Social network analysis of wild chimpanzees provides insights for predicting infectious disease risk



  1. Heterogeneity in host association patterns can alter pathogen transmission and strategies for control. Great apes are highly social and endangered animals that have experienced substantial population declines from directly transmitted pathogens; as such, network approaches to quantify contact heterogeneity could be crucially important for predicting infection probability and outbreak size following pathogen introduction, especially owing to challenges in collecting real-time infection data for endangered wildlife.
  2. We present here the first study using network analysis to quantify contact heterogeneity in wild apes, with applications for predicting community-wide infectious disease risk. Specifically, within a wild chimpanzee community, we ask how associations between individuals vary over time, and we identify traits of highly connected individuals that might contribute disproportionately to pathogen spread.
  3. We used field observations of behavioural encounters in a habituated wild chimpanzee community in Kibale National Park, Uganda to construct monthly party level (i.e. subgroup) and close-contact (i.e. ≤5 m) association networks over a 9-month period.
  4. Network analysis revealed that networks were highly dynamic over time. In particular, oestrous events significantly increased pairwise party associations, suggesting that community-wide disease outbreaks should be more likely to occur when many females are in oestrus.
  5. Bayesian models and permutation tests identified traits of chimpanzees that were highly connected within the network. Individuals with large families (i.e. mothers and their juveniles) that range in the core of the community territory and to a lesser extent high-ranking males were central to association networks, and thus represent the most important individuals to target for disease intervention strategies.
  6. Overall, we show striking temporal variation in network structure and traits that predict association patterns in a wild chimpanzee community. These empirically-derived networks can inform dynamic models of pathogen transmission and have practical applications for infectious disease management of endangered wildlife species.


Many pathogens spread through host populations via social interactions (Altizer et al. 2003); thus, knowledge of a community's social system and contact structure can provide crucial information for predicting infectious disease outbreaks (e.g. Nunn et al. 2008; Drewe 2010; Griffin & Nunn 2012). Interindividual contacts that lead to pathogen transmission can be represented using networks, where each node represents an individual, and edges between nodes represent interactions that allow for pathogen transmission. Contact networks for humans and animals are often heterogeneous (e.g. Lusseau 2003; Schneeberger et al. 2004), which violates the common assumption of many basic epidemiological models that contacts are random and individuals are well mixed (Anderson & May 1991). Network analysis provides a feasible (albeit data intensive) approach to mathematically formalize transmission pathways and host contact variation (Newman 2010). Further, network analysis can identify potential superspreaders, individuals with disproportionately high contact levels, that could be targeted for vaccination, treatment or isolation (Lloyd-Smith et al. 2005). Studies of human contact networks often detect heterogeneity and the presence of superspreaders, which has been extremely influential in our understanding of transmission dynamics for SARS and HIV/AIDS (Anderson, Gupta & Ng 1990; Lloyd-Smith et al. 2005; Meyers et al. 2005). Superspreaders have also been identified in a few wildlife populations (e.g. Porphyre et al. 2008); however, network analysis is rarely used to investigate the epidemiology and control of wildlife diseases (Craft & Caillaud 2011). Here, we present the first study to analyse empirical wild chimpanzee contact networks within a framework of predicting implications for infectious disease risk.

Endangered wild ape populations have recently experienced outbreaks of Ebola and respiratory viruses, making infectious disease a major threat to their survival (Caillaud et al. 2006; Köndgen et al. 2008; Ryan & Walsh 2011). Respiratory diseases in particular have resulted in outbreaks with up to 25% community-level mortality at several long-term chimpanzee research sites (Ryan & Walsh 2011). With low birth rates and late reproductive maturity, ape populations can take decades to recover in size after an outbreak. For example, using mathematical models and a range of parameters derived from published ape epidemics (i.e. mortality rates of 4–25%), Ryan & Walsh (2011) estimated that a mountain gorilla population would require 5–32 years to recover following an outbreak of respiratory disease. In accordance with these predictions, a Tanzanian chimpanzee community took 15 years to return to its pre-epidemic population size after a 1987 respiratory disease outbreak (Williams et al. 2008).

