Documenting and understanding sex-specific variation in migratory phenology is important for predicting avian population dynamics. In spring, males often arrive on the breeding grounds before females (protandry), though whether these patterns result from fitness benefits versus sex-specific constraints on arrival timing remains poorly understood. Sex-specific variation in the timing of fall migration is less well-documented than in spring, in part because documenting fall departures is often limited by cryptic behaviors, lower vocalization rates, and shifting territory boundaries during this time of year. We used two years of high-resolution encounter data from radio-frequency identification (RFID)-equipped bird feeders to monitor the daily presence of male and female Lazuli Buntings (Passerina amoena) throughout the breeding season at a high and a low elevation site in Cache County, Utah, USA. These encounter data were used to estimate daily arrival and departure probabilities and to investigate possible differences in migration timing in relation to sex and elevation. At low elevation, male arrival (n=15) preceded female arrival (n=16) by approximately one week, consistent with previous research that has documented protandry in other migratory songbirds. At high elevation, however, no significant differences were found between male (n=19) and female arrival (n=6). In fall, we found little difference in departure dates between elevation or sex, or between years.  Our observations are most consistent with constraint-based hypotheses explaining protandry, possibly relating to sex-specific constraints operating during the non-breeding period. We additionally emphasize the need for quantifying uncertainty in phenological estimates and importance of addressing potential differences across demographic groups. 

Study Site and Species

We monitored breeding Lazuli Buntings (Passerina amoena) within the Bear River Range of Cache County, UT, USA (41.8° N, -111.7° W). The Lazuli Bunting is a sexually dimorphic, generalist songbird that breeds throughout the western United States. This species is a moderate-distance migrant, travelling to west-coastal Mexico via the Sonoran region, where it is thought to complete its annual pre-basic molt (Greene et al. 2020). Though they are highly omnivorous, Lazuli Buntings readily visit bird feeders when stocked with white millet (Panicum miliaceum), making them an excellent candidate species for use of RFID resighting. We collected data from April to September, during 2019 and 2020 on two study plots established at low (1450 m) and high (1930 m) elevation, located 24.5 km apart within the same canyon. Both study sites provide suitable habitat for Lazuli Buntings, with no observable differences in sex ratios or breeding densities. The sites did vary in their vegetative composition, with the low-elevation site being dominated by non-native grasses and shrubs, while native grasses, sagebrushes, and quaking aspen (Populus tremuloides) characterize the high-elevation site.

Field Methods

At each site, we established six RFID-enabled bird feeders in a 2 x 3 grid, spaced 75 m apart. Each feeder assembly was mounted on a stationary pole and consisted of a feeder body, two antennas serving as perches, an electronics box housing a battery and circuit board, and a solar panel for power. All feeders were maintained with white millet from before spring arrival (mid-April) through fall departure (mid-September). 

We used a combined RFID reader and data-logger similar to that described by Bridge et al. (2019). Each printed circuit board contained an RFID module (UB22270, Atmel Corporation, San Jose, CA, USA), a microprocessor (PIC16F688, Microchip, Chandler, AZ, USA), a memory module (24LC512, Microchip, Chandler, AZ, USA), a real-time clock (DS1307, Maxim Integrated Products, Sunnyvale, CA, USA), and an SD card slot for data storage. The circuit board’s microprocessor emits a carrier wave signal via two loop antennas mounted as perches on the feeder. When passive RFID tags are within range (~3–5 cm), the carrier wave energizes copper coils within the tag, inducing emittance of a unique 10-ASCII character code. This code is then received and transmitted by the antenna to the RFID module, where it is interpreted and stored to the onboard SD card with the unique ID and time stamp.

Loop antennas were custom-built from coiled wire to produce a target inductance of 1.350 mH, the optimal inductance for RFID detection using this system. Each antenna coil was then wrapped in electrical tape to protect it from moisture and UV exposure. To ensure antennas and RFID components remained operational, we tested each feeder with a designated RFID tag two to three times per week throughout the season. 

We captured and subsequently recorded the presence of  Lazuli Buntings using non-invasive RFID detections from our feeders at both study sites from mid-April through late-September in 2019 and 2020. Buntings were captured using a combination of mist nets and feeder traps. Upon capture each bird was aged and sexed based on criteria from Pyle (1997) and fitted with a U.S. Fish & Wildlife Service band and a colored 2.6mm diameter plastic RFID band containing a passive integrated transponder (PIT) tag (Eccel Technology, Leicestershire, U.K.). Like conventional color bands, the RFID bands are lightweight (0.092g) and are not expected to influence bird behavior or performance. For each bird, we recorded standard morphological measurements, mass, fat, breeding characteristics, and molt status. Once banded, each RFID-marked bird was passively recorded each time it used any feeder at our sites, producing a nearly continuous-time encounter record throughout the breeding season.

Analysis

We used the package feedr (LaZerte 2020) in R (4.0.3; R Core Team 2019) to organize, clean, and isolate individual feeder visits from the raw RFID data. Single tag reads or reads of the same tag within 30 s of each other were considered a single visit. This threshold was selected based on field observations of 1) birds displaying weak aggression at feeders, 2) displaced feeding individuals being only displaced temporarily by others without leaving the immediate feeder location (i.e. perching on top or adjacent to the feeder), and 3) multiple feeders were closely spaced, allowing for most marked bird to select from multiple feeders if exclusion were to happen from territorial individuals. We then used the feedr output to create a daily detection history of number of observed feeder visits by each individual across the study sites. 

