## Similar protocols

## Protocol publication

[…] We have applied exploratory social network analysis [, ] in combination with Exponential Random Graph Models (ERGMs). ERGMs belong to the class of statistical inference models and are among the most popular and theoretically well-developed class of network models []. ERGMs are used for testing hypotheses about the social processes that might have led to the creation and development of an empirically observed network. The statistics in these models are based on the occurrence of certain micro-level patterns of ties that indicate specific mechanisms of tie formation at work. Examples are preferential attachment (to popular nodes), reciprocity between nodes (resulting in the formation of a double arrow), transitivity (friends of friends are likely to become friends) resulting in a local triangle structure and processes of homophily in which two nodes with the same trait are more likely to form a tie. ERGMs are used to test statistically whether the relative occurrence of such patterns is consistent with these underlying dynamic processes of network formation. For a more detailed introduction into ERGMs see Lusher et al., Harris, and Lubell et al. [–]. The analysis of network properties and **ERGM** specification was done in R, using the statistical ‘**statnet**’ package (version 2016.9) [], and the associated ‘ergm.ego’ package (version 0.3.0) []. See for an overview of the used analysis code. The ‘ergm.ego’ package was developed especially for ego-networks. In such ego-networks the collected data are considered to be a sample of a larger network of a known, or unknown size. In our case we did not know the total size of the population network of the AIS in all the three countries. In addition, the membership of the MSP is not fixed: it changes over time and the sample is therefore necessarily only capturing a snapshot picture of the ego-networks of MSP participants of what–in reality–is a dynamic process of collaboration and partnering. Nevertheless, the sample represents a reliable picture of the typical ego-networks at the national levels in Burundi and Rwandan and provincial level for DRC and as such can be used as input for ego network modelling.The ergm.ego package is based on the finding by Krivitsky et al. [] that it is possible to obtain a “per capita” size invariant parameterization for dyad-independent statistics by using an offset that preserves the mean degree (approximately equal to −log(n), where n is the number of nodes in the network). Simulations have suggested this is also possible for some dyad-dependent statistics. However, the processes of so-called ‘network self-organisation’ at the level of the entire network (like triadic closure, degree assortativity and 4-cycles) are not incorporated in the ergm.ego package. In their description of the package, Krivitsky and Morris [] state that if the population network is not overly large the parametrization of such higher order effects might not be necessary.Terms were added in consecutive blocks (node level and dyad level) to examine their relative contribution to enhancing the goodness-of-fit of the models []. Three models were tested and evaluated: starting with a simple random graph model (M0) (where all nodes have an equal chance to form a tie), and adding complexity in subsequent models by adding terms corresponding to our hypotheses at the node level (M1) and the dyad level (M2). We have scaled all our results to a”pseudo-population” size of 1,000 for all three countries, following the advice of Krivitsky and Morris [].At the node level we look at the degree (amount of ties) that organisations have within the knowledge network to test hypothesis 3. The knowledge degree serves therefore as an indication for the perception of other actors that an organisation possesses complementary knowledge. Following the AIS perspective we assume that such relevant knowledge is not limited to research and extension organisations, but is also possessed by farmers, NGOs, businesses, etc. To operationalise hypothesis 2 we take the indegree of organisations in the influence network as a measure of their perceived power. Again: not only government organisations are deemed to be powerful, but other types of organisations can also possess other forms of power []. At the dyad level we look at ties between different types of organisations. A typology was made in 6 different categories of actors: 1) business, 2) farmer, 3) government, 4) non-governmental organisations (NGOs), 5) research, training and extension, and 6) unknown. Hypothesis 1 is thus tested by looking at the tendency for different types of organisations to form collaborative ties. Finally, for scaling we look at the administrative level where organisations are (most) active: 1) local, 2) provincial, 3) national, 4 supranational, or 5) unknown. Hypothesis 4 is tested by investigating the tendency of actors working at different levels to form collaborative ties.Models were checked for potential degeneracy (see ) and goodness-of fit through visual inspection of the standard plots that the statnet package generates for this purpose, as suggested by Hunter et al. []. Since all the models underestimated the number of organisations with a degree of 1, we fixed this amount in the models to increase the fit. To ease the comparison of the plots, we have calculated a goodness-of-fit percentage following the example of Harris et al. []. The calculated percentage is based on the proportion of the relevant degree distribution that fall within the 95% confidence intervals of simulations based on the models. The term relevant here is not defined for all degrees, but only those degrees where either the results of the ergm.ego model, or the original measurement show a value unequal to 0. […]

## Pipeline specifications

Software tools | ergm, Statnet |
---|---|

Application | Protein interaction analysis |