Northeast Natural History Conference 2011: Selected Papers 2012 Northeastern Naturalist 19(Special Issue 6):143–158 Validation of a Citizen Science-Based Model of Site Occupancy for Eastern Screech Owls with Systematic Data in Suburban New York and Connecticut Christopher Nagy1,2,3,*, Kyle Bardwell1, Robert F. Rockwell2, Rod Christie1, and Mark Weckel1,2,3 Abstract - We characterized the landscape-level habitat use of Megascops asio (Eastern Screech Owl) in a suburban/urban region of New York and Connecticut using citizenscience methodologies and GIS-based land-use information. Volunteers sampled their properties using call-playback surveys in the summers of 2009 and 2010. We modeled detection and occupancy as functions of distance to forest and two coarse measures of development. AICc-supported models were validated with an independent dataset collected by trained professionals. Validated models indicated a negative association between occupancy and percent forest cover or, similarly, a positive association with percent impervious cover. When compared against the systematic dataset, models that used forest cover as a predictor had the highest accuracy (kappa = 0.73 ± 0.18) in predicting the occupancy observations in the systematic survey. After accounting for detection, both datasets support similar owl-habitat patterns of predicting occupancy in developed areas compared to highly rural. While there is likely a minimum amount of forest cover and/or maximum level of urbanization that Screech Owls can tolerate, such limits appear to be beyond the ranges sampled in this study. Future research that seeks to determine this development limit should focus on very urbanized areas. The high accuracy of the citizen-science models in predicting the systematic dataset indicates that volunteer-based efforts can provide reliable data for wildlife studies. Introduction As urbanized and suburban areas expand, many researchers are now investigating how more common and generalist wildlife species can—or cannot— tolerate and adapt to human presence and development. Information regarding “urban adapter” species can assist managers and developers to enhance biodiversity in developed areas or to design future developments with wildlife in mind. Megascops asio L. (Eastern Screech Owl) is a common raptor in eastern North America (Gehlbach 1995), and in the northeastern United States, it can be found in mixed young to middle-aged forest (Bosakowski and Smith 1997, Gehlbach 1995, Smith and Gilbert 1984). It is well-known for tolerating some amount of development—indeed there are many instances of Screech Owls preferentially selecting or having higher survival and reproductive rates in lightly developed areas and/or edge habitats, as compared to undeveloped, contiguous forest (Artuso 1Mianus River Gorge Preserve, Bedford, NY, 10506. 2Division of Vertebrate Zoology, American Museum of Natural History, Central Park West, New York, NY 10024. 3Biology Department, City College, City University of New York, 160 Convent Avenue, New York, NY 10031. *Corresponding author - cnagy@amnh.org. 144 Northeastern Naturalist Vol. 19, Special Issue 6 2009, Bent 1938, Gehlbach 1994, Smith and Gilbert 1984, Sparks et al. 1994). However, in extremely urbanized areas (e.g., New York City and the adjacent municipalities), it is not ubiquitous. Despite a wealth of general knowledge of the Eastern Screech Owl’s habitat, there has been little work (see Artuso 2009) done to quantify the response of Screech Owls to development over a large landscape that includes varying levels of development (e.g., how urban is too urban?). While Strix occidentalis Xantus De Vesey (Spotted Owl) and other less-than-common owls (i.e., Strix varia Barton [Barred Owl], Athene cunicularia hypugaea Molina [Burrowing Owl], and Aegolius acadicus Gmelin [Saw–Whet Owl]) have been the subject of numerous habitat modeling studies (Spotted Owl: Azuma et al. 1990, Carroll and Johnson 2008, Franklin et al. 2000; Barred Owl: Corbin 2007, Singleton et al. 2010; Burrowing Owl: Lantz et al. 2007, Stevens 2008; Saw–Whet Owl: Grose and Morrison 2010), the Eastern Screech Owl is not a species of concern and has not received similar management-oriented attention. Nevertheless, given the rapid spread of urbanized ecosystems, it seems prudent to better understand the impact of human development on habitat selection, population dynamics, and adaptation of Screech Owls. Studying a widespread yet cryptic species such as the Screech Owl across a fragmented and largely privately owned landscape presents substantial challenges for data collection. In suburban areas, most of the land is owned privately and would require immense logistical effort to obtain permission for access and to sample in a timely manner. Additionally, staff and funding resources for a non-game, non-threatened species are limited, and if one wishes to address landscape-level questions, traditional labor-intensive and costly methods (telemetry, mist netting, roost/nest box surveys) are simply not feasible. However, citizen-science methodologies are becoming increasingly well-developed by wildlife researchers who wish to obtain data over large areas and who recognize the importance of involving the local community in conservation and management decisions (Bonney et al. 2009, Dickinson et al. 2010, Silvertown 2009). While in some circumstances, citizen science certainly has its own limitations regarding sampling bias and feasible research objectives (Bonney et al. 2009, Lepczyk 2005, Nerbonne and Vondracek 2003, Webster and Destefano 2004), most concerns regarding the quality of data collected by citizen science observers vs. trained “experts” have proven relatively trivial if proper training is ensured (Cohn 2008, Galloway et al. 2006, Penrose and Call 1995). If researchers can properly frame their objectives, train their volunteers, and provide some independent validation of their volunteer data, citizen-science methodologies can be a very useful complement to more rigorous techniques, with the added benefit of involving stakeholders in local wildlife management. In this study, we sought to measure Screech Owl distribution