Year-round monitoring at a Pacific coastal campus reveals similar winter and spring collision mortality and high vulnerability of the Varied Thrush

Bird–window collisions are a leading cause of direct anthropogenic avian mortality, yet our state of knowledge regarding this threat relies heavily on eastern North American studies. Seasonal patterns of collision mortality may differ along the Pacific coast, and western North American species remain understudied. We therefore surveyed a stratified random sample of 8 buildings for collisions at the University of British Columbia, Vancouver, Canada over 45-day periods during 2 winters, 1 spring, 1 summer, and 1 fall season between January 22, 2015 and March 15, 2017. After accounting for the rate of scavenging and efficiency of observers in finding carcasses, we estimated that 360 collision fatalities (95% CI: 281–486) occurred over 225 days of collision monitoring. Collision mortality was highest in fall, but in contrast to most published research, collision mortality was intermediate in both winter and spring and was lowest in summer. In winter 2017, we performed point-count surveys to assess whether individual species are disproportionately vulnerable to collisions when accounting for population size and found that the Varied Thrush ( Ixoreus naevius ) was 76.9 times more likely to collide with buildings, relative to average species vulnerability in winter. To our knowledge, this is the first study to report the Varied Thrush as a species that is disproportionately vulnerable to collisions. Further studies are needed to assess the vulnerability of Western North American species and subspecies, and to determine whether similar patterns of seasonal collision mortality are found elsewhere. observateurs pour trouver des carcasses, nous avons estimé que 360 cas de mortalité par collision (IC 95%: 281–486) se sont produits au cours des 225 jours de suivi des collisions. La mortalité par collision était plus élevée en automne, mais contrairement à la plupart des recherches publiées, la mortalité par collision était intermédiaire en hiver et au printemps et plus faible en été. À l’hiver 2017, nous avons effectué des relevés par point d’écoute pour évaluer si certaines espèces sont disproportionnellement vulnérables aux collisions lorsqu’on tient compte de la taille de la population et nous avons découvert qu’ Ixoreus naevius était 76,9 fois plus susceptible d’entrer en collision avec les bâtiments, comparativement à la vulnérabilité moyenne des espèces en hiver. À notre connaissance, il s’agit de la première étude à signaler qu’ I. naevius est une espèce disproportionnellement vulnérable aux collisions. D’autres études sont nécessaires pour évaluer la vulnérabilité des espèces et sous-espèces de l’ouest de l’Amérique du Nord et pour déterminer si des patrons similaires de mortalité saisonnière par collision se retrouvent ailleurs.


INTRODUCTION
Bird-window collisions are a leading cause of avian mortality in North America (Calvert et al. 2013, Loss et al. 2015 with an estimated 16-42 million birds killed annually in Canada ) and 365-988 million killed per year in the United States (Loss et al. 2014). Collisions with glass affect adult and young birds of both migratory and resident species (Klem 1989, Hager et al. 2008) and typically occur for 2 main reasons: (1) birds mistakenly perceive the reflection of vegetation or open sky in glass windows as an extension of the environment, or (2) birds attempt to fly through transparent glass because they do not perceive it as a solid barrier (Banks 1976, Klem 1989, 1990, Martin 2011. Most collision research has been conducted during the fall and spring migratory seasons in central and eastern North America, creating both a seasonal and a regional bias in published results (Loss et al. 2014; but see Kummer and Bayne 2015, Kahle et al. 2016and Hiemstra et al. 2020).
Year-round collision studies found that collisions occurred most frequently during fall migration, followed by the spring migratory period, and were least frequent in winter (Hager et al. 2008, Borden et al. 2010, Schneider et al. 2018. Seasonal patterns of collision mortality may differ in coastal western North America, due to differences in vegetation, climate, topography, and associated bird movements and distribution. Western migrants can undergo looped migration, whereby they track greener coastal routes more strongly during spring migration, and switch to more direct interior routes and higher elevation stopover sites that are more productive in the fall (Wilson and Martin 2005, Carlisle et al. 2009, La Sorte et al. 2014. This tendency may put coastal migrants at greater risk of collisions at spring migratory stopover sites in low-elevation cities, compared to fall, when stopovers may occur in less urban montane forest and subalpine areas. Winter bird collisions are reported to occur where abundant food sources concentrate birds in proximity to buildings, including homes with bird feeders and buildings with nearby fruit-bearing trees (Klem 1989, Dunn 1993, Brown et al. 2020. Pacific coastal regions of North America support high densities of overwintering birds, in contrast to the north-east and north-central United States (National Audubon Society 2018) where most winter collision research has been conducted, and therefore collision mortality in winter may be higher than is currently appreciated.
Understanding differences in species' vulnerability to building collisions is central to predicting potential population-level effects of this source of mortality (Loss et al. 2012, Cusa et al. 2015, Elmore et al. 2020), but species with primarily western North American distributions comprised only 4 of 100 species reported to be more vulnerable to collision mortality in the United States (Loss et al. 2014, Elmore et al. 2020. Broad-scale analyses suggest that life history characteristics such as nocturnal migration, insectivory, and preference for forested habitat are most commonly associated with increased vulnerability to collisions (Arnold and Zink 2011, Wittig et al. 2017, Elmore et al. 2020). These traits are not unique to eastern North American species, therefore, the relative absence of western species from lists of species most vulnerable to collision mortality may be an artifact of sampling bias. Further, many continentally distributed bird species (e.g., Swainson's Thrush [Catharus ustulatus]) include distinct western populations or subspecies due to divergent migratory routes, molting schedules, behaviors, and ecological requirements (Kelly and Hutto 2005, Rohwer and Irwin 2011, Delmore et al. 2012. The potential for reproductive isolation, cryptic species, or discrete populations across east-west divides (Toews andIrwin 2008, Rohwer andIrwin 2011) further complicates our ability to evaluate population-level consequences of bird-window collisions, in the face of limited data (Loss et al. 2012(Loss et al. , 2015. To begin to address knowledge gaps in seasonal variation and species' vulnerability to window collisions, we conducted standardized collision surveys across all seasons at 8 low-to mid-rise buildings in Vancouver, British Columbia (BC). We included trials to assess searcher efficiency (SE) and carcass persistence (CP) to correct for biases in estimated collision rates, and performed point counts of birds during 1 winter season to estimate population sizes at our Ornithological Applications 123:1-15 © 2021 American Ornithological Society study buildings. Our goals were to (1) assess seasonal differences in collision mortality, including the winter season which is under-sampled across North America, and (2) estimate the vulnerability of bird species and populations that collide with windows at our study site along the Pacific coast of western North America.

