Conceived and designed the experiments: PC SA. Performed the experiments: PC CM. Analyzed the data: PC SA. Contributed reagents/materials/analysis tools: SA PC. Wrote the paper: PC SA CM. Contributed all the long-term elephant life history data of the Amboseli elephants: CM. Provided all the guidance necessary for executing this research: SA.
The authors have declared that no competing interests exist.
Factors that influence learning and the spread of behavior in wild animal populations are important for understanding species responses to changing environments and for species conservation. In populations of wildlife species that come into conflict with humans by raiding cultivated crops, simple models of exposure of individual animals to crops do not entirely explain the prevalence of crop raiding behavior. We investigated the influence of life history milestones using age and association patterns on the probability of being a crop raider among wild free ranging male African elephants; we focused on males because female elephants are not known to raid crops in our study population. We examined several features of an elephant association network; network density, community structure and association based on age similarity since they are known to influence the spread of behaviors in a population. We found that older males were more likely to be raiders than younger males, that males were more likely to be raiders when their closest associates were also raiders, and that males were more likely to be raiders when their second closest associates were raiders older than them. The male association network had sparse associations, a tendency for individuals similar in age and raiding status to associate, and a strong community structure. However, raiders were randomly distributed between communities. These features of the elephant association network may limit the spread of raiding behavior and likely determine the prevalence of raiding behavior in elephant populations. Our results suggest that social learning has a major influence on the acquisition of raiding behavior in younger males whereas life history factors are important drivers of raiding behavior in older males. Further, both life-history and network patterns may influence the acquisition and spread of complex behaviors in animal populations and provide insight on managing human-wildlife conflict.
Factors that influence learning and the spread of behavior in wild populations of social animals are of interest in understanding the response of species to changing environments and for species conservation
Life history milestones, being correlated with age, may influence crop raiding behavior in male elephants and possibly other large mammals in several ways. First, sexual maturation and attainment of a peak in reproduction represent life history milestones manifested by a rise in energy requirements
Social networks provide a way to measure patterns of exposure of individual animals to conspecifics. The behaviors of individuals in an animal's network can determine its probability of acquiring those behaviors through social learning
The structure of association networks can also influence the transmission and spread of socially learned behavior in populations
Studies of male elephant association network properties such as the density of pairwise associations, community structure and association based on age have received little attention
In this paper, we tested several predictions. First we tested the prediction that the probability of being a raider increased with age. Second we tested the prediction that the probability of being a raider was higher for males whose top associates were raiders than for males whose top associates were non-raiders. Third, we tested the prediction that the probability of a male being a raider rises with increase in the relative age of his associates who are raiders. Fourth, we examined elephant social network properties: density of associations, community structure, and association based on age similarity for their potential to influence the spread of raiding behavior in the population. After establishing the presence of community structure or social groups, we tested the prediction that (a) raiders are distributed non-randomly across elephant social groups, (b) the mean age of individuals in each social group is not different from a random age sample taken from the population.
The protocols used in this study were approved by the Institutional Animal Care & Use Committee (IACUC), Duke University and Medical Center under Registry Number: A333-05-12. In Kenya, permission to conduct this research in the Amboseli National Park and surrounding areas was approved by the Office of the President of the Government of Kenya through permit number MOEST 13/001/35C 225.
