Assessing the sociability of former pet and entertainment chimpanzees by using multiplex networks

The study examines whether multiplex social network analysis, which simultaneously considers multiple interaction types, provides a richer and more realistic assessment of sociability in small groups of chimpanzees than traditional single-layer or aggregated networks. Prior animal social network research has shown complex structures and the importance of multiple interaction dimensions. Chimpanzee sociality often centers on allogrooming, but other interactions (affiliative behaviours, spatial proximity) are also relevant, especially when grooming is rare or difficult to observe. Earlier work with former pet/entertainment or ex-laboratory chimpanzees indicates early life history (wild-caught status and being housed without conspecifics during infancy) can impair social behaviours, particularly grooming. The authors hypothesize that: (1) multiplex analysis increases information gain over single-layer and aggregated networks even in small groups; (2) specific layers will show similarities (e.g., allogrooming with passive close proximity; affiliative behaviour with allogrooming) while others will not (e.g., stationary vicinity vs. high-level contact layers); and (3) individual and dyadic differences in interaction occurrence will relate to biographical history of both partners. The four interaction types are conceptualized as a sociability gradient: stationary vicinity (low-level), affiliative behaviour (medium-level), and allogrooming and passive close proximity (high-level). The study aims to test interlayer similarity, information reducibility, and effects of origin, predominant housing during infancy, and sex on interaction patterns.

The paper situates multiplex network analysis within animal social network research that has elucidated learning, cooperation, disease spread, and group structure. Traditional single-edge analyses can miss multidimensional dynamics, motivating multilayer methods. For primates, especially chimpanzees, social grooming is a key behaviour often used as the sole edge, but spatial co-occurrence and other affiliative interactions can also inform social structure. Prior studies show early adverse experiences (wild-capture, social deprivation in infancy) lead to long-lasting effects on chimpanzee social skills, personality, stress physiology, and even brain structure. Early maternal care is critical; grooming emerges later in development than play, underscoring the potential sensitivity of grooming to early deprivation. Earlier work in this population found early life variables affect grooming centrality, suggesting broader sociability measures may also be impacted and warrant a multilayer approach.

Ethics: Behavioural observations only, compliant with institutional and national guidelines (Fundació MONA; ASAB/ABS; Spain RD 53/2013). Sample: Fourteen adult former pet/entertainment chimpanzees in two stable social groups (Mutamba: 5 males, 2 females; Bilinga: 4 males, 3 females) housed at Fundació MONA (Catalonia, Spain). Two outdoor enclosures (2,420 m² and 3,220 m²), visual access between groups but no physical contact. Standardized feeding and enrichment; water ad libitum. Biographical details provided (Table 1). Data collection: May 2018–January 2019, two-minute scan sampling during 20-minute sessions between ~10:30–18:30, evenly distributed over days/times. Nine trained observers passed a three-step reliability protocol (method test and video test with ≥85% agreement). Recorded for all individuals: behaviour (affiliative types, allogrooming), passive close proximity (within arm’s reach), positions (GPS on enclosure map), and vertical level (ground + four levels). Linear inter-individual distances computed every two minutes via QGIS 2.18 matrix distance; pairs within 5 m counted for stationary vicinity, corrected by subtracting scans with arm’s-reach proximity and when vertical level difference exceeded one. Data logged on tablets using ZooMonitor. Total scans: 67,997 (Bilinga 32,320; Mutamba 35,677). Edge definitions (indices): - Stationary vicinity: out of arm’s reach but within 5 m; index = scans at 5 m (excluding arm’s reach and >1 level difference) divided by scans with mutual access. - Affiliative behaviour: social play, socio-sexual, follow, embrace, feed together, touch, mouth-to-mouth, short body contact, extend arm (excluding allogrooming); directed; index = scans with A affiliative to B / scans with mutual access. - Allogrooming: directed grooming; index = scans with A grooming B / scans with mutual access. - Passive close proximity: within arm’s reach without other interaction; symmetric; index = scans within arm’s reach / scans with mutual access. Indices are mutually exclusive per dyad per scan. Access defined as both having access to the outdoor area (even if one was indoors at the moment). Indices expressed as proportions per directed dyad for directed behaviours; symmetric for proximity measures. Multiplex preparation: Indices normalized per layer by dividing by that layer’s maximum observed value across both groups, yielding weighted values in [0,1]. Separate multiplex constructed for each group. Network construction and analysis: Used MuxViz (R environment) to build 4-layer directed weighted multiplex per group. Analyses included: - Visualizations of each layer and annular plots of node properties. - Interlayer correlation via edge-overlap (fraction of shared edge weights across layer pairs). - Reducibility analysis using Von Neumann entropy and quantum Jensen–Shannon divergence with Ward hierarchical clustering to iteratively aggregate most similar layers, computing relative entropy at each reduction step. - Node measures: eigenvector centrality (per layer and aggregated network) and eigenvector versatility (multiplex) to rank individual importance. Statistical analysis (LMMs in R 3.5.0, lme4): Five models using indices as dependent variables: four for each interaction type and one for the aggregated sum of the four indices. Fixed effects: origin (wild-caught vs captive-born, modeled dyadically as sender→receiver categories), predominant housing condition during infancy (PHCinfant; with vs without conspecifics for >2.5 of first 5 years, dyadic sender→receiver categories), and sex (M/F dyadic categories). Random effects: group, with sender ID nested within group. Model selection by comparing full vs null (no fixed effects). Type III ANOVA (Satterthwaite) and post hoc Tukey tests with Holm–Bonferroni adjustment. VIFs <1.2 for all fixed effects. QQ plots inspected for residual normality.

