Seasonally frugivorous forest birds and window collision fatalities: novel integration of bird counts in fall improves assessment of species vulnerability to collisions

Bird-window collision surveys

We surveyed study building perimeters daily for collision evidence over two fall, one early winter, and three late winter 21-day periods over four calendar years (2019–2022), following the protocols described by Hager & Cosentino (2014) and De Groot et al. (2021). Fall survey dates were from September 16 to October 6, 2019, and September 19 to October 10, 2020, within the peak of latitudinal songbird migratory activity for the area, based on local banding station data (Leckie, 2008). To allow for regional comparisons, winter survey dates were chosen to coincide with those selected by De Groot et al. (2021) at the University of British Columbia campus located in Vancouver, B.C., on the traditional, unceded territory of the (Musqueam) People (NativeLand Digital, 2025), 90 km to the northeast of our study site, across the Strait of Georgia. Late winter surveys were therefore conducted from January 21 to February 10, 2020, January 19 to February 8, 2021, and January 17 to February 6, 2022. To explore potential within-winter differences in collision mortality, we added an early winter survey period between November 22 to December 12, 2021.

On the day prior to each of these six study periods, we performed a complete clean-up of all carcasses, feather piles, and feather smears at each study building, to ensure that evidence accumulated prior to the study period was not included in the collision tallies. For each daily survey, two observers simultaneously walked the perimeter of all study buildings in opposite directions in early to mid-afternoon, searching within two meters of the façade for evidence of bird-window collisions. Individuals varied direction of their surveys and did not share information about carcasses or other evidence found until the survey was completed. Collision monitoring across the six sampling periods was carried out by a total of three observers, with one observer performing all collision surveys across all study periods, one observer participating in all surveys in the first study period, and one observer participating in all surveys in all remaining five study periods. Collision mortality evidence included all bird carcasses, as defined by Hager & Cosentino (2014): intact carcasses, partially scavenged carcasses, and feather piles. We defined a feather pile as the remains of an entirely scavenged carcass with 10 or more feathers within a one meters radius, to avoid counting feathers shed naturally. We assumed that intact carcasses and remains of carcasses found within two meters of buildings were the result of birds killed outright by windows, or by predators that captured and consumed birds that were stunned by hitting windows. Additional window collision evidence included feather smears on windows, or stunned live birds found adjacent to building façades and taken to rehabilitation facilities, i.e., the survival outcome of these collision events was unknown. We assumed that two pieces of evidence represented different individuals only if they were detected at least two meters apart (e.g., if a carcass was found within two meters of a feather smear, only one collision was counted). To improve searcher efficiency, vegetation within the two meters survey zone was cleared by campus grounds staff prior to the commencement of surveys. Labor shortages related to the COVID-19 pandemic prevented trimming of the small amount of vegetation that regrew between our third and fourth survey year. However, this did not significantly affect searcher efficiency, as confirmed by analysis of our searcher efficiency trials (Table S1).

Carcass collection, storage, and identification

Bird carcasses and feather piles found by surveyors were collected and frozen for future confirmation of species identification, authorized under scientific permits issued by Environment and Climate Change Canada for possession of migratory bird carcasses and feathers (SC-BC-SC-BC-2019-0012SAL, SC-BC-SC-BC-2020-0012SAL, SC-BC-SC-BC-2021-0012SAL and SC-BC-SC-BC-2022-0012SAL). All feather piles were later independently identified to the lowest level of classification possible (species, genus, or family) by authors (RG and KLD) and two other experts in feather identification, using the available literature (Scott & McFarland, 2010; Piranga, 2015; US Fish Wildlife Service Forensics Lab, 2018), and feather samples from carcasses of known identity. Challenging identifications, i.e., those for which independently derived identifications did not agree, were assigned “unknown ID”.

Assessment of carcass persistence and searcher efficiency

To assess the efficacy of searchers in finding carcasses that are present, and the persistence of carcasses that could be removed by scavengers or humans prior to surveys (i.e., overall likelihood of surveyors detecting a window-killed bird), we conducted searcher efficiency (SE) and carcass persistence (CP) trials using previously window-killed birds. Eighteen trials were performed in each of the fall 2019 and 2020 study periods (three per building per season), and twelve trials were performed in each of late winter 2020, 2021, and 2022, and early winter 2021 (two per building per season) for a total of 84 SE and CP trials.

