV and other Flaviviruses like Usutu 3 A region intersecting the Siliana (13%) and Kairouan virus (Ledermann et al., 2011). The perfect agreement (30%) governorates; (100%) observed between ELISA and MNT results was an 4 A large part of Tozeur (52%) and the northern part of indication that most serologically positive equids had been Kebili (58%), corresponding to the Chott area. infected with WNV, indeed. However, we cannot exclude The map of variation coefficient (predicted standard the possibility that some other Flavivirus had also been error over predicted probability) showed that the predicted transmitted to equids, with a low prevalence rate.

Fig. 3. Human cases of West Nile fever in Tunisia, during the 1997 and 2003 outbreaks. Case locations were superimposed on the map of predicted West Nile serological prevalence rate in equids, 2009. Contour lines show the probability density of human cases (inner lines at higher density).

Moreover, the low mobility and turnover observed in equid northern parts of Tunisia, this sampling design might not populations showed that positive animals had been infected result in higher survey costs. close to the sampling sites. This was important to check The observed WNV serological prevalence rate (28%, before trying to identify environmental features associated n = 1189) was similar to previous observations made by with serological prevalence rate. Ben Hassine et al. (2011) in the north-western part of The lack of between-species difference in serological Tunisia (27%, n = 133). Indeed, equid sera collected and prevalence rates was expected. In most cases, individuals of tested as from 2005 have shown a strong WNV transmis- these species share the same habitat and housing conditions sion in the central region of Tunisia, as well as in other and are used for similar tasks in agricultural works. There- parts of the country: Monastir (74%), Mahdia (46%), Sfax fore, they are exposed to the same risk of WNV transmis- (44%), Bizerte, Jendouba and Kef (30%) (Bergaoui et al., sion. The observed design effect was high (D^ = 5.4), 2007). meaning that sampling size computed under the assump- In sub-Saharan Africa where WNV is considered as tion of independent observations should at least be doubled endemic, similar prevalence rates were observed in Cote^ (square root of 5.4 2.3) to achieve the same target preci- d’Ivoire (28%, n = 95) or Democratic Republic of Congo sion (Bennett et al., 1991). The intracluster correlation was (30%, n = 20) (Cabre et al., 2006). They were even higher also high (q^ = 0.29) – although not exceptional, as com- in Chad (97%, n = 30) or in Senegal (92%, n = 25). Other pared with values commonly observed in veterinary epide- results obtained in this latter country confirmed high prev- miology studies in developing countries (Otte and Gumm, alence rates in Senegal Valley (85%, n = 367) and fossil 1997). Because there is no between-equid WNV transmis- Ferlo Valley (78%, n = 120) (Chevalier et al., 2006, 2010). sion, this high intracluster correlation suggested that WNV Similar findings were also obtained in South Africa (75% in transmission was different between imadas because some mares, n = 243) (Guthrie et al., 2003). The seroprevalence unobserved factor (e.g. WNV introduction). This informa- rate observed in Tunisia might be consistent with the tion might be useful for future WNV serological surveys in assumption of endemic WNV activity. However, repeated Tunisia. Because the design effect linearly increases with introductions of WNV at the occasion of bird migrations, within-cluster sampling size (for a fixed intracluster corre- followed by local transmission and amplification of WNV lation coefficient), one should sample fewer animals in a in bird populations, might lead to similar serological pat- greater number of imadas to achieve a better precision with terns. In addition, these two processes (endemic transmis- the same overall sample size. Given the relatively small sion and repeated introduction) might occur together. country size and concentration of equids in the central and Therefore, this serological survey should be completed with

8 © 2013 Blackwell Verlag GmbH • Transboundary and Emerging Diseases. R. Bargaoui et al. West Nile Fever in Equids, Tunisia ecological and virological studies in birds, equids and mos- serological prevalence rate (relative importance of 0.96 and quitoes to decipher WNV transmission processes in Tuni- 0.91), with straightforward interpretation. The positive sia. interaction between NLST and the distance to the nearest The observed serological prevalence rate was comparable Ramsar site (relative importance of 0.79; see also Table 3) in equids and humans over the same study period on the was more difficult to understand. On the one hand, the north-eastern coast of Tunisia: 5% in equids versus 8% in proximity with a Ramsar site was a source of infected mos- humans in Sfax, 30% versus 28% in Kairouan (Bahri et al., quitoes; however, surface water decreased NLST. Consecu- 2010). However, different results were reported in the tively, mosquito activity might have been lower, with a northern area of Bizerte, with a higher seroprevalence rate longer extrinsic virus cycle, and finally a lower WNV trans- in equids (30%) than in humans (1%). This was probably mission rate. On the other hand, mosquitoes were probably related to a greater exposure risk in equids. Indeed, in the more active when temperature increased, that is farther Bizerte governorate, equids were mostly sampled around from Ramsar sites. However, their abundance rapidly Lake Ichkeul which is a breeding, resting, and stop-over site decreased when the distance from the Ramsar site for many local and migratory bird species. Thus, we did increased. not confirm the conclusions of Bahri et al. (2010) who con- High NDVI values are related to higher photosynthetic sidered the region of Bizerte as weakly endemic for WNV activity and enhanced food supply for animals, as well as infection in humans. High seroprevalence rate in equids, breeding and resting sites for mosquitoes. Hence, seasonal sampled in the vicinity of human habitat, provided evi- differences in NDVI are one of the best predictive factors dence that several human population clusters were strongly for mosquito abundance (Brownstein et al., 2002; Jacob exposed to WNV in northern Tunisia. et al., 2009).