We divided the birds into four groups for the purpose of exploring migratory phenology: males and females at high and low elevations. To isolate locally breeding birds and ensure transient individuals did not influence our phenology estimates, we only included individuals for departure analysis if they were of known sex, were marked prior to 15 July of a sampling year, and were recorded using a feeder ≥10 days during the core breeding period (15 June–15 July). For the arrival model, individuals were included in the analysis if they were marked the prior season and were detected ≥10 days between 15 May and 15 July of the current season. We used two seasons of data in the departure model and a single season of data in the arrival model. We modelled arrival and departure separately because we did not have sufficient data to estimate survival probabilities between seasons.

Observation Model

We modeled the daily number of visits recorded for each individual, denoted y_i,t, using a Poisson point process model:

y_i,t  ~ Poisson ( λ_i,t  z_i,t )                 (1)

where λ_i,t is the expected number of visits for individual i on day t and z_i,t is the true status (present or absent) of individual i on day t. We incorporated daily and individual-level random effects to account for temporal variation and individual heterogeneity in visitation rates:

log ( λ_i,t ) =  µ_λ  +  γ_t  +  ε_i               (2)

γ_t  ~ normal (0, σ_day)

Ε_i ~ normal (0, σ_ind)

State Process Models

Each individual i has a true status (present or absent) on each day t that is only partially observable, as individuals can be present at the study site but not using a feeder. To account for this form of imperfect detection, we modeled the latent state variable z as a Bernoulli trial. We modeled arrival as:

  z_i,t  ~ Bernoulli ([1 -  z_i,t-1] α_g[i],t-1 +  z_i,t-1)                               (3)

where α_g[i],t-1 is the cumulative daily probability that an individual in group g arrives between days t and t - 1. This formulation ensures that once an individual arrives (z_i,t = 1), it must remain present for the remainder of the study period. We modeled departure similarly:

z_i,t  ~ Bernoulli ([1 -  δ_g[i],t-1]   z_i,t-1)                        (4)

where δ_g[i],t-1 is the cumulative daily probability that an individual in group g departs between days t and t - 1. In this formulation, once an individual departs (z_i,t = 0), it cannot be detected at the feeders for the remainder of the season.

We modeled daily arrival and departure probability using survivorship retention curves. Following Pledger et al. (2009), we used a Weibull distribution to model the cumulative daily arrival and departure probability for each group:

α/δ_g,t  =  1 - e^( -[(t+1)/γ_g]^κ_g  +  [t/γ_g]^κ_g )                                         (5)

where ?_g is the scale parameter and ?_g is the shape parameter for each demographic group. The Weibull distribution provides a flexible approach for modeling migration timing, allowing for a wide range of curves among groups depending on the values of the shape and scale parameters. In particular, ?_g > 1 results in accelerating daily arrival or departures rates (i.e., arrival/departure probability are initially low and then increase rapidly towards the end of the study period), whereas ?_g = 1 implies constant arrival/departure rates and ?_g < 1 indicates decelerating arrival/departure rates (Pledger et al. 2009). The shape and scale parameters can be further used to quantify relevant summary statistics for each group, including daily arrival/departure probability densities and the mean, median, and variance of arrival/departure dates, based on the properties of the Weibull distribution. We modeled group-level variation in migration timing by treating the log of the Weibull parameters (log(?_g), log(?_g)) as a normally-distributed random effect with means µ_?/µ_? and standard deviations µ_?/µ_?.

We estimated posterior distributions for each parameter using Markov chain Monte Carlo (MCMC) methods implemented in R using NIMBLE (NIMBLE Development Team 2020). We used zero-centered normal priors with standard deviation = 1.75 for all mean parameters (µ_?, µ_?, and µ_λ) and half-normal priors with standard deviation = 1.75 for all standard deviation parameters (σ_?, σ_?, σ_day, and σ_ind). In our arrival model, we selected a more diffuse normal(0, 5) prior for µ_?, because initial inspection of the data indicated that timing of first detections differed significantly across the elevational gradient, indicating a potentially wider spread in arrival phenology than departure. For each model, we ran three MCMC chains for 30,000 iterations with 5,000 sample burn-in and thinning by four. For the departure model, we initially estimated separate 2019 and 2020 parameters for each group to assess annual variation in departure phenology. However, no significant differences were observed between years (Table S1), so we pooled data from both years in our final analysis.

Methods References:

• Bridge, E. S., J. Wilhelm, M. M. Pandit, A. Moreno, C. M. Curry, T. D. Pearson, D. S. Proppe, C. Holwerda, J. M. Eadie, T. F. Stair, A. C. Olson, et al. (2019). An Arduino-Based RFID Platform for Animal Research. Frontiers in Ecology and Evolution 7:257.
• Greene, E., V. R. Muehter, and W. Davison (2020). Lazuli Bunting (Passerina amoena). In Birds of the World (A. F. Poole and F. B. Gill, Editors). Cornell Lab of Ornithology, Ithaca, NY, USA.
• LaZerte, S. E. (2020). feedr: Transforming Raw RFID Data.
• NIMBLE Development Team (2020). NIMBLE: MCMC, Particle Filtering, and Programmable Hierarchical Modeling. Zenodo.
• Pledger, S., M. Efford, K. Pollock, J. Collazo, and J. Lyons (2009). Stopover Duration Analysis with Departure Probability Dependent on Unknown Time Since Arrival. In Modeling Demographic Processes In Marked Populations (D. L. Thomson, E. G. Cooch and M. J. Conroy, Editors). Springer US, Boston, MA, pp. 349–363.
• Pyle, P. (1997). Identification Guide to North American Birds, Part I: Columbidae to Ploceidae. Slate Creek Press, Point Reyes Station, CA, USA.
• R Core Team (2019). R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria.