Monitoring Survival Rates Of
Landbirds At Varying Spatial Scales: An Application Of The MAPS ProgramDaniel K. Rosenberg1,3,
David F. DeSante1, and James E. Hines2
INTRODUCTION Concern over reports of wide-spread declines of neotropical migratory songbirds (e.g., Robbins et al. 1989, Terborgh 1989) motivated the establishment of the Neotropical Migratory Bird Conservation Initiative, "Partners in Flight", an international cooperative program designed to reverse declines in migratory birds (Rogers et al. 1993) and landbirds in general (this volume). An important component of the initiative is to encourage establishment of monitoring programs, such as MAPS (Monitoring Avian Productivity and Survivorship [Desante et al. 1993]) to identify trends in species abundance and demographic parameters affecting these trends (Finch and Stangel 1993). Monitoring temporal trends of primary demographic parameters (fecundity, recruitment, and survivorship) that influence population size is especially important (Temple and Wiens 1989, DeSante and Rosenberg 1998), because environmental factors affect these parameters directly, and thus its impacts can be observed over a short time period. Because of buffering effects of transient individuals and density-dependent responses of populations, there may be substantial time lags between changes in primary demographic parameters and resulting changes in population size or density as measured by census or survey methods (Temple and Wiens 1989). Moreover, because of the vagility of most bird species, local variations in population size may often be masked (George et al. 1992) or accentuated (DeSante 1990) by varying levels of recruitment from a wider area. Thus, density of a species in a given area may not be indicative of population health, due to source-sink dynamics (Pulliam 1988). Primary demographic parameters are thus critical in understanding population dynamics and are directly applicable to population models that can be used to assess land-management practices (Noon and Sauer 1992). Demographic parameters of avian populations, such as density (Brown et al. 1995), productivity (DeSante and Burton 1994, Robinson et al. 1995), and survivorship (Burnham et al. 1996; Johnson et al. 1992), vary spatially and temporally. Estimating these sources of variation is important for understanding dynamics of populations and for detecting trends that may reflect reduced viability (Wilcove and Terborgh 1984). Lack of knowledge on the extent of temporal variation in demographic parameters often leads to incorrect conclusions regarding population health and makes it difficult to argue that specific population declines are noteworthy and deserve additional attention. Although detection of environmental influences on animal populations is difficult, especially considering the "noisy" nature of time-series data, such as estimated population trajectories (e.g., Botsford and Brittnacher 1992), the detection process can be considered a preliminary search for patterns to be tested in detailed field studies (Holmes and Sherry 1988, Botsford and Brittnacher 1992, Sherry and Holmes, this volume), or to provide information on system state (Nichols, this volume). In this sense, monitoring helps direct priorities for applied research. Furthermore, monitoring, if done in an experimental or quasi-experimental manner (Nichols, this volume), is necessary to determine the effectiveness of management actions designed to reverse population declines or bring about the recovery of small or threatened populations (Noon 1992). In 1989, The Institute for Bird Populations initiated MAPS, a cooperative effort among federal and state agencies, private organizations, and individual bird banders in North America to operate a continent-wide network of constant-effort mist-netting stations (DeSante et al. 1993, DeSante and Burton 1994). The MAPS Program was patterned to a large extent after the British Constant Effort Sites (CES) Scheme that has been operated by the British Trust for Ornithology since 1981 (Baillie et al. 1986) and that has become one of the cornerstones of the British Integrated Avian Population Monitoring Scheme (Baillie 1990). An important aspect of monitoring is the ability to detect demographic trends and investigate the scale at which they may be occurring. Regional trends in annual survivorship may occur due to large-scale weather changes or changes in the landscape that are large enough to affect many local populations similarly. Local changes or trends in annual survivorship, such as may occur in a specific national forest, may occur due to changes in the habitat quality, for example, from increased harvest. Understanding the scale over which demographic parameters are changing will thus be informative for identifying future research needed to isolate problems, and, once identified, determining management solutions. We evaluated the ability of MAPS to detect patterns of avian survivorship at different spatial scales. We view the results herein as a preliminary assessment, using a subset of the MAPS data. At the time of the symposia, only data through 1995 were available. Our analyses were restricted to Swainson's Thrush (Catharus ustulatus), a neotropical migrant that is a common breeder in western and northeastern North America in woodland environments, and winters in Central and South America (Ehrlich et al. 1988). We chose data from this species because it was one of the most commonly captured species in terms of both the number of stations in which it was captured and the numbers of individuals captured and recaptured. METHODS We used data from Swainson's Thrush populations in western North America to investigate