Study Area
We conducted our research at the 420-ha Vancouver campus of the University of British Columbia (hereafter referred to as UBC), along the Pacific coast of southern BC, Canada (49.2606°N, 49.2606°W). The campus is situated within the Fraser Lowlands (elevation: 0-152 m) bordered by the Strait of Georgia and the Coast Mountains ( Figure 1). UBC is located on a peninsula, separated from the City of Vancouver by Pacific Spirit Regional Park, a 874-ha park comprised mainly of 70to 130-year-old second-growth forest (Metro Vancouver 2019). Tree canopy at UBC is predominantly coniferous; dominated by western red cedar (Thuja plicata), Douglasfir (Pseudotsuga menziesii), western hemlock (Tsuga heterophylla), and grand fir (Abies grandis). About 58% of the campus is in park-like open condition, with mowed grass, a number of coniferous and evergreen broad-leaf plantings, and scattered deciduous trees; primarily red oak (Quercus rubra) and maple species (Acer spp.) (Dyck 2016). The campus and surrounding forest lie within the Coastal Western Hemlock (CWH) Biogeoclimatic Zone (DataBC 2018) characterized by warm dry summers and mild wet winters, with a climax forest dominated by western hemlock, as well as Douglas-fir and western red cedar (Pojar et al. 1987(Pojar et al. , 1991. Broadly, the region is part of the temperate rainforest biome, also known as the Northern Pacific Rainforest Bird Conservation Region, extending from the Gulf of Alaska through to Northern California, and sharing similar climate, habitat types, and bird species (Alaback 1996, North American Bird Conservation Initiative 2000, Rich et al. 2004).

Building Selection
We chose 8 buildings for bird-window collision monitoring at UBC, using a randomized design, stratified by building height and extent of surrounding vegetation (see Cosentino 2014, Hager et al. 2017; Figure 2). We used a map layer of 229 UBC building footprints and size-related attributes of all buildings >1 story, excluding parking garages (University of British Columbia 2015), to divide buildings into 2 categories: "short" buildings (2-4 stories) and "tall" buildings (5-12 stories). We used ArcMap 10.3.1 to generate a 50-m buffer around buildings and digitized all vegetation using the World Imagery base map (ESRI 2015) to calculate percentage of vegetation within 50 m of each building. Within each building height class, we then randomly selected 2 buildings with >50% vegetation cover within 50 m of the building and 2 buildings with <50% vegetation cover within 50 m, as defined in Hager and Cosentino (2014). Excluding 1-story buildings eliminated most maintenance sheds and utility buildings with few windows. We further specified that study buildings must be a minimum of 100 m apart, which resulted in 1 building being removed from the initial selection. We excluded a second building from our initial selection due to construction fencing which prevented access to the majority of the building perimeter.

