SPATIAL DISTRIBUTION MODEL OF A HANTAVIRUS RESERVOIR, THE LONG-TAILED COLILARGO (OLIGORYZOMYS LONGICAUDATUS), IN ARGENTINA

Transcription

1 Journal of Mammalogy, 88(6): , 2007 SPATIAL DISTRIBUTION MODEL OF A HANTAVIRUS RESERVOIR, THE LONG-TAILED COLILARGO (OLIGORYZOMYS LONGICAUDATUS), IN ARGENTINA ANÍBAL E. CARBAJO* AND ULYSES F. J. PARDIÑAS Unidad de Ecología de Reservorios y Vectores de Parásitos, Departamento Ecología, Genética y Evolución, Facultad de Ciencias Exactas y Naturales, Universidad de Buenos Aires, Ciudad Universitaria, Pab. II, C1428EHA - CONICET, Buenos Aires, Argentina (AEC) Centro Nacional Patagónico, Casilla de Correo 128, 9120 Puerto Madryn, Chubut, Argentina - CONICET (UFJP) A 1st step in understanding the ecology of rodents as reservoirs and their relation with the disease they transmit is to determine their geographical distribution. This distribution can be modeled as a function of environmental variables. We georeferenced an extensive database of records of the hantavirus reservoir Oligoryzomys longicaudatus (Cricetidae: Sigmodontinae) in Argentina and used generalized linear models to model the probability of the presence of this reservoir as a function of environmental variables. The variables used in the multiple logistic regression were temperature, precipitation, evapotranspiration, altitude, tree cover, grass cover, bare soil cover, and distance to rivers, to water bodies, and to roads; 2 phytogeographic classifications also were included. Spatial autocorrelation was considered in the model by including a spatial dependence covariate. The best model included temperature and precipitation as explanatory variables. External validation showed that the model without the space covariate correctly classified 95% of the sites with the rodent and 70% of the sites without it; the model including the spatial term correctly classified 100% of the sites with the rodent and 70% of the sites without it. A secondary model included days with frost and percent cover by bare soil as explanatory variables. O. longicaudatus was recorded in 97% of sites in the High Andean Subantarctic regions, 65% of sites in the Monte Espinal Patagonian regions, and 0% of sites in the Pampean region. Key words: Argentina, distribution, generalized linear model, hantavirus reservoir, long-tailed colilargo, Oligoryzomys longicaudatus, Patagonia A 1st step in understanding the ecology of rodents as reservoirs and their relation with the disease they transmit is to accurately determine their geographical distribution (Mills and Childs 1998). Also, epidemiological analysis and planning of preventive measures require knowledge of the geographic distribution and ecological conditions relevant to the circulation of a pathogen (Kosoy et al. 1997). For many South American rodents involved in zoonoses, basic aspects such as taxonomy and geographic distribution remain poorly known. The development of predictive habitat distribution models, or niche modeling, has rapidly increased in many fields such as biogeography, evolution, ecology, epidemiology, conservation, and invasive-species management. These models associate species and even community occurrences with environmental * Correspondent: manimal@ege.fcen.uba.ar Ó 2007 American Society of Mammalogists variables (biotic or abiotic) to predict their potential geographic distribution (Anderson et al. 2003; Guisan and Zimmermann 2000). The models can rely on presence data alone (Anderson et al. 2003) or include absence data (Guisan and Zimmermann 2000). The latter models comprise regression models, which have been widely used and become popular thanks to generalized linear models and generalized additive models (Guisan et al. 2002). Generalized linear models are parametric in nature and facilitate the development of simple equations relating the environmental variables to species distributions; this helps to understand the association between a species distribution and the environment and also to transfer the results to a geographic information system because the equation can be programmed for each cell of a map. General additive models are better at modeling nonlinear responses, but they give no explicit equation because they are semiparametric. The long-tailed colilargo (Oligoryzomys longicaudatus Bennet, 1832) belongs to a genus of small-sized mice classified in the New World tribe Oryzomyini (Cricetidae: Sigmodontinae). It is a widespread rodent primarily found in woods and 1555

2 1556 JOURNAL OF MAMMALOGY Vol. 88, No. 6 shrublands in Chile and southwestern Argentina (Palma et al. 2005). This sigmodontine rodent is of great importance because of its role as a major reservoir for the Andes Sout genotype of hantavirus, which produces hantavirus pulmonary syndrome in humans in southern South America (Lopez et al. 1996; Toro et al. 1998). Specimens testing positive for hantavirus antibody have been confirmed all along its distributional range, from 388S to 518S latitude (Padula et al. 2000; Torres-Pérez et al. 2004). The extensive database on O. longicaudatus and the diversity of environmental conditions present in southern Argentina suggest that the association between the rodent and the environment might provide an informative model for the distribution of this viral reservoir. An occurrence probability distribution would be crucial in mapping risk of hantavirus and an important step forward in knowledge of the distribution of this rodent. The main objective of our study is to build a spatial model for predicting the distribution of hantavirus reservoirs. This model was developed and tested with O. longicaudatus and is the 1st step in a series of distribution models for all species involved in hantavirus transmission in Argentina that will be used as a basis for the study of risk of hantavirus transmission. MATERIALS AND METHODS Geographical database. An exhaustive database of records