*2.2. Lynx-Killed Roe Deer*

The roe deer kills of very high frequency (VHF) radio-collared and snow-tracked lynx were routinely located in the period 1995–2008 [25,26]. Field procedures are described in Nilsen et al. 2009 [25]. Permissions were granted by the Norwegian Environment Agency and procedures were approved by the Norwegian Committee for Experimental Animal Welfare (permit numbers 08/127430, 07/81885, 07/7883, 2004/48647, 201/01/641.5/FHB, 127/03/641.5/fhb, 1460/99/641.5/FBe, 1081/97/641.5/FBe, and Norwegian Institute for Nature Research (NINA) 1/95). Most of the lynx-killed roe deer carcasses included in this study were found in winter (72 in winter and 7 in summer, Table S1). Kills were checked at various times. When studying predation behavior of collared lynx we only approached kills when the lynx had abandoned the kill. When snow-tracking unmarked lynx or trying to recapture already collared lynx, we would approach a new found kill straight away in order to place traps. When the kills were examined, any signs of the presence of scavengers were recorded (such as scats, footprints, feathers) as well as signs of the kill being covered by the lynx. However, we cannot exclude that some scavengers remained undetected. We knew lynx identity for 53 lynx-killed roe deer carcasses. For this subset of data, we explored individual variation in lynx caching behavior by comparing the proportion of covered carcasses for single individuals vs. females with kittens, and for males vs. females, by means of Chi-square tests. We had information on habitat type for 66 of the lynx-killed roe deer carcasses. For this subset of data, we compared the proportion of covered carcasses in open habitats and forest habitats, by means of Chi-square tests.

We used multivariate logistic regression models to explore which variables better explained the probability of a carcass being scavenged or not. The full model explaining the probability of a carcass being scavenged included number of days after being killed, the treatment (not covered/covered) and their interaction as explanatory variables. The final model was selected by model reduction, where only explanatory variables with significance *p* < 0.05 were retained in the model.
