Study site and population 

Kruger National Park (KNP) spans nearly 19,485km² (22.5° - 25.5°S, 31.0° - 31.57°E) and hosts a diversity of wildlife, including wild African buffalo. Our study population included a wild buffalo herd, which was captured in Northern KNP during the early 2000's and relocated to a 900-hectare enclosure in the center of the park, near Satara camp (Fig. S1). The buffalo sample size varied during our study (N = 60–70) due to births and deaths. On average, 25.60 ± 0.10 % of the herd were calves, 22.95 ± 0.09 % were juveniles and 51.45 ± 0.19 % were adults. Buffalo were free to graze and breed in the enclosure, and during extreme droughts, they had access to supplemental grass hay. Water was available to buffalo at a natural pan and a manmade water point (Fig. S1). This “nearly natural” enclosure included numerous other species typical of the ecosystem (e.g., giraffe, zebra, warthogs, and small predators), while excluding megaherbivores (e.g., rhinos, elephants) and large predators (e.g., lions, leopards). 

Buffalo captures and sedation procedures 

Data collection spanned six observation periods (OPs) that occurred during March 2014 through May 2015 (See Rushmore et al 2023, Table S1). We captured buffalo to collect biological data and download association data from proximity-logging collars at the end of each OP. Captures occurred five times per year, at two to three-month intervals. We performed three active captures, in which buffalo were darted from a helicopter, and four passive captures during the dry season in which researchers filled man-made water troughs that attracted buffalo into a fenced area with a remote-controlled gate closure (See Rushmore et al 2023, Fig. S1). At each capture, we darted small groups of buffalo using chemical immobilization procedures described by Couch et al. (2017).  

Data collection: behavioral association data   

In February 2014, we fitted the majority of buffalo aged over 6 months with Sirtrack proximity-logging collars (Sirtrack Tracking Solutions, Havelock North, New Zealand), which record the identity of collars in close proximity in addition to the date, time, and duration of each encounter. Percentage coverage across the herd is provided in Table 1 of Supplemental information. Calves < 6 months in age were not collared for ethical considerations, as their growth rate exceeded collar re-fitting schedules. We programmed collars with a UHF range coefficient of 20 and a separation time of 240s, which in a laboratory setting initiated an association when collars were within 1.22m ± 0.46m (mean ± SD), and terminated the association when collars exceeded a distance of 1.70m ± 0.67m for more than 240s. We deemed this a reasonable representation of transmission distances for pathogens spread via close contact (e.g., respiratory viruses and bacteria) (Wells 1934, Olsen et al. 2003). While proximity collars allowed for near-complete data collection without observation bias, we note that close contacts were inferred from data on physical proximity (Farine 2015, Farine and Whitehead 2015) rather than directly-observed interactions.  

Estimating association indices and social networks  

We analyzed association data for 69 buffalo (males = 22, females = 47), including 18 calves ( < 1.5y). We created a matrix of pairwise association indices for each of the six observation periods (OPs). The number of buffalo varied across OPs due to births, deaths, and collar malfunctions. Overall, matrices included an average of 39.33 buffalo (± SD: 14.11) and ranged from 17–53 buffalo (Table S1).  

When cleaning association data, we excluded capture days and a two-day buffer after each intervention, and we removed 1s encounters shown to skew results (Drewe et al. 2012). To reduce asymmetries in association matrices, we excluded encounters logged after one individual in a pair had a full proximity logger memory. For each pair of individuals i and j, we calculated an association index (AI) for a given OP as follows:  

When recording association data, ideally both collars in a pair would record identical information; however, previous studies demonstrate that collars vary in their abilities to transfer information (Drewe et al. 2012; Boyland et al. 2013). We reduced inter-collar variation biases, specifically conflation of individual ID and error/strength of proximity collar, using a method similar to that proposed by Boyland et al. (2013). In brief, we assessed variation in reciprocal AIs in a given matrix to evaluate each collar’s relative performance and to develop a measure of collar bias. We then corrected matrix AIs by scaling each collar’s data according to its average bias across collars, resulting in a nearly identical pre- and post-correction average AI for the matrix. Further details about collar corrections are provided in the Supplementary Information for the associated manuscript, Rushmore et al. 2023.