Collision Surveys
We followed protocols outlined in Hager and Cosentino (2014) and Hager et al. (2017) to conduct a total of 155 bird-window collision surveys between January 22, 2015 and March 15, 2017 at our 8 study buildings. Surveys were conducted over five 45-day periods consisting of 2 winter seasons (January 22, 2015to March 7, 2015and January 23, 2017to March 15, 2017, 1 fall season (September 9, 2015 to October 23, 2015), 1 spring season (April 15, 2016to May 29, 2016, and 1 summer season (June 16, 2016 to July 30, 2016). We aligned our spring and fall collision monitoring with the calendar dates when a local banding station records the highest numbers of migrants (WildResearch 2015). We performed a clean-up survey the day before each monitoring period to remove all evidence of bird-window collisions that accumulated prior to survey periods, including carcasses and feathers. In winter 2015, we surveyed every second day for 45 days. For the second winter season and all other seasons we conducted daily surveys for 21 days followed by surveys every second day for an additional 24 days; for a total of 45 days. We stopped surveys for 6 days during the second winter season due to a heavy snowfall that prevented us from reliably finding collision evidence. Surveys resumed after a clean-up survey to remove any evidence that accumulated while surveys were halted, and the survey period was extended accordingly.
We conducted collision surveys in mid-afternoon as recommended by Hager and Cosentino (2014) because window collisions often peak in the morning and continue at a lower rate later in the day (Klem 1989, Kahle et al. 2016. Two surveyors worked concurrently, walking around the perimeter of buildings in opposite directions, searching for carcasses within 2 m of the building façades. We defined a façade as the exterior face of a building, generally facing the same direction. In some cases, the complexity of a building footprint required us to consider several portions of a building footprint as 1 façade (e.g., several sides of an alcove) if we could not reliably discern which portion of the façade that a bird may have hit. One surveyor, A.N.P., remained constant throughout all 5 collision monitoring periods, whereas the second surveyor(s) differed seasonally. Evidence of a collision included whole carcasses and partial carcasses (e.g., wing, feet, bill, and/or feather piles) Cosentino 2014, Hager et al. 2017). We conservatively specified that feather piles must include a minimum of 10 tail, wing, or body feathers confined to 50 cm diameter circular area (Ponce et al. 2010). Only 2 stunned/ injured birds were found during surveys, and both died shortly after discovery. We removed all carcasses immediately following each survey to ensure double counting did not occur and identified all intact carcasses to species. We also collected partial carcasses, including all feathers, and where possible these were later identified to species, genus or family level, using online resources (U.S. Fish and Wildlife Service Forensics Lab 2018), hard copy guides (Scott and McFarland 2010), or by comparing to specimens from the UBC Beaty Biodiversity Museum.

