3 70 CHAPTER 3 3.5.3 NETWORK ANALYSIS Because the data set included both count and categorical data, we used mixed graphical models (mgm; Haslbeck & Waldorp, 2016) to estimate network structures. Three networks were constructed using regularized network estimation as described by van Borkulo and colleagues (2014). Regularization is used to control the Type I error rate; this technique has been shown to result in networks with high specificity and adequate sensitivity (van Borkulo et al., 2014). This means that connections present in the sample network are likely to be present in the population as well, whereas connections that are absent may either be absent in the population network or too weak to be picked up by the regularization technique. To control for spurious connections that may result from sampling error and to estimate sparser and more interpretable network models, we used extended Bayesian information criteria for L1 penalized regularization, which has become the standard in the network literature (Costantini et al., 2015; Epskamp & Fried, 2017). To compare the networks visually, we set the maximum value (1.00) and cutoff (0.08) for possible connections at the same level for all three networks. Furthermore, we used the layout of the network with sexual recidivism as a template for the network containing violent (including sexual contact) recidivism. In the first network, each node represents a dynamic risk factor as measured by the STABLE-2007 (Fernandez et al., 2012; n = 788). The nodes in the second and third networks were complemented with data on sexual and violent (including sexual contact) recidivism (n = 611). In all networks, green edges indicate positive statistical associations. The stronger an association, the more saturated and wider the edge. 3.5.3.1 Node centrality metrics In a weighted network, a node high in strength centrality has a relatively high number of edges with high magnitudes (Opsahl et al., 2010). In this study, we focused on strength centrality because it reflects the likelihood that activation of a dynamic risk factor will be followed by activation of other dynamic risk factors (McNally, 2016), which may ultimately result in recidivism. 3.5.3.2 Identification clusters of dynamic risk factors The spinglass algorithm, which is based on the principle that edges of nodes of the same community should connect and that edges of nodes belonging to different communities should not, was used to identify clusters of items in the networks (Reichardt & Bornholdt, 2006; Yang et al., 2016). With this procedure, each dynamic risk factor can be part of only one community. We simulated this process 1,000 times and extracted the number of communities with the highest frequency. In addition, following Briganti, Kempenaers, Braun, Fried and Linkowski (2018), we applied the walktrap algorithm (Demetriou et al., 2017; Golino & Epskamp, 2017), which is based on the principle that adjacent nodes tend to belong to the same community (Yang et al., 2016).
RkJQdWJsaXNoZXIy MjY0ODMw