patterns of survivorship among spatial scales. We compared variability of survival rates within and among scales. The spatial scales we investigated included: "locations", collections of neighboring stations, similar to the "cluster" (Rosenberg et al. 1999), biogeographic provinces, winter-migration scale (populations that winter in either Central America or South America as delineated by Marshall [1988]), and western U.S., the scale in which we pooled all stations. We used data from Swainson's Thrush that were marked and recaptured at 52 MAPS stations during 1992-1994, the only time period in which most stations had three years of data. Stations used in the analyses presented here were located in forest or mixed forest-scrub habitats. Typically, ten permanent net sites were distributed rather uniformly throughout the central 8 ha of the 20 ha study area, but were placed opportunistically at sites where birds could be captured most efficiently. One 12-m, 30-mm-mesh mist net was erected at each net site and the type and location of all nets were kept constant for the duration of the study. The operation of the nets was standardized; most stations were operated for six morning hours, beginning at sunrise, for one day per 10-day period, and for 8-12 consecutive 10-day periods, depending on latitude, from approximately May 1 to August 28. All unbanded birds captured were banded with USFWS bands and an attempt was made to identify all birds captured to age; sex could not be determined reliably for Swainson’s Thrush. DeSante and Burton (1994) provided details of establishment of net sites and their operation. Mark-recapture was used each year to allow estimation of survival rates. We fit
multinomial models to the capture-recapture data using program SURVIV (White 1983). To
select the most appropriate model, we used Akaike's information criterion (AIC); we tested
the significance between general and constrained models with likelihood ratio tests
(Burnham and Anderson 1992, Lebreton et al. 1992). We compared five models that allowed
survival rates ( Table 1. Comparison of models of annual survival probability for Swainson's Thrush populations in western North America.
a Goodness-of-fit (GOF) statistics were used to evaluate model fit; larger P-values represent better fit (Lebreton et al. 1992). b Likelihood Ratio Tests were used to determine significance of differences between the more reduced (simple) model (identified at the top of the column, and serves as the null hypothesis), and all other general submodels (listed under "Model," and serves as the alternate hypothesis). Higher P-values suggest little improvement in model fit with the more general model, and thus the reduced (simple) model is the most appropriate. c Models include survival probabilities that are allowed to vary among
locations ( d Akaike's Information Criteria (AIC); the minimum value represents the most appropriate model (Burnham and Anderson 1992, Lebreton et al. 1992). Table 2. Summary information on survival estimates among various spatial scales for Swainson's Thrush populations from western North America.
a Winter range was determined by Marshall (1988). b Sample size is the number of Swainson's Thrush banded and released in year one (1992). c Survival rates were estimated with a recapture probability (0.54) that was estimated from a model of common recapture probabilities among all subpopulations (Table 1). To evaluate the statistical power for detecting differences in survival rates and for
detecting trends among different spatial scales, we constructed hypothetical populations
using the approximate parameter estimates for the regional estimate of survival (0.45) and
recapture probability (0.54) from the Swainson's Thrush data (see Results). We used the
average number of individuals captured in the first year (1992) from the Swainson's Thrush
data as the number released each year (total of unmarked and marked individuals) for each
spatial scale in the simulated populations (Location: n=45, Biogeographic Zone: n=130,
Winter Range: n=395, and Regional: n=790 individuals). We created three hypothetical
populations, each with a different survival rate. Two of the three populations were
constructed such that they had survival probabilities  /2, ' =  - D /2, where D represented the effect size, that is, the difference
in survival rates between the populations. The third population was constructed to have
the mean survival rate, . These methods were similar to those in
Nichols et al. (1982). We investigated D from 0.05 to 0.40, with =0.45, and with 3
and 12 years of simulated data. We used the sample sizes (number of individuals banded in
year one; assumed to equal the number of releases in each year of the study) listed above
for each spatial scale, each reflecting the average number of individual Swainson's
Thrushes marked in the first year of sampling. Thus, statistical power comparisons among
spatial scales reflect the differences in population size attributed to each scale. We
estimated the statistical power of detecting differences among 3 populations at each
spatial scale; models permitting geographic variation in were compared to Model
{ , P}, in
which all stations were constrained to have equal survival probabilities. For all power
approximations we used an analytic approach rather than a Monte Carlo simulation approach.
We used the parameter values under the various models and protocols (e.g., number of
years) to create expected numbers of birds exhibiting each different capture history. The
expectations were then entered into SURVIV as data, and estimator bias and statistical
power were computed as described by Burnham et al. (1987:214-217).