Correcting for Biases Due to Scavenging and Missed Carcasses by Observers
Collision frequency estimates can be negatively biased due to carcass removal by humans and scavengers prior to their observation by surveyors, and imperfect detection of carcasses by surveyors (Klem et al. 2004, Bayne et al. 2012, Riding and Loss 2018. Therefore, it is important to design trials to estimate bias parameters when comparing building mortalities among studies, or among seasons and regions of interest (Loss et al. 2015). To account for carcass removal and imperfect detection, we conducted CP and SE trials, then incorporated the results into statistical models using the Generalized Mortality Estimator in R package Ornithological Applications 123:1-15 © 2021 American Ornithological Society GenEst (Dalthorp et al. 2018a(Dalthorp et al. , 2018b to correct our collision mortality estimates (details below). We tested both intercept-only and season-specific models, but had insufficient data for both CP and SE trials to allow us to apply bias corrections at the species level, by building perimeter substrate or by other potential covariates.
Carcass persistence. We conducted 103 CP trials to quantify bias due to scavenging of carcasses. Trial periods were: February 16 to March 7 in winter 2017 (n = 33); April 20-28 in spring 2016 (n = 28); July 12-30 in summer 2016 (n = 28); September 10 to October 16 in fall 2015 (n = 14). We used carcasses from previously window killed birds, ranging in size from 4 g (Anna's Hummingbird [Calypte anna]) to 128 g (Steller's Jay [Cyanocitta stelleri]). Carcasses were stored in deep freeze and held under our Canadian Wildlife Service Salvage Permit, as outlined in the Ethics Statement. We placed thawed carcasses at a range of times throughout the day at a randomly selected study building and at a randomly selected façade. The location along each façade was also chosen randomly but was no closer than 2 m from a corner or edge of an adjacent façade. Carcasses were placed on a variety of substrates, depending on the randomized location chosen, including concrete, asphalt, river rock, bark mulch, grass, or under bushes and ground cover. Collision monitors were alerted to the CP study and to the location of CP trial carcasses. However, as a precaution, data were checked carefully following all collision surveys to ensure that carcasses from CP trials were not inadvertently included in the collision mortality dataset. We checked carcasses every 5-6 hrs after initial placement on the first day until early evening, in early morning and mid-day on the second day, and at mid-day on subsequent days until carcasses were no longer present (presumably removed by scavengers or campus maintenance staff ), or until the end of the study period, whichever arose first. Carcass removal time was the interval between placement of a carcass at a façade until it was removed by a "scavenger" and no longer detectable by observers, as defined by Riding and Loss (2018).
We used bias-correction functions within the R package GenEst to fit a CP (or parametric survival) model to estimate the amount of time a carcass would persist, given the conditions under which it arrived (Dalthorp et al. 2018a(Dalthorp et al. , 2018b. We tested whether a persistence model that explicitly varies depending on the season, or one that depends only on the intercept, best described the detection probability of a carcass over time. We fit 1-parameter exponential and 2-parameter (location and scale) Weibull models, and tested whether persistence parameters varied categorically by season. The location parameter in the exponential and Weibull models allowed covariates to shift the mean of the probability density function, whereas the additional scale parameter in the Weibull model allowed covariates to affect the degree of spread of the distribution. Since the sample size was small (n/K < 40) we chose the model with the lowest second-order Akaike's Information Criterion corrected for small sample size (AIC c ) goodness-of-fit statistic across models.
Searcher efficiency. We estimated SE by having a nonsurveyor place a total of 29 thawed bird carcasses during the winter 2015 (n = 1), winter 2017 (n = 20), spring 2016 (n = 3), summer 2016 (n = 3), and fall 2015 (n = 2) study periods. Carcasses ranging in size from 4 g (Anna's Hummingbird) to 160 g (Northern Flicker [Colaptes auratus]) were placed at randomly selected buildings, façades, and façade locations as described above for CP trials. Carcasses were removed from the trial if they were detected by at least one of the surveyors. Carcasses that were not detected on the first day were checked prior to each subsequent survey until they were located by observers, removed by a scavenger, or until the end of the collision monitoring period, whichever arose first. We used the R package GenEst to model SE as a function of the probability that carcasses are located by observers on successive searches (Dalthorp et al. 2018a(Dalthorp et al. , 2018b. The probability of finding a carcass on the first search after carcass arrival was denoted by the parameter, p, and k was the change in probability of finding a given carcass on the second and subsequent searches until the end of the collision monitoring period (Dalthorp et al. 2018a). We fit 2 models to estimate the observer bias parameters, including a model that depended only on the respective intercepts, and another model that allowed SE to vary by season. We selected the best model by comparing the goodness of fit of models using AIC c .
Bias-corrected mortality estimation: a 2-stage parametric bootstrap approach. We used the function estM in the R package GenEst to produce a bias-corrected estimate of mortality (Dalthorp et al. 2018b). estM estimated each carcass' contribution to the total mortality in each search interval by accounting for the carcasses missed due to scavenging/imperfect CP, and imperfect detection/SE (Dalthorp et al. 2018a(Dalthorp et al. , 2018b. We assumed that most birds that collided with buildings fell within 2 m of building façades (Hager et al. 2013, Hager and Cosentino 2014, Riding et al. 2020, and therefore set the parameter density-weighted proportion (DWP) or the proportion of carcasses expected to arrive in the area searched by observers (Simonis et al. 2018) to 1. The model uncertainty from the 2 bias-correction models (CP and SE models) was propagated into estimates of total mortality using parametric bootstrap methods (Efron and Tibshirani 1986). Specifically, function estM in the R package GenEst drew 1,000 sets of estimated carcass counts from the asymptotic distributions of the maximum likelihood estimators of parameters from the 2 bias-correction models. The result Ornithological Applications 123:1-15 © 2021 American Ornithological Society was a matrix of carcass contributions where each row corresponded to a carcass, and each column corresponded to a simulated realization of that carcass' estimated contribution to total mortality. These simulated outputs resulted in bias-adjusted mortality estimates that included corrections for differences in search interval (time-increments between searches on the same building façade), corrections for seasonal difference in CP, and a correction for imperfect SE.