of O. longicaudatus in Argentina (Appendix I) was compiled from 3 main sources of information: voucher specimens housed in mammalogical collections; osteological remains recovered from analyses of owl pellets, conducted primarily by 1 of the authors (UFJP; material is housed at the Colección de Material de Egagrópilas y Afines del Centro Nacional Patagónico, Puerto Madryn, Chubut, Argentina); and literature records based on voucher specimens. Collections surveyed included: Colección de Mamíferos y Colección de Material de Egagrópilas y Afines Elio Massoia del Centro Nacional Patagónico, Puerto Madryn, Argentina; Colección de Mamíferos del Museo Argentino de Ciencias Naturales Bernardino Rivadavia, Buenos Aires, Argentina; Colección de Mamíferos del Museo de La Plata, La Plata, Argentina; and Museum of Vertebrate Zoology, Berkeley, California. Voucher specimens and osteological remains were directly examined to check their taxonomic identity through morphological traits following Carleton and Musser (1989), Gallardo and Palma (1990), and Osgood (1943). Several populations of Oligoryzomys magellanicus in southernmost Argentina were excluded from the analysis because is considered a full species (Gallardo and Palma 1990; see also Musser and Carleton 2005). In addition, putative records for O. longicaudatus in northwestern Argentina (e.g., Cabrera 1961; Díaz 2000; Díaz and Barquez 2002; Ojeda and Mares 1989) were excluded because examination of a large amount of data indicated that they belong to another species of Oligoryzomys (perhaps O. destructor see Espinosa and Reig 1991; Gonzalez-Ittig et al. 2002; Musser and Carleton 2005). Procedures followed guidelines approved by the American Society of Mammalogists (Gannon et al. 2007). Records of O. longicaudatus were incorporated into a geographic information system (GIS) using ArcView 3.2 (Environmental Systems Research Institute 1994). The distribution of O. longicaudatus was represented by a thematic point map with the sites where the rodent was present or absent. Presence was defined by the existence of a voucher specimen or osteological remains. Absence was defined exclusively based on sites where owl pellets with the remains of at least 100 rodents were examined without detecting any sign of O. longicaudatus. This constitutes a large number of pellets without sign, because the number of rodents per pellet is usually 2 with a maximum of 6. A total of 252 sites were examined, and O. longicaudatus was detected at 146 of them. Sites without the rodent that were,10 km from a site where it was present were excluded from the analysis (n ¼ 6). Ten randomly selected sites without long-tailed colilargos were separated and used along with 19 records of the rodent (Porcasi et al. 2005) as a validation data set. The analysis included southern Argentina from 338S to 518S latitude. Two phytogeographic classifications were used to characterize the broad climatic and vegetation characteristics of the sites. The Latin American biogeographical provinces (Cabrera and Willink 1973) were used as a broadscale classification. The study area lies in the neotropical region and included 6 phytogeographic provinces encompassed in the Andean Patagonian, Chaqueño, and Antarctic domains. Each province was considered as a level of the broadscale factor (Table 1). For a detailed classification, we used the classification of the vegetation forms of the Patagonian steppes (Leon et al. 1998). This classification only covers the Monte and Patagonian provinces of the broadscale classification, extending slightly to the west over the Subantarctic and High Andean broadscale provinces. The broadscale classification consisted of 3 provinces (Monte, Patagonian, and Ecotone) subdivided into 10 phytogeographic districts. The detailed classification consisted of 4 broadscale classification levels (Espinal, High Andean, Pampean, and Subantarctic) plus the 10 district subdivisions of the Patagonian and Monte broadscale provinces (Table 1). The Subantarctic and High Andean broadscale provinces were slightly superimposed on the Central, Occidental, and Subandean detailed-scale districts. To analyze the relation between the distribution of O. longicaudatus and environmental variables, thematic maps of the continuous variables were built in grid format covering the whole of Argentina. Mean annual temperature (temperature), mean annual cumulative precipitation (precipitation), and mean annual cumulative evapotranspiration (evapotranspiration) were available as isolines (Subsecretaría de Recursos Hídricos 2002); they were transformed to points and interpolated (inverse weighted distance method). Distance to permanent rivers (river distance), to any kind of river or water body (water distance), and to roads (road distance) were calculated from vector-based digital maps (Subsecretaría de Recursos Hídricos 2002) using Arcview 3.2. Other variables were available in grid format directly: annual frost-day frequency (frost days New et al. 1999), elevation above sea level (United States Geological Survey 2005), and percent cover by trees, grass, and bare soil