over a threecounty area in Westchester and Putnam, NY and Fairfield, CT counties in relation to human development to evaluate the relationship between Screech Owl occupancy and urbanization. We used measures of development and forest cover around each survey site. Small-scale site-specific variables certainly play a role in 2012 C. Nagy, K. Bardwell, R.F. Rockwell, R. Christie, and M. Weckel 145 Screech Owl habitat selection (Belthoff and Ritichison 1990, Sparks et al. 1994), but we hoped to develop models that would find landscape-level patterns and that could be easily implemented in new areas by natural resource managers. We used a citizen-science (CS) framework to collect information on Screech Owl distribution over a tri-county area and build our initial occupancy model. To evaluate the efficacy of this effort, we tested our CS models against a smaller, independent dataset collected systematically (SYS) in a section of the larger study area (the town of Ossining in Westchester County, NY). If the models developed with the CS dataset performed well in predicting the SYS dataset, then we could be confident that our citizen-science methodology provided reliable estimates. In general, testing models with independent data is the ideal method of performance validation (Fielding and Bell 1997, Verbyla and Litvaitis 1989), and in addition to testing occupancy patterns of Screech Owls, our approach could further assess the congruence between citizen science and more traditional sampling frameworks. Field–Site Description The citizen-science component of the study was conducted in Westchester and Putnam counties, in New York State, and Fairfield County in Connecticut. These three contiguous counties lay on the eastern side of the Hudson River. The landscape was generally a suburban mix of residential towns with light commercial development, a few large cities, and in the northern sections, larger areas of undeveloped forest. There was a distinct urban–rural gradient, with urbanization declining as one moves north through the counties and away from New York City and, to a lesser extent, the City of Stanford, CT in southern Fairfield County. Putnam County was the most northern and rural of the three counties, with population densities along US census tracts ranging from 42 to 755 people/ km2 (mean = 167 people/km2). The population density of Fairfield County ranged from 133 to 14,207 people/km2 (mean = 567 people/km2). Westchester County had the steepest urban to rural variation, with population densities ranging from 87 people/km2 in the northernmost sections of the County to 20,812 people/km2 in the city of Mt. Vernon on the border of New York City (mean = 733 people/ km2; US Census Bureau 2010). Methods Field protocol Screech Owls are readily found using call-playback surveys (Cavanagh and Ritchison 1987, Johnson et al. 1981, Ritchison et al. 1988). Such surveys are inexpensive and easy to learn and perform, and thus lend themselves to a volunteer- based study. Survey protocol was identical in both the CS and SYS surveys: recordings of Screech Owl calls were played with a pattern of 20 seconds of calls (alternating between “bounce” and “whinny” calls; see Cavanaugh and Ritchison 1987) and 20 seconds of silence for 10 minutes. If an owl responded, the calls were stopped. All surveys were performed after dusk. 146 Northeastern Naturalist Vol. 19, Special Issue 6 Citizen-science survey The citizen-science aspect of this study was performed by volunteers who conducted call surveys on their properties at least twice in April to September of 2009 and 2010. Volunteer data collectors from the suburban Westchester and Putnam Counties, NY and Fairfield County, CT were recruited at local nature preserves, county parks, and schools in the spring of 2009. Additional volunteers were enlisted via the Ossining, NY School District in 2010 as part of a multiple-school science project. All recruitment sessions consisted of information and training workshops delivered by one or more of the authors. Participants learned about owl life history and species identification, habitat modeling and occupancy analysis, and how to conduct playback surveys at their homes. Field demonstrations were also provided following each workshop. Detailed directions, information, and downloadable owl call tracks were made available at a project website. The majority of the citizenscientist participants came from central Westchester. We encouraged participants to perform 4–6 surveys in a six-month period of April to September. We used this time frame to conduct surveys because owls most readily respond during the spring through late summer (Ritchison et al. 1988), and we thought volunteers would most likely perform surveys during the warmer months. Participants submitted data via an online survey or by direct email to project staff. Required data included date, survey address, time of survey, whether an owl was seen, heard, vocalized, or failed to vocalize. All participants conducted their surveys at their property or at a previously agreed-upon location. Street addresses were converted to GPS coordinates using Batchgeo (Holmstrand 2010). Our analysis required a minimum of 2 surveys per site if no owls were detected; a site that was surveyed only once was usable if an owl was detected. To maintain volunteer interest, we updated all participants with current project information on a monthly basis. This information included the posting of an interactive sighting map (via GoogleMaps API) that showed the current distribution of positive and negative sightings on the project website. We also held periodic project meetings and “owl walks” for interested participants. The downloadable recordings, data submission form, and progress maps were password protected on the website and accessible only to project participants in an attempt to minimize spam and “unauthorized” surveys as well as to keep owl locations somewhat confidential. The website home page, survey instructions, and recruitment information were public. Systematic survey In the summer of 2010, we established 30 systematic sites in the village (3024 people/km2) and surrounding town (696 people/km2) of Ossining, Westchester County, NY, which lies along the Hudson River in Westchester County about halfway between New York City and the northern extent of Putnam County. We chose Ossining to test our model owing to its diversity of habitat types from dense human development in the village to more forested area in the surrounding town. Points were initially set in a 500-m grid across the entire city; we later adjusted three points in the field for safety and security reasons because they 2012 C. Nagy, K. Bardwell, R.F. Rockwell, R. Christie, and M. Weckel 147 were near a state prison. We alerted nearby residents and the Ossining Police Department prior to each survey night. Each of these 30 sites was surveyed three separate times between June and July 2010. Habitat measurement Our chosen habitat covariates were measures of urbanization and human development and forest cover. We wanted our model to be easily used by managers and others, and thus used only easily available GIS-based covariate data. Land-use and vegetative cover information was obtained from the 2006 National Land Cover Database (NLCD; Fry et al. 2011). The NLCD provides land-use categorizations and percent impervious (pavement and buildings) cover at 30- x 30-m resolution for the entire United States. We characterized the amount of urbanization and forest cover in a 200-m-radius buffer zone around each CS and SYS survey site. We characterized the amount of urbanization of each survey site by the average percent impervious cover (%I) in the raster cells within the 200-m zone (12.5 ha, slightly larger than the average size [6–11 ha] of Screech Owl home ranges in suburban areas; Gehlbach 1995) around each site. Forest cover (%F) was estimated by the average number of forested cells out of total cells in the 200-m survey zone. These measurements are negatively correlated (e.g., a site with 100% impervious cover in all cells would have no forested pixels). However, there is room for considerable variation between the two measures since a cell can be categorized as forested and still have up to 80% non-forested area, and we thought it worthwhile to test each separately. All geographic measurements and calculations were performed using ArcGIS 9.3 and the Spatial Analyst extension. Analysis and model validation Our analysis consisted of two main phases. We first modeled detection and occupancy of Screech Owls in the CS dataset as functions of percent impervious or percent forested cover. Models that fit the data well were then used to predict the occupancy states of the SYS sites. Then, the actual observations of the SYS data were compared to these predictions. We chose this method in order to simulate how wildlife managers would use a predictive model to determine occupancy probabilities for unknown sites. Accuracy of a model was defined as how often the model’s predictions were correct after accounting for chance. Occupancy and detection modeling of CS dataset. Predictive habitat mapping is a powerful tool for wildlife managers seeking to determine where a species is likely found (Austin 2002, Guisan and Thuiller 2005, Guisan and Zimmerman 2000), prioritize conservation or restoration sites (Cabeza et al. 2004, Guisan and Thuiller 2005), identify corridors (Clevenger et al. 2002, Corsi et al. 1999), and investigate patterns of species distribution in relation to environmental factors (Freeman and Moissen 2008a, Guisan and Thuiller 2005, Guisan and Zimmerman 2000). We used the occupancy modeling methodology developed by MacKenzie et al. (2006), where detection rate (p) and occupancy (ψ) are logistically modeled using maximum likelihood, and candidate models estimating both are evaluated 148 Northeastern Naturalist Vol. 19, Special Issue 6 with Akaike’s information criterion (AIC; Burnham and Anderson 2002). In this study, we adjusted AIC to AICc for use with small sample sizes (Hurvich and Tsai 1995). We modeled p first under an intercept-only model of ψ and then used the AICc-selected best model for p while modeling ψ with covariates to reduce the size of the candidate model set (MacKenzie 2006, Negroes et al. 2010) and to enable comparisons of detection rate across the CS and SYS methodologies. All occupancy and detection modeling was performed with program Presence 3.0 (Hines 2006). Eastern Screech Owls are often characterized as “edge” species and may occupy large undeveloped forest patches less often because of competition with and predation by larger owls and hawks (Artuso 2009, Craighead and Craighead 1956), lower nest success, or lower population density (Artuso 2009, Gehlbach 1994). Therefore, we thought it worthwhile to test relationships other than simple linear responses and included quadratic terms in some of our models. Our candidate model set included 5 models: an intercept-only model ([.]), two 1-parameter models ([%I] and [%F]), and two quadratic models ([Q%I] and [Q%F]). Both detection and then occupancy were estimated with these models. Model validation with SYS dataset. AIC is a relative ranking of the models under consideration; the best model out of a set can still be poor if the entire set is poor overall. A relevant and well-chosen model set is an assumption of using AIC (Burnham and Anderson 2002), and ideally, models should be validated by comparing model predictions to independent data. In the case of habitat models, model-predicted probabilities should be compared to the number of actual presences and absences in a validation set collected from other locations (Fielding and Bell 1997, Verbyla and Litvaitis 1989). Our objective of validating our CS occupancy models with new data from the CS survey became somewhat complicated because we sought to incorporate detection rate. Not all of the CS sites where we did not detect owls could be assumed to be truly unoccupied. To quantify detection rate and determine which unoccupied sites could be considered true “absences”, we modeled detection rate in the SYS dataset identically to the CS dataset. The AICc-selected best model for p was then used to calculate pi for each SYS site. The probability of at least one owl detection given the three visits