Network structure and densities: - Each group had 7 nodes and 4 layers (168 possible directed edges per multiplex). Present edges: Bilinga 139 (83%); Mutamba 166 (99%). - Overall multiplex density: Bilinga 0.83; Mutamba 0.99; layers ranged 0.57–1.00 (Table S1). - Stationary vicinity density = 1.00 in both groups; passive close proximity density = 1.00 in both groups. - Affiliative behaviour: fully connected in Mutamba; >25% edges missing in Bilinga. - Allogrooming density: 0.57 (Bilinga) vs 0.95 (Mutamba); Mutamba’s mean allogrooming index about twice Bilinga’s. Edge-overlap and interlayer similarity: - Mean global edge-overlap across all four layers: 10% (Bilinga), 19% (Mutamba) (Table S2). - Highest pairwise edge-overlap in both groups between stationary vicinity and passive close proximity: 65% (Bilinga) and 79% (Mutamba). - Bilinga: next highest affiliative vs passive close proximity 62%; affiliative vs allogrooming 49%; lowest allogrooming vs stationary vicinity 17%. - Mutamba: second highest allogrooming vs passive close proximity 61%; all other pairs 54%. Reducibility (Jensen–Shannon distances; lower = more similar): - Most similar layers in both groups: stationary vicinity vs passive close proximity (Bilinga 0.117; Mutamba 0.102). - Bilinga: next most similar allogrooming vs affiliative 0.172; then affiliative vs passive close proximity 0.291; greatest dissimilarity allogrooming vs stationary vicinity 0.358. - Mutamba: second most similar allogrooming vs passive close proximity 0.197; greatest dissimilarity affiliative vs passive close proximity 0.344 (Table S3). - Relative entropy decreased at each aggregation step; 4-layer multiplex retained the highest information, indicating that reducing layers loses information (Fig. 3c,d). Individual-level patterns (eigenvector measures): - Some individuals showed high centralities across most layers (e.g., Juanito, Waty, Africa in Mutamba; Tom, Bea in Bilinga), while others were low overall (e.g., Nico in Bilinga; Charly, Toni in Mutamba). However, individuals could be high in one layer and low in others (e.g., Charly and Toni high in affiliative but low elsewhere; Victor and Tico low in allogrooming but high in passive close proximity), underscoring the value of multilayer assessment. Linear mixed models: - Affiliative behaviour: significant effects of origin (F=4.272, p=0.007), PHCinfant (F=12.447, p<0.001), and sex (F=2.892, p=0.040). Post hoc: captive→wild and wild→captive > captive→captive (z=3.120, p=0.009; z=3.184, p=0.009). Dyads with both partners housed with conspecifics in infancy (with→with) > with→without (z=−4.328, p<0.001), without→with (z=−4.193, p<0.001), and without→without (z=−5.967, p<0.001). Females→females > females→males (z=−2.590, p=0.048) and > males→females (z=−2.673, p=0.045). - Allogrooming: significant effects of origin (F=3.085, p=0.032) and sex (F=6.032, p<0.001); PHCinfant showed a trend (without→without < with→with: z=−2.485, p=0.078). Post hoc: wild→wild < captive→captive (z=−2.904, p=0.022); trend wild→captive < captive→captive (z=−2.501, p=0.062). Males→males < males→females (z=−2.630, p=0.034), < females→females (z=−3.783, p<0.001), and < females→males (z=−3.092, p=0.010). - Passive close proximity: significant sex effect (F=6.527, p<0.001); PHCinfant and origin not significant. Females→females highest; females→males lower (z=−4.003, p<0.001); males→males (z=−4.142, p<0.001) and males→females (z=−3.656, p=0.001) lower than females→females. - Stationary vicinity and aggregated sum models: full models did not improve over null models. Overall, multiplex analysis revealed nonredundant information across layers and biographical effects particularly on medium- and high-level contact behaviours (affiliative and grooming), with sex consistently influencing interaction rates (female–female dyads highest).