In addition to the detrimental impact that pathogens can have on endangered apes, obtaining real-time infection data for wildlife is notoriously difficult. Collecting biological samples often requires risky interventions including darting and possibly anesthetizing immune-challenged individuals. Furthermore, the speed with which respiratory pathogens typically spread through ape communities (e.g. with a duration of roughly 2 weeks to 2 months: Hanamura et al. 2008; Williams et al. 2008; Köndgen et al. 2010) can limit researchers' abilities to collect comprehensive health data during an outbreak. Given these challenges, parameterizing realistic epidemiological models with association data (e.g. Davis et al. 2008; Hamede et al. 2012) is essential for developing strategies to reduce the risk and impact of infectious diseases. An underlying assumption of these models is that network edges represent possible pathogen transmission routes. Indeed, while relatively few wildlife studies have both host infection and host association data, there is a growing body of evidence that wildlife social networks strongly predict individual infection status (Otterstatter & Thomson 2007; Leu, Kappeler & Bull 2010; Bull, Godfrey & Gordon 2012) and that highly connected individuals tend to have greater parasite burdens than less connected individuals (Corner, Pfeiffer & Morris 2003; Godfrey et al. 2009; Leu, Kappeler & Bull 2010; but see: Otterstatter & Thomson 2007).

Great ape societies are highly structured and complex. Chimpanzees in particular live in permanent social groups termed communities and have a fission–fusion social structure, whereby individuals within the community frequently break off into subgroups, called parties, that vary in size and composition (Goodall 1986). A chimpanzee mother and her offspring travel together in a family unit, and sociality can vary greatly among adult females (Boesch & Boesch-Achermann 2000). In fact, Goodall (1986) noted that eastern chimpanzee females ranging in the core of the community's territory encountered other individuals on a daily or weekly basis, whereas females ranging on the periphery of the territory might encounter community members only a few times per year. Compared to females, males follow a linear dominance hierarchy (Muller & Wrangham 2004) and tend to be more gregarious (Gilby & Wrangham 2008). Additionally, other studies showed that party size tends to increase when females are in oestrus or when ripe fruits are available (e.g. Wrangham 2000).

In this study, we use network analysis to examine association patterns among individuals in a community of wild chimpanzees at Kibale National Park, Uganda. In particular, we quantify how association patterns that represent potential pathogen transmission routes vary over time, in response to factors such as fruit availability or the number of oestrous females. We also examine individual traits that contribute to high levels of association, and predict that high-ranking males and oestrous females will have disproportionately high levels of association with community members owing to increased rates of grooming and mating (e.g. Goodall 1986; Emery Thompson & Wrangham 2008). Investigating the dynamics and drivers of contact variation in wild apes is a necessary step for simulating pathogen spread and evaluating the success of much needed disease intervention strategies for this highly threatened primate clade.

Materials and methods

Study site and population

We studied the habituated wild Kanyawara chimpanzee community at Kibale National Park (0°34′ N, 30°21′ E) in Uganda. The site is dominated by moist deciduous forest interspersed with secondary forest, grassland and swamp (Chapman & Wrangham 1993). Weather data for the site were provided by C. Chapman. Further details on the ecology of Kibale are discussed in Struhsaker (1997). The Kanyawara chimpanzee community occupies roughly 37·8 km2 of forest (Wilson, Hauser & Wrangham 2001), and during the time of the study the community included 48 chimpanzees with 12 adult males (aged > 14), 14 adult females (aged > 13), eight immature males and six immature females (aged between 5–14 and 5–13, respectively; hereafter referred to as juveniles) and eight dependent offspring (aged ≤ 4). For additional information on the Kanyawara community (Appendix S1, Supporting information).

Data collection

We collected data on chimpanzee association patterns over 9 months between December 2009–August 2010 for 4–6 days per week between 6:00 am and 7:30 pm. Each morning, we randomly selected a focal chimpanzee from a party (typically at a nest site) to follow for 10 h. Every 15 min, we scanned the focal individual's party and recorded the identity of all party members based on individuals within a 50-m radius, a common criterion for estimating chimpanzee party sizes (Clark & Wrangham 1994). As an index for assessing patterns of close association within parties, at the same 15-min intervals, we also recorded pairs of individuals that were within 5 m of each other, which is a measure that commonly contributes to identifying close associations among primates (e.g. Gilby & Wrangham 2008). We limited our focal follows and party composition data to chimpanzees > 4 years of age (i.e. excluding dependent offspring, which remain in close contact with their mothers); we also excluded two adult females and a juvenile male on the periphery of the community who were observed only twice during the study. Our total sample size was 37 individuals (12 adult males, 12 adult females, seven juvenile males and six juvenile females). We recorded days when parous females had maximal sexual swellings and noted ripe fruit species on which focal animals foraged.

