Habitat Relations of Shrubsteppe Birds: A 20-Year Retrospective

Abstract.

Documenting bird—habitat relationships by statistical modeling has been a cornerstone of avian ecology for decades, but rarely is the predictive capacity of such models tested. To evaluate how well quantitative models of habitat relationships developed during an initial survey period predicted species distributions and/or abundances in a later period, in 1997 we revisited 13 shrubsteppe sites that we had previously surveyed from 1977 through 1983. Using multiple regression (linear and logistic) and classification and regression trees (CART), we developed habitat models for each species based on the “historic” period. R² values for these models ranged from 0.45 to >0.95. We then predicted bird species distributions and abundances by using the 1997 habitat attributes as inputs for the models derived from the earlier data. These models generally failed to predict 1997 bird distributions and abundances accurately; only 1 of 14 multiple regressions and 2 of 14 CARTs explained a statistically significant amount of variation in the target species. Thus, although the models may capture relationships between a species and environmental variables when aggregated over multiple years, they may not adequately predict the subsequent distribution and abundance of populations over shorter time scales. This result may limit the usefulness of multivariate habitat models in resource management.

Resumen.

La documentación de las relaciones entre h´bitat y aves a través de modelos estadísticos ha sido un pilar importante en la ecología de aves durante varias décadas. Sin embargo, raramente se ha puesto a prueba la capacidad predictiva de estos modelos. Para evaluar la capacidad de los modelos cuantitativos, desarrollados con datos de un periode de muestreo inicial, de predecir distribuciones y/o abundancias de especies en un periode posterior, en 1997 visitamos 13 sitios de estepa arbustiva que habíamos muestreado previamente desde 1977 a 1983. Utilizando regresiones multiples (lineales y logísticas) y dendrogramas de clasificación y regresión (CART, por sus siglas en inglés), desarrollamos modelos de hábitat para cada especie basados en el período “histórico.” Los valores de R² para estos modelos variaron entre 0.45 a >0.95. Luego, predijimos las distribuciones y abundancias de especies de aves utilizando los atributos de hábitat de 1977 como entradas para los modelos que derivaron de los datos antiguos. Estos modelos generalmente fallaron en predecir de forma exacta la distribución y abundancia de las aves en 1997; sólo una de 14 regresiones múltiples y 2 de 14 CARTs explicaron una cantidad significativa de la variación en una especie determinada. Por lo tanto, a pesar de que los modelos pueden capturar las relaciones entre una especie y variables ambientales cuando agrupan varios años, éstos no pueden predecir de forma adecuada las distribuciones y abundancias subsecuentes de poblaciones a escalas temporales más cortas. Este resultado puede limitar la utilidad de los modelos de hábitat multivariados en el manejo de recursos.

Introduction

A fundamental premise of avian community ecology and habitat management holds that the distribution of bird species is closely linked with habitat and that quantitative measures of habitat features can therefore be used to predict occurrence and abundance (Heglund 2002). This premise is based on the notion that habitat selection has been molded by natural selection, so individuals will seek out and remain in that portion of the environment in which their ability to survive, attract a suitable mate, and successfully reproduce is maximized (Cody 1985).

Understanding bird—habitat relationships has both theoretical and practical applications. Many theoretical models that seek to explain adaptive variation in animal behavior include as one of their essential elements the relationship between the number of individuals in a habitat and various aspects of environmental “quality” relating to the fitness of the individual within that habitat (Fretwell 1972, Emlen and Oring 1977, Mock 1983, Rosenzweig 1985, 1991). From a practical perspective, if we can identify habitat attributes that directly or indirectly influence bird population numbers, then we can target habitat management to enhance the persistence of species of concern (Morrison et al. 2006). Indeed, most efforts at managing or conserving bird populations involve some form of habitat management.

In 1977, we initiated a study in which we annually visited a series of sites in eastern Oregon and northern Nevada where we assessed bird abundances and measured floristic and physiognomic habitat variables (e.g., Rotenberry and Wiens 1980, Wiens and Rotenberry 1981). The resulting data (1977–1982) formed the basis of analyses that explicitly examined bird—habitat relationships. In 1997 we revisited and resampled a large subset of the original sites, using the same methods and same (albeit older) observers. Our goals were to (1) document any regional changes in bird abundances or habitat features that may have occurred in the intervening 14–20 years and (2) test how well quantitative models of habitat relationships developed during the initial survey period predicted species distributions and/or abundances in the later period.