was calculated for each site based on the model-specific detection rates (Pfinali|pi, x surveys = 1 – [1 - pi]x). A site was assumed not occupied if the Pfinali was >0.85, i.e., we were comfortable with a 15% chance at most that we would include a site with no detections that was really occupied. If this probability was <0.85, the site was removed, as we concluded that there was a substantial chance that the site was occupied despite no detections. This process gave us a subset of presences and absences we were confident in using as a validation set. All CS occupancy models that performed better (i.e., had lower AICc scores) than the intercept-only models were considered in validation with the SYS dataset. Occupancy probabilities were calculated for each of the SYS sites; these predictions would be compared to the SYS presence and absence observations. In addition, based on the Akaike weights (wi), we included a set of predictions based 2012 C. Nagy, K. Bardwell, R.F. Rockwell, R. Christie, and M. Weckel 149 on the wi-averaged occupancy predictions of the top models. Model-averaged parameter values were not possible because the varying response types that we modeled (linear and quadratic) would confound parameter averages (Blums et al. 2005, Burnham and Anderson 2002, Wilson et al. 2007). To assess the predictive accuracy of a habitat model, the predicted probabilities must be converted to binary predictions of “presence” or “absence”, which are then compared to the observed presences and absences. A threshold probability must be set that groups predictions above the threshold as predicted presences and those below as predicted absences. Traditionally, this threshold is set at 0.5 (Freeman and Moisen 2008a). However, the choice of 0.5 is arbitrary, and model accuracy can vary greatly across threshold values (e.g., 0.4, 0.6, etc.). In addition, accuracy can vary greatly according to the species’ rarity or commonness across the study area (termed “species prevalence” in most habitat modeling studies; Allouche et al. 2006, Freeman and Moisen 2008a, Wilson et al. 2005). We calculated optimized threshold values so the observed prevalence—after accounting for detection—was maintained in the final predictions, as recommended by Freeman and Moisen (2008a). Lastly, there are numerous methods of scoring model accuracy, beyond the simple proportion of correct predictions, which take chance correct predictions into account (kappa: Freeman and Moisen 2008a, Landis and Koch 1977; true skills statistic (TSS): Allouche et al. 2006), and the difference between correctly predicting presences (sensitivity) and absences (specificity; Fielding and Bell 1997, Freeman and Moisen 2008a). We calculated the proportion correctly classified (PCC; correct predictions/total sites), sensitivity (correctly predicted presences/all predicted presences), specificity (correctly predicted absences/all predicted absences), kappa (model accuracy corrected for chance), and the TSS for each occupancy model as measures of accuracy. Finally, receiver-operating characteristics (ROC) plots allow for thresholdindependent measures of model performance (Freeman and Moisen 2008a, Manel et al. 2001; but see also Lobo et al. 2008). An ROC plot displays true positives (sensitivity) vs. false positives (1– specificity) across a large number of threshold values. A model that performs well will asymptote at 100% sensitivity at low levels of 1–specificity (see Fielding and Bell 1997 for examples). Thus, the area-under-curve (AUC) proportion shows how a model performs compared to randomly assigning observations (i.e., AUC = 0.5) independent of threshold, as generally better models will have larger AUC’s (e.g., >0.8 or >0.9). As additional measures of model performance, ROC graphs and AUC proportions were generated for the validation model set. These measures of accuracy would evaluate the agreement of the 2 datasets, and thus give a measure of congruence between the citizen-science data and the systematic data. Site– specific detections and occupancy predictions were estimated using Presence 3.0. Optimized thresholds, model accuracy indices, and ROC plots were calculated using the PresenceAbsence library (Freeman and Moisen 2008b) for R statistical software. 150 Northeastern Naturalist Vol. 19, Special Issue 6 Results One hundred ninety-seven individuals attended workshops or registered online to conduct the citizen-science survey. At the end of August 2010, 63 sites were usable in our analysis. Effort of these 63 participants ranged from 1 survey (12 sites) to as many as 7 surveys (2 sites) per site. Percent forest and impervious cover were similar in both studies (Table 1). Detection during the citizen-science surveys was best modeled by [p(Q%I)], a quadratic model of percent impervious cover (AICw = 0.71; β1 = 0.0978 ± 0.0142 (mean ± SE); β2 = -0.0027 ± 0.0005). Detection ranged from approximately 0.4 to 0.6, until it fell below 0.4 at 35% impervious cover. Both percent forest and impervious cover appeared in the top ψ models (Table 2). There was no evidence of overdispersion in any of the models (c-hat ≤ 1.0). In the SYS dataset, detection was best modeled linearly using percent forest cover, [p(%F)] (AICw = 0.72). Detection increased rapidly with percent forest cover (β1 = 0.1105 ± 0.0403); most sites had a derived p > 0.45. Out of 30 sites, an owl was detected at least once in 14 sites. Based on the modeled detection probabilities, 5 out of the 16 sites with no detections had ≥85% chance of being true absences after three visits and thus were included in the model validation phase. Optimized threshold values were calculated based on 5 absent sites and 14 present sites. Three models—[ψ(%F)], [ψ(Q%F)], and [ψ(%I)]—performed better than the intercept-only model and were compared to the SYS validation set. These models correctly predicted the occupancy status of 89% of the sites. However, indices other than PCC should be used as the final measure of model performance. Originally developed to assess agreement between observers (Cohen 1960, Landis Table 1. Model covariate means and sampled ranges used for occupancy estimation of Eastern Screech-owls in Westchester and Putnam, NY and Fairfield, CT counties, 2009–2010. % impervious % forest Citizen science Systematic survey Citizen science Systematic survey Average 16.1 26.2 33.5 15.6 SD 14.9 17.0 31.2 24.0 Minimum 0.0 0.2 0.0 0.0 Maximum 54.3 68.0 99.3 87.9 Table 2. AICc results for occupancy (ψ) models of Eastern Screech-owl observations using citizen science- based call-playback surveys in Westchester, Putnam, and Fairfield counties, NY and CT, 2009. Model* k ΔAICc wi –2Log(L) ψ(%F),p(Q%I) 5 0.00 0.45 208.07 ψ(Q%F),p(Q%I) 6 1.88 0.18 207.50 ψ(%I),p(Q%I) 5 2.17 0.15 210.24 ψ(.),p(Q%I) 4 2.29 0.14 212.72 ψ(Q%I),p(Q%I) 6 3.51 0.08 209.13 *Detection (p) modeled as [p(Q%I)]. 