Assessment of Relative Vulnerability to Collisions in Winter 2017
After 4 seasons of collision monitoring, we noted that some species were particularly prevalent in the collision mortality data. However, assessment of relative vulnerability to collisions requires an estimate of population size of species within the study area, as well as species-specific mortality counts. Therefore, in winter 2017, we conducted avian point-count surveys during our collision monitoring period, and corrected these counts for imperfect detection. We used raw collision mortality counts to calculate relative vulnerability of species to window collision mortality on UBC campus, as we did not have sufficient data to generate species-specific bias-corrected mortality counts as described above. To determine how population size influenced the likelihood of collision by species, we applied a 2-step approach: (1) we estimated the change in a total number of birds by species over time in response to population-level processes of mortality, immigration, and emigration; and (2) we regressed mortality estimates on the corrected population counts to determine a speciesspecific vulnerability rating.
Winter 2017 point-count surveys. We conducted 5-min point counts at a randomly chosen façade of each building approximately weekly between January 25, 2017 and March 12, 2017, for a total of 72 point counts over 9 survey days throughout the winter 2017 collision monitoring period. We minimized factors that could contribute to detection error by (1) recording all birds seen or heard by call notes within a small area to maximize the probability of detecting all individuals (i.e. a 50-m radius semi-circle, from the edge of building façades), and excluding birds that were only seen flying over; (2) conducting surveys during the morning hours when human foot traffic was low, and only in weather conditions favorable to maximum detection of all individuals (i.e. with no precipitation and winds below 13 km hr -1 ); and (3) using the same observer (K.L.D.) to conduct all point-count surveys to minimize variation in observer bias.
Population size estimates. We calculated the population size for each species using the Dail-Madsen model (Dail and Madsen 2011), which is a generalized form of the Royle (2004) binomial N-mixture model developed for open populations (i.e. the closure assumption is relaxed, allowing abundance to change over time). We chose this model for our winter point-count data because it (1) assumes that abundance patterns are determined by an initial territory establishment process followed by gains and losses resulting from mortality, movements into the population (recruitment), and movements out of the population and (2) accounts for imperfect detection probability. For all species, we assumed that individuals could move across all sites (building locations) and all days within the winter 2017 collision monitoring and point-count period. Thus, the number of gains across days was directly proportional to the number of survivors (individuals occupying winter territories), so we applied an auto-regressive model to recruitment, γ × N s [i, t -1]; where γ = recruitment; N = the number of individuals per site, i, for each species, s, on survey day, t. As a result of the relatively short detection distance, we used a starting value of 1 for probability of detection. We applied a Poisson distribution to all species in which the total number of individuals detected across all 9 surveys, N stot , was greater than 10. For species with N stot < 10, we applied either a negative-binomial or zero-inflated Poisson distribution to a simple intercept-only model and restricted immigration to zero. We then used the parameter estimates from this simple model as starting values for the more complex, open count model that allowed immigration across survey locations and days. For all models, we set the upper bound for discrete integration, K, at the greater value of N stot + 1, or 10. We summed across all sites the resulting site-specific population size estimates corrected for detection probability and movement into the study area, divided by the number of survey days (9), and multiplied by the number of buildings surveyed to obtain population estimates for each species, N corr . We calculated the root mean squared error on the residuals of each model to estimate residual standard error for each species' estimate of N corr . We estimated all parameters and obtained corrected population size estimates using the function, pcountOpen and back-transformed the estimates for detection probability from each model using the function plogis in the R package, unmarked (Dail and Madsen 2011, Fiske and Chandler 2011, Hostetler and Chandler 2015, R Development Core Team 2018. Calculation of relative vulnerability to collisions by species. Next, we used our raw mortality counts per species for winter 2017 and species population estimates at our 8 study buildings (N corr ) to assess the relative vulnerability of species to collisions during the winter 2017 season. We fit a simple linear regression to log 10 (X + 1) of the raw winter 2017 mortality counts on log10 of the population size estimates, N corr (Arnold andZink 2011, Loss et al. 2014). The use of the (X + 1) transformation on mortality counts allowed us to account for zero mortality for Ornithological Applications 123:1-15 © 2021 American Ornithological Society species that were observed in point counts but not in the mortality counts. We set regression coefficients for population size, N corr to 1.0, calculated residuals, then raised 10 to the power of the absolute value of the residuals, for all species observed in our point-count surveys. A slope of 1.0 assumes that mortality due to collisions for a given species is proportionate to their population size, as described by Arnold and Zink (2011). Vulnerability values represent the factor by which a given species is more (+) or less (-) vulnerable to collisions relative to the expected value for a species with average collision risk (residual = 0).  Table S1).