Methods

Sites

The “historical” analyses of avian habitat relationships we present in this paper use data from 13 sites (transects) surveyed for birds two to six times (years) between 1977 and 1982 that were then resurveyed in 1997 (68 site-years;  Appendix A).

We use the term “site” to refer to a specific transect; although several transects might be within a few kilometers of each other, they were selected to be representative of different vegetation types and were not intended to be replicates. We did not select sites randomly. Within an area we chose a location dominated by big sagebrush (see Table 1 for scientific names of plants) and that contained Sage Sparrows (see Table 2 for scientific names of birds). If we conducted a second survey in the area, we chose a site dominated by a different shrub species characteristic of the area.

Bird Surveys

During a site visit, we surveyed a 600-m-long belt transect two to four times (85% four times) within a 3-day (90% within a 2-day) period, usually evenly split between the same two observers (JTR and JAW) and between morning and evening. Surveys consisted of a single observer slowly traversing the transect and noting all birds seen or heard within 250 m of the centerline. Methodological details have been presented elsewhere (Rotenberry and Wiens 1980, Wiens and Rotenberry 1981).

Our previous analyses of the historical data relied on estimates of absolute densities derived from Emlen-type distance methodology (Emlen 1971, Buckland et al. 2000). In each year we averaged the two to four runs of a transect to generate a single density value. Our index of a bird species' abundance for the present analysis was the maximum number of that species seen on any one of the surveys repeated at a site during a sampling period. We chose this method mainly to reduce the depressing effect of evening surveys, which invariably resulted in fewer detections. To ensure that our change in indices was unlikely to be responsible for any differences between current and previous analyses, we examined correlations between the two indices for each of the five species for which we expected to use regular multiple regression. (Those species for which we employ logistic regression to assess habitat relationships are unaffected, as they are simply scored as present or absent at a site regardless of abundance.) To be conservative, we calculated the correlations by using only those transects at which the species was detected; including all site-years would imply adding in 0,0 observations (if a species is absent both the density estimate and the maximum counted yield zero), thus inflating the correlations. These correlations were quite high (r = 0.79–0.99), and all were highly statistically significant (all P < 0.001), leading us to conclude that the results of our analyses for those five species are unlikely to be affected by the choice of index.

We deleted from analysis all species detected on fewer than 10% of the 55 historical surveys, yielding a total of 14 species (Table 2).

Vegetation Measurements

Details of the measurement of vegetation structure and composition have been given elsewhere (Rotenberry and Wiens 1980, Wiens and Rotenberry 1981). In brief, we used a stratified random design to locate 100 points within a 100-m-wide band centered on the transect. At each point we placed a 4-mm-diameter rod (“pin”) vertically into the vegetation and recorded (1) the identity of all plant species contacting the pin, (2) the physiognomic or structural attributes at the pin, and (3) the number of vegetation contacts (“hits”) in each decimeter height interval along the pin. We estimated coverage of plant species or physiognomic types as the proportion of pin-drops on which a species or physiognomic type contacted the pin. We calculated average number of hits in different height classes along the pin, as well as variation in the distribution of those hits among height classes, by using a diversity index (Hill's index; Hill 1973). We estimated richness and diversity (Hill's index) of shrub species as well. After eliminating all shrub species that were detected on five or fewer surveys, we had 12 species (or species groups for bunchgrasses) and 17 structural variables (8 based on physiognomic types, 6 based on vegetation contacts, and 3 others; Table 1).

Precipitation

Monthly precipitation totals for 1975–1997 were acquired from the Western Regional Climate Center, Historical Climate Information web site (http://www.wrcc.dri.edu/cgi-bin/cliMAIN.pl?orhart) for the weather station at Hart Mountain Refuge, Oregon. Of candidate weather stations throughout the region, only Hart Mountain was centrally located with respect to the distribution of our survey sites, was at an elevation similar to that of our sites, and had an uninterrupted series of monthly records for the period of interest. To characterize precipitation in a manner relevant to biological production in this ecosystem, we plotted the annual monthly accumulation of precipitation beginning in July and extending through the following June (“bioyear” precipitation; Daubenmire 1970).