Estimating association indices

We calculated monthly pairwise association indices between individuals at two spatial scales: (i) party-level association indices were based on the frequency of monthly co-occurrence in the same party and (ii) close-contact association indices (i.e. within-party and overall 5-m associations, described below) were based on the frequency with which two individuals were seen within 5 m of each other during a given month. We examined associations at the party level as a proxy for the transmission of pathogens spread by non-close contact (e.g. via fomites, aerosol transmission, or faecal-oral routes). To estimate party-level associations, we calculated a monthly ‘twice weight index’ (Cairns & Schwager 1987), hereafter referred to as a monthly party association index (PAI), from party membership scans. This parameter calculates the ratio of scans in which chimpanzees A and B were observed in the same party relative to the total number of scans in which either A or B was observed in any party as follows:

display math(eqn 1)

where SAB represents the number of scans where A and B were observed in the same party, SA represents scans where A was observed in a party without B and SB represents scans where B was observed in a party without A. PAIs and subsequent indices described below could range from 0 (i.e. individuals in a pair were never observed associating in the given month) to 1 (i.e. individuals in a pair were observed to be associating during 100% of the observations for the given month).

Close-contact interactions were examined as a proxy for pathogens requiring direct contact or respiratory droplets to spread. As one close-contact measure, within-party association indices (WPAI) represent the proportion of scans in which chimpanzees A and B were observed within 5 m of each other, given that they were within the same party:

display math(eqn 2)

where SAB5 represents the number of scans where A and B were observed within 5 m of each other. To examine which individuals were most central to the 5-m networks, we calculated an overall 5-m association index (5mAI), which incorporated the probabilities that individuals A and B would be both within the same party and within 5 m of each other:

display math(eqn 3)

Thus, this index estimates the overall proportion of time that individuals A and B were within a 5-m distance.

To examine host interactions at a temporal scale that reflects the transmission biology of real-world pathogens, we analysed association patterns at both 2-week and monthly intervals, as respiratory diseases common to chimpanzees and humans have infectious periods that range from a few days to 1 month (e.g. influenza: 2–3 days, measles: 6–7 days, chicken pox: 10–11 days, Streptococcus spp.: 14–30 days; Anderson & May 1991; Ekdahl et al. 1997) and published reports of wild chimpanzee respiratory illnesses suggest that epidemic durations often range from roughly 2 weeks to 2 months (Williams et al. 2008; Köndgen et al. 2010). Because associations across both time steps were significantly correlated for both PAIs and 5mAIs (Table S1, Supporting information), and other results were robust across time steps, we present results for monthly associations in the main text (see Tables S2 and S3, Supporting information for 2-week time step results).

Visualizing networks

We constructed monthly party and 5-m association networks in r version 2.15.1 (R Core Development Team 2010) with the igraph package version 0.5.5–4 (Csardi & Nepusz 2006). Party and 5-m network edges were weighted according to the monthly pairwise PAIs and 5mAIs, respectively, such that pairs with higher association indices had thicker edges.

Individual trait data

In all analyses, we categorized chimpanzees based on their age, sex, dominance rank, family size (Table S4, Supporting information) and for pairwise analyses, whether two individuals were related to each other. Chimpanzee rank, based on dominance interactions for adult males, was categorized such that high-, medium- and low-ranking adult males, respectively, occupied the rank categories of Male 1 (M1, N = 5), Male 2 (M2, N = 4) and Male 3 (M3, N = 3). By grouping individuals in this way, all males stayed within their respective rank categories throughout the study period, despite minor reshuffling in the linear hierarchy. Female chimpanzees rarely show dominance interactions; however, females occupying and foraging in the core area of the territory (at Kanyawara) tend to be higher ranking than those occupying the peripheral areas (Kahlenberg, Emery Thompson & Wrangham 2008). Thus, we assigned core-area adult females and their juvenile offspring to rank categories Female 1 (F1, N = 6) and Juvenile 1 (J1, N = 9) and edge-ranging adult females and their offspring to Female 2 (F2, N = 6) and Juvenile 2 (J2, N = 4). Additional details on rank categorization are in Appendix S2 and Table S4 (Supporting information).

Lastly, we defined a family unit as a mother and her noninfant offspring, such that an individual's family size was the total number of noninfant chimpanzees in this family unit. In one unique case, a young adult male and his juvenile sibling were considered a family unit (Table S4, Supporting information), as their mother was deceased. Chimpanzees who travelled without a family unit (e.g. adult males, females with infants only) were assigned a family size of one. We considered mother–offspring pairs and maternal siblings to be related, based on long-term records from the field site.

Monthly changes in network density

To compare PAIs and 5mAIs over time, we calculated monthly network density as the sum of the network's observed edge weights divided by the sum of the maximum possible edge weights (Hanneman & Riddle 2005). To examine how stable party and 5-m networks were over time, we assessed correlations between monthly association index matrices using a quadratic assignment procedure (see Appendix S3, Supporting information for details) in ucinet version 6.343 (Borgatti, Everett & Freeman 2002).