Corrections for Bias Due to Scavenging and Searcher Error
Carcass persistence models. Using AIC model selection, the best fit was a 2-parameter Weibull model that specified that location and scale both depended on the season (Supplementary Material Table S2). Median CP time was shortest in the fall (0.81 days) and longest in summer (6.54 days; Supplementary Material Table S3).
Searcher efficiency models. Of the 2 SE models compared, the intercept-only model (p ~ 1; AIC c = 73.7, Akaike weight [w i ] = 0.93) was better supported by the data than the model that allowed efficiency to vary by season (p ~ season; AIC c =78.9). SE was then calculated at a constant median rate of p = 0.57 (95% confidence interval [CI]: 0.40-0.73), with the fractional loss in efficiency over time at a constant median rate of k = 0.57 (95% CI: 0.33-0.79).

Bias-Corrected Collision Mortality Estimates, Seasonal Patterns
Our bias-corrected mortality estimate for our 8 study buildings was 360 collision fatalities (95% CI: 281-486) across the 225 days that corresponded to our 45-day winter 2015, fall 2015, spring 2016, summer 2016, and winter 2017 sampling periods. Corrected bird-window mortality estimates showed a similar pattern among seasons to raw collision counts, with the highest collision mortality in the fall, intermediate levels of mortality in winter and spring periods, and the lowest collision mortality in summer ( Figure 3).

Species Affected
We identified 119 of the 152 birds that collided at our study buildings to species or family level (23 species from 10 families; Supplementary Material Table S3). Averaging mortality counts for the 2 winter sampling periods (to avoid biasing collision data toward species present in winter), 38% of 98 carcasses that could be identified to Family were thrushes (Turdidae), 32% were sparrows (Passerellidae), and 15% were kinglets (Regulidae). Varied Thrush (Ixoreus naevius) was the most common species identified from raw mortality counts; when winter counts were averaged, they comprised 13% of all fatalities, with 88% of these fatalities detected in either winter or spring (only 2 fatalities in fall and none in summer). American Robin (Turdus migratorius), Fox Sparrow (Passerella iliaca), and Darkeyed Junco (Junco hyemalis) were also prevalent (>5 individuals identified when counts for the 2 winter sampling periods were averaged) among carcasses located over the entire study period (Supplementary Material Table  S3). Kinglets (Regulus spp.) were the most frequent colliders during the fall monitoring period, with a total of 15 fatalities recorded in fall, of which 7 could be identified as Golden-crowned Kinglets (Regulus satrapa).

Species Vulnerability to Collisions in Winter 2017
Winter 2017 population sizes. We observed a total of 641 individuals of 28 species over 72 point counts, with an average of 9 individuals per count during the winter 2017 survey period (January 25, 2017 to March 12, 2017; Supplementary Material Table S1). Population size  Table S1.

DISCUSSION
We observed 152 collision mortalities in 225 days of collision monitoring, however when corrected for biases due to scavenging and searcher error, we estimated that 281-486 fatalities occurred at our 8 study buildings across our 5 seasonal periods of study. The highest rate of collision mortality occurred in fall, but equivalent collision mortality occurred during winter and spring, at our study site. Among species detected on point counts in winter 2017, Varied Thrush, Spotted Towhee, and American Robin showed the greatest vulnerability to collisions, whereas American Crow, Pine Siskin (Spinus pinus), Glaucous-winged Gull (Larus glaucescens), and European Starling (Sturnus vulgaris) were least vulnerable (Table 1).

Seasonal Patterns of Collision Mortality
Consistent with all other studies that investigated collisions year-round in temperate regions of North America, we observed a pattern of peak collision mortality during the fall migratory period (Klem 1989, Kahle et al. 2016, Schneider et al. 2018. We therefore did not find evidence that looped migration, or use of higher elevation stopover sites by western migrants in the fall, resulted in relatively lower collision mortality in fall at our low-elevation campus. There may be considerable use of coastlines by fall migrants in both eastern and western North America either by choice or due to the inexperience of young passerines (coastal effect; Ralph 1978). Further, fall migrants are known to stay longer at migration stopover sites compared to spring migrants (de Zwaan et al. 2019, Covino et al. 2020, thereby potentially increasing their risk of window collisions during stopovers. Our study represents just 1 of 2 year-round collision studies on the Pacific coast (see Kahle et al. 2016). Therefore, further study is needed to determine whether the patterns of mortality that we observed in fall relative to the spring migratory period would emerge consistently in other western cities.
In contrast to other year-round collision studies performed primarily in central and northeastern United States, we found that collision mortality during winter was as high as mortality during the spring migratory period and mortality was lowest in summer (Hager et al. 2008, Borden et al. 2010. Hager et al. (2008) observed that bird collision mortality was influenced by a combination of local bird density and vegetation that attracts birds to the vicinity of building façades. Further, buildings with extensive glass that reflects vegetation, generally pose a higher collision risk than buildings with lower glass areas and little surrounding vegetation , Klem et al. 2009, Borden et al. 2010. In south coastal BC, bird densities are higher in winter, when residents are joined by short-distance latitudinal and altitudinal migrants (Boyle and Martin 2015), compared to summer when birds are dispersed on breeding territories (Campbell et al. 1997(Campbell et al. , 2001. These patterns may be attributed to the characteristics of the temperate rainforest zone, including year-round lush evergreen coniferous and broad-leaf vegetation, mild year-round temperatures, and abundant rainfall in winter, with drier conditions in summer (British Columbia (BC) Ministry of Forests and Range 2018). By contrast, TABLE 1. Species vulnerability to collisions during winter 2017 at the UBC, Canada. Relative vulnerability values account for population size, and indicate the factor by which a species has a higher (+ values) or lower (-values) vulnerability to collisions in winter relative to a species with average vulnerability. Vulnerability analysis followed methods outlined in Arnold and Zink (2011) and Loss et al. (2014).  (Kuchler 1964). The combination of high bird densities and potential for reflections of green foliage at our building façades in winter (Basilio et al. 2020) may have contributed to the differences in the magnitude of winter collision mortality observed between our study and other studies conducted at similar latitudes further east. Similar patterns of high winter collision mortality may occur in other Pacific coastal and southern cities with comparable vegetation characteristics (abundant evergreen foliage) and high densities of overwintering birds.