We also evaluated statistical power for detecting trends indicating an exponential
decline (0.5, 1.0, and 3.0% annual declines) in survival rates with 12 and 20 years of
simulated data. An exponential decline would be one in which the survival rate for a given
year was a constant fraction of the previous year's survival rate, such that The effect of pooling data when differences among locations are not detected but
present was investigated by computing the percent relative bias of the regional survival
estimate. We used the estimates for each spatial scale, weighted by number of individuals
captured in year one at the scale being investigated to provide a regional (weighted)
estimate of survivorship and compared this value to that obtained by pooling all the data
and estimating a single regional estimate. If we take the estimates of survival for each
Swainson's Thrush population within a geographic scale and treat the estimates as the true
survival rate, then percent relative bias can be computed as: 100[
where , among locations using
hypothetical populations. The 3 "populations" were given the same parameter
values and years as was done for the power analyses, and percent relative bias was
computed as described above.
RESULTS Survival rate estimates varied widely in both point estimates and coefficients of
variation for the 52 stations for which Swainson's Thrush data were obtained, and for
which there was at least one recapture. Numbers of individuals captured and marked in year
one ranged from 1-76 ( There was a total of 17 locations used in the analyses, with 3.1 ± 0.5 stations/location. Number of individuals captured in the first year of banding ranged from 1-230 (46.4 ± 14.6 individuals) per location. Estimates of annual survival varied considerably among locations (Table 2). The coefficient of variation for each location was also fairly large, generally >20%. We identified 6 biogeographic provinces in which Swainson's Thrush were captured in western North America (Figure 1). Central Alaska included 2 locations, coastal areas in California and the Coast Ranges included 3 locations each, 4 locations were in the mountains of the Cascades and Sierra Nevada, 1 was in the Basin-Range Province, and 4 were in the Rockies. Within the scale of biogeographic provinces, number of individuals captured in the first year ranged from 40-282 (131 ± 35.9 individuals). Survival rate estimates varied from 0.39 to 0.49 (Table 2). Coefficients of variation ranged from 9-19% (13.1 ± 1.5%), reflecting considerable improvements in precision over the results at the scale of the location. Based on work by Marshall (1988), we identified locations which could be separated by the subpopulations assumed to winter in either Central America or South America (Figure 1). Central American wintering populations included 10 locations (556 individuals marked in year 1) and those from South America included 7 locations (235 individuals marked). We estimated 45% and 41% annual survival rates for the Central and South American wintering populations, respectively. Precision was good for these estimates (Table 2): the coefficients of variation was <10%. Figure 1. Sampling stratagy for estimating survival rates of Swainson's Thrush populations in western North America. Populations that winter in either Central or South America (Marshall 1988) were pooled for comparisons; these populations were pooled from a collection of biogeographic provinces that included coastal California, Coast Ranges, Cascades and the northern Sierra Nevada, Basin and Range, Rockies, and central Alaska. The subpopulations in the biogeographic provinces were pooled from a collection of "locations," which consist of 1 - 6 stations, each of which consist of a network of mist nets. For example, Swainson's Thrush in the Cascades/Sierra biogeographic province were part of the central America wintering population, and consisted of four locations (Mt. Baker, Wenatchee, Willamette, and Tahoe). Each of these locations consisted of one or more stations where birds were actually captured (Table 2).
The largest spatial scale was that of the entire sampled area of western North America. A total of 791 individuals were captured in the first year of banding. We estimated a survival rate of 44% for this pooled population (Table 2), with a 7% coefficient of variation. We estimated a recapture probability of 0.54 ± 0.05; this represented the "pooled" recapture probability estimated in all the previous analyses as well. The model tests, using both AIC and likelihood ratio tests, suggested the most
parsimonious, adequate model was that with a common survival rate among all locations
(Model { Statistical power to detect spatial differences in survival rates increased with number of years (length of study), with greater differences in survival rates between groups (effect size), and with larger spatial scales (sample size). The ability to detect different survival rates between geographic areas was greatest for a scale similar to winter-ranges; with 12 years of simulated data, high power was achieved for small D for the winter-range
scale (Figure 2). At the spatial scale of the individual location, adequate power
(>80%) was not achieved with 3 years of simulated data until D was approximately 30%.
With 12 years of simulated data, adequate power was achieved when D was approximately 10%
(Figure 2).