2012 C. Nagy, K. Bardwell, R.F. Rockwell, R. Christie, and M. Weckel 151 and Koch 1977), kappa has been widely used to validate confusion matrices of species presence-absence data. It is considered superior to PCC because it takes chance agreement between observed and predicted results into account. Kappa itself has been criticized for being biased at low and high levels of prevalence, and Allouche et al. (2006) recommended TSS as a prevalence-unbiased measure of accuracy. In this study, kappa and TSS agreed on the most accurate models (kappa and TSS = 0.73), the two forest-cover models ([ψ(%F)] and [(ψ(Q%F)]). The model-averaged predictions also had high kappa and TSS, while the [%I] model had the lowest kappa of the four (Table 3). ROC curves were also similar, and the four predictor models all had AUC scores > 0.85. Discussion Implications regarding Eastern Screech Owl ecology The two models with the most accurate predictions (as measured by kappa and TSS) both indicated that Screech Owl occupancy declined around >50% forest cover. These patterns were observed in the CS dataset as well (Fig. 1). We had expected to find a “humped” relationship between occupancy and our chosen covariates, such that Screech Owls would tend to occupy sites with moderate levels of development and forest cover. The advantageous characteristics of suburbia include more stable climate, larger and/or concentrated food sources, and fewer large raptors (Artuso 2009, Gehlbach 1994, Smith and Gilbert 1984), although the suburban landscape has its own perils as well, e.g., secondary poisoning, vehicles and more Procyon lotor L. (Raccoon), Didelphis virginiana Kerr (Opossum; which prey on eggs and nestlings), and Sciurus carolinensis Gmelin (Eastern Grey Squirrel; which compete for nest cavities). Forested areas may also provide more natural nest sites and more invertebrate and/or amphibian prey. Table 3. Accuracy of citizen science-based occupancy models on predicting systematic data of Eastern Screech-owl distribution in Ossining, NY, 2010. ψ Model OTA SensB SpecC PCCD KappaE (SE) TSSF (SE) AUCG (SE) ψ(Q%F) 0.740 0.93 0.80 0.89 0.73 (0.18) 0.73 (0.20) 0.89 (0.08) ψ(%F) 0.775 0.93 0.80 0.89 0.73 (0.18) 0.73 (0.20) 0.88 (0.09) wi–avg model 0.780 0.93 0.80 0.89 0.73 (0.18) 0.73 (0.20) 0.87 (0.09) ψ(%I) 0.760 0.86 0.80 0.89 0.62 (0.20) 0.66 (0.18) 0.87 (0.10) AOptimized threshold for categorizing a predicted absence or presence (Freeman and Moisen 2008a). BSensitivity = proportion of correctly classified presences out of all observed presences (Fielding and Bell 1997). CSpecificity = proportion of correctly classified absences out of all observed absences (Fielding and Bell 1997). DProportion of correctly classified observations (Freeman and Moisen 2008a). EProportion of correctly classified observations after accounting for the probability of chance agreement (Freeman and Moisen 2008a). FTrue skill statistic: (sensitivity + specificity) - 1 (Allouche et al. 2006). GArea under curve of receiver operator plot (Manel et al. 2001). 152 Northeastern Naturalist Vol. 19, Special Issue 6 Instead, our models seemed to predict a largely monotonically increasing relationship among Screech Owl occupancy and decreasing forest/increasing development. Taken literally, the model curves for the best supported model, [(ψ(%F)], seems to suggest that Screech Owl occupancy is nearly assured at <10% forest cover (Fig. 1). The prediction curve of [(ψ(Q%F)] was nearly identical to the simpler linear model of percent forest. The [(ψ(%I)] model also suggests that occupancy approaches 1 at >50% impervious cover (Fig. 2). However, the forestcover covariate we used is derived from satellite-based reflectance images and thus quantifies the landscape at a coarse level, and a forest cover of 0% by this measure does not necessarily mean the cell is devoid of trees. A given cell must have at least 20% forest cover to be classified as such, so there is ample room for some amount of overstory in a cell that is not classified as forested. Additionally, the maximum percent impervious cover we sampled was 67%, so all sites had at least some (>30%) non-impervious cover. It seems unlikely that Screech Owl populations could persist in areas that are at the farthest end of the rural-to-urban spectrum without access to any vegetated cover, and the monotonic relationships we observed are somewhat due to the fact that the maximum impervious cover sampled in our suburban Westchester, Putnam, and Fairfield sites was around 65–70%. However, since the models developed from the CS dataset were still able to predict the SYS dataset well, the functional pattern of high Screech Owl Figure 1. Predicted occupancy of Eastern Screech-owls as a function of percent forest cover under the [ψ(%F)] model. Curvilinear lines represent predictions, lower, and upper 95% confidence intervals (β1 = -0.0259 ± 0.0113). “+” marks along the prediction line represent citizen science sites. Filled and hollow diamond marks at ψ = 0 and ψ = 1 represent observed occupied and unoccupied SYS sites, respectively. Filled diamonds at ψ = 0 are occupied sites that were incorrectly predicted as unoccupied, and hollow diamonds at ψ = 1 are unoccupied sites that were incorrectly predicted as occupied. 