Data acquisition and filtering
The Pacific ocean off western North America is an ideal region to study rarely observed offshore passerines due to its extensive spatial and temporal coverage from birdwatching pelagic trips, repositioning cruises, and research vessels. Observations from these vessels are often submitted by reliable observers to eBird [24], a community science platform for birdwatchers to submit observation data. eBird data is entered in georeferenced checklists containing tallies of species linked to effort data including transect distance, number of observers, and time.
I chose common coastal migratory passerine species to study factors driving offshore vagrancy. Non-migratory and rare species were omitted because there are too few offshore records for analysis. I selected study species by visually inspecting eBird records from reference counties encompassing the West-facing slope of the coastal range in California (Orange, Santa Cruz, Humboldt, and Del Norte), Oregon (Curry, Coos, Lincoln, Tillamook, and Clatsop) and Washington (Pacific and Grays-harbor). I ignored interior counties to avoid introducing extraneous, non-coastal species to the analysis. I selected species reported in > 2% of eBird checklists in any county, a threshold chosen to include all regular species in the region based on prior knowledge. Additionally, I only included species if they have different breeding and non-breeding range in Birds of the World [25] range maps, thus being migratory. This resulted in a set of 49 common costal migratory passerines with a variety of migration distances, diel migration timings, and morphology.
I downloaded the “basic” eBird dataset containing all eBird checklists submitted prior to November 2023, and filtered it only include the 49 study species using a custom AWK script. I then “zero filled” records for all species using the package auk in R v4.3.2 [26, 27]. “Zero-filling” interpolates absences of species left unreported in complete eBird checklists. The function also removes duplicate observations produced when eBird users share their checklists with fellow observers. The resulting dataset contained eBird checklists with each focal species marked as either present “1” or absent “0”.
I thoroughly filtered the raw eBird data to produce a dataset appropriate for studying offshore migration. I selected offshore eBird checklists submitted from southern California to southern British Columbia (32.5–48° latitude) from 10 to 470 km offshore (Western boundary − 130° longitude), removing records within 10 km of land because observers occasionally combine harbor and nearshore data into one eBird checklist. I also selected onshore records up to ~ 10 km inland as a comparison. Within 2.5° latitude slices of the onshore data (width based on weather covariate scales), I calculated observation frequency for each species on each day. Observation frequency is a standard eBird statistic calculated as the percent of all checklists in an area that contain a report of at least one individual of a species [28,29,30]. I then excluded all offshore records (either presence or absence) of species that had an on-land frequency of zero. This ensured I was studying the conditions during the migratory period of that species, avoiding inflating my dataset with uninformative absences when a species isn’t migrating. Additionally, I only included data submitted in March–May and August–November to omit the breeding and non-breeding season of most passerines.
Lastly, I acquired descriptive aspects of the dataset. eBird checklists hold valuable descriptive data entered by users, including various tags and a free text field where users may enter observation comments [31, 32]. I extracted keywords from observation comments and age tags to determine the relative proportion of adults and juveniles offshore. These are only available when noted by the observers, so the relative proportions of each are not necessarily representative of the population.
External drivers of offshore vagrancy
I selected predictors to evaluate four hypotheses for offshore vagrancy. (Hypothesis 1) Vagrancy may be more common in the fall, possibly because naïve juveniles navigate poorly. I included season as a predictor. (Hypothesis 2) Wind-drift may drive birds offshore. I included u-wind (east/west wind vectors) to test if strong east winds drive birds offshore and an interaction between v-wind (north/south wind vectors) and season to account for the different migratory directions during spring and fall migration. I also inlcuded an interaction between u-wind and v-wind to assess the impacts of wind direction and speed simulatenously. (Hypothesis 3) Low visibility caused by fog or low cloud cover may prevent birds from seeing landmarks like the coastline. I used relative humidity as a proxy for fog and low cloud cover because they are generally correlated [33]. I extracted wind and relative humidity NCEP/NCAR weather Reanalysis 1 data from offshore locations using the R package RNCEP [34]. I averaged hourly weather data for an entire 24-h day for each checklist, because weather may have affected birds before the time of observation. (Hypothesis 4) Global changes in geomagnetism or radiation may interfere with magnetoreception. Following Tonelli et al. [8], I included two predictors relating to interference with magnetoreception: geomagnetic disturbance and solar activity. I included geomagnetic disturbance because geomagnetic events may cause misorientation [8, 19, 20]. I downloaded Kp (an average magnetic index from 13 globally distributed geomagnetic observatories) geomagnetic data between 1968-01-01 and 2023-01-01 from the International Service of Geomagnetic Indices (https://isgi.unistra.fr/data_download.php). I then converted Kp to ap, a numeric magnetic index more suitable for statistics. Additionally, radiomagnetism from solar activity may disrupt magnetoreception [8]. I used daily sunspot number as a proxy for solar activity [35]. I downloaded ‘American Relative Sunspot Number-Daily’ data between 1968-01-01 and 2023-02-05 from Laboratory for Atmospheric and Space Physics (https://lasp.colorado.edu/lisird/data/american_relative_sunspot_number_daily/). I also included two predictors to control for other aspects of the data. I included latitude as a predictor because vagrancy may vary geographically. Additionally, the number of offshore vagrants likely varies with the overall number of birds migrating on a given day. To control for migrant abundance, I included daily on-land species frequency, which should scale with the number of migrants. I inspected all possible explanatory variables for outliers and multicollinearity with VIF and plot-correlation matrices. I also checked for missing values.