The findings support the primary hypothesis that multiplex analysis increases information gain over single-layer and aggregated approaches, even in small groups. Interlayer analyses showed that stationary vicinity and passive close proximity were most similar, contrary to the initial expectation that high-contact layers (grooming and close proximity) would align most strongly. Reducibility demonstrated that each layer contributes unique information; aggregating layers leads to information loss and can obscure biographical effects. Individual-level results indicated that sociability is multidimensional: individuals can be central in some interaction types but peripheral in others, so single-behaviour networks can mischaracterize an individual’s social role. The LMMs confirmed that early life experiences shape sociability: affiliative behaviour was highest when both partners had social rearing (with conspecifics), and allogrooming was most frequent among captive-born dyads and least among wild-caught dyads. These patterns suggest social learning and comfort with tactile contact depend on early social environments. Sex differences were robust across behaviours, with female–female dyads most affiliative and most in close proximity, and male–male dyads least grooming and least proximate, consistent with captive conditions diminishing male competition-related structuring and revealing female social potential. Passive close proximity was not significantly affected by origin or infancy housing, indicating that tolerance of adjacency may recover or be less sensitive to early adversity than tactile interactions. For captive management, multiplex insights can guide group composition by considering individuals’ early histories and the specific interaction types they are likely to express, enhancing welfare by fostering compatible, socially functional groups.

Multiplex social network analysis provided a realistic, information-rich assessment of sociability in two small groups of former pet/entertainment chimpanzees by simultaneously modeling stationary vicinity, affiliative behaviour, allogrooming, and passive close proximity. Layers were partly similar but nonredundant; reducibility analyses showed that aggregating layers sacrifices information. Early life factors (origin and social housing during infancy) and sex influenced medium- and high-contact behaviours, while proximity tolerance was less affected by biography. These results underscore the value of multilayer approaches for diagnosing social dynamics and informing management decisions in captivity. Future work should apply multiplex analyses to larger groups, incorporate additional layers (e.g., agonistic interactions), consider personality and finer-grained biography (maternal vs hand-rearing), and extend to longitudinal datasets to track developmental and rehabilitative changes over time.

- Small group sizes (n=7 per group) limit the use of certain network tools (e.g., community detection, triadic analyses) and reduce generalizability. - Behavioural layers were limited to frequently occurring interactions; rarer behaviours (e.g., aggression) were not included, potentially omitting relevant dimensions. - Aggregated vs multiplex distinctions become more informative in larger, sparser networks; in dense small networks, some measures may be less sensitive. - Biographical categories were coarse (origin and predominant infancy housing); small samples precluded stratifying layers by finer rearing histories or personalities. - Observations occurred during outdoor access periods; indoor behaviours might be underrepresented. - Cross-sectional observation window (May 2018–January 2019) limits inference about temporal dynamics or long-term changes.