2012 C. Nagy, K. Bardwell, R.F. Rockwell, R. Christie, and M. Weckel 153 occupancy in areas of less forest cover was supported across the sampled extent of forest and impervious cover. Our measure of percent forest cover seemed to fit suburbia well (ranging 0–99% fairly evenly), but a measure with a broader scale will be needed in more urbanized areas. We expected that our sampling area, particularly within Westchester County, would find the upper level of urbanization at which Screech Owl occupancy would decline, as previous studies have characterized areas with tree canopy cover of 71.1 to 96.5% (Gehlbach 1994) and percent greenspace cover of 46–70% (Artuso 2009) as “suburban” (although in Texas, Gehlbach [1994] also reported that Screech Owls can successfully nest in areas with tree densities as low as 53/ha). We had just two usable CS sites in very urbanized cities, an occupied site in White Plains (2048.5 people/km2) and a non-detection site in New Rochelle (2692.5 people/km2) (US Census Bureau 2010). Screech Owls are quite generalist in their habitat selection (Craighead and Craighead 1956, Gehlbach 1994), diets (Craighead and Craighead 1956, Gehlbach 1995, Marti and Hogue 1979, VanCamp and Henny 1975), and nest-site selection (Belthoff and Ritchison 1990), and some developed sites likely have features on a smaller scale that make them suitable despite their higher levels of urbanization. While further Figure 2. Predicted occupancy of Eastern Screech Owls as a function of percent impervious surface cover under the [ψ(%I)] model. Curvilinear lines represent predictions, lower, and upper 95% confidence intervals (β1 = -0.0569 ± 0.0402). “+” marks along the prediction line represent CS sites. Filled and hollow diamond marks at ψ = 0 and ψ = 1 represent observed occupied and unoccupied SYS sites, respectively. Filled diamonds at ψ = 0 are occupied sites that were incorrectly predicted as unoccupied, and hollow diamonds at ψ = 1 are unoccupied sites that were incorrectly predicted as occupied. 154 Northeastern Naturalist Vol. 19, Special Issue 6 research regarding the ability of dispersing owls to penetrate an urban matrix and occupancy patterns in highly urbanized areas is needed, it nevertheless appears that moderately developed urban areas (i.e., less than 70% impervious cover and/ or 20% forested cover) can be suitable for Screech Owls. We would still expect occupancy rate to drop off at some point with increasing urbanization, and for managers interested in making predictions in suburban or urban areas, we would therefore recommend using a model that includes a measure of forest cover in a nonlinear relationship, similar to [(ψ(Q%F)]. Implications regarding citizen science A key finding of this study was the substantial agreement among the citizenscience and the systematic methodologies. In addition to the ability of our CS occupancy models to predict the SYS data, average detection rates in the CS study were almost identical (CS p = 0.46 ± 0.02 ; SYS p = 0.46 ± .06), which illustrated that our volunteer observers were as effective as our trained staff. Most avian studies require expertise in identifying species either by sound or sight. Our methodology avoided this problem first by training and supporting volunteers, then—perhaps more importantly—employing a technique where the target species is easily identified (the call of the real owl the observer is listening for sounds the same as the call being broadcast). A number of volunteers also reported that the owl calls and pictures on the project web site helped them identify on the spot any other questionable birds they happened to hear (Barred Owl or Bubo virginiensis virginiensis Gmelin [Great Horned Owl] in a few cases, Zenaida macroura L. [Mourning Doves] in others). Similarly, the occupancy analysis developed by MacKenzie et al. (2006) lends itself well to a citizen-science program because it can incorporate varied numbers of surveys per site, missed surveys, multiple seasons, and incomplete covariate information. Still, at least 3–4 surveys per site are ideal, and it was a challenge to find an adequate number of volunteers to perform multiple surveys. Many points in our systematic study fell on private single-family residences. While most property owners allowed us on or near their property, there were several occasions where we were asked not to conduct owl calls anywhere in the neighborhood. Citizen science offers the possibility of accessing private properties by enlisting property owners as partners. Despite a potential sampling bias of using volunteers—who disproportionately may have lived near owls or were interested in owls and nature, joined because they already knew they had owls on their property, or perhaps were more likely to fail to bother to report negative results—we were able to sample a similar range of habitats in the CS survey as in the SYS survey (Table 1). The reliability of a citizen-science approach to sample species distributions is particularly helpful for biologists looking to draw inferences across large scales of urbanization but who may not have the resources or access they need (Dickinson et al. 2010). Citizen science certainly has implicit pitfalls, and potential sampling biases must be taken into account. Researchers may not be able to find volunteers to sample in particular areas of interest; in this study, we were 2012 C. Nagy, K. Bardwell, R.F. Rockwell, R. Christie, and M. Weckel 155 only able to enlist a few volunteers who lived in the extremely urbanized areas of southern Westchester County, likely due to a combination of less interest and access. As well, volunteer recruitment and training, continuing correspondence, and maintenance of volunteer interest are challenging and time-consuming tasks that researchers must not take lightly if they hope to have a successful program. Employing methods that require only limited technical expertise can help greatly, as was the case in this study. All methodologies have limitations, and the advantages of citizen science may include greater amounts of data over a large geographic area (Devictor et al. 2010, Dickinson et al. 2010), access to private properties that make up the majority of the landscape in urban landscapes and are often the habitats specifically of interest to (sub)urban ecologists (Webster and Destefano 2004, Weckel et al. 2010), and partnership amongst researchers