3 December 2007 CARBAJO AND PARDIÑAS SPATIAL DISTRIBUTION OF O. LONGICAUDATUS 1557 TABLE 1. Phytogeographic classifications used in the study. Characteristic climatic conditions and vegetation forms are summarized from the Latin American biogeographical provinces (Cabrera and Willink 1973) and the classification of the vegetation forms of the Patagonian steppes (Leon et al. 1998). The percentage of sites with Oligoryzomys longicaudatus (total number of sites within parentheses) are shown for the broad and detailed classifications. The initial factor column shows the starting level for the detailed classification factor (identified by numbers); regions with fewer than 6 sites were grouped together into level 11. The simplified factor column shows the resulting levels (identified with letters) to which regions were assigned after the merging of the initial factor levels. Source Phytogeographic province Phytogeographic district Precipitation (mm)/altitude (m)/temperature (8C) Cover (%) by vegetation forms % Sites broad % Sites detailed Leon et al Patagonian Subandean.300/NA/NA a 60% dense grasses b 97 (33) 7 B Central,200/NA/NA 30 60% shrubs 42 (19) 8 C Occidental 50% tall grasses and shrubs 59 (34) 9 C Payunia NA/600 2,000/NA Shrubs/scrubland 100 (4) 11 C San Jorge Gulf NA/0 700/NA Shrubs and scrubland in valleys, 25 (4) 11 C 80% grasses in hilltops Magellan /NA/NA 50% xeric to 90% humid grasses (0) Monte Austral 200/NA/NA 60% shrubs and scrubland 67 (12) 10 C Oriental.250/200/NA 50 80% shrubs and short trees 00 (1) 11 C Ecotone Río Negro 200/ /NA 30 50% scrubland and 100 (3) 11 C cushionlike shrubs Península Valdés.200/NA/NA 40 60% shrubs (0) Cabrera and Pampean 600 1,200/NA/13 17 Herbaceous steppe c 00 (44) 00 (44) 1 A Willink 1973 High Andean Snow and hail/na/,5 Xeric grassland 92 (24) 95 (20) 2 B Subantarctic 800 5,000/NA/5 9 Perennial and deciduous forest 100 (41) 100 (33) 3 B Patagonian /NA/5 13 Low cover shrubs and grass 70 (96) 91 (11) 4 B steppes Monte /NA/ Shrub grass steppes 50 (20) 43 (7) 5 C Espinal /NA/15 20 Scrubland and low trees 46 (11) 46 (11) 6 C a NA ¼ not available. b Shrubs grow in areas with higher precipitation and cattle grazing. c Replaced by crops and cattle. Initial factor Simplified factor (Hansen et al. 2003). The spatial resolution of these grids ranged from to km (Table 2). The environmental variable value at each site was obtained with the geographic information system; for vegetation cover variables the modal value in a 5-km-radius circle also was calculated. Distribution model. Preliminary analysis included comparing environmental variables between sites with and without O. longicaudatus with a 2-sample Wilcoxon test. The presence or absence of O. longicaudatus was modeled as a function of environmental variables with generalized linear models (McCullagh and Nelder 1989; Nelder and Wedderburn 1972). Because the model uses maximum-likelihood estimators, the fit is measured by the reduction in deviance instead of variance (typical of least-squares estimation). The explanatory variables (x 1,x 2,...) are related to the response variable through a linear predictor (LP). LP ¼ a þ bx 1 þ cx 2..., where a, b, c,... are parameters to be estimated. We assumed a binomial distribution of errors and applied the logistic function as a link between the response variable and the LP. This link constrains the predicted values to lie between 0 and 1. The response variable used was presence or absence of O. longicaudatus. The probability of a site having O. longicaudatus (p) follows an S-shaped curve when LP is a 1st-order polynomial: p ¼ elp/(1 þ elp). This can be linearized as ln[p/(1 p)] ¼ LP (Crawley 1993); values lower than 0.5 indicate the absence of O. longicaudatus, whereas values equal to or higher than 0.5 indicate its presence. To account for overdispersion, the dispersion parameter was calculated by quasilikelihood methods (McCullagh and Nelder 1989). A manual upward stepwise multiple regression procedure was used with alpha ¼ 0.01 for retention because of the large number of variables considered (Donazar et al. 1993). The significance of continuous variables and each factor level were evaluated with a t-test (parameter/ standard error [SE] with null model degrees of freedom [d.f.]). To deal with collinearity between explanatory variables, we computed a pairwise Pearson correlation coefficient; when it surpassed 0.45 the variable responsible for the greater change in deviance was retained, whereas the other was excluded from further analyses. To simplify the models, when 1 of the levels in a factor was not significant, the level was merged to another with similar parameters (Nicholls 1989). This procedure was stopped when the merging implied a significant decrease in total explained deviance (chi-square test for the change in deviance with 1 d.f.). Vegetation cover variables where used in 2 forms, the exact value at the site and the modal value in a 5-km radius around the site; both forms were tested and the 1 that explained the higher deviance was retained. When a model could not be improved any further, interaction terms between the significant variables were added to check if they contributed to a better fit of the model. Afterward, absolute position (spatial coordinates) were fitted to discard remnant spatial trends (Legendre 1993). Because it might be possible to have spatial dependence without a trend (autocorrelation), an extra covariate representing

4 1558 JOURNAL OF MAMMALOGY Vol. 88, No. 6 TABLE 2. Univariate statistics of the explanatory variables used to model distribution of Oligoryzomys longicaudatus in Argentina. The Wilcoxon 2-sample test (column Z) was used to compare the variables between sites grouped according to presence (OP) or absence (OA) of O. longicaudatus. Generalized linear models parameter (B), standard error (SE), and explained deviance (Dev) are given for each significant univariate fit (t-test B/SE and 235 d.f.) for continuous variables; chi-square test on the deviance and 2 d.f. for the phytogeographic broadscale factor (broad factor) and detailed-scale factor (detailed factor). Null model deviance was a Variable Description Units Cell side b (km) Z OP median (LQ UQ) OA median (LQ UQ) BSE Dev t Altitude Elevation above sea m *** 826 (686 1,075) 128 (26 755) ** level Temperature Mean annual temperature 8C *** 7.6 ( ) 14.0 ( ) ** Precipitation Mean annual cumulative mm *** 969 (446 1,726) 677 ( ) ** precipitation Frost days Annual frost days days *** 126 ( ) 47 (34 107) ** frequency Evapotranspiration Mean annual cumulative mm ** 292 ( ) 445 ( ) ** evapotranspiration Tree cover Percentage of surface % *** 8.5 ( ) 3.0 ( ) ** with