Study site and subjects
We tested innovativeness in four neighboring spotted hyena clans within the Maasai Mara National Reserve, Kenya between June 2016 and November 2017. These clans ranged in size from 30 to 55 adult hyenas. Spotted hyena clans represent distinct social groups that are made up of multiple unrelated females, their offspring, and adult immigrant males. Clans are structured by strict linear dominance hierarchies, with an alpha female and her offspring at the top, followed by lower-ranking females and their offspring, with adult immigrant males occupying the lowest rank positions. Births occur year-round and unrelated females raise their offspring together at a communal den. Female hyenas stay in their natal clan throughout their lives, whereas male hyena usually disperse to join new clans when they are 24–60 months old, after they reach sexual maturity43,44.
All subjects were identified by their unique spot patterns and ear damage. Hyenas of all age classes and both sexes were included in the study. All subjects were sexed within the first few months of life based on genital morphology40. Age classes were based on life history stage45. Cubs were defined as hyenas that were still dependent on the communal den for shelter; on average, Mara cubs become den-independent around 9–12 months of age45. Subadults were hyenas who were den-independent but had not yet reached sexual maturity. Adults were hyenas that had reached sexual maturity. In females, sexual maturity was determined by the observation of mating, visual evidence of first parturition, or the female reaching three years of age, whichever came first46. In males, sexual maturity was determined by dispersal status, males who were still present in their natal clan at testing were classified as subadults and immigrant males were classified as adults.
Multi-access box paradigm for measuring repeated innovation
We tested innovativeness in wild spotted hyenas using a multi-access box designed for use with mammalian carnivores24. The multi-access box (hereafter, ‘the MAB’) is a problem-solving paradigm, also known as an artificial or novel extractive foraging task, where subjects must solve a novel problem to obtain a food reward. In contrast to traditional problem-solving tasks, MAB paradigms typically offer multiple solutions to the same puzzle, each requiring its own unique behavior pattern. As a condensed battery of tasks, the MAB paradigm allows researchers to measure innovation, not just once, but multiple times across different solutions47. We chose to use a MAB paradigm because it allowed us to compare reliability across repeated trials within the same solution to reliability across different solutions. Reliable success with the same solution across trials may be a result of individual learning rather than a result of a stable cognitive trait. However, if individuals reliably innovate by opening multiple unique solutions to the MAB this would suggest that innovativeness is a stable cognitive trait. The MAB in the current study was a steel box, measuring 40.64 × 40.64 × 40.64 cm (length × width × height), with four unique doors, each requiring a different motor behavior, that could be used to access a common interior baited with a food reward (Fig. 1. We used this MAB previously to test repeated innovation in captive hyenas; for more information about the design specifications see Johnson-Ulrich et al.24.
Test protocol
We conducted all testing between 0630 to 1000 h and 1700 to 1830 h, the daylight hours during which hyenas are most active. We deployed the MAB anytime a suitable group of hyenas was located within the territories of our study clans. A suitable group was defined as one containing five or fewer hyenas within 100 m or within visible range that were either walking or resting but not engaged in hunting, border patrol, mating, courtship, or nursing behaviors. We used our research vehicle as a mobile blind to shield the researchers from the view of hyenas while we baited and deployed the MAB on the opposite side of the vehicle from hyenas. We baited the MAB with approximately 200 g of either goat or beef muscle, skin, or organ meat. During familiarization trials we also used full cream milk powder in addition to, or in place of, meat. We deployed the MAB approximately 20 m away from the nearest hyena and after MAB deployment we drove the research vehicle to a distance of 20–50 m away from the MAB. We began videotaping immediately after we deployed the MAB and we ended videotaping when we collected the MAB.
During familiarization trials we deployed the MAB with the top removed to acclimate subjects to the MAB and allow them to learn to associate the MAB with bait. Familiarization trials began when a hyena came within 5 m of the MAB and ended upon successful food retrieval (a “feed” trial) or when the hyena moved more than 5 m away from the MAB for more than 5 min. We recorded hyenas that approached the MAB, but did not make contact, as not participating in the trial. Average duration of familiarization trials was 11.7 ± 12.3 min.