Environmental risk factors for WNV serological Proposals for WNV prevention and surveillance prevalence rate In countries where WNV is endemic, local breeds of horses Geographic variations of WNV serological prevalence rate and donkeys rarely show clinical signs. However, ‘new’ were predicted by the set of selected models (Fig. 2a). High strains of WNV with higher virulence might be introduced predicted values occurred (i) in the northern parts of Jen- and cause severe cases. This happened in Morocco in 2003 douba, Beja and Bizerte governorates, (ii) in the coastal with a lineage-1 WNV (Schuffenecker et al., 2005), or in regions of Sousse and Monastir governorates, (iii) in a South Africa with a lineage-2 WNV: viruses from this latter region intersecting the Siliana and Kairouan governorates lineage were previously considered as non-pathogenic and (iv) in the Chott area (Tozeur, Kebili). Moreover, pre- (Venter et al., 2009). Therefore, it is necessary to imple- dicted prevalence rate in equids was in good agreement ment and maintain a comprehensive WNF surveillance sys- with the location of human cases during 1997 and 2003 tem in Tunisia, including advanced laboratory diagnostic WNF outbreaks (Fig. 3 and bootstrap test). This was an for virus isolation and fine characterization. If a pathogenic indication for the possible interest of this model for public virus is identified, commercial vaccines against WNV are health. However, the WNV transmission risk in humans available and can be used to vaccinate valuable animals might be somewhat different, in particular because human (Dauphin and Zientara, 2007; Beasley, 2011). population density might be proportionally different from In Tunisia, WNF surveillance in equids mostly consists equid population density. in passive surveillance. Syndromic surveillance of neurolog- The distance to the closest Ramsar site was the most ical cases in equids might improve WNF surveillance. This important explanatory variable, present in 97% of the is already partly done in the frame of rabies surveillance: investigated models of serological prevalence rate. A further samples are taken from horses dying with neurological evidence of the importance of this risk factor was the signs and are tested in Tunis Institut Pasteur. These occurrence in 2011 of 3 human cases of WNF in Kebili, in samples should be tested for WNV when they come from the Chott area, near one of the Ramsar sites (National high-risk area and time period for WNV transmission. Health Authorities of Tunisia, 2011). Because wetlands pro- Also, specific training on WNF epidemiology, diagnosis vide suitable habitats for mosquitoes and birds, they are and surveillance should be organized for veterinarians and also favourable places for the emergence of WNV (Bengis technicians. et al., 2004; Kramer et al., 2008). Ramsar sites are protected Regarding the entomological aspects, the identification areas with a higher animal abundance and higher biodiver- of mosquito species responsible for WNV transmission (i) sity than elsewhere, including migratory and resident birds. from infected to sensitive birds (enzootic vectors), and (ii) The arithmetic mean of NLST and the biannual phase of to equids and humans (bridge species vectors) is a pre-req- NDVI were also important explanatory variables of WNV uisite before implementing entomological surveillance.