tree cover Grass cover Percentage of surface % *** 61.5 ( ) 70.5 ( ) ** with grass cover Bare soil cover Percentage of surface % ( ) 21.5 ( ) 1.13 with bare soil cover River distance Distance to nearest km *** 1.4 ( ) 4.2 ( ) ** river Water distance Distance to nearest km *** 1.0 ( ) 2.0 ( ) ** water body or course Road distance Distance to nearest km ( ) 1.0 ( ) 0.56 road Spatial dependence Spatial dependence *** 0.00 ( ) ** covariate Broad factor Broadscale VP 132.0*** phytogeographic factor Detailed factor Detailed-scale phytogeographic factor VP 156.1*** a LQ ¼ lower quartile; UQ ¼ upper quartile; VP ¼ vectorial polygon layers. b Square cells. *** Significant at P, 0.001; ** P, the response autocorrelation was added to the logistic model (Augustin et al. 1996). The covariate was built by kriging interpolation with a modification of the method described in Miller and Franklin (2002). This method requires the building of a presence probability surface for O. longicaudatus by indicator kriging, which was impeded by a strong largescale trend in the response variable. To bypass this problem the trend was removed by regressing the response variable on latitude and longitude (Bailey and Gatrell 1998) with generalized additive models following Kaluzny et al. (1998). To obtain the spatial dependence covariate (SDC), the detrended residuals were analyzed with variograms and interpolated with kriging (Cressie 1993; Jongman et al. 1987) to a grid covering the whole study area (approximately a 9 9-km cell). In the multivariate regression the spatial trends are modeled by the environmental variables and the remnant autocorrelation by the addition of this covariate at the last step. The standardized residuals were plotted against normal quantiles to check for normality. The explanatory power of the model was estimated with D 2, the ratio of the residual to null deviance (equivalent to R 2 in least-square models); accuracy and error measures were calculated from the confusion matrix of the predicted and observed values on the validation data set. The kappa index for unbalanced number of positive and negative cases (Titus and Mosher 1984) was used to measure the model improvement over a random classification. The models were resampled by jackknife (refitted excluding 1 observation at a time) to smooth the effect of influential observations; the mean parameter and SE were used to retest significance (t-test as explained above). The potential distribution maps were built by applying the final jackknifed generalized linear model formula pixel to pixel in the geographic information system. S-plus 6.1 with Sþ SpatialStats and Arcview GIS add-ons (Insightful Corp.2002) was used for modeling and mapping. RESULTS Univariate comparison of sites with and without O. longicaudatus showed significant differences for most of the environmental variables considered, as did the univariate

5 December 2007 CARBAJO AND PARDIÑAS SPATIAL DISTRIBUTION OF O. LONGICAUDATUS 1559 TABLE 3. Generalized linear models for the presence or absence of Oligoryzomys longicaudatus in Argentina as a function of mean annual temperature and annual precipitation with (TPS) and without (TP) the fitting of the spatial dependence covariate (SDC). Parameters and SE were obtained by jackknife resampling. Model TP Parameter SE d.f. t Parameter SE d.f. t Intercept Temperature Precipitation Temperature precipitation SDC Residual deviance Null deviance TPS generalized linear models (Table 2). In the latter, phytogeographic regional factors explained the highest deviance, with the detailed classification being slightly better than the broadscale classification. The detailed-scale factor levels were reclassified because there were 2 classes without sites and 4 classes with fewer than 5 sites (Table 1). The number of levels of both phytogeographic factors was reduced to 3 after model simplification. The resulting percentages of sites where O. longicaudatus was present according to the broadscale levels were High Andean Subantarctic (97%), Pampean (0%), and Monte Espinal Patagonian (65%). The best models described the distribution of O. longicaudatus as a function of temperature and precipitation (model TP [Table 3]), and as a function of frost days and percentage of bare soil (FB [Table 4]). Both models with the spatial dependence covariate (TPS and FBS) explained higher percentages of the deviance than their counterparts without the spatial term (Table 5). The models TP and TPS are more parsimonious than FB and FBS because they retained more degrees of freedom and had no quadratic terms (Table 5). In the validation analysis, the TPS model presented the highest correct classification rate and kappa index, the highest sensitivity (matched with FB model), and the highest specificity (matched with TP and FBs models [Table 5]). Considering these results, we would choose the TPS model to predict the presence of O. longicaudatus. However, this model only allows interpolation along the sampling area because the spatial dependence covariate cannot be extrapolated. The building of a potential distribution map covering all Argentina requires the exclusion of the spatial term. In this case, the TP model is preferred to FB; it has more degrees of freedom, explains more deviance, and FB shows low specificity. Regarding the grain of the explanatory variables, models TP and TPS present a cell side ranging from 4 to 8 km, whereas FB and FBS present a cell side ranging from 0.5 to 45 km. Similar grains are preferred because dissimilarity would make the grain of the prediction map irregular. The map of potential distribution of O. longicaudatus according to temperature and precipitation (TP) shows a higher occurrence probability along the western side of the Andean range south of 368S latitude, narrowing beyond 508S (Fig. 1). The Patagonian central plateaus show low probability (Chubut and Santa