Study system
San Jose, Lambayeque, Peru (6°46' S, 79°58' W) is home to 168 small-scale commercial gillnet skippers that fish throughout the year. We surveyed 165 fishers representing 98.2% of the gillnet skippers at the site between July–September 2017 (Fig. S2b, Table S1). Gillnet skippers in San Jose are known to capture sea turtles in high numbers [30, 31, 36]. Green turtles (Chelonia mydas) are captured most frequently, followed by olive ridley turtles (Lepidochelys olivacea), and leatherback turtles (Dermochelys coriacea) [33]. Five gillnet skippers and their crew are currently involved in a trial community co-management bycatch reduction scheme operating from San Jose that requires fishers to use light-emitting diodes on their nets to reduce sea turtle bycatch [34]. Skippers were deemed active if they fished from the San Jose port with gillnets in the winter of 1 July – 30 September 2017. The network was surveyed during winter as skippers actively fishing during these months are established fishers in the San Jose community throughout the year. We define gillnets as encompassing surface drift gillnets and fixed bottom gillnets in single or trammel net configurations. The total San Jose gillnet skipper population (n=168) was determined using a combination of membership lists of the two main fishing groups in San Jose, lists of boats towed in and out of the water with tractors, and key informant interviews (Supplementary Information).
Data collection
Detailed social network data was collected using a structured questionnaire with a fixed choice survey design. Respondents were asked to consider up to ten individuals with whom they exchange useful information about fishing and whom they considered valuable to their fishing success. We classified nine fine-scale information-sharing types about which we expect gillnet skippers to exchange fishing related information (Table 1). As each nominee was given by the respondents, they were asked to highlight which fishing-related information they discussed with each nominee. For each fishing-related information network, respondents were asked to consider relationships that they have had with other skippers, vessel owners, crew members, other fishery leaders, fishery management officials, members of the scientific community, boat launching/landing support, fish sellers/market operators, family members, and any other stakeholders they fished or shared information with about fishing. Respondents were not asked who they receive information from. Interviews were undertaken verbally and respondents were not shown the questionnaire where responses were written (Supplementary Information). Questionnaires were trialed with fishers (n=8) in the Santa Rosa fishing community 17 km down the coast from San Jose (Fig. 2a). Pilot study data were not included in this study’s analysis. Fishers were interviewed in their native language (Spanish). Documented, free, prior, and informed consent was sought before respondents could take part in the study. This research has Research Ethics Approval (CUREC 1A; Ref No: R52516/RE001 and R52516/RE002).
Statistics and Reproducibility
Social network construction
A social network was created for each fishing-related information type (Table 1). In each network, the nodes were the individuals, and the binary directed edges were the nominations by one node (sender) of another node (receiver) for this information type. All analysis was carried out in R [37] using the igraph package [38] for visualising and processing the analysis and carrying out the network comparisons using the null models.
Structural differences across information-sharing networks
To investigate whether networks of information-sharing between individuals were similar across different information types, we examined the networks' structural properties in terms of their degree assortativity and the variance and mean of individual centrality (Table 2). To account for the effect of basic characteristics of the networks (e.g., number of links, degree distributions) we compared these observed summary statistics to null models, which allowed inference of structural differences and similarities over and above that expected from these simple differences using null models (Fig. 1). While the null model methods applied in the current study were developed in ecology, they are beginning to be used in human network analysis. For example, in the fields of epidemiology for assessing human contact tracing disease control measures [39]. In the supplementary information we explain additional analysis detail and include discussion of the reason for the null models applied.