Little is known in Tunisia on this topic. Moreover, studies sions on the results: Dr W Mhiri, Dr S Sghier, Dr S El Ha- on distribution and population dynamics of mosquito spe- douchi, Dr S Ben Hassan, Dr E Ayari, Dr R Mansourri, Dr cies would be useful for this purpose. A Ben Salem, Mr K Ben Hmida, Mr A Dkhil, Mr A Krichen, Wild-bird mortality monitoring, associated with the test- and Mr Y Ayari. We also thank IRVT administrative staffs ing for WN, Usutu and avian influenza viruses would pre- and drivers. Prof DJ Rogers (Curator, Hope Entomological sumably allow the identification of highly pathogenic virus Collections), Dr D Benz, and Dr WG Wint (Environmental strains before they infect humans. Mass mortality of crows Research Group) from Oxford University, Department of was recorded during the 1997 outbreak (Feki et al., 2005). Zoology, provided us with the MODIS data processed with This could not certainly be linked with WNV infection temporal Fourier analysis used in this paper. The first because no investigation was made in dead birds. However, author of this paper benefited from an EC-funded AVER- WNV isolated from an infected patient in Tunisia at that ROES Erasmus Mundus 1 grant. This study was partially time was similar to an Israeli strain isolated from a dead funded by EU grant FP7-261504 EDENext and is cata- goose, as well as to the strain that was introduced in New logued by the EDENext Steering Committee as EDE- York in 1999, and caused alarming dye off in many bird Next068 (http://www.edenext.eu). The contents of this species, particularly corvids (Lanciotti et al., 1999). Moti- publication are the sole responsibility of the authors and vated field workers and strict procedures for necropsies and don’t necessarily reflect the views of the European Com- laboratory analyses are needed for an effective monitoring mission. of wild-bird mortality. Given the rarity of bird die-offs in Tunisia, it looks unrealistic to propose its implementation for routine WNF surveillance. It might be activated in case References of such event. Bahri, O., I. Dhifallah, N. B. Alaya-Bouafif, H. Fekih, J. Gargo- Serological monitoring of peridomestic bird species like uri, and H. Triki, 2010: Etude seroepidemiologique de la cir- the Eurasian magpie (Pica pica) is a sensitive indicator of culation du virus West Nile chez l’Homme en Tunisie [Sero- WNV transmission (Jourdain et al., 2008). Also, follow-up epidemiological study of West Nile virus circulation in human surveys of sentinel chickens are widely used to detect low- in Tunisia.]. Bull. Soc. Pathol. Exot. 104, 1–5. noise transmission of WNV. In the Mediterranean basin, sen- Barton, K. (2013) MuMIn: Multi-model inference. R package ver- tinel chickens have recently provided good results in Egypt sion 1.9.3/r207. http://R-Forge.R-project.org/projects/mumin/ (Soliman et al., 2010) and Greece (Chaskopoulou et al., (accessed March 7, 2013). 2011). In Tunisia, it might be implemented in the high-risk Beasley, D. W. C., 2011: Vaccines and immunotherapeutics for areas identified in this study: Jendouba, Bizerte, Nabeul, the prevention and treatment of infections with West Nile Sousse, Monastir, Sfax, Medenine and Djerba (Fig. 2a). At virus. Immunotherapy 3, 269–285. last, virus screening might be used to detect WNV in wild Ben Hassine, T., S. Hammami, H. Elghoul, and A. Ghram, 2011: birds. It was already implemented in Tunisia, in the frame of Detection of circulation of West Nile virus in equine in the avian influenza surveillance (Gaidet et al., 2007). However, north-west of Tunisia. Bull. Soc. Pathol. Exot. 104, 266–271. € this method is expensive and cannot be widely used. Bengis, R. G., F. A. Leighton, J. R. Fischer, M. Artois, T. Morner, and C. M. Tate, 2004: The role of wildlife in emerging and re- emerging