Statistical Analyses

Region-wide changes in birds and habitat. To test for region-wide changes in the abundance index of a species, we used a paired t-test to compare the mean value of the index for each site for the period 1977–1982 (the number of years surveyed varied by site;  Appendix A) to the observed value for that site for 1997. If there were no region-wide changes between 1997 and the historical period, then the mean difference should be zero, nonsignificant by a t-test; positive values represent regional increase, negatives denote a decrease. We applied the same approach to detecting region-wide changes in vegetation attributes.

Habitat relationships—multiple regression. To assess distributional correlates of bird abundances (i.e., “habitat relationships”), we first used multiple regression of each species' abundance indices on plant species' coverage values and physiognomic structural indices. We used linear models for abundance of those five species occurring on at least half of the total number of surveys (56%–100%) and logistic models of occurrence for those nine species occurring on fewer surveys (11%–31%; Table 2).

Because of a relatively low ratio of observations to variables (55 to 29), we were particularly concerned about the possibility of overfitting multiple regression models, i.e., developing a relatively high R² value with the inclusion of a large number of explanatory variables in a model at the expense of the generalizability of the results. Statistical methods of producing reduced-rank models (i.e., models containing only a subset of the original explanatory variables), such as stepwise regression, are particularly prone to this problem (Harrell 2001, Whittingham et al. 2006). Instead, we choose to use an information-theoretic approach to model selection, which takes the number of parameters into account as well as the goodness of fit of a model (Burnham and Anderson 2002). This trade-off between number of variables and fit is quantified by Akaike's information criterion (AIC; Akaike 1973). Of a set of competing models, the best approximating model for the data is the model with the lowest AIC value; minimum AIC values are attained for models that fit the data well with the fewest parameters. The difference between the minimum AIC and the AIC of any other model is termed “ΔAIC;” absolute values of AIC are unimportant. Most often, this approach is employed with an a priori specification of a set of competing models (Burnham and Andersen 2002). As is true for most studies of habitat associations based on surveys, however, no unambiguous set of a priori models presents itself. Instead, we follow the approach of Young and Hutto (2002), in which we generate a set of candidate models based on statistical criteria, from which we select a “best” or most parsimonious model by using AIC.

For each species, we used SAS 9.1 (SAS Institute 2003) PROC REG (for those species for which linear multiple regression was appropriate) or PROC LOGISTIC (for those for which logistic multiple regression was appropriate) to generate “best subsets” of candidate models based on R² values (PROC REG) or chi-square score (PROC LOGISTIC). From these we extracted the model with lowest AIC plus all those with ΔAIC < 2.0 (a criterion suggested by Burnham and Anderson 2002). This procedure usually resulted in a large subset of candidate models, many of which often shared many of the same explanatory variables. To reduce this number further, we began deleting variables (and any model in which they appeared) beginning with the variable that appeared in the fewest candidate models, then the one appearing in the next fewest, and so forth, until only one model remained. In most cases this was the model with the lowest AIC.

Habitat relationships—CART models. As an alternative to standard regression, we also employed CART (classification and regression trees; Breiman et al. 1984) to generate habitat-association models of reduced rank. In general terms, CART employs a tree-building algorithm to determine a set of “if-then” logical conditions (splits) based on a set of explanatory variables that permit accurate prediction (if the dependent variable is continuous) or classification (if the dependent variable is categorical) of observations. CART models are intrinsically nonparametric and nonlinear and usually involve only a small subset of the initial set of explanatory variables. Splits were determined by the Gini criterion, which has a minimum value when all observations within a node belong to the same class or have the same value of the dependent variable (Breiman et al. 1984, De'ath and Fabricius 2000). To reduce the chance of overfitting, we required that each terminal node of a branch must contain at least two observations and that any node containing six or fewer observations could not be split. These latter two restrictions also serve as stopping rules for tree-building. CART has been shown to perform well in modeling bird—habitat relationships (O'Connor et al. 1999, Dettmers et al. 2002, Fearer et al. 2007).

Habitat relationships—site medians. As an alternative to using values based on modeling response to the details of vegetation, we also evaluated whether differences among sites in a species' observed abundances in 1997 were predictable simply by considering differences among the median abundances of that species among the sites during the historical period.