This study was conducted in the Amboseli National park and adjacent areas, which constitute part of the 8,000 km2 of the larger Amboseli ecosystem located in southern Kenya and at the northern slopes of Mt. Kilimanjaro. Rainfall in the ecosystem varies spatially and temporally with an average annual rainfall of 340 mm recorded within the Amboseli National Park
Human agriculture and settlements occur 10 km to the east in Namelok, and about 20 km to the east and south east of Amboseli National Park in the Kimana and Loitokitok farming areas
This study focused on the Amboseli elephant population, currently consisting of about ∼1400 elephants. Of these, ∼365 males and ∼510 females were 10 years or older by August 2007. This population has been intensively studied since 1972 by the Amboseli Elephant Research Project (AERP). All elephants in the Amboseli population are individually known and are identified using natural tears, notches, holes and vein patterns on ear pinnae
We collected association data during sightings of all-male elephant groups from June to December of 2005 to 2007; these observations were carried out opportunistically because locating elephants was not predictable. We searched for male elephants daily by driving to areas where elephants were likely to be sighted. When we sighted elephants in all-male groups, we recorded the identities of individuals in the group. We defined an elephant group as a spatially cohesive and behaviourally coordinated aggregation of two or more elephants. An elephant group was defined as spatially cohesive if individuals were aggregated within a radius of 100 m and if they were orientated or moving in the same direction. Elephants were considered to be behaviourally coordinated if they had similar activity patterns or interacted during a 10–30 min observation window.
After data collection was complete, we strove to obtain a realistic and an unbiased representation of male association patterns by choosing individuals for whom we had a minimum of 15 sightings when they were in all-male groups during the study period. This produced a sample of 58 individuals (15% of the male population 10 years and older) defined by the following measures: mean of 36 sightings, median of 18 sightings, mode of 46 sightings and a maximum of 107 sightings. We also chose individuals whose frequent associates were sighted at least 15 times, to eliminate individuals from the sample whose major associates were not sampled intensively. From these data we estimated the association of dyads using a simple association ratio or Association Index (
Data on crop raiding was collected independently of the association data using a different sampling protocol. We identified 43 individual crop raiders from direct observations and from genetic analysis of feces collected from raided farmland over a three year period starting in May 2005 and ending in November 2007. We followed elephant tracks from raided fields during the day until we located and identified all individual raiders. When we were not able to locate raiders, we collected elephant fecal samples from raided crop fields and used molecular-genetic techniques to identify individuals
A: The age distribution of all male elephants in the Amboseli population that were 10 years and older (n = 365). B: The age distribution of all the male crop raiders that were detected from the Amboseli population (n = 43). C: The age distribution of all elephants in our study sample (n = 58 males). D: The age distribution of crop raiders in our study sample (n = 21). The Age distribution of our sample of 58 individuals (C) was not significantly different from the age distribution of the entire Amboseli population of 365 individuals (A) (Kolmogorov-Smirnov test, D = 0.800, P = 0.079). Similarly, the age distribution of raiders in our study sample (D) was not different from that for all the raiders detected in the Amboseli elephant population (B) (Kolmogorov-Smirnov test, n = 43, D = 0.4, P = 0.810).
In all the statistical analyses that follow, subjects were designated as raiders (represented by the numeral 1) or non-raiders (represented by zero). The relative age of an associate was defined as the age of a focal male subtracted from the age of his associate. Values of relative age or age difference were negative for associates who were younger than the focal male and positive for associates who were older than the focal male. All probability values reported in the results are for two tailed statistical tests and significance was assessed at P>0.05.
To predict the crop raiding status of a male from his age and association data, we employed logistic regression models. In a logistic model, the probability of being a raider (
To test the prediction that the probability of being a raider increased with an increase in age, we performed logistic regression analyses using data for the entire male population of 365 Amboseli elephants that were 10 years and older. In addition, we also performed a similar analysis on the subset of only 58 males for whom we had association data for and who thus formed the focus of the rest of the analyses. For the larger dataset, we grouped elephants into age classes corresponding to major life history milestones and then performed a logistic regression on the raiding status of elephants as a dependent variable and their age classes as an independent categorical variable. We grouped males into five age categories. Age classes 10–14 years, and 15–19 years correspond to early and late dispersal phases respectively
To test whether the raiding status of a male's top five associates influenced his probability of being a raider we first ranked each male's associates from most close associate (1) to least close associate (57) for associates with the largest and the smallest AI values respectively. We then performed single variable logistic regression analyses with the raiding status of the focal male as a dependent variable and the raiding status for each of the five closest associates as independent variable. We evaluated the significance of our logistic coefficients for each of the five closest associates using randomization tests. We generated 1000 datasets. In each dataset, each male retained their raiding status but their top associates raiding status was randomly assigned a raiding status from one of his 57 potential associates. We performed logistic regression analyses on these randomized datasets and extracted the coefficients for the intercept and raiding status of the randomly chosen closest associate. We used the coefficients from the randomized data to evaluate the hypothesis that a male's probability of being a raider was not predicted by the raiding status of their associate regardless of the associate's rank.