Analyses of pairwise associations

To examine how social factors (e.g. rank status) and ecological factors (e.g. fruit availability) affect temporal pairwise associations at party and 5-m levels, we fit two models (for PAI and WPAI data, respectively) to Bayesian logistic mixed-effect models using a Markov chain Monte Carlo (MCMC) framework. We tested for significant relationships between monthly pairwise associations and the following predictor variables: age (adult–adult, adult–juvenile, juvenile–juvenile), sex/oestrus (i.e. pairwise combinations of males, nonoestrous females and oestrous females, Note: parous females were categorized as oestrous during months in which they were observed to be in oestrus; nulliparous females were never categorized as oestrous), relatedness (related, unrelated), difference in family size (range: 0–3) and difference in rank category (scored as 1/0 where a pair in the same rank category scored a 0 and a pair in different ranks scored a 1). Because, we expected mothers and their juveniles to associate frequently, for this analysis we collapsed the adult female and juvenile ranks into FJ1 (core-ranging individuals) and FJ2 (edge-ranging individuals).

We also included two key parameters that could affect associations over time. First, we included a parameter for the number of parous oestrous females observed during each month, as males prefer mating with parous over nulliparous females (Muller, Thompson & Wrangham 2006). Additionally, research at some sites shows that increased fruit availability is linked to larger parties (e.g. Wrangham 2000). We did not have fruit abundance data; however, we included parameters for the monthly presence/absence of preferred ripe fruit species (Mimusops bagshawei, Pseudospondias microcarpa, Uvariopsis congensis: Wrangham et al. 1996) according to our focal data, as eating of preferred fruits is strongly associated with fruit availability for Kanyawara chimpanzees (Wrangham et al. 1991). We also included a parameter for mean daily rainfall from 2 months prior as a proxy for current fruit availability.

To account for autocorrelation from repeated measures, we assessed model fit with random effects of chimpanzee ID, chimpanzee pair and month. One difficulty with including a random effect for individual ID was that an individual could appear interchangeably as individual A or individual B in the observed pairwise associations described in eqns 1–3. This interchangeability was due to the fact that the associations were not directed, meaning they did not have a specific ‘sender’ and ‘receiver’. We resolved this issue by using the multimembership modelling capabilities of the MCMCglmm package (Hadfield 2010) in r. Additional analysis details are in Appendix S3 (Supporting information); r code is available upon request.

Individual traits associated with network centrality

To identify individual traits associated with increased contact, we used ucinet to calculate three weighted network centrality measures for each chimpanzee: degree, eigenvector and flow-betweenness. Weighted degree centrality (hereafter referred to as degree) for a node is the sum of the node's edge weights (Newman 2010). Eigenvector centrality is based on an individual's connectedness and the connectedness of an individual's associates, where an individual with high eigenvector centrality is connected to well-connected associates (Newman 2010). Lastly, flow-betweenness centrality is defined as the proportion of times an individual lies along the shortest path between pairs in the network (Freeman, Borgatti & White 1991). Previous theoretical and empirical work in human and wildlife systems has shown that individuals with high degree, eigenvector or flow-betweenness centrality are more likely to contract and transmit pathogens than individuals with low centrality (e.g. Corner, Pfeiffer & Morris 2003; Salathé et al. 2010).

Using node-level permutation-based regressions, we fit individual centrality data in r with 30 000 permutations per test to investigate relationships between each centrality measure and the following predictor variables: rank, oestrous-status, family size, continuous age and sex (while controlling for month effects). We controlled for sampling effort across individuals by weighting the model variance structure according to the number of scans in which each individual was a focal subject. To account for comparisons of three centrality measures, we applied a Bonferroni correction and considered relationships where < 0·017 (i.e. < 0·05/3) to be significant. Age and sex were excluded from the final models because they were confounded with rank (which was already separated by age and sex groups), explained < 1% of the variation (as determined by adjusted R2) and were never significant after Bonferroni correction. Additional analysis details are in Appendix S3 (Supporting information).


Association patterns and social network descriptions

On average, each chimpanzee was followed as a focal subject for 27·79 (±3·6) h (Fig. S1, Supporting information), comprising a total of 1028 focal observation hours and 4114 fifteen-min scans for all individuals combined. Our analysis included 306 212 pairwise party associations and 14 673 pairwise 5-m associations over 9 months. When averaged across months and individuals, randomly selected chimpanzee pairs were observed associating at the party level c. 26% of the time (mean PAI: 0·255, range: 0·0–1·0, SE: 0·003) and at the 5-m level 4% of the time (mean 5mAI: 0·041, range: 0·0–1·0, SE: 0·001). Three parous females came into oestrus at different points in the study; the number of oestrous females per month was low (range: 0–2) owing to a high proportion of lactating females in the study population.