Evaluation of habitat-model predictions. If our historically based models accurately capture consistent habitat relationships of the target species, then these models should predict the abundance and/or distribution of species observed in 1997 from the vegetation measurements in the same year. To test this assumption, we applied the historic multiple-regression and CART models to 1997 vegetation data from each site to generate expected values for each bird species. We then compared expected values to observed values by calculating the correlation coefficient between the two. If the historical models do indeed capture consistent relationships, we should expect a strong positive correlation between expected values and observed values, and much of the variation in the observed 1997 abundances should be explained by the predicted values (i.e., have a high R² value). For those models that were significant in their predictions, we further examined the relationship between observed and predicted to see if it was consistent with our expectation of a slope of approximately 1 and an intercept of zero. Similarly, we used correlation to examine the relationship between median abundance in the historical data and 1997 observed abundances for each species.

All analyses were performed with SAS 9.1 (SAS Institute 2003) on a personal computer.

Results

Regional Changes

There were no substantive differences between survey periods in precipitation. The pattern of cumulative precipitation during the bioyear preceding the 1997 resurvey differed little from the average observed during the 1976–1982 historical period, although both tended to be slightly higher than the long-term average (1955–1999) or the average during the intersurvey interval (1983–1996) ( Appendix B).

In 1997 both grass and shrub coverage increased over that in the historical period, although small bunchgrasses (mainly Poa spp.) were the only category of plant to change significantly (Table 1). The shrub increase appeared to be associated with increasing sagebrush and bitterbrush, although neither alone was significant. Coverage of the invasive exotic cheatgrass in the two periods did not differ, even though it had appeared to have been increasing at all sites between 1977 and 1982. Litter coverage increased while average litter depth decreased and bare ground increased, whereas forb coverage was substantially less in 1997 than the historical average. There were no significant differences in any of the vegetation's vertical-contact statistics, nor in shrub species richness and diversity.

Apart from a statistically significant decrease in the distribution of the Loggerhead Shrike and marginally significant increases in the Sage Sparrow and decreases in the Rock Wren, there were no marked regional changes between the periods in the distribution and abundance of bird species (Table 2).

Habitat Relationships

Rather than present the details of each bird-habitat model for either the historical period (see Rotenberry and Wiens 1980, Wiens and Rotenberry 1981, Rotenberry 1986) or 1997, we are interested here in examining the temporal consistency of habitat relationships of each species. To do this, we consider how well habitat models derived from the historical data predict abundances observed in 1997.

By conventional criteria such as the amount of variation accounted for (R²), the regression-based habitat models derived from the historical data are adequate to excellent (Table 3). Of the 14 species analyzed, all had apparent P-values (deflated because of dependence between surveys taken in consecutive years at a site) of <0.001. These models account for 45% to 87% of a species' variation in abundance in the sample, with an average of around 70%. The CART models appeared equally adequate to excellent, with R² values ranging from 47% to 96% with an average just over 70% as well. In four cases the CART R² value was substantially greater (>0.10) than that for a species' regression models, in three cases it was substantially less (<0.10), and in the remainder it was largely the same.

These models (“good” by conventional criteria) did not predict abundances in 1997 at all well, however. Consider, for example, the results for the Sage Sparrow, one of the most widely distributed shrubsteppe birds. The model derived from the 1977–1982 data fit remarkably well (R2 = 0.76), yet the abundances this model predicted from the 1997 habitat data failed miserably (Fig. 1). In general, correlations between observed and expected values were equally poor for other species (Table 4) and were significant only for the Western Meadowlark. This latter relationship appeared to be driven mainly by our observation of many meadowlarks at a single site where they were predicted to be abundant (Fig. 2A); deleting that value drops the R² value to 0.13 with P > 0.22. Likewise, only two of 14 CART models based on historical relationships appeared to account accurately for distribution or abundance in 1997. Although the model for Brewer's Sparrow predicted increasing values for the species in an appropriate manner, the variance of observed values about that prediction also increased substantially (Fig. 2B). (Figures showing relationships between observed and predicted values for presence/absence, such as that for the Black-throated Sparrow, are not particularly informative as all data points are clustered at 0,0, 0,1, 1,0, or 1,1).