Trials were carried out by an observer who did not participate in collision monitoring surveys, without alerting the two bird collision observers to the timing or location of trials. We thawed trial carcasses overnight and clipped their entire hallux so they were identifiable as trial carcasses. These were deployed in a manner that mimicked the fall of a bird from a window, within two meters of randomly selected façades of study buildings, starting at approximately 8:00 on the initial carcass placement day (Day 1). Carcasses were then checked starting at 12:00 and in late afternoon (17:00 in fall and 16:00 in winter) on Day 1, 8:00 a.m., 12:00 noon and late afternoon on Day 2, at 12:00 noon and late afternoon on Day 3 and late afternoon on Days 4 and Day 5, or until trial carcasses or remains were no longer present. Monitoring ended on Day 5. Following scheduled collision monitoring surveys, collision mortality data files were reviewed to ensure trial carcasses were not recorded as collision events. The number of hours from carcass placement until it was found by at least one collision monitoring surveyor was recorded for SE model input (or ‘0’ if trial carcasses were never found). We did not assess inter-individual searcher efficiency, as our model corrects for biases due to carcasses not being located by at least one surveyor. The number of hours from carcass placement to the midpoint between trial checks when the carcass was present and when it was discovered to have disappeared without a trace was recorded and used as CP model input.

Bird counts for abundance and vulnerability estimates

To assess bird species’ abundance on campus relative to individuals killed in window collisions, we conducted five-minute point counts at each study building twice per week through each 21-day survey period (i.e., six point counts per building for each survey period). For each count, the observer was positioned immediately adjacent to the building façade, mid-way along the façade. The sampled area consisted of a semicircle with a 50 m radius, centered at the midpoint of the façade. All birds seen or heard within the semicircle or perched on the edge of the façade (i.e., at distance 0) were counted. To reduce observer effects in our bird counts (Farmer, Leonard & Horn, 2012), all point counts were performed by a single observer in fall 2019, and all point counts for the subsequent five seasons were performed by a second single observer.

Banding station data as independent data source for species abundance estimation

To account for detection probability issues, particularly of species that are quiet or otherwise cryptic during fall migration, we used mist net capture data collected from September 1 to October 18 in 2019 and 2020 from a migration monitoring station (Rocky Point Bird Observatory; RPBO) situated approximately 30 km southwest of the University of Victoria, on the unceded, traditional territories of the Scia’new and T’Sou-ke First Nations and other and WSÁNEĆ Peoples (NativeLand Digital, 2025; Rocky Point Bird Observatory, 2025), as an independent source of regional abundance data. The banding station is located within the same Coastal Douglas-fir (Pseudotsuga menziesii) ecological zone as the University of Victoria, but within a mix of old- and second-growth forest and open fields, with virtually no buildings, and restricted access to the public. The observatory’s standardized sampling design consisted of 13 mist nets deployed for a maximum of six hours each day (78 net hours/day). Mist net capture data were summarized as the number of individuals of each species captured per mist net per day.

Correcting biases due to imperfect carcass detection and carcass removal by scavengers and humans in bird-window collision and mortality estimates

We used the Generalized Mortality Estimator (GenEst) R package (Dalthorp et al., 2018) to estimate total collision mortality for our six study buildings during each of the six 21-day surveyed periods. Estimates were corrected for imperfect carcass detection probability by searchers (SE; see Methods) and carcass removal by scavengers or people (CP; see Methods). We also generated total collision estimates using all collision evidence (including feather smears and live bird collisions) as an indication of potential total collision mortality, i.e., the outcome for birds that struck buildings and left a feather smear or were stunned and flew away was unknown, but these collisions may have resulted in mortality or sublethal effects.

For both CP and SE estimates, we fit a null model and combinations of models that allowed both parameters of SE (intercept and slope) or CP (location and scale) to vary by season. For the CP models, we also fit different models using exponential, Weibull, lognormal, and loglogistic distributions to determine best fit for our data. Models were ranked by the Akaike Information Criteria (AIC) and the model with the lower value was selected (Tables S2 and S3). Models with AICc < 2 were considered well supported by our data. We then used the best models’ CP and SE estimates to produce bias-corrected estimates of collision mortality by applying the estM function of GenEst to collision observations, and collision monitoring schedule to propagate uncertainty of estimates using parametric bootstrap methods.

Campus-wide collision mortality estimates across fall migration and winter

To estimate average mortality and collision numbers that occur campus-wide at the University of Victoria during the peak fall migratory period and the wintering period, we extrapolated our estimates by propagating the uncertainty to the total number of buildings on campus and to the total number of days in each sampled season. Specifically, we extrapolated the season-specific estimated collision mortality and collision counts and associated uncertainty from our 21-day seasonal study periods from our stratified random sample of six buildings to: (1) the 31-day peak fall migratory period between September 15 and October 15 according to the nearby Rocky Point Bird Observatory banding station records (Leckie, 2008); (2) the 67-day early winter period between October 16 and December 21 (winter solstice in the Northern Hemisphere) and (3) the 51-day late winter period between December 22 and February 10 (our last day of surveys) to derive an estimate of collision mortality during the sampled fall and winter non-breeding seasons on campus (September 15 to February 10) for our six study buildings; and (4) to the 45 additional institutional buildings on the University of Victoria campus, omitting windowless utility and storage buildings, and houses.