I fit binomial generalized linear mixed models (GLMM) with the R package lme4 [36] to evaluate the relative role of external drivers and species differences in offshore vagrancy. A binomial GLMM was chosen to fit the binary presence and absence response variable. I chose a mixed effects modeling framework for two reasons. First, I wanted to evaluate the relative contribution of external factors and species differences to offshore vagrancy. To meet this aim, I included species as a random effect. Second, there is inherent non-independence in the eBird dataset. Each checklist has presence and absence of all studied species, so those observations have the same predictor values and are not independent. Additionally, birdwatchers on a single vessel will often submit multiple checklists throughout the day, so passerines riding on board may be reported multiple times. Lastly, some days may have high levels of vagrancy, so checklists from that day may have inflated numbers of offshore passerines. To test the best random effect structure to remedy non-independence, I fit five global models (all fixed effects included) with differing random effect structures (Supplementary Materials). A random intercept for observation date compensated for all three levels of non-independence and had clear support when compared to other random effect structures with AICc. Thus, I opted for a crossed random effect structure with intercepts for species and observation date.
I fit 17 binomial GLMMs with different combinations of eight predictors chosen a prioi to test different hypotheses for offshore vagrancy. To ensure good model fit, I randomly subsampled absences to have the same number as presences. Resampling techniques are commonly used to correct for class imbalance in GLMM and machine learning models across disciplines [12,13,14]. Subsampling was an effective approach to correct for class imbalance when estimating species distribution in a similar community science dataset [37]. I scaled all continuous predictors to ensure model convergence and allow effect size comparison. I fit univariate models to evaluate the individual contribution of all seven predictors and an intercept only model to act as a null hypothesis. I then fit combinations of my overall migration abundance predictor, on-land species frequency, and the most likely predictors based on previous research [17, 21], season, v-wind, and an interaction between v-wind and season. After evaluating model performance, I used the model with these three predictors and interaction as my base model from which to build models testing all combinations of the other five predictors. Models with more than 6 fixed effects were overfit and would not converge. This resulted in 21 candidate models which I compared with AICc. I chose the best model as having the lowest AICc, but any model within 2 values of the best model was considered in the results. I also conducted model averaging with the full model set to assess the importance of each predictor and similarity to estimates in the best model with MuMIn [38]. In the best model, I evaluated the importance of fixed and random effects with conditional and marginal R2. As a model diagnostic check, I confirmed the centrality and normality of model residuals by comparing mean residuals to 0, plotting simulated residuals with the R package DHARMa [39], and plotting random effect residuals with the R package sjPlot [40]. I also plotted binned residuals against all predictors, both those included and excluded from the model. I evaluated influential points and dates with cooks distance.
Species characteristics affecting offshore vagrancy likelihood
To evaluate traits that may increase vagrancy I studied interspecific differences in vagrancy likelihood. The frequencies of each species may be heavily influenced by their overall population size. To generate a metric of offshore vagrancy likelihood corrected for differences in relative commonness, I calculated species average of offshore frequency weighted by one minus the daily on-land frequency. Thus, common species on land would have a lower adjusted offshore frequency.
I tested whether morphological and life history traits drive species differences in offshore vagrancy likelihood. I downloaded hand-wing index (HWI) and mass data from the AVONET dataset [41]. HWI is a measure of wing pointedness, calculated as the factor of the longest primary and secondary feather length. I included HWI because it is widely considered a proxy for aerodynamic efficiency [42,43,44], thus possibly impacting migratory ability. I included mass to consider an additional morphological trait. I downloaded Elton traits to extract foraging habitat trait data [45]. Species occupying habitats with different levels of exposure may be differently impacted by wind and other weather events. I included foraging preference for understory, midstory, canopy, and aerial strata, which are expressed as a percentage of time spent foraging in that habitat. Ground foraging preference was omitted as it was highly negatively collinear with understory and midstory foraging preference. I retrieved migration traits from Birds of the World [25]. I recorded if birds migrated during the day, night, or both. To estimate migration distance, I averaged both breeding and non-breeding latitude and longitude on range maps, and then calculated spherical distance in km between the points. I included migration distance because birds that migrate further tend to be more prone to vagrancy [21]. I fit 15 multiple linear regression models using adjusted vagrancy likelihood as a response variable, normalized with a log transformation. In the initial modeling phases, I tested all four combinations of HWI and migration distance, including an interaction term between HWI and migration distance to test if birds with greater aerodynamic efficiency may be more able to compensate with the vagrancy causing events. This initial phase revealed the importance of including the interaction between HWI and migration distance, which I used as the base model for future models. Later, I added the other five variables in univariate models and combined with the base model. I verified the normality of the best model residuals and checked influential points with Cook’s distance.
I conducted a phylogenetic analysis to test if related species had similar vagrancy likelihood due similar life history characteristics. I downloaded 1000 phylogenetic tree subsets with a Hackett backbone from birdtree.org and generated a 50% majority-rule consensus tree [46]. I made the tree ultrametric and dichotomous to meet the assumptions of standard phylogenetic analyses and tested for phylogenetic signal of offshore vagrancy likelihood with Blomberg’s K using the R package phytools [47, 48].
Add Comment