Monthly party networks were dynamic over time (Figs 1 and 2; Figs S2 and S3, Supporting information) and network density ranged from 0·14 (January) to 0·42 (April). Party networks for consecutive months were highly correlated (Fig. S3, Supporting information), but correlation coefficients decayed as the time lag increased, indicating that party networks were locally stable within 2–3 month periods but were dynamic on a longer time scale. The 5-m network density ranged from 0·03 (March) to 0·06 (January) (Fig. 2; Fig. S4, Supporting information). There was no significant relationship between monthly party network density and monthly 5-m network density (R2 = 0·17, = 0·270; Fig. 2 inset). Variance-to-mean ratios of total edge weights per individual (i.e. weighted degree centrality) for party and 5-m networks were low across months (party network: 3·52, 5-m network: 0·73; Fig. S5, Supporting information), and while monthly party networks were significantly more aggregated than 5-m networks (t8·3 = 4·38, = 0·002), degree distributions indicated that networks were not highly aggregated at either scale (Figs S6 and S7, Supporting information).

Figure 1.

Monthly party association networks for a month with (a) no oestrous females (March), (b) one oestrous female (June) and (c) two oestrous females (August). Nodes (circles) represent individual chimpanzees (N = 37) and edges (lines) represent observed associations, where edge thickness corresponds to the pairwise party association indices (PAIs). All networks are displayed with identical layouts and only edges with PAIs > 0·35 are shown. Dark red nodes have at least one edge above the PAI cut-off whereas light red nodes do not have any edges above the PAI cut-off. All nine monthly party association networks are shown in Fig. S2.

Figure 2.

Density of monthly party networks (blue solid line) and 5-m networks (red dashed line) with standard error bars. The inset shows that there is no significant relationship between monthly party network density and monthly 5-m network density (Spearman Rank Test: rho = −0·4, = 0·291). Circled numbers show the number of oestrous females in each month.

Effects of social and ecological factors on pairwise associations

The number of oestrous females in a given month significantly increased pairwise associations at the party level, where for each additional oestrous female, the odds of a pair associating were roughly twice as high (Table 1, Fig. 3). There was a significant interaction between the number of oestrous females and age, such that adult–adult pairs experienced the largest increase in associations as the number of females in oestrus increased. Similarly, of all the pairwise sex combinations, pairs that included one oestrous female associated the most frequently. The odds of related pairs being in a party together were over 20 times greater than the odds for unrelated pairs, and chimpanzees were significantly more likely to associate with individuals of their own rank category. Family size difference negatively affected association indices, indicating that individuals with large families (i.e. 3–4 members, Table S4, Supporting information) tended to associate with other large families, and individuals without family units tended to associate with each other (Table 1).

Table 1. Effect of social factors on pairwise associations in party networks. The posterior mean, 95% credible interval, P-value based on Markov chain Monte Carlo sampling and odds ratios (OR) are shown for fixed effect parameters
FactorPosterior mean95% CI P OR
  1. Bolded relationships are significant at < 0·05.

  2. Sex/oestrus and age categories are abbreviated as follows: age (adult:adult, AA; adult:juvenile, AJ; juvenile:juvenile, JJ), sex/oestrus [pairwise combinations of male (M), female in oestrus (Fe) and female not in oestrus (F)].

Intercept−3·58(−4·90, −2·22) <0·001  
Related3·01(2·63, 3·39) <0·001 20·20
Sex (M:F)0·73(−0·11, 1·57)0·0872·07
Sex (M:M)1·30(−0·38, 2·92)0·1193·67
Sex (F:Fe)1·76(1·24, 2·28) <0·001 5·83
Sex (M:Fe)2·67(1·72, 3·65) <0·001 14·44
Difference in family size−0·13(−0·20, −0·06) <0·001 0·88
Difference in rank−1·04(−1·21, −0·86) <0·001 0·35
Age (AJ)0·69(−0·23, 1·55)0·1251·99
Age (JJ)1·16(−0·59, 2·92)0·1913·18
Number of oestrous females0·98(0·84, 1·12) <0·001 2·65
Number of oestrous females:Age (AJ)−0·22(−0·40, −0·02) 0·025 2·14
Number of oestrous females:Age (JJ)−0·44(−0·72, −0·16) 0·003 1·70
Figure 3.

Estimated effect of oestrous events on pairwise party associations. Model estimates of average association indices are shown for the three age-pair combinations with 95% credible intervals. The x-axis shows the number of females in oestrus for a given month. Age combinations of adult–adult, adult–juvenile and juvenile–juvenile pairs are represented by squares, circles and triangles, respectively. Figure estimates were calculated from the Markov chain Monte Carlo posterior distributions, while holding the presented parameters constant and allowing all other parameters to range across their possible values.

The final model for pairwise party associations included random effects of chimpanzee ID and pair ID. Month was not included as a fixed or random effect, as the number of oestrous females per month was a better predictor of monthly pairwise associations than month per se, based on the relative deviance information criterion, DIC (ΔDIC > 50). Rainfall lag and fruit availability parameters were removed because their exclusion increased model fit (rainfall: ΔDIC > 30, fruit: ΔDIC > 20, see Appendix S4, Supporting information for discussion of fruit availability and network structure). The final model had R2 values that ranged from 0·32 to 0·58 for the amount of variation explained in each of the monthly networks, with the exception of August (R= 0·07; Fig. S8, Supporting information).

Results for 5mAIs were similar to party-level results, although several variables in the 5-m model were significant in some but not all months (Table S5, Supporting information). A major difference between these two levels of association was that pairs including an oestrous female were often less likely to associate within 5 m (as compared to pairs including an oestrous female being more likely to associate at the party level). As a second key difference, month was included as a fixed effect variable that interacted with every other fixed effect variable (age, sex/oestrus, relatedness, family size difference and rank category difference), allowing the coefficients of these variables to vary for each monthly network (Table S5, Supporting information). While more challenging to interpret, this final model fitted the data much better than the model including month as a random (and hence, additive) effect (ΔDIC > 100) or excluding month and including the number of oestrous females to describe monthly change (ΔDIC > 350). This indicates that the number of oestrous females was not as good of a predictor for 5-m associations as it was for party associations. The incorporation of month as a fixed effect precluded testing temporal variables (i.e. fruit availability, rainfall and number of oestrous females per month). Monthly R2 values for the final 5-m association model ranged between 0·18 and 0·53 (mean: 0·34; Fig. S8, Supporting information).

Predictors of individual centrality

Family size and dominance rank were the most important predictors for individual centrality at both the party and 5-m levels after controlling for the month of observation (Table 2; Fig. 4; Fig. S9, Supporting information). Adult females and juveniles with large families (i.e. 3–4 members) had significantly higher degree and eigenvector centrality; however, family size was not an important predictor for an individual's flow-betweenness centrality. This indicates that chimpanzees with large families had more edges and associated with other well-connected individuals, but were not more likely than random to connect two other individuals in the community.

Table 2. Effect of social factors on party and 5-m association network centrality measures. Coefficients (β) and P-values are presented. P-values for rank post hoc significance tests are in Table S6. Coefficients and P-values for month parameters are presented in Table S7
 Party association networks, N = 2945-m association networks, N = 294
β P β P β P β P β P β P
  1. Bolded values indicate significant relationships after Bonferroni correction. R2 values are shown for each test.

Intercept17·36 <0·001 0·170·26933·410·0581·350·0960·130·32435·520·109
Rank: M2−1·430·108−0·030·052−1·410·212−0·230·153−0·040·085−0·750·449
Rank: M3−1·540·112−0·020·078−2·110·135−0·65 0·004 −0·050·028−0·150·496
Rank: F1−1·190·150−0·010·230−3·510·023−0·180·209−0·020·271−5·080·163
Rank: J1−0·820·2530·000·389−3·740·019−0·280·112−0·020·221−3·490·255
Rank: F2−5·30 <0·001 −0·09 <0·001 −6·91 <0·001 −1·20 <0·001 −0·12 <0·001 −7·430·074
Rank: J2−4·47 <0·001 −0·08 <0·001 −5·82 0·002 −1·13 <0·001 −0·13 <0·001 1·570·389
Family size0·73 0·008 0·01 <0·001 0·570·1140·18 0·001 0·02 0·001 −0·500·356
R 2 0·618 <0·001 0·347 <0·001 0·195 <0·001 0·406 <0·001 0·251 <0·001 0·0370·835
Figure 4.

Estimated effects of rank and family size on average degree for (a) party and (b) 5-m networks. There was a significant positive relationship between an individual's family size and degree centrality. Black, white and cross-symbol circles represent model estimates for individuals with family sizes of one, three and four members, respectively (by definition of family unit, adult male ranks are only presented with a family size of one). Letters on plots show which rank categories were significantly different (where overlap in letters between two rank categories indicates no significant difference after Bonferroni correction), after controlling for family size and oestrous status.

Regarding rank, in both the party and 5-m association networks edge-ranging females and juveniles (F2 and J2) had significantly lower degree and eigenvector centrality than all other ranks in both party and 5-m networks (with the exception that J2 did not have significantly lower degree centrality than low-ranking males in 5-m networks after Bonferroni correction; Fig. 4; Fig. S9, Table S6, Supporting information). F2 and J2 also had significantly lower flow-betweenness centrality in party networks than high- and medium-ranking males. Altogether, these results indicate that edge-ranging adult females and juveniles were less connected to others and had fewer well-connected associates than all other ranks. They were also less likely than adult males to connect two random individuals in the party networks.

