4 92 CHAPTER 4 4.2.3.3 Identification communities of dynamic risk factors The spinglass algorithm was used to identify clusters of dynamic risk factors in the networks (Reichardt & Bornholdt, 2006; Yang et al., 2016). In network analyses, these clusters are referred to as communities. The spinglass algorithm is based on the principle that edges of nodes of the same community should connect while edges of nodes belonging to different communities should not. We ran this process 1,000 times and extracted the number of communities with the highest frequency. Following Briganti et al. (2018), we also applied the walktrap algorithm (Demetriou et al., 2017; Golino & Epskamp, 2017). This algorithm is based on the principle that adjacent nodes tend to belong to the same community (Yang et al., 2016). 4.2.3.4 Statistical network comparison To statistically compare the networks from the two independent data sets on invariance of global strength and network structure, we used the network comparison test (NCT; van Borkulo et al., 2019). Global strength concerns the overall connectivity across the network, calculated by summing the absolute values of all edges. Network structure concerns the structure as a whole, defined by the maximum deviation between two edges in a network. NCT is a permutation test in which the difference between networks of two groups is calculated repeatedly for randomly regrouped individuals in three stages. First, the NCT estimates the networks of both groups. Second, different networks are estimated based on randomly regrouped cases from these different networks. Third, the test statistics of global strength and network structure are calculated which generates a reference distribution and a p value (null hypothesis is that network structures of both samples are identical on the statistics of interest; α = .05 is the threshold to be significant). Regarding the NCT, power analyses were performed by van Borkulo et al. (2022). They concluded that whether the power of the NCT is high enough (> 0.8) depends strongly on a combination of the sample sizes, density of the networks, and number of variables included in the networks. When sample sizes are unequal, the smallest sample size should be used to determine the power. In their simulation study, the power to detect invariance of network structure in dense networks with 10 variables and a sample size of 500 versus 750 is, respectively, 0.787 and 0.915. While the power in a dense network with the same sample sizes consisting of 20 variables is, respectively, 0.160 and 0.262. However, the accuracy of the approximation increases with the number of iterations (Ernst, 2004; van Borkulo et al., 2022). Considering our networks constructed with the lowest sample sizes (DSP; n = 611 and 14 variables; n = 788 and 13 variables) are dense, we assume that the power to detect invariance of network structure is close to 0.8. To ensure that we have sufficient power, we performed the analyses with 10,000 iterations instead of 1,000 which is the default in NCT. When investigating the invariance of global strength, van Borkulo et al. (2022) conclude that the power is acceptable under all circumstances in their simulations with networks with 20 nodes or more. With
RkJQdWJsaXNoZXIy MjY0ODMw