As a class, historical medians were clearly better predictors, although they still did not lead to significant observed/predicted correlations in the majority of cases (Table 4). The high values associated with the Green-tailed Towhee, Vesper Sparrow, and Rock Wren were driven by the absence of each species from most sites. We observed a relatively large number of Western Meadowlarks at a site where we had consistently observed them in the past (Fig. 2C), as in the regression-model results. The high value for Brewer's Sparrow is somewhat surprising, given that this species was the most variable in its abundance at a site from year to year. In keeping with this high variance, there was substantial scatter about the 1:1 predicted:observed line, increasing at higher predicted values (Fig. 2D).

Some insight into why the 1997 data fit poorly to the models derived from the historic data can be gained by considering the abundance estimates for each species for individual years at each of the six sites that were sampled in all 6 historical years (Fig. 3). The 1997 abundance estimates fell within the range of estimates for 1977–1982 for all except 10 of the 30 site—year combinations for the five focal species (supporting the statistical conclusion of no significant differences in abundances of these species between the time periods; Table 2). Nonetheless, many of the 1997 abundance estimates fell close to the extreme values recorded during the historic period (10/30; 20/30 if the values falling outside the historic range are included; Fig. 3). These 1997 extreme estimates were evenly distributed among site-years in which abundance was high or low relative to the historic range, but there were no systematic patterns of highs and lows for individual species. In short, abundances of the five species varied among years at each of the sites as well as among sites during 1977–1982 and were, if anything, even more variable in 1997, but in no systematic way in relation to the sites (and therefore the habitat measures from those sites) or the historic range of variation.

Discussion

Much of our conventional wisdom about species and their relationships to habitat is based on an assumption of animals in equilibrium with their environment: individuals make the right choice when confronted with environmental variation, and do so promptly. Changes in the environment, whether natural fluctuations or induced by human activities, whether for better or worse from the species' perspective, are assumed to be tracked faithfully and appropriately. To the extent that this is so, we expect to see generally repeatable relationships between species' abundances and/or occurrences and features of the habitat. This repeatability should enable us to construct models based upon relationships identified during one time period that can be used to predict abundances observed in another, not too distant period.

The models of bird—habitat relations that we developed on the basis of 6 years of data (1977–1982) were, by all accounts, good models. Yet these models generally performed poorly in predicting bird abundances in 1997 from concurrent habitat measures. The failure of such “good models” to generate accurate predictions is disturbing, for it calls into question both our understanding of bird—habitat relationships in these systems and, more importantly, the practice of basing habitat-management decisions on such models (Rotenberry 1986, Wiens 2002). What might be the causes of this poor performance?

Modeling Issues

Because we are dealing with a relatively low ratio of observations to variables, overfitting of regression models is always a concern (James and McCullough 1990). We minimized this problem by optimizing the trade-off between increasing variance accounted for and increasing bias associated with adding parameters to a multivariate model (Burnham and Anderson 2002). Perhaps we did not succeed. However, we also used an alternative approach, classification and regression trees, that is less prone to overfitting (at least if trees are properly “pruned”) and has been successfully used in previous analyses of avian habitat relationships (O'Connor et al. 1999, Fearer et al. 2007). Although the details are not presented here, those models invariably required fewer variables than did regression models to achieve similar R² levels. Nonetheless, they were not notably superior to regression models as predictors.

As an alternative, there has emerged a set of techniques called “niche-based” models (Soberón 2007), including Mahalanobis D², Pearson's planes of closest fit, GARP (genetic algorithm for rule-set production), and maximum entropy (Clark et al. 1993, Stockwell and Peters 1999, Rotenberry et al. 2002, Phillips et al. 2006). These approaches all seek to identify a set of variables that best describes the habitat in which a species is most likely to occur and to which an area newly sampled for those variables may be compared. However, these techniques have been used primarily to analyze species distributions over broad spatial scales, usually on the basis of land-use/land-cover variables generated from remote imagery and geographic-information systems, and with variable predictive success (Elith et al. 2006, Tsoar et al. 2007).

Variable Selection

Perhaps we measured the wrong variables. We recognize, as do most who develop bird-habitat models, that the variables we use are usually surrogates for the “truly important” environmental variables, those that are directly rather than indirectly linked to fitness. We justify this approach on the basis of the assumption that organisms also often select habitats by using surrogate predictors rather than direct variables such as food supply or predation risk (Hildén 1965). This shortcoming may apply to any statistically based modeling exercise, of course, but more often it is invoked when models fit poorly during their initial construction, to provide an explanation of why the models were not significant in the first place. On the other hand, perhaps we have identified good surrogates but have omitted important nonlinear relationships and/or interactions. Unfortunately, without any compelling reason to expect a particular interaction or nonlinearity, simply expanding the roster of variables to include their interactions and, say, quadratic transformations only serves to decrease further the ratio of observations to variables ratio and enhance the prospect of overfitting.