Cruz provinces), whereas the eastern plateaus and the north (Río Negro Province) exceed the probability of 0.5. The probability of occurrence falls toward the northeast in Buenos Aires and Cordoba provinces. The map shows zones of high probability of presence outside the study area, to the north along the Andes (from 338S latitude to the north) and in Tierra del Fuego. The TPS map shows a similar pattern inside the study area with the addition of 2 high-probability zones (Fig. 2), 1 in the southeast, between 488S and 508S latitude, and the other in the northwest, extending north from Neuquén through Mendoza provinces. These latter zones should be regarded with care, because there are few data sites within them. The collinearity observed between the explanatory variables retained in the model and the variables that were excluded TABLE 4. Generalized linear models for the presence or absence of Oligoryzomys longicaudatus as a function of annual frost days and percentage cover by bare soil with (FBS) and without (FB) the fitting of the spatial dependence covariate (SDC). Parameters and SE were obtained by jackknife resampling. Model FB Parameter SE d.f. t Parameter SE d.f. t Intercept Frost Frost Bare soil Bare soil SDC Residual deviance Null deviance FBS

6 1560 JOURNAL OF MAMMALOGY Vol. 88, No. 6 TABLE 5. Characteristics of distribution models for Oligoryzomys longicaudatus and prediction accuracy measured by the external validation data set. The terms included in each model are indicated: temperature (T), precipitation (P), temperature precipitation interaction (T P), frost days (F), frost days quadratic term (F 2 ), bare soil (B), and spatial dependence covariate (S). Model TP TPS FB FBS Terms T þ P þ T P T þ P þ S þ T P F þ F 2 þ B þ B 2 F þ F 2 þ B þ B 2 þ S d.f % deviance explained Validation Sensitivity Specificity Correct classification Kappa index should be taken into account to interpret the results (Table 6). They show possible secondary variables related to the rodent distribution that were excluded because other covariates were better for modeling that distribution. However, there is no way to distinguish how much effect corresponds to each of the correlated variables. For example, the variables temperature and precipitation in model TP are highly correlated with frost and bare soil (Table 6); these last 2 might also be related to the distribution (as was verified by the selection of the FB and FBS models). Grass cover also was correlated to temperature, frost days, and bare soil, and although it is not a good explanatory variable (it was not retained in any model), it might have some relation with the distribution of the rodent. DISCUSSION Our distribution map for O. longicaudatus shows the 1st records in the Espinal phytogeographic province and in the political provinces of Mendoza, La Pampa, and Buenos Aires. The distribution of this hantavirus reservoir also extends to the Patagonian central plateaus through Río Negro and Chubut provinces and up to the Atlantic coast (Fig. 1), exceeding the previous known distribution (Redford and Eisenberg 1992). Oligoryzomys longicaudatus is known to inhabit subantarctic forests, where woods are abundant, and also extend into the steppe along scrublands adjacent to streams and roads (Murua and Gonzalez 1982). In a wood steppe ecotone, captures were associated with shrub cover and abundance of spiny shrubs (Lozada et al. 2000). The distribution model correctly predicted higher presence probabilities for O. longicaudatus in the west, where woods are found, and also showed a gradient decreasing toward the east, which might follow the decline in abundance of these habitats. Temperature and precipitation were the variables that best described the distribution of O. longicaudatus at the regional scale; interaction between the variables makes interpretation difficult. At high temperatures, the rodent would probably be associated with lower precipitation (i.e., northeastern Patagonia) and at low temperatures with higher precipitation (i.e., western Andes, close to the Chilean border). The available literature concerns mainly the western part of the study area, which averages lower temperatures than the northeastern pampas. In this area the relation of the distribution of O. longicaudatus to higher precipitation already has been noted (Monjeau 1989), and although some studies disregard its importance (Murua et al. 2003), they have tried to associate precipitation with temporal and not to spatial abundance. Although the methodology chooses some variables and excludes others because of collinearity and lower explanatory power, we cannot reject a possible secondary relation of the excluded variables to the distribution of O. longicaudatus. The relation to higher precipitation in the TP and TPS models might correspond to higher tree cover (Table 5); in Chile, association to vegetation cover was observed previously, although it was at low rodent densities and a very local scale (Gonzalez et al. 2000). Frost days and bare soil also were found to describe the distribution of O. longicaudatus. In these models, the rodent was associated with higher frost frequency and lower cover by bare soil. The inverse relation of the presence of the long-tailed colilargo to bare soil was previously found in a transition zone between the Andean forests and the Patagonian steppes (Lozada et al. 2000). The distribution of O. longicaudatus is related primarily to the presence of favorable habitat and not to absolute geographic location (Monjeau et al. 1998). The coarse-grain model presented herein modeled probability of rodent presence as a function of environmental conditions. Phytogeography might be an easily available indicator of these conditions, but its explanatory power was surpassed by combinations of climatic and vegetation cover variables. These latter variables probably provided more detailed