We next used logistic regression to test whether a male's probability of being a raider was predicted by the number of his five closest associates who were raiders. We compared this model with one that included the raiding status of closest and second closest associates as two separate independent variables in a multiple logistic regression framework using Akaike's Information Criterion (AIC). The model with the lowest AIC value was selected as the most likely model. We evaluated the statistical significance of the observed logistic coefficients of the selected model by comparing them with the null distribution of logistic coefficients obtained from analyses of random data generated by shuffling covariates from the original data.
To test the prediction that males were more likely to be raiders if their associates were older raiders than if their associates where younger raiders, we performed a logistic regression analysis on a males raiding status as a dependent variable, and (1) a males own age in years (2) the raiding status of his first closest associate (3) the raiding status of his second closest associates, and (4) the relative age of the top associate (i.e. age of associate minus the age of focal male) (5) the relative age of the second associate (6) the interaction between an associates relative age and raiding status of his top associate and (7) the interaction between an associates relative age and raiding status of his second associate as independent variables. We evaluated the significance of the logistic coefficients we obtained using a null distribution of coefficient values generated by performing logistic regression analyses on 1000 randomized data sets. We generated each of these random datasets by randomly assigning a raiding status and relative age to the associates of each focal individual from his pool of 57 potential associates.
We predicted a positive coefficient of the interaction between an associate's raiding status and relative age on the focal male's raiding status. A positive interaction means that the probability of a focal male being a raider increases with the increase in relative age of his associates who are raiders more than with an increase in relative age of associates who are not raiders. A statistically significant difference in slopes of raiding probability suggests that younger individuals are more likely to be learning from older individuals rather than older males learning from younger males.
We employed exponential random graph (ERG) analyses to examine the density of male elephant associations, the strength of clustering in these associations and to determine if males associated based on age similarity. We also performed ERG to confirm that males with similar raiding status were more associated than males with dissimilar raiding status as predicted by randomization tests. ERG analyses were performed on a directed binary elephant association network generated by connecting individuals (or nodes) with closest associates whose attributes we found to predict his raiding behaviour using randomization tests. An ERG model is used to express the probability (
For the ERG analyses, we incorporated parameters for structural configurations (endogenous effects) known to capture salient features of social networks such as association density (relationship between the number of observed associations and the possible number of associations among individuals in the network), reciprocity (the tendency for mutual associations to occur in the network such that if A is top associate of B, then B is also top associate of A), the distribution of the number of associates per individual (or degree distribution) and transitivity (the tendency for associates of two associated individuals to also be associated)
We fitted ERG model with above parameters to the elephant association network using the Markov Chain Monte Carlo Maximum Likelihood Estimation procedure employing the Robbins–Monro algorithm
When we obtained zero as an estimate of a fitted parameter, we interpreted that as an indication that the effect being modelled occurred by chance. We interpreted a positive parameter as an indication that the effect modelled was more prevalent and occurred more often than expected by chance. A negative parameter, on the other hand, indicated that the effect was less prevalent than expected by chance alone. However, for the age covariate, a significant negative parameter estimate indicated a tendency for associated elephants to be closer in age more than predicted by chance because we used absolute age differences to fit our model.