Figure 2. Statistical power to detect differences in survival rates of simulated populations in relation to maximum difference in survival ( D), number of years of
monitoring, and spatial scale. Spatial scale differences in statistical power reflect
differences in sample sizes among the spatial scales equivilent to that of location,
biogeographic province, and winter range. We estimated the statistical power of detecting
differences among 3 populations at each spatial scale using methods of Burnham et al.
(1987: 215); models with geographic variation were compared to Model {, P} in which
all stations were constrained to have equal survival probabilities.
When differences are not detected, but are present, the bias due to pooling heterogeneous samples must be considered. We found that bias of regional estimates is small when pooling data from within spatial scales in which heterogeneity of survival rates is low (Figure 3). However, bias increased with greater differences in survival rates (Figure 2). Fortunately, when large differences existed, the power to detect them was high (Figure 2). Percent relative bias was very low (<0.7%) with the Swainson's Thrush data; if differences existed, but were not detected, pooling these heterogeneous subpopulations would have negligible effects on regional survival estimates. Figure 3. Relationship of percent relative bias due to pooling simulated
populations with heterogeneous survival rates. When populations are pooled, but
when they have different survival rates, the regional estimate may be biased.
Percent relative bias was computed as 100[
 i
and Ri are the estimated survival rate and number of individuals marked and
released at year one, respectively, for a given ith population (i=1,k)
within a spatial scale (e.g., "location"). We found small bias with
relatively small differences in survival rates among simulated populations, but this bias
increased rapidly with greater differences among populations.
Although detecting spatial differences in survival rates is interesting and important, one of the primary goals of a demographic monitoring program is to detect trends in these rates. The ability to detect trends at different scales is related to three factors: number of years, effect size (magnitude of decline), and spatial scale (sample size of banded birds). Statistical power was inadequate to detect 0.5% annual declines (i.e., survival rates each year are 0.5% lower than they were the previous year) in survivorship with both 12 and 20 years of data for all sample sizes examined (Figure 4). Statistical power was >80% for detecting 3% annual declines with 12 years of data for the larger spatial scales (Figure 4). Power was increased substantially with 20 years of data. Nevertheless, 20 years of data were necessary to detect 3% annual declines for the spatial scale of location; this demonstrates the difficulty of detecting local trends in survivorship. Figure 4. Statistical power to detect exponentially declining
survivorship in relation to sample sizes (e.g., number of birds released at year 1),
length of monitoring, and percent annual decline for simulated populations.
Power was based on the likelihood ratio test of the negative trend model (
DISCUSSION Survival rates of landbirds may vary spatially and temporally. Examining survival rates in both time and space allows interesting and important management questions to be addressed. Our results suggest that these questions can be explored using a modeling approach with data collected from a monitoring program such as MAPS. We found weak evidence of spatial differences in survivorship for Swainson's Thrush populations in western North America; the most parsimonious model was that of equal survival rates within western North America. The statistical power analyses we performed suggested that geographic differences in survival rates would be difficult to detect with three years of data collected from 3 locations. Power would likely have been higher with a larger number of locations, though many locations typically had lower number of captures than we used in the simulations. The results of the power analyses simply demonstrate the difficulty in detecting differences in survival rates with the sample sizes that are typical at small spatial scales with MAPS data (DeSante and Rosenberg 1998). In cases where survival rates vary geographically, but power to detect such differences is low, the regional estimate from a model of a common survival rate among locations would be chosen as the most appropriate. If geographical differences in survivorship exist (e.g., among locations, biogeographic provinces, etc.), but are not detected, then pooling data may lead to a biased regional estimate. Fortunately, we found that only negligible bias will result when geographic differences in survivorship are low. When differences in survival rates between groups were sufficiently high to result in biased estimates (due to pooling) statistical power was sufficiently high to detect such differences. Similar results were reported by Nichols et al. (1982), although they reported lower power to detect geographic differences in survival rates than we did; this is likely due to the much higher capture probabilities used (and estimated from our field study) in our power analyses and the differences between band-recovery data vs capture-recapture models and data. In large-scale studies, geographic differences in survival rates should be evaluated prior to pooling data for a single regional estimate. If geographical differences exist, regional estimates could then be computed as averages of estimates from each geographic area. There was sufficient evidence to warrant concern over pooling all locations with the Swainson's Thrush data; however, bias resulting from this was negligible. Although we explored only bias resulting from