information than the former. On the other hand, it is generally climate (together with soil and history, not considered in this study) that determines the distribution of phytogeographic regions. It would be desirable to predict the distribution of biotic entities on the basis of ecological parameters that are believed to be the causal, driving forces for their distribution and abundance (Guisan and Zimmermann 2000). In this respect, the predictors could be considered as indirect, direct, or resource variables (Austin 1980). Resource variables address matter and energy consumed by the rodents (e.g., food and water). Direct variables are environmental parameters that have physiological importance, but are not consumed (e.g., temperature and humidity). Indirect variables are variables that have no direct physiological relevance for a species performance (e.g., slope, aspect, elevation, topographic position, and geology). In this

7 December 2007 CARBAJO AND PARDIÑAS SPATIAL DISTRIBUTION OF O. LONGICAUDATUS 1561 FIG. 1. Potential distribution of Oligoryzomys longicaudatus in Argentina according to the temperature and precipitation (TP) model. Probability of presence is shown as a function of mean annual temperature and annual precipitation (gray scale). Presence is predicted in areas with probability higher than 0.5. The rectangle approximates the area covered by the study sites. Sites used for model development are indicated by squares and validation sites are indicated by triangles (filled for presence and empty for absence of O. longicaudatus). Hantavirus pulmonary syndrome occurrences (Andes Sout genotype) are plotted as circles. regard, our model used resource variables (river distance), direct variables (climatic and vegetation cover), and indirect variables (altitude, phytogeography, and road distance). In our model, the better predictors of the distribution of the long-tailed colilargo might be encompassed as resource or direct variables. This ensures that the model is more general and applicable over larger areas because it is based on what is supposed to be more physiologically mechanistic (Guisan and Zimmermann 2000). The extrapolation of the model to the entire country of Argentina shows that the probability of finding O. longicaudatus extends north along the Andes range. Although the distribution of O. longicaudatus from Tierra del Fuego up to the Argentinean Bolivian border (Redford and Eisenberg 1992) is repeatedly mentioned in the literature, genetic studies FIG. 2. Potential distribution of Oligoryzomys longicaudatusin Argentina according to the temperature and precipitation model with the spatial dependence covariate fitted (TPS). Probability of presence is shown as a function of mean annual temperature, annual precipitation, and spatial dependence covariate (gray scale). Presence is predicted in areas with probability higher than 0.5. Sites used for model development are indicated by squares and validation sites are indicated by triangles (filled for presence and empty for absence of O. longicaudatus). have shown differences between the specimens classified as O. longicaudatus in northern and southern Argentina (Espinosa and Reig 1991; Gonzalez-Ittig et al. 2002). Even though these populations were not a single species, they might both have similar habitat requirements, suggesting recent speciation or weak differentiation. Both models (with and without spatial covariate) failed to predict accurately the presence in the northwest and central-eastern Patagonia. In the northwest, there were few sites, and in the east, the sites with presence and absence of long-tailed colilargos were relatively close to each other, diminishing predictive power at the scale of this study. If favorable habitats were very small and rare in the area, the coarse spatial resolution of the model would not allow them to be detected properly; thus more intensive sampling or higher spatial detail would be required. For example, in 1 of the largest Patagonian plateaus, Meseta de Somuncurá (.25,000 km 2

8 1562 JOURNAL OF MAMMALOGY Vol. 88, No. 6 TABLE 6. Collinearity between explanatory variables and variables retained in the model. Pairwise correlation coefficients higher than 0.45 or lower than 0.45 are shown. Temperature Precipitation Frost days Bare soil cover Evapotranspiration Precipitation 0.59 Elevation Frost days 0.89 Tree cover 0.81 Grass cover of tableland basalts at an average altitude of 900 m), trapping efforts indicate that O. longicaudatus is restricted to small (,500-m 2 ) and isolated patches. These patches are composed of dense graminoids (Cortaderia) associated with permanent freshwater springs in or near basalt plateau margins at m elevation. The hantavirus pulmonary syndrome cases recorded in Patagonia are encompassed in the higher presence probability area (Fig. 1). In this area the Andes Sout genotype is responsible for hantavirus cases (southwestern Patagonia). In central Patagonia, no records of hantavirus exist for O. longicaudatus (Cantoni et al. 2001; Piudo et al. 2005). However, a case of hantavirus pulmonary syndrome whose reservoir was not identified occurred near the border between Río Negro and Buenos Aires provinces; according to our maps O. longicaudatus is found in this area suggesting that serological studies should be intensified along all the distribution range. In the northeast, the probability of presence of O. longicaudatus decays toward Buenos Aires Province, where Oligoryzomys flavescens is the reservoir responsible for the hantavirus pulmonary syndrome (Maciel genotype). Similar maps for other reservoirs such as O. flavescens and O. chacoensis in the north and the east might help as preliminary tools to distinguish the relative roles in the transmission of hantavirus and the construction of a map of transmission risk for Argentina. The present model might be a good starting point. It proved