To visualize the elephant association network and to identify social clusters in the elephant network, we used the Girvan-Newman modularity maximization algorithm implemented in NetDraw
After establishing that the modularity of the elephant association network was significantly different from a random expectation, we tested whether the proportion of raiders and the mean age of males in each cluster were significantly different from a random sample from the population. We determined the expected proportion of raiders and mean age of males in each cluster by randomly shuffling the raiding status and ages of individual elephants across clusters while holding the number of individuals per cluster constant. We created 1000 randomized datasets and for each data set, we estimated the proportion of raiders and the mean age of individuals in each clusters. We calculated the proportion of raiders in each cluster and the mean age of individuals in each cluster from 1000 randomized datasets that were equal to or more extreme than the proportion of raiders and the average age in each cluster estimated from the original data. We used this ratio as an estimate of the probability that our observed proportion or mean age in years in each cluster was significantly different from a random expectation.
The probability of being a crop raider was positively predicted by a males' own age (intercept = −3.925, P<0.001; age = 0.0730, P<0.001, n = 365). This pattern did not change when we used only the subset of the dataset from the Amboseli population for which we had adequate association data and which formed the focus for the rest of the analyses (intercept = −2.468, P = 0.002; age = 0.0740, P = 0.010, n = 58). When we used age class as an independent categorical variable with age classes corresponding with major life history milestones, we detected a dramatic increase in the odds of being a raider for males initiating reproduction and a near doubling of these odds when males attain their reproductive prime age or 45+ years (
Independent variable | Number of males (Raiders) | Coefficient ± Standard error | Chi-Square | Odds ratio | Probability value |
|
−5.209±1.426 | 13.349 | 0.000 | ||
|
91 (0) | 0.000±0.000 | |||
|
81 (3) | 2.099±1.528 | 1.887 | 8.159 | 0.170 |
|
123 (22) | 3.703±1.445 | 6.568 | 40.566 | 0.010 |
|
49 (12) | 4.111±1.464 | 7.890 | 61.000 | 0.005 |
|
21 (6) | 4.340±1.504 | 8.329 | 76.742 | 0.004 |
The probability values show whether the coefficients were significantly different from zero for each age class.
Age class 10–14 years was used as a baseline age class and as result its' coefficient is set to zero. We used data for the all males over 10 years of age from Amboseli National Park (n = 365 elephants). The total number of males in each age class is indicated and the number of crop raiders in each age class is shown in parenthesis. The odds ratio is the exponent of the coefficient and provides a measure of how more likely a male from a named age class is to be a raider compared to the raiding status of the baseline age class.
The raiding status of top associates was a strong predictor of a male's raiding status. Specifically, the probability of being a raider was higher than predicted by chance for a male whose closest associate was a raider (intercept = −1.19, P = 0.017; raiding status of first associate = 2.379, P = 0.001, n = 58 elephants) or whose second closest associate was a raider (intercept = −1.513, P = 0.030; raiding status of second associate = 2.113, P = 0.003, n = 58 elephants). However, a male's raiding status was not predicted by the raiding status of his third (intercept = −0.389, P = 0.670; raiding status of third associate = 0, P = 1.000, n = 58 elephants), fourth (intercept = −0.188, P = 0.910; raiding status of fourth associate = 0, P = 1.000, n = 58 elephants), or fifth closest associate (intercept = −0.747, P = 0.151; Raiding status of fifth associate = 0.636, P = 0.195, n = 58 elephants). Although a male's raiding status was predicted by the number of his five top associates who were crop raiders (intercept = −1.889, P = 0.001; proportion of top five associates who were raiders = 0.650, P = 0.021, n = 58 elephants), the model that included the raiding status of a males' top two associates as two separate independent variables (
Independent variables | Observed coefficient | Mean±Standard error of expected coefficient | Probability value |
|
−1.740 | −0.454±0.012 | 0.002 |
|
1.783 | −0.102±0.014 | 0.005 |
|
1.596 | −0.133±0.016 | 0.010 |
Independent variables | Estimated value | Standard error | t-ratio | Probability value |
|
−5.297 | 1.343 | 0.025 | 0.000 |
|
0.606 | 0.105 | 0.068 | 0.000 |
|
4.781 | 0.564 | −0.029 | 0.000 |
|
6.357 | 1.322 | −0.049 | 0.000 |
|
1.745 | 0.681 | 0.034 | 0.011 |
|
−5.243 | 0.722 | −0.044 | 0.000 |
|
1.678 | 1.037 | 0.076 | 0.106 |
|
0.241 | 0.971 | 0.054 | 0.804 |
|
0.672 | 0.319 | 0.019 | 0.035 |
|
−0.027 | 0.010 | −0.024 | 0.007 |
|
−0.251 | 0.670 | −0.022 | 0.708 |
Transitivity parameter indicates clustering and a schematic representation of this and other parameters in this table are shown in
Data on age differences between dyads was used to test whether associations in the network are based on age proximity in this analysis.
