Since the first recognized outbreak of HPS in the southwestern United States in 1993, approximately 240 cases have occurred, with a death rate of approximately 40% (J. Mills, pers. comm.) (16). Information about SNV host-virus-environment relationships is limited (16,17). No simple relationships have been found between host density and antibody seroprevalence (16-18), but more complex nonlinear relationships appear to exist (17). SNV infections also appear to be less frequent in relatively high- or low-elevation habitats (16,17).

Types of data

RS data are commonly used to generate maps of vegetation types. Vegetation types can be useful indicators of environmental characteristics, including moisture, soil type, and elevation. However, transforming RS images into vegetation maps can be subjective and imprecise (19,20); therefore, we supplemented our vegetation maps with other RS/GIS data, including elevation, slope, vegetation density, and hydrology.

We sampled rodents over four field seasons (June to October during 1995 to 1998). However, in 1997, population densities of deer mice in our study area averaged approximately 25% of 1995-96 and 1998 levels (unpub. data, Boone et al.). Most of the 47 sites sampled in 1997 had three or fewer deer mice, and 14 sites had none. Simple t-tests (SAS ver. 6.10) showed that the mean number of animals per site was statistically equivalent in 1995, 1996, and 1998 (11.1, 10.3, and 8.9 animals per site, respectively; p>0.10 for all comparisons), but differed significantly in 1997 from all other years (2.4 animals per site; p<0.0001 for all comparisons). In 1997, antibody-positive animals were significantly less likely to be positive by reverse transcription-polymerase chain reaction (RT-PCR) for viral RNA in the blood than in any other year, suggesting unusual infection patterns (25% were PCR positive in 1997 and >50% to 70% in other years; chi-square test, p <0.002 for all comparisons of 1997 to other years; p >0.20 for all comparisons of years excluding 1997). On the basis of these tests, we pooled data from 1995, 1996, and 1998 and excluded 1997 data from all analyses because host density and infection dynamics appeared atypical and likely to obscure the baseline spatial infection patterns we sought to identify (21).

Infection Status of Sites

Presence of SNV infections is commonly inferred by determining antibody seroprevalence in a host population (14,16-18,21-23). However, antibody prevalence at the same site may vary considerably (<5% to >60%) over relatively brief periods of <1 year (17,18,22,23), probably because of rapid turnover of rodent populations through death, reproduction, dispersal, and migration. We focused on the presence or absence of SNV infections inferred from antibody data, a more stable measure than antibody prevalence. However, determining infection status is complicated by several factors: animals may remain antibody-positive well after the transmissible phase of an infection (17); noninfectious but antibody-positive deer mice may migrate to a site where no active SNV infection is present; and detectable antibody response requires at least 1 to 2 weeks to develop in newly infected animals (17).

Because of these uncertainties, we used two criteria to demonstrate the effect of classification on analytical outcome. "Status 1" classified sites with one or more antibody-positive animals as positive (active infection present). This criterion may have falsely assigned positive status to some sites where no active, transmissible infections were present. "Status 2" required two or more antibody-positive deer mice or an overall antibody seroprevalence of at least 10% for a site to be classified positive. This criterion may have falsely assigned negative (active infection absent) status to some sites that had a single infectious animal.

Site Selection

Figure 1

Figure 1. Location of Walker River Basin (17) and its eight major vegetation types, as well as developed areas. Piñon-juniper woodland and montane shrubland tend to be highly interspersed and were...

Our study area was the Walker River Basin, a 10,200-km2 region in western Nevada and east-central California northeast of Yosemite National Park (Figure 1). At least nine cases of HPS have occurred in the area since 1993. Major vegetation types in the river basin along an increasing elevational gradient (1,200 m to 3,760 m) are salt desert scrub, sagebrush-grass scrub, piñon-juniper woodland, coniferous forest, montane shrubland, and alpine tundra, with riparian habitat and meadows at a wide range of elevations (24).

We compiled a GIS database for the study area, including a second-generation map of vegetation types (Figure 1) (25). The vegetation map, which was generated from Landsat Thematic Mapper images and digital elevation data, had a 100-hectare mapping unit. We aggregated the 36 vegetation subtypes on the GAP map into the eight general vegetation types described above. To estimate vegetation density, we used the normalized difference vegetation index (NDVI), a transformation of near infrared (TM band 4) and red wavelengths (TM band 3) correlated with the amount and productivity, or rate of plant growth, of vegetation (5,26,27). The standard deviation of NDVI within a local area was calculated to estimate the uniformity of vegetation density at each field site. Elevation and slope (i.e., steepness) data were derived from the 2-arcsecond digital elevation model of the U.S. Geological Survey. Because riparian zones could influence rodent population densities and facilitate rodent dispersal across arid regions, we calculated proximity to streams and bodies of water on the U.S. Geological Survey's 1:100,000-scale digital line graph datasets.