and stakeholders (Bonney et al. 2009, Cooper 2007, Dickinson et al. 2010, Evans et al. 2005, Kransy and Bonney 2005). Acknowledgments We are greatly thankful to the numerous residents, students, and teachers who volunteered their time and energy to this study, and to J. Peltz, X. Tartter, and K. Caswell for their help with fieldwork. We would like to acknowledge the assistance from our numerous partner organizations that helped plan and run recruitment workshops: Teatown Reservation, Greenburgh Nature Center, Westchester Parks Department and Ward Pound Ridge Park, Westmoreland Sanctuary, Bedford Audubon, Audubon Greenwich, Friends of the Great Swamp, Putnam County Land Trust, Rockefeller State Park, and Saw Mill River Audubon. We’d like to thank the Ossining School District, the students and staff of Claremont and Roosevelt Elementary Schools, and M. Lockwood and G. Velardo. We are grateful for the thoughtful comments from two anonymous reviewers. This research was made possible by contributions from the AE Charitable Foundation. Literature Cited Allouche, O., A. Tsoar, and R. Kadmon. 2006. Assessing the accuracy of species distribution models: Prevalence, kappa, and the true skill statistic (TSS). Journal of Applied Ecology 43:1223–1232. Artuso, C. 2009. Breeding and population density of the Eastern Screech Owl Megascops asio at the northern periphery of its range. Ardea 97:525–533. Austin, M.P. 2002. Spatial prediction of species distribution: An interface between ecological theory and statistical modelling. Ecological Modelling 157:101–118. Azuma, D.L., J.A. Baldwin, and B.R. Noon. 1990. Estimating the occupancy of Spotted Owl habitat areas by sampling and adjusting for bias. USDA Forest Service General Technical Report PSW–124. Pacific Southwest Research Station, Berkeley, CA. 9 pp. Belthoff, J.R., and G. Ritchison. 1990. Nest-site selection by Eastern Screech-owls in central Kentucky. Condor 92:982–990. Bent, A.C. 1938. Life histories of North American birds of prey, Part 2. US National Museum Bulletin 170, Washington, DC. Blums, P., J.D. Nichols, J.E. Hines, M.S. Lindberg, and A. Mednis. 2005. Individual quality, survival variation, and patterns of phenotypic selection on body condition and timing of nesting in birds. Oecologia 143:365–376. 156 Northeastern Naturalist Vol. 19, Special Issue 6 Bonney, R., C.B. Cooper, J. Dickinson, S. Kelling, T. Phillips, K.V. Rosenberg, and J. Shirk. 2009. Citizen science: A developing tool for expanding science knowledge and scientific literacy. Bioscience 59:977–984. Bosakowski, T., and D.G. Smith. 1997. Distribution and species richness of a forest raptor community in relation to urbanization. Journal of Raptor Research 31:26–33. Burnham, K.P., and D.R. Anderson. 2002. Model Selection and Multimodel Inference: A Practical Information-Theoretic Approach. 2nd Edition. Springer–Verlag, New York, NY. 488 pp. Cabeza, M., M.B. Araujo, R.J. Wilson, C.D. Thomas, M.J.R. Cowley, and A. Moilanen. 2004. Combining probabilities of occurrence with spatial reserve design. Journal of Applied Ecology 41:252–262. Carroll, C., and D.S. Johnson. 2008. The importance of being spatial (and reserved): Assessing Northern Spotted Owl habitat relationships with hierarchical Bayesian models. Conservation Biology 22:1026–1036. Cavanagh, P.M., and G. Ritchison. 1987. Variation in the bounce and whinny songs of the Eastern Screech–owl. Wilson Bulletin 99:620–627. Clevanger, A.P., J. Wierzchowski, B. Chruszcz, and K. Gunson. 2002. GIS-generated, expert-based models for identifying wildlife habitat linkages and planning mitigation passages. Conservation Biology 16:503–514. Cohen, J. 1960. A coefficient of agreement for nominal scales. Educational and Psychological Measurement 20:37–46. Cohn, J.P. 2008. Citizen science: Can volunteers do real research? Bioscience 58:192–197. Cooper, C.B. 2007. Citizen science as a tool for conservation in residential ecosystems. Ecology and Society 12:11 Available online at http://www.ecologyandsociety.org/ vol12/iss2/art11/. Accessed 24 January 2011. Corbin, C.E. 2007. An assessment of Barred Owl (Strix varia) habitat suitability in Ocean Township, New Jersey, USA. Research Journal of Biological Sciences 24:413–418. Corsi, F., E. Dupre, and L. Boitani. 1999. A large-scale model of wolf distribution in Italy for conservation planning. Conservation Biology 13:150–159. Craighead, J.J., and F.C. Craighead, Jr. 1956. Hawks, Owls, and Wildlife. Stackpole, Harrisburg, PA. 443 pp. Devictor, V., R.J. Whittaker, and C. Beltrame. 2010. Beyond scarcity: Citizen-science programmes as useful tools for conservation biogeography. Diversity and Distributions 16:354–362. Dickenson, J.L., B. Zuckerberg, and D.N. Bonter. 2010. Citizen Science as an ecological tool: Challenges and benefits. Annual Review of Ecology, Evolution, and Systematics 41:149–172. Evans, C., E. Abrams, R. Reitsma, K. Roux, L. Salmonsen, and P.P. Marra. 2005. The neighborhood nestwatch program: Participant outcomes of a citizen-science ecological research project. Conservation Biology 19:589–594. Fielding, A.H., and J.F. Bell. 1997. A review of methods for the assessment of prediction errors in conservation presence/absence models. Environmental Conservation 24:38–49. Franklin, A.B., D.R. Anderson, R.J. Gutierrez, and K.P. Burnham. 2000. Climate, habitat quality, and fitness in Northern Spotted Owl populations in Northwestern California. Ecological Monographs 70:539–590. Freeman, E.A., and G.G. Moisen. 2008a. A comparison of the performance of threshold criteria for binary classification in terms of predicted prevalence and kappa. Ecological Modeling 217:48–58 Freeman, E.A., and G.G. Moisen. 2008b. PresenceAbsence: An R package for presence absence analysis. Journal of Statistical Software 23:1–31. 2012 C. Nagy, K. Bardwell, R.F. Rockwell, R. Christie, and M. Weckel 157 Fry, J., G. Xian, S. Jin, J. Dewitz, C. Homer, L. Yang, C. Barnes, N. Herold, and J. Wickham. 2011. Completion of the 2006 National Land Cover Database for the Conterminous United States. Photogrammetric Engineering and Remote Sensing 77:858–864. Galloway, A.W.E., M.T. Tudor, and W.M. Vander Haegen. 