useful in Patagonia, with a correct classification higher than 86% and more than 95% of the sites with O. longicaudatus correctly predicted. It uses direct variables such as temperature and precipitation as predictors, which gives a good potential for generalization of the model to test it with other species outside the study area. Actually, the model projected a potential habitat area to the northwest of the country that might be tested for other reservoirs such as O. chacoensis or O. flavescens. Our results show that the distribution of a hantavirus reservoir can be modeled as a function of environmental variables, obtaining a map of the probability of presence at the regional scale. This kind of approach might also be extended to other regions, because the variables used are easily available, and because the probability of presence of different reservoir species might be related to occurrences of hantavirus pulmonary syndrome. The association between hantavirus pulmonary syndrome cases and the highest rodent probabilities in our study area also suggests that this kind of model might be used as an early tool to estimate transmission risk until more comprehensive models, considering the reservoir, the virus, and the human population, are developed. ACKNOWLEDGMENTS We thank J. Polop for facilitating the validation data. This work was partially funded by PICT 2004 no from the Agencia Nacional de Promoción Científica y Tecnológica. LITERATURE CITED ANDERSON, R. P., D. LEW, AND A. T. PETERSON Evaluating predictive models of species distributions: criteria for selecting optimal models. Ecological Modelling 162: AUGUSTIN, N. H., M. A. MUGGLESTONE, AND S. T. BUCKLAND An autologistic model for the spatial distribution of wildlife. Journal of Applied Ecology 33: AUSTIN, M. P Searching for a model for use in vegetation analysis. Vegetatio 42: BAILEY, T. C., AND A. C. GATRELL Interactive spatial data analysis. 3rd ed. Longman, Harlow, United Kingdom. CABRERA, A. L Catálogo de los mamíferos de América del Sur. Parte II. Revista del Museo Argentino de Ciencias Naturales Bernardino Rivadavia Ciencias Zoológicas 4: CABRERA, A. L., AND A. WILLINK Biogeografía de América Latina. Monograph 13. Serie Biologia. Organización de los Estados Americanos, Washington, D.C. CANTONI, G., ET AL Seasonal variation in prevalence of antibody to hantaviruses in rodents from southern Argentina. Tropical Medicine and International Health 6: CARLETON, M. D., AND G. G. MUSSER Systematic studies of oryzomyine rodents (Muridae, Sigmodontinae): a synopsis of Microryzomys. Bulletin of the American Museum of Natural History 191:1 83. CRAWLEY, M. J GLIM for ecologists. Blackwell Scientific Publications, Oxford, United Kingdom. CRESSIE, N. A Statistics for spatial data. John Wiley & Sons, Inc., New York. DÍAZ, M. M Key to the native mammals of Jujuy Province, Argentina. Occasional Papers of the Oklahoma Museum of Natural History 7:1 29. DÍAZ, M. M., AND R. M. BARQUEZ Los mamíferos de Jujuy Argentina. Literature of Latin America, Buenos Aires, Argentina. DONAZAR, J., F. HIRALDO, AND J. BUSTAMANTE Factors influencing nest site selection, breeding density and breeding success in the bearded vulture (Gypaetus barbatus). Journal of Applied Ecology 30: ENVIRONMENTAL SYSTEMS RESEARCH INSTITUTE ArcView 3.1. Environmental Systems Research Institute, Redlands, California. ESPINOSA, M. B., AND O. A. REIG Cytogenetics and karyosystematics of South American oryzomine rodents (Cricetidae, Sigmodontinae). Zeitschrift für Säugetierkunde 56: GALLARDO, M. H., AND E. PALMA Systematics of Oryzomys longicaudatus (Rodentia: Muridae) in Chile. Journal of Mammalogy 71: GANNON, W. L., R. S. SIKES, AND THE ANIMAL CARE AND USE COMMITTEE OF THE AMERICAN SOCIETY OF MAMMALOGISTS Guidelines of the American Society of Mammalogists for the use of wild animals in research. Journal of Mammalogy 88: GONZALEZ, L. A., R. MURUA, AND C. JOFRE Habitat utilization of two muroid species in relation to population outbreaks in southern

9 December 2007 CARBAJO AND PARDIÑAS SPATIAL DISTRIBUTION OF O. LONGICAUDATUS 1563 temperate forests of Chile. Revista Chilena de Historia Natural 73: GONZALEZ-ITTIG, R. E., G. R. THEILER, AND C. N. GARDENAL A contribution to the subgeneric systematics of Oligoryzomys (Rodentia, Muridae) from Argentina by means of PCR-RFLP patterns of mitochondrial DNA. Biochemical Systematics and Ecology 30: GUISAN, A., T. C. EDWARDS, AND T. HASTIE Generalized linear and generalized additive models in studies of species distributions: setting the scene. Ecological Modelling 157: GUISAN, A., AND N. E. ZIMMERMANN Predictive habitat distribution models in ecology. Ecological Modelling 135: HANSEN, M., R. DEFRIES, J. R. TOWNSHEND, M. CARROLL, C. DIMICELI, AND R. SOHLBERG m MODIS vegetation continuous fields. The Global Land Cover Facility, College Park, Maryland. INSIGHTFUL, CORP S-plus 6.1 Profesional; S-plus for ArcView GIS 1.1; SþSpatialStats 1.5, Seattle, Washington. JONGMAN, R. G. H., C. J. F. TER BRAAK, AND O. F. R. VAN TONGEREN Data analysis in community and landscape ecology. Pudoc, Wageningen, The Netherlands. KALUZNY, S. P., S. C. VEGA, T.P.CARDOSO, AND A. A. SHELLY SþSpatialStats user s manual for Windows and Unix. Springer, New York. KOSOY, M. Y., R. A. SLONOVA, J. N. MILLS, E. MANDEL, AND J. E. CHILDS Community structure and prevalence of hantavirus infection in rodents: a geographic division of the enzootic area in far eastern Russia. Journal of Vector Ecology 22: LEGENDRE, P Spatial autocorrelation: trouble or a new paradigm? Ecology 74: LEON, R., D. BRAN,M.COLLANTES,J.PARUELO, AND A. SORIANO Grandes unidades de vegetación de la Patagonia extra Andina. Ecología Austral 8: LOPEZ, N., P. J. PADULA, C. ROSSI, M. E. LAZARO, AND M. T. FRANZE- FERNANDEZ Genetic identification of a new hantavirus causing severe pulmonary syndrome in Argentina. Virology 220: LOZADA, M., N. GUTHMANN, AND