A model predicting the probability of a male's raiding status from his age, the raiding status and relative ages of his two closest associates and the interaction between the raiding status and relative ages of his closest associates had the lowest AIC compared to a model with raiding statuses of the two closest associates only (Δ AIC = 15.312) or a model with age of focal male and the raiding status of his associates (Δ AIC = 14.923).
In the model with the lowest AIC, we found a strong positive interaction coefficient between the raiding status and relative age of his second closest associate but not his first closest associate (
Independent variables | Observed coefficient | Mean ± Standard Error of expected coefficient | Probability value |
|
−5.232 | −0.506±0.029 | 0.001 |
|
0.141 | −0.001±0.001 | 0.002 |
|
0.053 | −0.031±0.002 | 0.224 |
|
0.013 | −0.027±0.001 | 0.267 |
|
2.550 | −0.104±0.018 | 0.004 |
|
0.033 | −0.117±0.017 | 0.379 |
|
−0.029 | 0.000±0.002 | 0.367 |
|
0.188 | −0.003±0.001 | 0.009 |
The density of association in the elephant social network was significantly sparse (
The nodes represent 58 individual male elephants and the size of the node is proportional to age of an individual male. Black circles (nodes) indicate raiders and the white circles indicate non-raiders. Nodes are grouped into six clusters using the Girvan-Newman algorithm in NetDraw. Clusters in the top row from left, center and right are identified as A, B and C respectively and clusters in the bottom row are identified as D, E, F from left, center and right respectively.
At the cluster level, the distribution of individuals by raiding status and by age across association clusters was as expected by chance in five of six clusters constituting our network (
Cluster identity |
Cluster size | Observed proportion of raiders | Mean±Standard error of the expected proportion of raiders | Probability value |
|
8 | 0.125 | 0.359±0.005 | 0.230 |
|
10 | 0.600 | 0.358±0.004 | 0.152 |
|
12 | 0.417 | 0.369±0.004 | 0.964 |
|
12 | 0.083 | 0.360±0.004 | 0.038 |
|
9 | 0.333 | 0.370±0.005 | 0.906 |
|
7 | 0.714 | 0.352±0.005 | 0.076 |
Corresponding cluster identities are shown as a diagram of a network of association clusters in
Cluster identity |
Cluster size | Mean ± Standard error of observed age in years | Mean ± Standard error of expected age in years | Probability value |
|
8 | 17.875±1.529 | 25.201±0.106 | 0.022 |
|
10 | 31.500±2.861 | 25.206±0.099 | 0.064 |
|
12 | 24.833±3.649 | 25.202±0.083 | 0.902 |
|
12 | 20.583±2.624 | 25.211±0.085 | 0.078 |
|
9 | 25.667±3.444 | 25.149±0.100 | 0.846 |
|
7 | 32.429±3.213 | 25.149±0.116 | 0.060 |
Cluster identities shown here correspond to identities shown in
Our results show that the probability of crop-raiding by male elephants increased as a function of age such that the odds of being a crop raider rose dramatically at the age at which reproduction is initiated and the odds nearly doubled for males at their reproductive peak. This result strongly suggests that crop raiding is linked to the increasing energetic costs associated with reproduction or increased risk taking behavior associated with attainment of a peak in reproduction at age 45 years in male elephants
Our results also indicate that raiding by males was predicted by the raiding status of their closest associates and this effect was stronger if their associates were older. These results suggest that raiding is acquired through social learning from older males that are raiders. Our observations support results from experimental studies of social learning in other vertebrates showing that animals show a bias towards learning from individuals that are older, more experienced or familiar. For example in one study, young female guppies learned mate preferences from older females
Although a shared spatial environment could allow associates to adopt similar foraging behaviors through independent exploratory learning irrespective of social learning, this is a very unlikely explanation for the pattern we observed. Elephants are not territorial, and males particularly range widely and have large home ranges
Crop raiding is a high-risk and high-gain foraging strategy: successful raiders in Amboseli
For crop-raiding elephants, minimizing risk and avoiding detection by farmers may entail raiding late in the night or raiding mainly on moonless nights