pooling heterogeneous samples, an additional source of bias that may be present in the estimates is that resulting from the presence of transient birds in the sampled population (Pradel et al. 1997). Transients will negatively bias survival estimates of resident birds because they, by definition, will not be present at the sampling site in subsequent years (i.e., probability of surviving and being recaptured = 0). For example, in a coastal California site monitored for 7 years, 72% of unmarked Swainson's Thrush were estimated as "transients", resulting in an estimated survival probablity for resident birds of 0.61, compared to an estimate of 0.36 when all birds were assumed to be residents (D. Rosenberg, unpubl. data). The "transient" model (Pradel et al. 1997) requires four sampling periods to provide estimates of annual survival rates of residents, and thus was not used in the analyses presented in this paper because at the time of analyses, only three years of data were available. Use of this model, however, will likely increase the accuracy of survival estimates for resident landbirds at a site when a significant portion of the individuals are transients. In the case of Swainson's Thrush, the transient model fit the data much better (P = 0.20) than the model that assumed all individuals were residents (P = 0.0001) at a scale equivalent to the "location" in analyses of 1992-1995 MAPS data (Rosenberg et al. 1999). The transient model improved the rather poor fit of the non-transient models used in the analyses we reported here (Rosenberg et al. 1999). The estimated percent of transients for Swainson’s Thrush was 56% (Rosenberg et al. 1999). The survival rates for Swainson’s Thrush under the transient model(0.63) was much higher than reported here (0.44). The large difference in AIC values between transient and non-transient models (Rosenberg et al. 1999) provided strong evidence that transients affected survival rates. Another source of bias of survival rates is emigration (DeSante and Rosenberg 1998). Because we were not able to estimate emigration rates, we do not know how the small study areas of MAPS stations may have biased survival rates. Although models have been developed that allow estimation of survival when movements occur among study areas (e.g., Hestbeck et al. 1991), the low probability of movements between stations under the current MAPS study design will likely prohibit the use of these models. Detecting trends in survivorship will require long-term monitoring. Our results suggest that local trends of even relatively large annual changes will be difficult to detect. Trends that are regional in scope, however, will likely be detected, especially when the number of years of monitoring is large (e.g., >12). Although we were able to detect relatively large annual declines (3%) of survivorship with moderate sample sizes, such declines are unlikely to persist for long periods of time. What may be more important to investigate is the probability of detecting when survival has reached an a priori threshold value reflecting a decline (J. Sauer, pers. commun.). For example, based on a specific productivity, there exists a value of survivorship below which the population can be considered a "sink" rather than a "source" or stable population (Pulliam 1988). The ability to detect when a threshold value is attained may be more important, operationally, than detecting trends. This topic deserves further attention. The approach we used to evaluate competing hypotheses of variation and trends in survival rates among spatial scales can be a powerful tool to detect environmental variation that directly affects survivorship and thus population dynamics of landbirds. The results highlight limitations and strengths of the sampling methods used in MAPS, the only North American program for assessing survivorship of landbirds. The MAPS program allowed for large-scale (regional) estimates, but did not adequately estimate rates for local scales with the 3 years of data used in the analyses presented here. Additional years of data will increase power considerably. Large sample sizes, high capture probabilities, and many years of sampling are key to estimating and comparing survivorship among sub-regional spatial scales. ACKNOWLEDGMENTS We thank the Biometrics group, especially J. Nichols, at Patuxent Environmental Science Center for their collaboration on methods of analyses for MAPS data, and to the many individuals, organizations, and agencies that have contributed data to the MAPS Program, to K. Burton, E. Feuss, D. O'Grady, E. Ruhlen, H. Smith, P. Velez, and B. Walker for preparation of data, and to J. Nichols, C. Conway, B. Cooper, J. Gervais, and N. Nur for constructive comments on earlier drafts of this manuscript. Financial support for the MAPS Program has been provided by the National Fish and Wildlife Foundation, The U.S. Fish and Wildlife Service, and Regions 1 and 6 of the U.S.D.A. Forest Service. This is contribution No. 57 of The Institute for Bird Populations. LITERATURE CITED
1 The Institute for Bird PopulationsP.O. Box 1346 Point Reyes Station CA 94956 2 U.S. Geological ServicePatuxent Wildlife Research Center Laurel, MD 20708 3 Present Address:Oregon Cooperative Fish and Wildlife Research Unit Department of Fisheries and Wildlife Oregon State Unversity Corvallis, OR 97331 |
|||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
individuals) per station. The coefficients of
variation (
,
a measure of relative precision) for these estimates were typically >50%; this suggests
that these (3 year) estimates were not generally very useful. Therefore, we did not
consider the spatial scale of the individual station in further analyses, except in cases
where the individual station was equivalent to the "location" because there were
no other neighboring stations over which to pool data.
