Movement Analysis
We calibrated the filtered data from the day after capture to the end of
July (i.e., when birds were at the breeding grounds). Each bird’s
capture location was used as the known location for calibration to
estimate the error in sunset/sunrise times. Limited summer darkness in
the subarctic can make sunset time challenging to estimate with
light-loggers; the number of calibration days was increased by extending
the calibration period to the end of July to augment the total number of
days. A threshold model estimated latitude and longitude after
correcting for twilight bias during the calibration period. Data
collection began too late in the summer for one Arctic Warbler (USGS
band 1780-53921) to effectively calibrate the geolocator, so we used
calibration data from the other Arctic Warbler (1760-53520) as the best
estimate of twilight detection bias for that individual.
The Markov Chain Monte Carlo (MCMC) modeling process used to estimate
position was implemented using the SGAT package in R (Sumner et
al. 2009; Lisovski & Hahn 2012; R Core Team 2020). These models
combine a position estimation model with a movement model to determine
the animal’s path in a Bayesian framework. We used vague priors on the
movement speed and bearing parameters of the movement model. The
distribution of terrestrial habitat contributed to the posterior
estimates of position. We created a raster that categorized each cell as
land or water and built a probability mask where overland travel was
given a higher prior probability of use (log(2) vs. log(1), e.g. (Hill
& Renfrew 2019). However, position estimation was unrestricted across
the globe. While the probability mask increases the likelihood of
terrestrial position estimates, it does not prevent overwater position
estimates. Three MCMC chains were run with a 5,000 iteration burn-in,
then a 10,000-iteration posterior sample. Posterior estimates of
locations were visually checked for chain convergence.
Final location estimates are the mean of the three-chain posterior for
each position, and the uncertainty in each is displayed using point
intensity of the location estimate MCMC posteriors across a grid system.
We estimated arrival and departure dates from the breeding and wintering
grounds using multiple changepoint analysis to detect shifts in modeled
latitude with a segmented neighborhood method (Killick & Eckley 2014).
We used a cumulative probability test for non-normal data, with a
penalty value of 0.9 to identify changes in mean longitude (for
departing the breeding grounds) and latitude (for arriving at the
wintering grounds). Estimated transition dates were visually assessed to
determine accuracy. R scripts are documented in Appendix B.