N. BACCALA Microhabitat selection of five sigmodontine rodents in a forest steppe transition zone in northwestern Patagonia. Studies on Neotropical Fauna and Environment 35: MCCULLAGH, P., AND J. A. NELDER Generalized linear models. Chapman & Hall, London, United Kingdom. MILLER, J., AND J. FRANKLIN Modeling the distribution of four vegetation alliances using generalized linear models and classification trees with spatial dependence. Ecological Modelling 157: MILLS, J. N., AND J. E. CHILDS Ecologic studies of rodent reservoirs: their relevance for human health. Emerging Infectious Diseases 4: MONJEAU, J. A Ecología y distribución geográfica de los pequeños mamíferos del Parque Nacional Nahuel Huapi y áreas adyacentes. Doctoral dissertation, Universidad Nacional de la Plata, La Plata, Argentina. MONJEAU, J. A., E. C. BIRNEY, L. GHERMANDI, R. S. SIKES, L. MARGUTTI, AND C. J. PHILLIPS Plants, small mammals, and the hierarchical landscape classifications of Patagonia. Landscape Ecology 13: MURUA, R., AND L. A. GONZALEZ Microhabitat selection in two Chilean cricetid rodents. Oecologia 52: MURUA, R., L. A. GONZALEZ, AND M. LIMA Population dynamics of rice rats (a hantavirus reservoir) in southern Chile: feedback structure and non-linear effects of climatic oscillations. Oikos 102: MUSSER, G. G., AND M. D. CARLETON Superfamily Muroidea. Pp in Mammal species of the world: a taxonomic and geographic reference (D. E. Wilson and D. M. Reeder, eds.). 3rd ed. Johns Hopkins University Press, Baltimore, Maryland. NELDER, J. A., AND W. M. WEDDERBURN Generalized linear models. Journal of the Royal Statistical Society 135: NEW, M., M. HULME, AND P. D. JONES Representing twentieth century space time climate variability. Part 1: development of a mean monthly terrestrial climatology. Journal of Climate 12: NICHOLLS, A. O How to make biological surveys go further with generalised linear models. Biological Conservation 50: OJEDA, R. A., AND M. A. MARES A zoogeographic analysis of the mammals of Salta Province, Argentina: patterns of community assemblage in the Neotropics. Special Publications 27, The Museum, Texas Tech University, Lubbock. OSGOOD, W. H The mammals of Chile. Field Museum of Natural History, Zoology Series 30: PADULA, P. J., ET AL Genetic diversity, distribution, and serological features of hantavirus infection in five countries in South America. Journal of Clinical Microbiology 38: PALMA, R. E., ET AL Phylogeography of Oligoryzomys longicaudatus (Rodentia: Sigmodontinae) in temperate South America. Journal of Mammalogy 86: PIUDO, L., M. MONTEVERDE, S.GONZALEZ CAPRIA, P.J.PADULA, AND P. CARMANCHAHI Distribution and abundance of sigmodontine rodents in relation to hantavirus in Neuquén, Argentina. Journal of Vector Ecology 30: PORCASI, X., G. E. CALDERÓN, M. LAMFRI, M. SCAVUZZO, M. S. SABATTINI, AND J. J. POLOP Predictive distribution maps of rodent reservoir species of zoonoses in southern America. Mastozoología Neotropical 12: REDFORD, K. H., AND J. F. EISENBERG Mammals of the Neotropics. Vol. 2. The Southern Cone. University of Chicago, Chicago, Illinois. SUBSECRETARÍA DERECURSOS HÍDRICOS Atlas digital de los recursos hídricos superficiales de la República Argentina. Subsecretaria de Recursos Hídricos, Presidencia de la Nación, Buenos Aires, Argentina. TITUS, K., AND J. A. MOSHER Chance corrected classification for use in discriminant analysis: ecological applications. American Midland Naturalist 111:1 7. TORO, J., ET AL An outbreak of hantavirus pulmonary syndrome, Chile, Emerging Infectious Diseases 4: TORRES-PÉREZ, F., ET AL Peridomestic small mammals associated with confirmed cases of human hantavirus disease in southcentral Chile. American Journal of Tropical Medicine and Hygiene 70: UNITED STATES GEOLOGICAL SURVEY Center for Earth Resources Observation and Science (EROS). United States Geological Survey. Submitted 16 June Accepted 6 March Associate Editor was John A. Yunger.

10 1564 JOURNAL OF MAMMALOGY Vol. 88, No. 6 APPENDIX I Localities used in the study. The column O.l. indicates presence (y) or absence (n) of Oligoryzomys longicaudatus. Provinces and coordinates also are provided. PN ¼ Parque Nacional; MN ¼ Monumento Nacional. Code Locality Province O.l. West longitude South latitude 1 Villa Cacique Buenos Aires n Estancia La Casualidad Buenos Aires n Camping Casa Amarilla Buenos Aires n Diego Gaynor Buenos Aires n Cordón Leleque Chubut y Estancia El Maitén Chubut y Estancia Leleque Chubut y Estancia Tecka Chubut y Punta Delgada Chubut n Meseta Lehman Chubut n Río Corintos Chubut y Cañadón Largo Chubut n a Estancia San Pedro Chubut n Laguna Verde Chubut n Sierra Apas Chubut n Sierra de Talagapa Chubut n Sierra de Talagapa Chubut n km E Las Chapas Chubut n km W Los Altares Chubut n km W Las Plumas Chubut n Cañadón del Loro Chubut n Caolinera Dique Ameghino Chubut n b Dique Ameghino Chubut y Laguna Blanca Chubut n Lle Cul Chubut y Los Altares Chubut n a Estancia Monira Chubut n Puerto Lobos Chubut n Casa de Piedra La Pampa n b Bajo Giuliani La Pampa n Junín de los Andes Neuquén y Junín de los Andes Neuquén y Junín de los Andes Neuquén y Paraje La Querencia Neuquén y Estancia Calcatreo Río Negro n Estancia Maquinchao Río Negro n PN Nahuel Huapi Río Negro y PN Nahuel Huapi Río Negro y Canteras Comallo Río Negro n Cerro Castillo Río Negro y Estancia Pilcañeu Río Negro n Paraje Leleque Río Negro n Paso de los Molles Río Negro n Lago Cardiel Santa Cruz y Puerto Ensenada Santa Cruz n Estancia Corcel Negro Neuquén y Chos Malal Neuquén y Riscos Bayos Neuquén n Cerro Casa de Piedra Santa Cruz y Gobernador Duval La Pampa n Puente Carreri Neuquén y Barda Negra Neuquén y Villa Regina Río Negro y La Rinconada Neuquén y Cerrito Piñón Neuquén y Cañadón del Tordillo Neuquén y Confluencia Neuquén y Estancia El Abra Buenos Aires n km NNW rutas 40 y 237 Neuquén y Cerro Castillo Gastre Chubut n b Cañadón Las Coloradas Río Negro y