If a hyena had a “feed” familiarization trial or successful test trial, and if it had moved at least 5 m away from the MAB, we immediately rebaited the MAB for successive testing. We gave hyenas successive trials as long as the testing conditions remained suitable, as described above, or until researchers ran out of bait. We did not administer successive trials following trials where every hyena that participated was unsuccessful because unsuccessful hyenas were those that had moved beyond 5 m from the MAB for more than five minutes without opening the MAB and these hyenas were extremely unlikely to spontaneously re-approach the MAB for another trial. On average, hyenas received 1.53 ± 1.25 trials per testing session and completed testing across 6.31 ± 2.58 separate sessions (Supplementary Fig. S1). Most sessions were separated by a median of 1 day (mean ± SD = 24.18 ± 60.30 days, range = 0–321 days).
We divided test trials into four different phases of testing. During Phase 1, we presented the MAB to hyenas with all four doors accessible. After a hyena had reached completion criterion for Phase 1, defined by success with the same door in three out of four consecutive trials, it progressed to Phase 2. During Phase 2, we blocked the door used in Phase 1 by bolting it shut. The same criteria for progression applied to subsequent phases until a hyena reached the criteria for progression with all four doors. We gave hyenas trials until they either reached criterion for all four doors or failed five consecutive trials during any phase of testing. We did not include hyenas that participated in fewer than five trials, of which none were successful, in our analysis. On average, hyenas participated in 9.64 ± 5.61 trials. Hyenas completed Phase 1 in 7.43 ± 2.93 trials (N = 72) either by reaching the criterion for progression or by failure, Phase 2 in 3.67 ± 1.11 trials (N = 15), Phase 3 in 4.08 ± 1.32 trials (N = 13) and Phase 4 in 4.25 ± 1.96 trials (N = 12).
We aimed to give every hyena two familiarization trials prior to being given the option to participate in test trials. On average we gave hyenas the opportunity to participate in 1.60 ± 1.54 (mean ± standard deviation) familiarization trials prior to their first Phase 1 trial, but hyenas only fed from the MAB on an average of 0.94 ± 1.11 familiarization trials prior to their first Phase 1 trial.
When we presented a group of hyenas with the MAB, we configured the MAB for the hyena at the most advanced phase of testing. For example, if one hyena in the group had progressed to Phase 3, but all the others were still on Phase 1, we would configure the MAB for the hyena on Phase 3 and block the doors that hyena had used in Phases 1 and 2. Overall, there were only five trials in total in which a hyena solved the MAB in a trial during the ‘wrong’ phase of testing by joining a trial where we configured the MAB for a group mate rather than itself. The average ‘trial group size’ per hyena per trial was 3.89 ± 3.71 hyenas (median = 3, range = 1–29). We calculated trial group size as a count of all hyenas that participated in a trial by contacting the MAB at any point in time during the trial. Overall, trial group size had a positive and significant effect on hyena participation; hyenas were slightly more likely to contact the box if there were other hyenas contacting the box (Binomial GLMM: z = 9.19, P < 0.001; Supplementary Table S1). We also examined the effect of ‘overall group size’ which we calculated as a count of all hyenas present within 20 m of the MAB. Overall group size had slightly negative effect on participation (Binomial GLMM: z = -9.81, P < 0.001; Supplementary Table S1); hyenas were slightly less likely to contact the box if there were more hyenas present within 20 m of the MAB.
Overall, we deployed the MAB 483 independent times including both familiarization trials and test trials to 280 different hyenas for a total of 2,891 observations. The dataset used in the present analysis only includes test trials from subjects that completed testing by reaching criterion for failure or subjects who had solved the MAB at least once (N = 72 hyenas, n = 694 observations). Of these 72 hyenas, 23 opened the MAB at least once (n = 318 trials). On average, we deployed the MAB 120.75 ± 25.80 times to each of our four study clans). In order to identify the 23 solvers, we deployed the MAB an average of 88.5 ± 34.72 (N = 4) times in each of our four study clans. In other words, by the 90th deployment on average, we had no new subjects solve the MAB that had not already solved it at least once.
