Because data were collected at varying intervals and over 26 years, several steps were taken to quality control the data. Mercury tip-switch data were quality controlled by stipulating that for each observation at least 2 or more identical activity values (redundancy) were received during any given satellite overpass to exclude data that were corrupted by one or more bit errors during transmission. Accelerometer data were explicitly quality controlled using a checksum calculation that accompanied every message sent to the satellite (Argos user manual; www.argos-system.org/manual). There was evidence within the mercury tip-switch data that the glass sensor housings were occasionally compromised, allowing oxygen to enter the casing and oxidize the surface of the tip-switch, thereby reducing the switch’s sensitivity to detect state changes (D. Douglas, pers. comm.). A compromised mercury tip-switch sensor typically exhibited a slow decline in activity over time that was not followed by increases, and was not associated with denning. Using a threshold based on each individual’s median activity level during the first 3 months of deployment, data were removed if they fell below the threshold and never recovered to their median activity levels. Each activity measurement was designated as occurring on land, ice over shallow continental shelf waters (on-shelf) (defined as depth less than 300 m; Durner et al. 2004, 2009; Rode et al. 2010), or ice over waters off the continental shelf (off-shelf; depth greater than 300 m). A bathymetry grid (2.5 km resolution;
http://www.ngdc.noaa.gov/mgg/global/etopo2.html; accessed 10 Sep 2014) was used to identify whether locations occurred on ice on- or off-shelf. All activity observations collected when an individual was on ice (on- and off-shelf) were spatially and temporally associated with the respective daily ice concentration (National Snow and Ice Data Center, 25 km resolution;
http://nsidc.org/data/NSIDC-0051; accessed 10 Sep 2014; Cavalieri et al. 1996). Polar bear location data were not always collected at intervals that matched collection of the activity sensor data. Therefore, using the observed locations and associated accuracies, we used the R statistical computing (R Core Team 2014) package ‘crawl’ (Johnson 2013) to predict polar bear locations at the same time as activity sensor measurements from 1 Jul through Oct as described in Rode et al. (2015). The CRAWL model accounts for variable location quality and sampling intervals and allows for location estimates to be obtained at user-defined intervals. A bear was classified as on land if its predicted location was within 5 km of land as identified by the Global Self-consistent, Hierarchical, High-resolution, Geographic Database (GSHHG version 2.3.4;
http://www.soest.hawaii.edu/pwessel/gshhg/), as described in Rode et al. (2015). For more details of the CRAWL model and associations of locations with habitat refer to Rode et al. (2015).Denning events were identified and removed from the dataset to control for suspected changes in activity associated with den entry. A control chart based algorithm was applied to collar temperature data to identify denning events (Olson 2015). This method is 95% effective at identifying denning when compared to direct observations of denning via radio-tracking of collared female polar bears. Additionally, because activity levels may be reduced during the days to weeks prior to den entry (Messier et al. 1994), we classified the two weeks prior to den entry as predenning.We designated bear locations within 5 km of known whale remains as ‘near remains’. Remains of bowhead whales were available in the fall during most years for Southern Beaufort bears at three locations: Kaktovik, AK, Cross Island, AK, and Barrow, AK (Suydam and George 2004; Koski et al. 2005). We are not aware that carcasses are predictably available from subsistence harvest in other parts of the bears range.