Figure 2

Figure 2. Environmental strata within the mapped extent of the sagebrush-grass scrub vegetation type. Each color represents a unique combination of high or low vegetation density index, standard deviation of vegetation density index,...

In 1995, we sampled rodents at 42 sites before the GAP map became available. These sites were selected as representative of the five most common vegetation types in the Walker River Basin. In 1996 and 1998, full GIS datasets and the GAP map were used to distribute 102 new field sampling sites systematically across the widest possible range of environmental conditions. We categorized each GIS variable according to its relevance to each of the eight vegetation types. For example, 'distance to streams' was a meaningful distinction within salt desert but not within riparian habitat; elevation varied substantially within sagebrush scrub but not within alpine tundra. For each vegetation type, the relevant variables were divided into high and low ranges. The resulting binary classes for each variable were then intersected in GIS to produce distinct environmental "combinations", or strata, for each vegetation type (Figure 2). Randomly located sample sites were selected within each stratum so that they were within 0.5 km of a passable road and at least 1 km from any other sample site (Figures 1,3). The number of replicates within each stratum (including 1995 sites, which were included retroactively) was proportional to its spatial extent, with a minimum sample size of two. This GIS-based stratification is a more objective and randomized variation of the gradsect sampling method (28, 29).

All samples were collected from early June to early October to minimize seasonal effects on host density and antibody prevalence (17). Seasonal influences were minimized by sampling the replicate sites within each environmental stratum at different times throughout the field season.

Field and Laboratory Procedures

Deer mice were live-trapped at all field sites according to a fixed protocol (17). Each site had 48 live-traps in place for 3 days. A blood sample was collected from each deer mouse by retroorbital puncture with a heparinized capillary tube or Pasteur pipette. Blood samples were placed on dry ice and returned to the laboratory for enzyme-linked immunosorbent assay testing for immunoglobulin G antibody to SNV, which indicates current or past infections (14). Relative population density was estimated by counting the number of animals captured during a trapping session.

Vegetation Types

Figure 3

Figure 3. Seroprevalence of IgG antibody to Sin Nombre virus at sample sites in the Walker River Basin (According to Status 1, sites with 0% prevalence were negative and all other sites were...

The proportion of positive sites in salt desert scrub (34% of 29 sites by Status 1, 14% by Status 2) was significantly lower than in any other vegetation type (p = 0.05 criteria for significance). No significant differences were found among any of the other seven vegetation types, where positive sites were more common by both Status 1 (50% to 100%) and Status 2 (50% to 83%) (Figure 3). By assigning the predominant infection status to all sites within a given vegetation type, overall classification accuracies of 76% (Status 1) and 59% (Status 2) could be achieved (Table 1). The Status 1 criterion resulted in better classification accuracy (for negative sites and for all sites combined) than Status 2. For both Status 1 and Status 2, positive classification was more accurate (88%) than negative classification (50%).

DFAs for both Status 1 and Status 2 produced significant canonical correlations showing that negative sites were associated with low elevations and sparse vegetation (Table 2). These qualities most often occur in salt desert scrub (24). In contrast, positive sites were higher and generally had more dense but less uniform vegetation. Slope and distance from streams were relatively unimportant factors. For both Status 1 and Status 2, negative classification was more accurate than positive classification (Table 1). Positive classification was more accurate in Status 2 than in Status 1.

Spatial Versus Temporal Disease Patterns

Because the RS and GIS maps summarize relatively fixed spatial properties of the environment, we focused on investigating the corresponding spatial patterns of SNV infections. SNV infections also exhibit temporal dynamics (13,16-18,22,23) superimposed on the baseline spatial pattern. However, a robust temporal study would require many years of replicated, longitudinal field data, as well as real-time RS data describing temporally variable environmental characteristics (such as climatic variables) for the corresponding period. We did not incorporate weather or climate data into GIS because weather monitoring stations are widely scattered throughout most of the study area, preventing meaningful extrapolations to most of the field sites.

Sampling Design

Because characterizing large-scale spatial disease patterns requires a large sample size, we maximized the number of sites sampled rather than visiting fewer sites on multiple occasions. This cross-sectional approach captured substantial ecologic diversity and provided statistical replicates of sites with similar characteristics. The disadvantage of the approach was a degree of uncertainty in determining the actual infection status at each site. However, when generalization of results is an important goal, a large, replicated, and diverse dataset that has a modest degree of measurement error is statistically preferable to a smaller, more precisely measured but poorly replicated dataset (32).

Comparison of Methods (Table 1)

The vegetation type approach was based on possible relationships between infection status and a preexisting vegetation classification that might or might not be relevant to deer mice and SNV infections. DFA, in contrast, generated a linear function that best distinguished the properties of positive and negative sites. Our results suggest that DFA yields a better balance between classification accuracies for positive and negative status (especially for Status 2).

The vegetation type method could not classify negative status as effectively as the DFA, and balance between error rates for positive and negative classifications was poor. This could be a result of using predefined vegetation types (rather than making environmental distinctions from actual infection patterns) or inaccuracies in identifying and mapping vegetation types. Site visits suggested that the DFA identified sites with pronounced salt desert features more effectively than the vegetation map. The substantial environmental variability within the mapped extent of salt desert scrub was easily captured by the set of RS and GIS variables but was analytically "invisible" to our aggregated GAP map. Some variability might have been captured by the GAP map's 36 original vegetation subclasses, but using all these in our analysis would have presented serious statistical problems.

Other analytical approaches are possible that were not presented here. For example, decision tree analysis (33,34) offers advantages if nonlinear relationships exist; hierarchical information on the effects of each predictor variable is desired; or ease of interpretation is important (29).

Classification and Prediction Accuracy

Classification accuracy varied significantly between the Status 1 and Status 2 criteria (Table 1), with Status 2 giving better classification balance for DFA and Status 1 producing better results for the vegetation type analysis. Unfortunately, the biological significance of these analytical differences is difficult to determine. However, the infection status of 73% of the sites was classified similarly by the two criteria. The remaining 32 sites of ambiguous infection status might represent regions where infection status changes with relatively high frequency. If so, this produces an intrinsic limitation in the capabilities of the methods we present. The choice of technique might be based on the relative risks and costs of false-negative versus false-positive predictions.

Both methods may have occasionally been unable to detect positive sites because of failure to capture positive deer mice. The likelihood of this error would be proportional to the number of resident animals not captured at a site. Our longitudinal data (17; unpub. data, Boone et al.) suggest that the 3-day sampling sessions captured most of the animals present at a site. During four 7-day trapping sessions, 86% ± 9% of animals were captured during the first 3 days of trapping. Additionally, examination of 123 3-day sessions within the context of their extended longitudinal infection timelines (17; unpub. data, Boone et al.) suggested that infection status was classified with 85% (Status 1) and 81% (Status 2) accuracy.

A cross-sectional, replicated, and randomized sampling approach should capture most sites while they exhibit their most typical infection status. However, a 'background' rate of classification errors is to be expected regardless of analytical method, given the temporally dynamic nature of SNV infections (17,18,22,23). For instance, even where infection status is predominantly positive, some sites may be sampled during atypical periods when infection is temporarily absent; the reverse could also occur. Additionally, a subset of sites might frequently change their infection status and not exhibit primary infection status. Thus it might be difficult to improve upon the highest overall classification success we achieved (with DFA and Status 2 criterion), unless temporal infection dynamics are incorporated into the predictive model. Another option would be to omit sites from analysis if they fail to meet unambiguous criteria for positive or negative status; however, this might result in the loss of biological insight.

Future Directions

We explored the ability of RS and GIS data to predict the baseline spatial patterns of SNV infections across an ecologically variable landscape. Our findings should be at least somewhat relevant to a number of other regions in the arid western United States, especially if infection dynamics are ultimately driven by host connectivity patterns. To expand these findings, we developed methods to filter environmental data to remove statistical noise and a computer simulation model to explore infection dynamics on a variety of virtual landscapes. Further work will focus on the role of landscape structure in producing spatial patterns of disease (35). For instance, deer mice in small patches of salt desert scrub within a matrix of more desirable habitat types might be more likely to be infected than mice living in large contiguous regions of salt desert scrub. Finally, it would be useful to test other types of RS and GIS data as possible indicators of SNV infections.

Further work is needed to identify possible climatic correlates of periodic outbreaks and the degree to which useful indicators of these outbreaks can be derived from RS and GIS data sources. In contrast to predictions in large-scale outbreaks, specific a priori predictions of temporal SNV infection dynamics in local sites may remain difficult. Once infections are initiated at a site (presumably by random dispersal events), changes in antibody and virus prevalence cannot be easily explained by changes in host density or environmental factors (17,18). However, it should be possible to estimate the frequency (if not the specific timing) of new infections as a function of a site's local environment. Additionally, extended longitudinal studies could identify typical infection trajectories of sites based on their environmental characteristics or demographic profiles of their host populations. When combined, these approaches should advance our ability to quantify and predict disease dynamics and human risk.