Statistical analysis
All statistical analyses were performed using the statistical software R48. Here, R values were calculated for subject ID in generalized linear mixed models (GLMMs). The rptR package also allowed us to estimate uncertainty around each point estimate for R via parametric bootstrapping (n = 1,000), in which we estimated a standard error, a 95% confidence interval, and a P value for each estimate of R. P values were generated using likelihood ratio tests where model fit was compared to a null model with no grouping factor. Here, we report both adjusted R values, calculated as a ratio of the variance explained by subject ID over the residual variance, and conditional R values, calculated as the ratio of the variance explained by subject ID over the total variance, including fixed effects.
Before calculating R values for innovative problem-solving ability across doors, we created a global GLMM that included door, age class, sex, and rank as fixed effects and subject ID as a random effect in order to identify factors that might influence innovativeness. We used the glmmTMB package to create all global models49. To identify fixed effects of importance, we used the ‘dredge’ function in the R package MuMIn50. We built our final model including only the factors with large or significant effects on innovative problem-solving as fixed effects. Dredge identified nine top models for our global model on problem-solving success across doors (Δ AICc < 4). None of the included effects were estimated as important in all nine top models, but age class was included in the most top models (N = 5) and had a large effect that tended towards significance (Supplementary Table S2). Therefore, we included only age class as a fixed effect in our final model.
Likewise, before calculating R values for innovative problem-solving ability across trials, we created a global model that included age class, sex, rank, trial number, and phase of testing as fixed effects, and subject ID as a random effect. To identify fixed effects of importance we used the ‘dredge’ function in the R package MuMIn50. We fixed trial number and phase of testing for inclusion in all models because we wanted to control for the effects of experience and task difficulty. Dredge identified eight top models with a delta AICc of less than four (Supplementary Table S3). Here, both trial number and phase had significant effects. Once again, age class was only marginally significant but also had the largest effect size (Supplementary Table S3). Therefore we included test trial, phase of testing, and age class in all subsequent models examining problem-solving success across trials (Models 2–5.9). In Model 2, phase of testing had a significant negative effect on the likelihood of solving the MAB (GLMM: β = -0.62, SE = 0.31, z = -2.01, P = 0.04) which suggests that later phases of testing, where solutions that hyenas used previously were blocked, were indeed more demanding for hyenas. After controlling for the effect of phase, overall test trial number had a significant positive effect on the likelihood of solving the MAB (GLMM: β = 0.11, SE = 0.06, z = 1.96, P = 0.05), indicating that hyenas were more likely to solve the MAB in later than earlier trials (Fig. 6A). The positive effect of trial number could indicate that hyenas were learning how to improve their performance across trials, but this effect might also be biased by some subjects reaching the criterion to end testing (five unsuccessful trials in a row) and dropping out of the subject pool. To test this possibility we created another model where we restricted our dataset to the first ten test trials only for hyenas that had at least 10 trials, and found that test trial still had a significant positive effect on the likelihood of solving the MAB (GLMM: β = 0.30, SE = 0.10, z = 2.91, P = 0.004, n = 139 trials, N = 14 hyenas; Fig. 6B).
Before calculating R values all models were checked for collinearity by examining variance inflation factors (VIF). Test trial number and phase of testing consistently had VIFs > 4 in most of our models, however, we chose to include both because the high collinearity here is a result of our test protocol; hyenas only progressed to Phase 4 of testing after completing a relatively large number of trials. The main concern with high VIFs is that the estimates error and P values for the collinear factors will be increased; however, both test trial and phase of testing had consistent significant effects, which suggests that this was not a problem in our models. Next, we also examined QQ plot residuals and a histogram of the residuals using the R package DHARMa to confirm that model assumptions about the normality of residuals were not violated.
Ethics statement
This work was conducted under research permit no. NACOSTI/P/16/35513/10422, issued by the Kenyan National Commission on Science, Technology and Innovation. The data collection procedure followed here was also approved by the Michigan State University Institutional Animal Care and Use Committee (IACUC): AUF #04/16-050-00. All research procedures were designed to adhere to the American Society of Mammalogists (ASM) Guidelines for the use of wild mammals in research and education51 and to the Association for the Study of Animal Behaviour (ASAB) Ethics Committee and the Animal Behaviour Society (ABS) Animal Care Committee Guidelines for the treatment of animals in behavioural research and teaching52.
Source: Ecology - nature.com