Species vulnerability to collisions

To estimate the per-capita vulnerability to collisions, we followed an approach similar to Arnold & Zink (2011), using the residuals of a log₁₀ − log₁₀ regression (Eq. (1)) of estimated total abundance ($N_k^{Total}$) on the number of collisions (y_k). Species with positive residuals from the regression are more vulnerable to collisions than average species (i.e., they have more collisions than expected after controlling for abundance), and species with negative residuals are generally less vulnerable to collisions than others. We limited our inference to species for which the 95% credible interval (CI) of the vulnerability measure (residual) excluded 0. We transformed both the total abundance and the number of collisions using log₁₀ to respect the non-negative scale of each value, and we added 1 to the number of collisions so that we could include species that we had detected via our local abundance data, but for which we had zero recorded collisions. (1)${\displaystyle {log}_{10}\left(y_k+1\right)\sim Normal\left({\mu }_k,\tau \right),{\mu }_k={\theta }_0+{\theta }_1\ast {log}_{10}\left(N_k^{Total}\right).}$ We fit this linear regression within a hierarchical Bayesian model that simultaneously estimated the total abundance using our point count data, or for the fall migration season using both point count data and local mist net capture data. Combining the full analysis into a single model allowed us to account for and propagate the uncertainty in the number of collisions recorded, and the estimated total abundance, into our final estimate of vulnerability. We conducted this vulnerability analysis separately for each season to account for the seasonal changes in the community of birds that may be exposed to collision risk in our study area, although for simplicity, we have not included the season-specific indexing in the equations.

For the fall migration season, the model integrated our replicated point count data from the University of Victoria with the mist net capture data from the RPBO banding station, to estimate the total number of individuals on the campus ($N_k^{Total}$, the predictor in the vulnerability calculation of Eq. (1)). For the four winter survey seasons (one early winter, three late winters), the model used only our University of Victoria point count data. The total number of individuals was estimated as a derived parameter, based on a re-scaling of the mean expected counts within the surveyed area of each point count.

Our model applied an N-mixture approach (Royle, 2004) where for each point count (i), we estimated the species (k) and site (i) abundance ($N_{ik}^{Point}$) as: (2)${\displaystyle N_{ki}^{Point}\sim Poisson\left({\lambda }_k\ast are{a}_i\right),}$ where the area_i is the sampled area of each point count (0.004 km²), which is given as data. For each mist net sampling n, we also estimated the species-specific abundance $N_{kn}^{Band}$ as: (3)${\displaystyle N_{kn}^{Band}\sim Poisson\left({\lambda }_k\ast area\text{\_}ne{t}_n\right),}$ where area_net_n is the hypothetical area that each net is sampling and is a variable estimated by the model. By sharing a λ_k parameter with both point counts and mist net capture data, we were able to correct the abundance estimates for cryptic species that are captured in mist nets and detected in collision surveys but not detected on point counts.

For the point count observation model, we defined the counts of each species (k), point (i), and visit (j), counts_kij as following a binomial distribution, with sample size $N_{ki}^{Point}$ and species-specific detection probability $p_k^{Point}$: (4)${\displaystyle count{s}_{kij}\sim Binomial\left(N_{ki}^{Point},p_k^{Point}\right).}$ We assumed that the detection probability varied among species and that it could be influenced by flocking behavior (i.e., individuals from a flocking species are more likely to be detected when compared with non-flocking species). We modeled $p_k^{Point}$ as: (5)${\displaystyle logit\left(p_k^{Point}\right)={\alpha }_k+{\beta }_{\left[floc{k}_k\right]},}$ setting flock_k as 0 for non-flocking species and 1 for flocking species. For the mist net capture data observation model, we defined the counts of each species k, net n, and visit j, counts_band_knj as following a Binomial distribution, with sample size $N_{kn}^{Band}$ and detection probability $p_k^{Band}$, which is weighted by the effort of each visit, effort_j: (6)${\displaystyle counts\text{\_}ban{d}_{knj}\sim Binomial\left(N_{kn}^{Band},p_k^{Band}\ast effor{t}_j\right).}$ For effort, we calculated the proportion of effort of a given visit j relative to the maximum effort (the maximum effort at RPBO is 78 mist net hours per visit).

Finally, the total abundance for each species N_k was defined as a draw from a Poisson distribution with a species-specific mean λ_k: (7)${\displaystyle N_k^{Total}\sim Poisson\left({\lambda }_k\right),}$

where λ_k is the relative abundance of each species k.