zoonoses. Rev. Sci. Tech. 23, 497–511. Competing Interests Bennett, S., T. Woods, W. M. Liyanage, and D. L. Smith, 1991: A simplified general method for cluster-sample surveys of health The authors declare that they have no competing interests. in developing countries. World Health Stat. Q. 44, 98–106. Bergaoui, R., S. Sghaier, S. Ben Hassen, and S. Hammami 2007: Acknowledgements La maladie de West Nile chez les Equid es: enquête sero-e pidemiologique dans six regions de la Tunisie. In Troisiemes We thank the anonymous reviewers for their helpful com- Journees Scientifiques de Microbiologie. Monastir, Novembre ments on the first version of this article. Mr H Guessmi 2007, p. 124. (Ministry of Agriculture), Dr H Haj Ammar, Dr S Berrbia, Bivand, R., E. Pebesma, and V. Gomez-Rubio, 2008: Applied Dr I Wannes, and Dr C Sghier (DGSV, Tunis) have helped Spatial Data Analysis with R. Use R!. Springer, New York. us in designing and organizing the field survey and Brownstein, J., H. Rosen, D. Purdy, J. Miller, M. Merlino, F. provided access to the equid data. Dr L Bougzela (CRDA, Mostashari, and D. Fish, 2002: Spatial analysis of West Nile Monastir), Dr J Fkih (CRDA, Mahdia), Dr AA Djayet virus: rapid risk assessment of an introduced vector-borne (CRDA, Sousse), Dr M Bayouth (Zoological Park of Frigui- zoonosis. Vector Borne Zoonotic Dis. 2, 157–164. a, Sousse), Dr E Arfaoui, and Dr S Innoubli (CRDA, Bryant, R., and M. Rainey, 2002: Investigation of flood inunda- Jendouba) have sampled equids. Our colleagues from IRVT tion on playas within the Zone of Chotts, using a time-series have been involved in the analysis of equid sera and discus- of AVHRR. Remote Sens. Environ. 82, 360–375.

Burnham, K. P., and D. R. Anderson, 2002: Model Selection and Guthrie, A. J., P. Howell, I. Gardner, R. Swanepoel, J. Nurton, C. Multimodel Inference: A Practical Information-Theoretic Harper, A. Pardini, D. Groenewald, and C. Visage, 2003: West Approach, 2nd edn. Springer-Verlag, New-York. Nile virus infection of thoroughbred horses in South Africa Cabre, O., M. Grandadam, J.-L. Marie, P. Gravier, A. Prange, Y. (2000–2001). Equine Vet. J. 35, 601–605. Santinelli, V. Rous, O. Bourry, J.-P. Durand, H. Tolou, and B. Hachfi, W., I. Bougmiza, F. Bellazreg, O. Bahri, N. Kaabia, F. Davoust, 2006: West Nile Virus in horses, sub-Saharan Africa. Bahri, and A. Letaief, 2010: Une deuxieme epidemie de Emerg. Infect. Dis. 12, 1958–1960. meningo-encephalite a virus West Nile en Tunisie [Second Calistri, P., A. Giovannini, Z. Hubalek, A. Ionescu, F. Monaco, G. epidemic of West Nile virus meningoencephalitis in Tunisia]. Savini, and R. Lelli, 2010: Epidemiology of West Nile in Europe Med. Mal. Infect. 40, 456–461. and in the Mediterranean basin. Open Virol. J. 4, 29–37. Hamdi,N.,andF.Chard-Cheikhrouha,2011:Estimation dunom- Chaskopoulou, A., C. I Dovas, S. C. Chaintoutis, I. Bouzalas, G. bretotaldesoiseauxaquatiqueshivernantenTunisie:periode Ara, and M. Papanastassopoulou, 2011: Evidence of enzootic 2001/2002a2006/2007.Rev.Ecol.(TerreVie) 66,231–253. circulation of West Nile virus (Nea Santa-Greece-2010, Hastie, T. J., and R. J. Tibshirani, 1990: Generalized Additive lineage 2), Greece, May to July 2011 . Euro. Surveill. 16, Models, Volume 43 of Monographs on Statistics and Applied pii=19933. Probability. Chapman & Hall, London. Chevalier, V., R. Lancelot, A. Diaite, B. Mondet, B. Sall, and X. Hubalek, Z., 2008: Mosquito-borne viruses in Europe. Parasitol. De Lamballerie, 2006: Serological assessment of West Nile Res. 103, 29–43. fever virus activity in the pastoral system of Ferlo, Senegal. Jacob, B. J., W. Gu, E. X. Caamano, and R. J. Novak, 2009: Devel- Ann. N. Y. Acad. Sci. 1081, 216–225. oping operational algorithms using linear and non-linear Chevalier, V., A. Dupressoir, A. Tran, O. Diop, C. Gottland, M. squares estimation in Python for the identification of Culex pi- Diallo, E. Etter, M. Ndiaye, V. Grosbois, M. Dia, N. Gaidet, A. piens and Culex restuans in a mosquito abatement district Sall, V. Soti, and M. Niang, 2010: Environmental risk factors (Cook County, Illinois, USA). Geospat. Health 3, 157–176. of West Nile virus infection of horses in the Senegal River Jourdain, E., M. Gauthier-Clerc, P. Sabatier, O. Grege, T. Green- basin. Epidemiol. Infect. 