2006. The reliability of citizen science: A case study of Oregon White Oak stand surveys. Wildlife Society Bulletin 34:1425-1429. Gehlbach, F.R. 1994. The Eastern Screech-owl: Life History, Ecology, and Behavior in Suburbia and the Countryside. Texas A&M University Press, College Station, TX. 302 pp. Gehlbach, F.R. 1995. Eastern Screech-owl (Megascops asio). In A. Poole (Ed.). The Birds of North America Online. Cornell Lab of Ornithology, Ithaca, NY. Available online at http://bna.birds.cornell.edu/bna/species/165. Accessed 24 January 2011. Grose, J.E., and M.L. Morrison. 2010. Habitat use by Saw-whet Owls in the Sierra Nevada. Journal of Wildlife Management 74:1523–1532. Guisan, A., and W. Thuiller. 2005. Predicting species distribution: Offering more than simple habitat models. Ecology Letters 8:993–1009. Guisan, A., and N.E. Zimmermann. 2000. Predictive habitat distribution models in ecology. Ecological Modeling 135:147–186. Hines, J.E. 2006. PRESENCE—Software to estimate patch occupancy and related parameters. USGS-PWRC. Available online at http://www.mbr-pwrc.usgs.gov/software/ presence.html. Accessed 24 January 2011. Holmstrand, P. 2010. Batchgeo. Available online at http://www.batchgeo.com. Accessed 24 January 2011. Hurvich C.M., and C.L. Tsai. 1995. Model selection for extended quasi–likelihood models in small samples. Biometrics 51:1077–84. Johnson, R.R., B.T. Brown, L.T. Haight, and J.M. Simpson. 1981. Playback recordings as a special avian censusing technique. Studies in Avian Biology 6:68–75. Kransy, M.E., and R. Bonney. 2005. Environmental education through citizen science and participatory action research. Pp. 292–320, In E. Johnson and M. Mappin (Eds.). Environmental Education and Advocacy: Changing Perspectives of Ecology and Education. Cambridge University Press, Cambridge, UK. Landis, J.R., and G.G. Koch. 1977. The measurement of observer agreement for categorical data. Biometrics 33:159–174. Lantz, S.J., C.J. Conway, and S.H. Anderson. 2007. Multiscale habitat selection by burrowing owls in Black-tailed Prairie Dog colonies. Journal of Wildlife Management 71:2664–2672. Lepczyk, C.A. 2005. Integrating published data and citizen science to describe bird diversity across a landscape. Journal of Applied Ecology 42:672–677. Lobo, J.M., A. Jimenez–Valverde, and R. Real. 2008. AUC: A misleading measure of the performance of predictive distribution models. Global Ecology and Biogeography 17:145–151 Lynch, P.J., and D.G. Smith. 1984. Census of Eastern Screech-owls in urban open-space areas using tape-recorded song. American Birds 38:388–391. MacKenzie, D.I. 2006. Modeling the probability of use: The effect of, and dealing with, detecting a species imperfectly. Journal of Wildlife Management 70:367–374. MacKenzie, D.I., J.D. Nichols, J.A. Royle, K.P. Pollock, L.L. Bailey, and J.E. Hines. 2006. Occupancy estimation and modeling: Inferring patterns and dynamics of species occurrence. Academic Press, San Diego, CA. 324 pp. Manel, S., H.C. Williams, and S.J. Ormerod. 2001. Evaluating presence–absence models in ecology: The need to account for prevalence. Journal of Applied Ecology 38:921–931. 158 Northeastern Naturalist Vol. 19, Special Issue 6 Marti, C.D., and J.G. Hogue. 1979. Selection of prey size by Eastern Screech-owls. The Auk 96:319–327. Negroes, N., P. Sarmento, J. Cruz, C. Eira, E. Revilla, C. Fonseca, R. Sollmann, N.M. Torres, M.M. Furtado, A.T.A. Jacomo, and L. Silveira. 2010. Use of camera-trapping to estimate Puma density and influencing factors in Central Brazil. Journal of Wildlife Management 74:1195–1203. Nerbonne, J.F., and B. Vondracek. 2003. Volunteer macroinvertebrate monitoring: Assessing training needs through examining error and bias in untrained volunteers. Journal of the North American Benthological Society 22:152–163. Penrose, D., and D. Call. 1995. Volunteer monitoring of benthic macroinvertebrates: Regulatory biologists’perspective. Journal of the North American Benthological Society 14:203–209. Ritchison, G., P.M. Cavanagh, J.R. Belthoff, and E.J. Sparks. 1988. The singing behavior of Eastern Screech-owls: Seasonal timing and response to playback of conspecific song. Condor 90:648–652. Silvertown, J. 2009. A new dawn for citizen science. Trends in Ecology and Evolution 24:467–471. Singleton, P.H., J.F. Lehmkuhi, W.L. Gaines, and S.A. Graham. 2010. Barred Owl space use and habitat selection in the Eastern Cascades, Washington. Journal of Wildlife Management 74:285–294. Smith, D.G., and R. Gilbert. 1984. Eastern Screech-owl home range and use of suburban habitats in southern Connecticut. Journal of Field Ornithology 55:322–329. Sparks, E.J., J.R. Belthoff, and G. Ritchison. 1994. Habitat use by Eastern Screech-owls in central Kentucky. Journal of Field Ornithology 65:83–95. Stevens, A.F.J.M. 2008. Identifying potential critical habitat for western burrowing owls (Athene cunicularia hypugaea) in the Canadian prairies. M.Sc.Thesis. University of Alberta, Edmonton, AB, Canada. US Census Bureau. 2010. Available online at http://www.census.gov. Accessed 9 May 2011. VanCamp, L.F., and C.J. Henny. 1975. The Screech-owl: Its life history and population ecology in northern Ohio. North American Fauna 71. Verbyla, D.L., and J.A. Litvaitis. 1989. Resampling methods for evaluating classification accuracy of wildlife habitat models. Environmental Management 13:783–787. Webster, C.M., and S. DeStefano. 2004. Using public surveys to determine the distribution of Greater Roadrunners in urban and suburban Tuscon, Arizona. Proceedings of the 4th International Urban Wildlife Symposium. University of Arizona, Tucson, AZ. 359 pp. Weckel, M.E., D. Mack, C. Nagy, R. Christie, and A. Wincorn. 2010. Using citizen science to map human-Coyote interaction in suburban Westchester County, NY. Journal of Wildlife Management 74:1163–1171. Wilson, S., K. Martin, and S.J. Hannon. 2007. Nest survival patterns in Willow Ptarmigan: Influence of time, nesting stage, and female characteristics. Condor 109:377-388. Wilson, K.A., M.I. Westphal, H.P. Possingham, and J. Elith. 2005. Sensitivity of conservation planning to different approaches to using predicted species distribution data. Biological Conservation 122:99–112.