Alternatively, perhaps it is not the explanatory variables that are the problem, but the dependent ones. For the widespread species, our dependent variables are based on counts of individuals (C), which we assume are correlated with the number of individuals actually present (N). But we also know that the relationship between the number counted and the number present is mediated by the detectability of each species (p) (i.e., C = Np; e.g., Rosenstock et al. 2002). Thus, to analyze all our surveys collectively, we must assume that p for a species does not vary markedly by observer, site, or time. We think that these assumptions are robust for this data set. For example, we could detect no systematic differences between observers in abilities to detect any of the species, and for all species one observer was as likely to have yielded the maximum count of that species (the dependent variable we used) as the other (unpubl. data). Systematic differences in a species' detectability by site (or year) are most frequently due to differences in ambient environmental conditions, such as noise, inclement weather, and vegetation structure, but among sites these differences were minimal within a year and across years, including between the historic period and 1997. The relatively low-statured vegetation is visually open, it is generally easy to both hear and see these species (i.e., their detectability is high), sites are remote with essentially no anthropogenic noise, and we consistently conducted our surveys under good weather conditions. Given that 20 years elapsed between the first surveys in 1977 and the last ones in 1997, it is possible that the observers' skills had eroded; for example, hearing acuity tends to decline with age and may affect one's ability to detect certain species (Cyr 1981). One would expect this to be manifest in lower counts of species, but this was not the case (except for the Loggerhead Shrike, most often seen rather than heard in any event; Table 2). Thus, overall we feel much more confident about meeting the assumptions associated with using raw count data as a simple index than we do about meeting the assumptions necessary to produce adjusted abundances (Johnson 2008).

Scale Issues

It is possible that the scale of our analysis may contribute to the lack of predictive ability of our models. Our present analysis uses only local-scale variables, those measured over a 30-ha area (500 × 600 m2) where the survey was conducted. We have discovered, however, that at least some of these species' distributions vary as a function of landscape-scale attributes integrated over hundreds of hectares (Knick and Rotenberry 1995, 1999). For common shrubland species in particular, attributes such as size of the shrubland patch or homogeneity of vegetation within the patch are strongly associated with distribution and year-to-year persistence at a sampling point. Moreover, the relationship between a species' occurrence and local-scale attributes (e.g., percent cover of sagebrush) varies as a function of the landscape-scale attributes (e.g., size of the shrubland patch) that surround the local survey point (Rotenberry and Knick 1999). Nevertheless, we do not believe that omission of landscape-level variables is likely to have led to the failure of our local-scale models, mainly because of the relatively scant change apparent in the landscapes surrounding our survey points between the historical period and 1997. Indeed, we omitted from analysis two sites visited during the historic period that had undergone extensive landscape-level alteration by 1997.

Time Lags

Many shrubsteppe bird species demonstrate site tenacity, the tendency for individuals to return to the same breeding territory used in previous years. We have observed this to be the case even under circumstances where the local environment has undergone substantial habitat alteration between breeding seasons (e.g., Rotenberry and Wiens 1978, Wiens and Rotenberry 1985, Wiens et al. 1986). This tenacity suggests that the abundance of birds at a site could reflect habitat conditions existing when each individual first settled rather than the current conditions. This scenario seems to apply in southwestern Idaho, where the distributions of the Sage Sparrow, Brewer's Sparrow, and Horned Lark reflect a “memory” of previous habitat conditions (Knick and Rotenberry 2000). This “ghost of habitats past” seems to persist for at least 10 years. That such ghosts appear implies that some species may not rapidly and efficiently adjust to changed conditions at a local level and that populations may not be in equilibrium with habitat conditions. In the present case, the failure of older models to predict more recent distributions might be attributed to the inertia of distributions associated with site tenacity. However, changes in vegetation between the historic period and recent surveys are considerably less than those recorded in south-western Idaho, where landscapes are undergoing substantial large-scale changes associated with high fire frequencies driven by the invasion and expansion of exotic annuals, particularly cheatgrass (Van Home et al. 1997, Knick and Rotenberry 2000). Given the apparent lack of habitat dynamics of such magnitude in either the period prior to our initial surveys in 1977 or prior to resurveys in 1997, the effects of time lags in response of birds to changed conditions seem unlikely either to have influenced the models generated for the historic period or to have the eroded the fit of distributions to those models in the recent resurveys.