138, 1601–1609. land, A. Leblond, M. Lafaye, and H. G. Zeller, 2008: Magpies Dauphin, G., and S. Zientara, 2007: West Nile virus: recent as hosts for West Nile virus, southern France. Emerg. Infect. trends in diagnosis and vaccine development. Vaccine 25, Dis. 14, 158–160. 5563–5576. Komar, N., 2003: West Nile virus: epidemiology and ecology in Direction generale des ressources en eau 2007: Annuaire North America. Adv. Virus Res. 61, 185–234. pluviometrique de Tunisie. Annee 2006-2007. Technical Kramer, L. D., L. M. Styer, and G. D. Ebel, 2008: A global per- report, Ministere de l’agriculture et des ressources spective on the epidemiology of West Nile virus. Annu. Rev. hydrauliques, 285 p. Entomol. 53, 61–81. Eisen, L., and R. J. Eisen, 2011: Using geographic information Lanciotti, R. S., J. T. Roehrig, V. Deubel, J. Smith, M. Parker, K. systems and decision support systems for the prediction, pre- Steele, B. Crise, K. E. Volpe, M. B. Crabtree, J. H. Scherret, R. vention, and control of vector-borne diseases. Annu. Rev. A. Hall, J. S. MacKenzie, C. B. Cropp, B. Panigrahy, E. Ostl- Entomol. 56, 41–61. und, B. Schmitt, M. Malkinson, C. Banet, J. Weissman, N. Ko- El Ghoul, H. 2009: Renforcement de la surveillance et des mar, H. M. Savage, W. Stone, T. McNamara, and D. J. systemes d’alerte pour la fievre catarrhale ovine, la fievre du Gubler, 1999: Origin of the West Nile virus responsible for an Nil Occidental et la rage au Maroc, en Algerie et en Tunisie - outbreak of encephalitis in the Northeastern United States. Fievre du Nil Occidental: historique et situation epidemiolog- Science 286, 2333–2337. ique en Tunisie. Projet GCP/RAB/002/FRA, FAO. 24 p. Lebreton, J.-D., K. P. Burnham, J. Clobert, and D. R. Anderson, Feki, I., C. Marrakchi, M. Ben Hmida, F. Belahsen, M. Ben 1992: Modeling survival and testing biological hypotheses Jemaa, I. Maaloul, F. Kanoun, S. Ben Hamed, and C. Mhiri, using marked animals: a unified approach with case studies. 2005: Epidemic West Nile virus encephalitis in Tunisia. Neu- Ecol. Monogr. 62, 67–118. roepidemiology 24, 1–7. Ledermann, J. P., M. A. Lorono-Pino, C. Ellis, K. D. Saxton- Gaidet, N., T. Dodman, A. Caron, G. Balancßa, S. Desvaux, F. Shaw, B. J. Blitvich, B. J. Beaty, R. A. Bowen, and A. M. Pow- Goutard, G. Cattoli, F. Lamarque, W. Hagemeijer, and F. Mo- ers, 2011: Evaluation of widely used diagnostic tests to detect nicat, 2007: Avian influenza viruses in water birds. Africa. west nile virus infections in horses previously infected with st. Emerg. Infect. Dis. 13, 626–629. louis encephalitis virus or dengue virus type 2. Clin. Vaccine Garcıa-Bocanegra, I., J. A. Jaen-Tellez, S. Napp, A. Arenas-Mon- Immunol. 18, 580–587. tes, M. Fernandez-Morente, V. Fernandez-Molera, and A. Lesnoff, M., 2008: Evaluation of 12-month interval methods for Arenas, 2012: Monitoring of the West Nile Virus epidemic in estimating animal-times at risk in a traditional African live- Spain between 2010 and 2011. Transbound. Emerg. Dis. 59, stock farming system. Prev. Vet. Med. 85, 9–16. 448–455. Murgue, B., S. Murri, H. Triki, V. Deubel, and H. G. Zeller, Goldstein, H., W. Browne, and J. Rasbash, 2002: Partitioning 2001: West Nile in the Mediterranean basin: 1950–2000. Ann. variation in multilevel models. Understand. Stat. 1, 223–231. N. Y. Acad. Sci. 951, 117–126.