In 5-m networks, high-ranking males (M1) had significantly higher degree centrality than low-ranking males (M3) (Fig. 4; Fig. S9, Table S6, Supporting information). Additionally, there was a strong trend (P < 0·05) for M1 to have higher eigenvector centrality than M3 in 5-m networks and higher flow-betweenness than core-ranging adult females and juveniles (F1, J1) in party networks; however, these differences were not significant after Bonferroni correction. Thus, while edge-ranging females and juveniles were nearly always the least central to the community, the relationship between high-ranking males and centrality was weaker and less consistent, with high-ranking males being significantly more central than other community members (e.g. M3, F2, J2) for some but not all centrality measures (Fig. 4; Fig. S9, Table S6, Supporting information). Lastly, oestrous status was never significantly related to centrality in party or 5-m networks (Table 2).


Association patterns and insights for disease transmission

Our results demonstrate interindividual and temporal variation in association patterns of wild chimpanzees, which should have profound effects on pathogen transmission dynamics. A main advantage of network analysis over more traditional connectivity measures, such as party size, is that network analysis explicitly quantifies how connectivity varies in relation to demographic and behavioural traits and among individuals in a community. Degree distributions demonstrated that neither party nor 5-m networks were highly aggregated (i.e. most individuals had moderate centrality as opposed to a few superspreaders accounting for a majority of contacts); yet certain types of individuals had significantly higher association rates than others.

Adult females and juveniles with large families (i.e. 3–4 family members) were significantly more central than expected by chance in both party and 5-m networks, and individuals in core-ranging families were significantly more central than those in edge-ranging families. Thus, core-ranging adult females and juveniles with large families were the most central to the community. Additionally, chimpanzees associated more frequently with related individuals and individuals that had similar family sizes. Therefore, it seems that core-ranging chimpanzees with large families associated frequently with family members and also formed what Goodall (1986) referred to as ‘nursing parties’, where mothers and juveniles of different family units socialize together. Notably, there is evidence in West African chimpanzees (Taï Forest) that young juveniles maintain respiratory diseases in the community through play or close contact (Kuehl et al. 2008), a dynamic that has been demonstrated among human children for various childhood diseases (e.g. Fine & Clarkson 1982). Edge-ranging families were nearly always the least central to the study community. In fact, the average degree centrality between a core-ranging adult female with a large family and an edge-ranging adult female without any juvenile offspring differed roughly by a factor of 2 in party networks and 2·5 in 5-m networks. Thus, individuals from edge-ranging families were the least likely to contribute to or be affected by pathogen transmission (although peripheral individuals could be exposed to pathogens from other communities or human settlements that overlap with forest edges).

Among core-ranging individuals, the average centrality of an adult female chimpanzee with three juveniles was roughly 2·5 degrees higher than that of an adult female with no juveniles. Previous wildlife network studies have demonstrated that even small differences in centrality can be linked to key differences in individual infection status. For example, a study examining parasites in gidgee skinks (Egernia stokesii) determined that while network centrality was an effective predictor of parasite burden, the average difference in centrality between skinks with and without ticks was only c. 0·35 degrees (Godfrey et al. 2009). Thus, while we recognize that the magnitude of centrality metrics (which depend on network size and system-specific association definitions) should not be directly compared across systems, the significant increase we observed in chimpanzee centrality due to family size (even if modest in magnitude) could have a crucial impact on individual infection status.

While not as consistently central as core-ranging adult females and juveniles with large families, high-ranking males also had high centrality. Past work on the same study community showed that high-ranking males tend to have increased levels of immunosuppressing testosterone (Muller & Wrangham 2004), and work in a nearby chimpanzee community (Ngogo) recently demonstrated that high-ranking males had both increased testosterone levels and greater helminth burdens (Muehlenbein & Watts 2010). Thus, in combination with the well-established immunosuppressive effects of sex hormones, their moderately central location in the network should make high-ranking males susceptible to contracting and transmitting a variety of pathogens. Taken altogether, we expect that core-ranging adult female and juvenile chimpanzees with large families, and to a lesser extent high-ranking males, should play an important role in pathogen transmission.

Contrary to our predictions, oestrous females were not significantly more central than expected by chance in party or 5-m networks. This is surprising considering that among party networks, pairs including oestrous females had higher levels of association and oestrous females significantly increased association patterns across the community. Because a majority of adult females in our study community were nursing infants, the sample size for oestrous females was limited (N = 3). Furthermore, one oestrous female was frequently absent from the community and was presumed to be engaging in consortships, in which a mating pair travels away from the community (Goodall 1986). In future studies of centrality with larger samples of oestrous females, it may be necessary to develop networks that span shorter time frames (i.e. the length of maximal swelling or roughly 1 week), as examining longer time steps includes intervals when the female does not have an oestrous swelling and is potentially experiencing lower centrality.

