Cooperative Fish and Wildlife Research Units Program: Oklahoma
Education, Research and Technical Assistance for Managing Our Natural Resources

Carnivore diet‐selection studies based on scat analyses are frequently used to elucidate predator ecology, predict potential effects on prey populations, and inform management decisions. However, accuracy of results and the following inference are contingent on multiple sources of sampling error including missed detections and pseudoreplication in statistical comparisons that assume independence within scat samples. We compared a repeated‐sampling occupancy framework intended to estimate detection and occurrence rates for diet items with a multinomial modeling approach intended to estimate diet selection while accounting for nonindependence of diet items within samples. Both methods allowed for multimodel inference to specifically test hypotheses about differences in diet. We applied each method to 2 example data sets, a bobcat (Lynx rufus) scat data set (n = 101) collected in western Virginia, USA, from 2011 to 2013 with morphological identification of diet items, and a coyote (Canis latrans) scat data set (n = 50) collected in Tooele County, Utah, USA, in 2014 with molecular identification of diet items, and compared results with those commonly implemented in diet studies (frequency of occurrence calculations). We found imperfect detection of diet items was not a major source of bias in either the morphological or molecular data set results, but grouping similar or indistinguishable diet items in the morphological data set affected estimates when there was heterogeneity in detection among items. Using the occupancy approach on the morphological data set demonstrated that presence or amount of some diet items could decrease detection of other items and bias occurrence estimates. Furthermore, comparing multiple models of bobcat diet using Akaike's Information Criterion with either approach revealed no support for seasonal differences, even though traditional frequency of occurrence calculations differed by almost 10%. Thus, we suggest even moderate trends in diet based on frequency of occurrence calculations without incorporating measures of uncertainty may represent sampling error, and not true differences in diet. When detection is not conditional on other diet items, comparison of multinomial models will typically be sufficient to make accurate inference about carnivore diets without requiring additional processing of scat samples.