National Health Authorities of Tunisia, 2011: West Nile virus – 2011: Outbreak of West Nile virus infection in humans, Greece. Tunisia. EpiSouth Weekly Epi Bulletin 2011, 2. Romania, July to October 2010 . Euro. Surveill. 16, pii=19762. Otte, M. J., and I. D. Gumm, 1997: Intra-cluster correlation Soliman, A., E. Mohareb, D. Salman, M. Saad, S. Salama, C. Fay- coefficients of 20 infections calculated from the results of clus- ez, H. Hanafi, I. Medhat, E. Labib, M. Rakha, N. El-Sayed, S. ter-sample surveys. Prev. Vet. Med. 31, 147–150. Yingst, J. Tjaden, and K. Earhart, 2010: Studies on West Nile Papa, A., K. Danis, A. Baka, A. Bakas, G. Dougas, T. Lytras, G. virus infection in Egypt. J. Infect. Public Health 3, 54–59. Theocharopoulos, D. Chrysagis, E. Vassiliadou, F. Kamaria, Sørensen, R., U. Zinko, and J. Seibert, 2006: On the calculation A. Liona, K. Mellou, G. Saroglou, and T. Panagiotopoulos, of the topographic wetness index: evaluation of different 2010: Ongoing outbreak of West Nile virus infections in methods based on field observations. Hydrol. Earth Syst. Sci. humans in Greece, July - August 2010. Euro. Surveill. 15, 10, 101–112. pii=19644. Sotelo, E., J. Fernandez-Pinero, F. Llorente, A. Vazquez, A. Mo- R Core Team, 2012: R: A Language and Environment for Statis- reno, M. Aguero,€ P. Cordioli, A. Tenorio, and M. A. Jimenez- tical Computing. R Foundation for Statistical Computing, Clavero, 2011: Phylogenetic relationships of Western Mediter- Vienna, Austria. ISBN 3-900051-07-0. ranean West Nile virus strains (1996–2010) using full-length Randolph, S. E., and D. J. Rogers, 2010: The arrival, establish- genome sequences: single or multiple introductions? J. Gen. ment and spread of exotic diseases: patterns and predictions. Virol. 92(Pt 11), 2512–2522. Nat. Rev. Microbiol. 8, 361–371. Triki, H., S. Murri, B. Le Guenno, O. Bahri, K. Hili, M. Sidhom, Scharlemann, J. P. W., D. Benz, S. I. Hay, B. V. Purse, A. and K. Dellagi, 2001: Meningo-encephalite a arbovirus West J. Tatem, G. R. W. Wint, and D. J. Rogers, 2008: Global Nile en Tunisie [West Nile viral meningo-encephalitis in data for ecology and epidemiology: a novel algorithm for Tunisia]. Med. Trop. 61, 487–490. temporal Fourier processing MODIS Data. PLoS One 3, Venables, W., and B. Ripley, 2002: Modern Applied Statistics with e1408. S. Statistics and Computing, 4th edn. Springer, New York. Schuffenecker, I., C. N. Peyrefitte, M. el Harrak, S. Murri, A. Le- Venter, M., S. Human, D. Zaayman, G. H. Gerdes, J. Williams, J. blond, and H. G. Zeller, 2005: West Nile virus in Morocco, Steyl, P. A. Leman, J. T. Paweska, H. Setzkorn, G. Rous, S. 2003. Emerg. Infect. Dis. 11, 306–309. Murray, R. Parker, C. Donnellan, and R. Swanepoel, 2009: Sirbu, A., C. Ceianu, R. Panculescu-Gatej, A. Vazquez, A. Teno- Lineage 2 West Nile virus as cause of fatal neurologic disease rio, R. Rebreanu, M. Niedrig, G. Nicolescu, and A. Pistol, in horses, South Africa. Emerg. Infect. Dis. 15, 877–884.