The integrated abundance model described above was applied for the fall data, while the early and late winter abundance estimates were performed using a simpler model, which uses only point count data (i.e., same model described, excluding Eqs. (3) and (6)). Models were fit in a Bayesian framework using JAGS in an interface with R (R Core Team, 2025) using the jagsUI package (Kellner, 2015; see GitHub page for entire R code). Estimates were produced by a Markov Chain Monte Carlo (MCMC) algorithm with four chains, 200,000 iterations, burn-in phase of 100,000, and thinned to every 100 iterations. We used vague priors for all parameters and assessed the convergence of the models by checking the MCMC trace plots and using the Gelman–Rubin statistic ($\hat{R}<1.1$; Gelman et al., 2013).

Breeding habitat association, non-breeding season diet and bird collision vulnerability

There are two main similarities among the most vulnerable species to collisions at our study site. All are forest-breeding birds, with the American robin inhabiting the broadest range of forest and parkland conditions (Dellinger et al., 2020; George, 2020; Swanson, Ingold & Galati, 2020; Vanderhoff et al., 2020). Forest birds are considered particularly prone to collisions around buildings surrounded by a lot of urban tree cover (Cusa, Jackson & Mesure, 2015), due their propensity to make short, high-speed flights through gaps within dense vegetation (Klem, 2014; Samuels et al., 2022). However, Tan et al. (2024) found that forest cover surrounding buildings was a risk factor primarily for non-migrants in Singapore. Moreover, Riding, O’Connell & Loss (2020) reported clear differences between the American robin (typically a short-distance or partial migrant) and Swainson’s thrush (long-distance migrant) in the degree to which vegetation and land cover variables affected collision mortality in Oklahoma, U.S. This suggests that breeding habitat association and migratory strategy of species may not reliably predict collision outcomes across all geographic regions and site-specific contexts.

Frugivory is also emerging as a diet-related risk factor for bird-window collisions in cold weather months. The mechanisms associated with this risk are not known, but may be influenced by the consumption of ethanol in fermented fruits (Kinde et al., 2012; Ocampo-Peñuela et al., 2016; Brown et al., 2019; De Groot et al., 2021) and/or behavioral factors associated with tracking fruit resources. Golden-crowned kinglets include fruit in their diets during fall migration, but they are not known to preferentially target fruit (Carlisle et al., 2012). However, four of our five most vulnerable species are thrushes (family Turdidae) that are highly frugivorous during the non-breeding season (Dellinger et al., 2020; George, 2020; Mack & Yong, 2020; Vanderhoff et al., 2020). Frost and other cold temperature events can accelerate fruit fermentation in fall and early winter, potentially causing ethanol toxicity and disorientation in birds that consume them in high abundance (Eriksson & Nummi, 1983; Fitzgerald, Sullivan & Everson, 1990; Kinde et al., 2012), although future studies are needed to confirm whether this phenomenon is widespread and contributes to collision deaths. Many species of birds, including those that track ephemeral resources such as fruits post-breeding, are highly mobile beyond the migratory period in tropical regions (Faaborg et al., 2010; Ruiz-Gutierrez et al., 2016). The winter movement patterns of the species most vulnerable to collisions at our study site are not well known; however, they may also be nomadic and non-territorial in temperate wintering areas. Large-scale or frequent movements throughout the non-breeding season may result in higher vulnerability to collisions, due to lack of familiarity with threats in continuously new surroundings (Kahle, Flannery & Dumbacher, 2016).

Few clear patterns in vulnerability to collision mortality have emerged within higher-order taxonomic groups of passerines. Nichols et al. (2018) noted that collision vulnerability varied considerably within the families Parulidae (New World warblers) and Passerellidae (New World sparrows). Colling et al. (2022) also found that individual warbler and sparrow species were both among the most vulnerable and the least vulnerable species in their Toronto, Canada-based study. However, some convergence in individual species assessments of vulnerability occurs in studies that include both migratory periods and winter in their analyses. Kahle, Flannery & Dumbacher (2016) also found that Swainson’s thrush and Hermit thrush were among the 14 species that collided at a rate that was disproportionate to their abundanc during migration, but overall, hummingbirds were the most highly overrepresented in mortality data at their study building on the Pacific coast. Varied thrush, American robin, and Hermit thrush were reported as highly vulnerable to collisions in late winter at the University of British Columbia (De Groot et al., 2021). The American robin and Swainson’s thrush were among the most numerous species recorded killed by window collisions in 20 and 17 studies respectively, of 40 studies conducted in the Americas (Basilio, Moreno & Piratelli, 2020). Swainson’s thrush is also the most numerous species by an order of magnitude, comprising over five percent of all records, within a comprehensive published dataset of bird window collisions in the Neotropics (Piratelli et al., 2025). However, the latter two data sets do not account for relative abundance. Further patterns may emerge in individual species’ vulnerability to collisions, through additional studies across a wide range of seasons, geographic areas, and urban to rural contexts. Species that are disproportionately vulnerable across multiple stages of their annual cycle and across their range will be more likely to suffer population-level effects of collision mortality.