Species Vulnerability to Collisions
Similar to Loss et al. (2014) and Wittig et al. (2017), we found that highly urban-adapted species such as the American Crow, Glaucous-winged Gull, and European Starling were among the least vulnerable species to collisions. This may be due to a greater familiarity with façade features at our study site, as a result of their tendency to perch atop campus buildings. The Varied Thrush had the highest vulnerability score in winter (more likely to collide relative to species with an average collision vulnerability by a factor of 76.9), followed by the Spotted Towhee and American Robin. Although there is anecdotal evidence that the Varied Thrush is prone to collisions with windows (Ching 1993, George 2020, this species has only been reported as an infrequent collider in 2 prior collision studies (Johnson andHudson 1976, Dunn 1993). Forestdwelling birds, such as the Varied Thrush and American Robin, may be more likely to collide with reflected vegetation in windows as a consequence of adaptations to living and foraging in forested environments, such as differences in visual perception and the ability to fly quickly through small gaps in foliage (Klem 1989, Wittig et al. 2017, Elmore et al. 2020). In addition, Varied Thrush flush easily at the approach of humans (George 2020), and panicked flights, e.g., from feeders, is a mechanism thought to increase fatality rates, whereby birds are forced into sudden flight by the abrupt appearance of a predator (Dunn 1993, Bracey et al. 2016. Finally, Varied Thrush, Spotted Towhee, and American Robin are omnivorous birds that switch to a diet composed primarily of seeds and fruit in winter (Bartos Smith and Greenlaw 2020, George 2020, Vanderhoff et al. 2020. Although seed-and fruit-eaters have not been previously identified as a foraging guild with high vulnerability to collisions (Wittig et al. 2017, Elmore et al. 2020, there are documented observations of extensive collision mortality of fruit-eating birds in winter (Ocampo-Peñuela et al. 2016, Brown et al. 2019, potentially linked to intoxication due to the ingestion of fermented fruit (Fitzgerald et al. 1990). The highest wintering abundances of Varied Thrush and the Spotted Towhee are within the southern portion of the Northern Pacific Rainforest Bird Conservation Region (eBird 2020). Pacific coastal wintering birds have smaller wintering ranges compared to other western and eastern North American wintering birds (Bock 1984), therefore species confined to these heavily urbanized regions during the non-breeding season may be more vulnerable to population-level effects if winter mortality is also sizable at other sites along the Pacific coast.
Of the 5 continentally distributed species with higher than average vulnerability to collision in winter (American Robin, Song Sparrow, Hermit Thrush, American Goldfinch, and Dark-eyed Junco; Table 1), all have Pacific coastal and/ or western intermountain subspecies variation (Arcese et al. 2020, Dellinger et al. 2020, McGraw and Middleton 2020, Nolan et al. 2020, Vanderhoff et al. 2020. It is unknown whether this variation is genetically based, however, subspecies population structure is important to consider when evaluating potential population-level effects of collision mortality (Loss et al. 2015). Future research should examine whether western species and subspecies are similarly affected by high collision vulnerability throughout their winter range, and whether this elevated vulnerability extends beyond the winter season.