Degree assortativity
The degree assortativity (or homophily) coefficient [40] measures the extent to which central nodes are connected to other central nodes, and peripheral nodes are connected to other peripheral nodes based on a particular trait. The level of degree assortativity [40, 41] in a network is known to have important social implications for the operation and emergence of competition and cooperation (e.g., fishers will work with others like them). Positive values demonstrate degree assortativity, with perfectly homophilous networks scoring 1, and negative values representing disassortment. When nodes of similar centrality are randomly distributed in a network (i.e., fully disassorted), those networks do not always score -1 due to the minimum value depending on the number of node types and the relative number of links within each group [40]. For each of the information-sharing networks, we first calculated the assortativity by in-degree (the number of nominations each interviewed skipper received). Degree assortativity measures the extent to which ‘individuals that are highly nominated are disproportionately connected to others that are highly nominated’ and ‘individuals that are rarely nominated are disproportionately connected to others that are rarely nominated’. This is the primary assortativity measure of interest as in-degree provides the measure of which individuals provide information to others. However, as individuals differed in the number of nominations they made within each information-sharing network, we also calculated the assortativity by out-degree (the number of nominations each interviewed skipper made) to examine whether individuals were also disproportionately connected to others who make a similar number of nominations as themselves. As social networks often show assortativity by degree, we predicted that all the information sharing networks would be positively homophilous by nominations made and nominations received (i.e., highly nominating and nominated individuals would be closely associated with highly nominating and nominated individuals, whilst peripheral individuals would be more likely to be connected).
Eccentricity
We aimed to consider node-level properties that depend on the structure of the social network (Table 2). For this purpose, we used node eccentricity (igraph package [38]) that measures how far a node is from the furthest other [42]. Although this metric describes a node’s position within the wider network, the range of potential values it can take is not overly affected by permutations of the network structure in comparison to other more vulnerable metrics (e.g., betweenness, clustering coefficient) which are innately dependent on multiple aspects of the set structure of the network and are intuitively expected to differ largely from permutations by default. Finally, this metric is also relatively fast to compute; this is particularly useful when calculating it for many iterations of null networks. As such, we computed the variation in eccentricity in ‘received nominations’ (in-eccentricity) for each of the information sharing networks.
Null models for structural differences
Drawing comparisons of network structure, correlations, and node positions across different networks requires particular consideration because the general structure of the network (such as the number of links or degree distributions) has a large effect on the observed values obtained from standard summary statistics. This structure can be taken into consideration by comparing networks to null permutations (controlled randomisations) of themselves and recalculating the same summary statistics on the null networks. Through comparing the observed values of the summary statistics to the distribution of those statistics generated from the null networks, insight can be gained into the actual differences between observed networks across other networks, over and above what is expected from simple properties such as the number of links.
When calculating summary statistics (in-/out-degree assortativity, eccentricity) of each of the information-sharing networks, we also compared these to the values generated from permuting each of the networks separately. Specifically, we carried out edge permutations. The first edge permutation simply allowed the randomisation of all in-going links, while maintaining the number of nominations (out-going links) each individual made within this information-sharing network (termed edge null model 1 - Fig. 1a). The second edge permutation was a more conservative version of this, allowing swaps of links (which individuals nominated which other individuals in this information-sharing network) but maintaining the number of nominations each individual made in this information-sharing network (termed edge null model 2- Fig. 1b). Separately, for each of the information-sharing networks, 1000 permuted networks (of both of these permutation types) were generated and the distribution of the summary statistics were calculated for them.
Cross-network correlations
To reveal the extent to which the sea turtle bycatch information-sharing networks can be predicted from the other networks evaluated, we examined the dyadic similarity between the different information-sharing networks. We used cross-network null models to compare the expected correlation between each network and subsequently determined how the observed correlation between each network was driven by fine-scale structure over-and-above that expected from the system's general social structure. To examine the relationship between each network of dyadic information-sharing nominations, we calculated the correlation between the dyadic nominations on the unfolded network matrices. This approach is somewhat analogous to the Mantel test [43] (that tests the correlation between two matrices), yet as the networks were directed (and non-symmetrical), this was applied to the entire matrix rather than the lower triangle part (but excluding the diagonals because ‘self-nominations’ were not possible). The calculated correlation statistic represented the similarity/dissimilarity in the directed dyadic nominations amongst networks (who nominates whom), and these were compared to the distribution of the correlation statistic generated from the null models. To infer the extent to which networks are more, or less, similar than expected under the general dyadic social structure, we carried out a cross-network null model: For each dyadic nomination across any of the networks, we randomised the networks that these nominations were made within (termed ‘cross-network null model 1’ – Fig. 1c). As an even more conservative version of a cross-network null model, we created a new version of these permutations and controlled for the number of nominations that took place overall within each network (termed cross-network null model 2 – Fig. 1d; Fig. S7, S8).