In Kenya, the Kenya Wildlife Service policy is to frighten elephants from farms by shooting into the air, but farmers may spear crop raiding elephants illegally. During this study, we observed some crop-raiders with spear injuries presumably incurred during crop raiding. A study examining stress associated with crop raiding in the Amboseli and Maasai-Mara ecosystems in southern Kenya, found that raiders had higher levels of stress than non raiders
We also observed features of the elephant association network that may limit the spread and hence reduce the prevalence of raiding behavior in the population. First, the male elephant network had a significantly low density of associations indicating that males may have limited exposure to raiding behavior from other males within the network. Secondly the elephant social network had strong clustering and community structure. These features of the social network will hinder the transmission of information between elephant social groups because of sparse associations between individuals in different groups, but will enhance transmission within social groups because of the dense associations between individuals in the same social cluster.
Thirdly, we observed a weak association of males based on age similarity. Because age influenced the motivation to raid, this suggests that some males may have initiated raiding through social facilitation among similarly aged individuals, perhaps driven by energetic demands of reproduction. The association based on age similarity was driven by older male elephants
Cluster level network analyses revealed that, in five of the six clusters in the network, raiders were distributed randomly across clusters. These results were discordant with the results of pairwise association from Monte Carlo and ERG analyses, which showed that elephants associated with other males who were similar to them in raiding status. The discordance in cluster level and pairwise level analyses suggest that elephant association based on similarity in raiding status was driven by local dyadic effects rather than by network cluster effects. The lack of association based on raiding status at the cluster level and its presence at the dyadic network level could be a result of a low density of associations in the network. If the network is less dense within a cluster, most individuals within a cluster will not be connected to each other and this may limit the spread of raiding information even within a cluster. Individual variation in risk taking can also limit the homogenization of clusters resulting in a random distribution of raiders across clusters. If some individuals are risk averse, they may not adopt raiding as a foraging strategy; this in turn will result in clusters with both raiders (risk takers that have been exposed to raiding) and non raiders (risk-averse individuals who have been exposed or not exposed to raiding or risk-prone individuals that have not been exposed to raiding).
Male social clusters had a heterogeneous age distribution. This heterogeneous age distribution and association based on age similarity from ERG network analysis suggest that most males prefer to associate with age peers and some prefer to associate with individuals younger or older than them. The heterogeneity in age distribution within clusters suggests that social learning and life history are likely to simultaneously influence raiding behavior within clusters.
Our findings have implications for the management of human-elephant conflict particularly for populations in which males cause most of the conflict. First, our results demonstrate how social networks and life history milestones may jointly influence the prevalence of learnt behavior. This in turn suggests that maintaining heterogeneous age structure may be important in promoting adaptive learning and response to changing environments in animal populations. Related to this, our results suggest that aversive conditioning techniques, such as pepper spray canisters
(TIFF)
We thank the Kenya Wildlife Services and the Amboseli National Park staff for their hospitality during the study. We also thank Norah Njiraini, Katito Sayialel and Soila Sayialel for sharing their invaluable knowledge of Amboseli elephants, and to Timothy Saigilu, John Teete, and Elijah Olekeni for their invaluable support with monitoring of crop raiding elephants.