Caveats: Sampling Design, Detectability of Carcasses, and Location of Study
We could not adjust for all potential biases in detectability of carcasses, and we were unable to assess intra-seasonal variation beyond our 45-day monitored periods. Kahle et al. (2016) noted an increase in juvenile bird mortality due to building collisions in late summer. However, our study did not include the period when juvenile birds would likely be at their highest densities. Although our best fit model for SE did not include season as a predictor variable, our models may have failed to detect seasonal variation due to low sample sizes in spring, summer, and fall. However, in an analysis of multiple studies that incorporated SE and CP measures, Barrientos et al. (2018) found that carcass size far outweighed searcher experience, habitat, and season in explaining the variance in detectability of carcasses. Smaller-bodied birds may be less detectable due to both reduced visibility and shorter CP times over which searchers might find carcasses, compared to larger-bodied birds (Bracey et al. 2016, Riding andLoss 2018). Therefore, we may have underestimated the relative vulnerability of small-bodied birds to collisions in winter.
Finally, we conducted our study at a university campus with large buildings surrounded by abundant evergreen vegetation and nearby forested greenspace. These factors are also known to increase collision risk (Hager et al. 2017, Basilio et al. 2020) and therefore collision frequencies may be higher at our site and other sites that share these features, compared to surrounding urban landscapes. However, collision risk may increase throughout urban areas as cities recognize the value of conserving forest fragments, increasing tree canopy, and enhancing gardens for birds and other biodiversity (Amaya-Espinel and Hostetler 2019, Baldock et al. 2019). Collision risk may also increase as more cities adopt nature-based solutions such as planting trees, and integrating wetlands and other green infrastructure into heavily urbanized zones to combat the effects climate change (Frantzeskaki et al. 2019, Portland Parks and Recreation 2019, McManus and de Hoog 2020).

CONCLUSIONS
Winter mortality due to collisions has been largely overlooked in North America, other than at homes with bird feeders, and was considered to be negligible at low-rise buildings in Canada (Calvert et al. 2013). However, winter collision mortality may be a threat to birds in other Pacific coastal cities and in southern temperate and tropical cities that host high densities of overwintering passerines (Brisque et al. 2017, Santos et al. 2017. As cities grow throughout the hemisphere, the potential for increased winter mortality has implications for Neotropical migrants with life history strategies that require high overwinter survival to maintain stable populations, and for temperate wintering birds that may require compensatory increases in recruitment on the breeding grounds (Dokter et al. 2018). To our knowledge, our study is the first to report Varied Thrush and Spotted Towhee as highly vulnerable to collisions with glass. Further intra-and inter-annual studies are needed to assess the vulnerability of western North American species and subspecies to collisions for the development of population models that incorporate a range of threats across the full annual life cycle of birds (Loss et al. 2015).
There are a number of solutions available to address bird collisions with glass, and attention to this issue is growing in importance as urbanization increases across the globe. Bird-friendly building guidelines outline cost-effective strategies to address collisions at the building design phase (City of Toronto 2016, Canadian Standards Association 2019), and targeted mitigation at problem façades has proven effective in reducing collision mortality (FLAP Canada 2018, Brown et al. 2019. Mitigation typically involves the application of a pattern of closely spaced markers or physical obstacles to the outside surface or in front of glass, ensuring that markers are highly visible from a distance, so that birds are able to change course in time to avert a collision (Klem 2010). While a number of commercially available and do-it-yourself options are available (American Bird Conservancy 2020, FLAP Canada 2020), other creative options exist, including public art installations that also serve to create community engagement and communicate the issue to the public (e.g., Figure 4). Colleges and universities should take a leadership role in conducting local research, demonstrating innovation in architecture and building design, and mitigating collisions.

SUPPLEMENTARY MATERIAL
Ornithological Applications 123:1-15 © 2021 American Ornithological Society to specimens, and assistance verifying feather identification. Graphical abstract was designed by Elena Gleiberman. We thank Kristina Cockle for assistance with manuscript editing, as well as 3 anonymous reviewers for their comments that helped to substantially improve the paper. Funding statement: Research was funded by Environment and Climate Change Canada (ECCC). R.J. was under contract by ECCC and all other authors were or are current employees of ECCC. ECCC did not influence the content of the manuscript nor require approval of the manuscript prior to publication. Ethics statement: Carcasses from previously window-killed birds were used for carcass persistence and searcher efficiency trials, under Canadian Wildlife Service Salvage Permits BC-SA-0020-15, BC-SA-0020-16, and BC-SA-0020-17. Author contributions: K.L.D. conceived of the study and secured funding; K.L.D. wrote the paper with contributions from all authors; A.N.P. managed field operations throughout the study, designed the carcass persistence study, and provided GIS analysis; A.N.P., A.C.H. and K.L.D. collected the data, with field assistants noted above; A.R.N. and R.J. performed statistical analysis; all authors contributed detailed review and edits. Data availability: Analyses reported in this article can be reproduced using the data provided by De Groot et al. (2021).