Concluding Remarks

One might be tempted to conclude that the bird-habitat models we generated for 1977–1982 simply did not “work”—certainly they predicted the distribution or abundance of most species in 1997 poorly. None of several reasons for the poor performance of these models (modeling issues, variable selection, scale issues, or time lags) seems satisfactory. At another level, however, the models worked quite well: by conventional criteria, most were statistically significant and accounted for a substantial amount of variation in the abundance or distribution of the target species. Therefore, we think that the key issue is not why these or other bird-habitat models might or might not work, but why models that did work over one set of years performed much more poorly when applied to the same areas, at the same spatial scale, with the same methods and observers, in the absence of major habitat changes, in another year.

Huston (2002) noted that the range of conditions sampled in a study will determine the accuracy, precision, and generality of any resulting predictions. Models based on a narrow range of environmental variation are likely to have low predictability because the random, unpredictable effects of dispersal/migration, disease, or other factors not correlated with environmental conditions will be much greater than any predictable responses to conditions. This might be the case for the Sage Sparrow, which occurred at all sites in all years; certainly, the range of environmental variation encompassed by our surveys did not exceed the tolerance of this species. The same may also be true for the Sage Thrasher, which was nearly as ubiquitous. For other species (and for the Sage Sparrow and thrasher, for that matter), the general lack of prediction may have resulted from our aggregation of surveys over multiple years to create models that were then applied to a single year. The probabilistic nature of habitat selection modified by dispersal/migration, disease, or other nonhabitat factors may yield high predictability (hence high R² values) of the distribution of a species among suitable habitats at a coarse temporal resolution but may also lead to low predictability of presence or absence at a fine temporal scale. In other words, the models may be “real” in the sense of capturing general relationships between a species and a set of environmental variables but be incapable of representing the natural variation associated with populations measured over short temporal scales. To a considerable degree, this situation describes the ecological paradox noted by Wiens (1981) over 25 years ago: ecological patterns are discerned on broad spatial or temporal scales of resolution, but when examined at a local scale these patterns can disappear, swamped by local variability. Although, in our historical surveys, variation in a species' abundance from year to year within a site and from site to site were both high (solid circles, Fig. 3), we still detected significant patterns of association between bird abundance and habitat attributes when we aggregated our surveys to produce our models (Table 3). The fact that one third of the site—species combinations depicted in Fig. 3 for 1997 (open circles) fell outside the historical ranges may indicate that our 6-year historical surveys did not capture the full range of variation in this system.

Because of profound changes in the structure and function of shrubsteppe ecosystems that have led to substantial (and continuing) loss and fragmentation of native vegetation, persistence of several shrubsteppe bird species is now a critical management concern (Paige and Ritter 1999, Knick et al. 2004). Despite the failure of our models to predict species' abundances or distributions, our results should not necessarily cause despair among resource managers attempting to protect and to conserve these species. What these results do suggest is that managers should not focus on the details of local habitat composition and configuration, and in particular not attempt to achieve some structure and composition “optimal” for a species. Instead, attention should be directed to preserving large tracts of relatively undisturbed shrublands, including those that may contain a mosaic of different shrubsteppe vegetation types. Doing so will capture more of the spatial variability and heterogeneity that characterizes these landscapes, providing the resiliency to accommodate shifting habitat associations.

Acknowledgments

Our initial research into shrubsteppe bird ecology was funded by grants from the National Science Foundation. Subsequent support for field travel was provided by academic institutions with which we have been affiliated: Oregon State University, University of New Mexico, Bowling Green State University, Colorado State University, and University of California-Riverside. We appreciate the companionship in the field of J. Daniels, D. Wiens, and M. Zuk and the long-term commitment of all members of the original Shrubsteppe Habitat Investigation Team. Patricia Heglund and Douglas Johnson provided helpful comments on the manuscript.

Literature Cited

Appendix A