While often overlooked in epidemiological analysis, temporal changes in behavioural interactions can affect outbreak timing (Altizer et al. 2006), as demonstrated by peaks in measles transmission in children during school sessions (Fine & Clarkson 1982) or by phocine distemper outbreaks coinciding with the haul-out behaviour of seals (Swinton et al. 1998). Chimpanzee pairs were twice as likely to associate, and party networks were denser when females were in oestrus, suggesting that oestrous events represent times of high vulnerability to infectious disease outbreaks. This result confirms findings from long-term field studies showing that chimpanzee party size increases with the number of oestrous females (e.g. Wrangham 2000). Notably, there was no significant relationship between party and 5-m network density, and the number of oestrous females did not significantly affect 5-m-level associations. Thus, our network analyses suggest that the potential risk of outbreaks from pathogens that require very close contact for transmission might not increase with oestrous events.

Implications for conservation and infectious disease management

Epidemiological modelling studies in humans have shown that targeting central individuals for control efforts is significantly more effective in mitigating disease than applying control efforts randomly (Lloyd-Smith et al. 2005; Salathé et al. 2010). In a handful of cases, vaccination has been used to reduce the impact of emergent epidemics in endangered wildlife populations (gorilla measles and chimpanzee polio: Woodford, Butynski & Karesh 2002; Ethiopian wolf rabies: Haydon et al. 2006). Given the detrimental impacts of pathogens on great ape communities (e.g. Bermejo et al. 2006; Caillaud et al. 2006; Köndgen et al. 2008), some wildlife biologists have called for vaccinating great apes prophylactically for high-risk pathogens (Ryan & Walsh 2011). To effectively plan control strategies and minimize human interference, network models can indicate the minimum number of well-connected individuals that should be vaccinated to reduce outbreak sizes (as per: Salathé et al. 2010). Importantly, using coarser connectivity metrics such as party size or group membership to parameterize infectious disease models would only capture a fraction of the contact heterogeneity observed in the networks described here. Our next steps include using susceptible–infected–recovered (SIR) bond percolation models (Newman 2002; Meyers 2007) to simulate pathogen transmission on the observed monthly chimpanzee networks to assess the effectiveness of different intervention strategies in mitigating epidemics (such as targeting core-ranging individuals with large families for vaccination). This work is already underway with results from these simulations showing that moderately infectious pathogens (e.g. influenza) starting in core-ranging adult females and juveniles with large families are likely to generate significantly larger outbreaks than infections starting in other individuals (J. Rushmore, unpublished data).

Our findings are limited by examining a single chimpanzee community, and we recognize the need for similar analyses at additional field sites to provide a more comprehensive framework for designing disease management plans. Notably, the association data necessary for network analyses are likely available in long-term data bases for many habituated wild ape communities. We encourage additional researchers to analyse such association data with a focus on potential pathogen transmission routes. In conclusion, our findings demonstrate temporal and interindividual variation in association patterns for a wild chimpanzee community and highlight how such behavioural variation could be incorporated into the development of disease management strategies for an endangered wildlife population.


We thank the Kibale Chimpanzee Project (including directors: R. Wrangham and M. Muller, field manager: E. Otali and field assistants: S. Bradford, C. Irumba, J. Kyomuhendo, F. Mugurusi, S. Musana, W. Tweheyo), the Uganda Wildlife Authority, the Uganda National Council for Science and Technology and the Makerere University Biological Field Station for research permission and field support. We also thank the Kibale Chimpanzee Project for access to unpublished Kanyawara demographic data. We thank C. Chapman for providing weather data for Kibale National Park and L. Meyers and the Meyers laboratory for thoughtful discussion on analyses. We acknowledge the Training Workshops on the Ecology and Evolution of Infectious Diseases (supported by NSF DEB-0722115) and the University of Georgia Statistics Consulting Center for helpful instruction on analysis methodology. We also thank C. Chapman, R. Hall, R. Wrangham and the Altizer and Ezenwa Lab Groups for comments on manuscript drafts. This research was supported by the US Fish and Wildlife Service Research Awards (96200-9-G250 and 96200-1-G183) to SMA, RMS and JR. Additional support to JR included a Morris Animal Foundation Fellowship (D10ZO-401), a Fulbright Fellowship, a Margot Marsh Biodiversity Foundation Grant, a Primate Action Fund Grant, a Graduate Women in Science Fellowship and an ARCS Foundation Award. DC was supported by NSF grant DEB-0749097 to L. Meyers. The University of Georgia Institutional Animal Care and Use Committee approved animal use protocols (protocol #A2009-10062).