Centrality is a function of the adjacency matrix. I’m not aware of combining—but you could e.g. plot, per node, in a histogram, edge coefficients (and then would see that node A has high centrality because it has many moderate connections, while node B has high centrality because it has one large and no other connections). There should also be some sort of coefficient to summarize that between 0 and 1 (0: one very large and no other, 1: all neighbor connections equal) or so.

]]>Glad to hear it was useful Marcin!

]]>The question I have however is – does this method work on networks using binary variables or mixed (binary and ordinal) variables?

All the best!

Marcin

]]>Thanks a lot for sharing this!

]]>There are forks of NCT on Github (I believe by Sacha Epskamp, also Payton Jones I think) that allow you to install NCT and then use e.g. polychoric or Spearman, I believe. So just google NCT Github and these names and you should find it. Maybe there are also easier ways to do that using NCT itself but I wouldn’t know—worth asking the maintainer (Claudia van Borkulo). Or just post the question in our “psychological dynamics” facebook group where the authors answer questions like that.

]]>Appreciate your quick response! I can imagine that the change of correlation matrices can affect the NCT results accordingly, But I didn’t observe any arguments involved in correlation methods in NCT, listed as follows:

NCT (data1, data2, gamma, it = 100, binary.data=FALSE, paired=FALSE, weighted=TRUE, AND=TRUE, abs=TRUE, test.edges=FALSE, edges=”all”, progressbar=TRUE, make.positive.definite=TRUE, p.adjust.methods= c(“none”,”holm”,”hochberg”,”hommel”,

“bonferroni”,”BH”,”BY”,”fdr”), test.centrality=FALSE, centrality=c(“strength”,”expectedInfluence”),nodes=”all”, communities=NULL,useCommunities=”all”, estimator, estimatorArgs = list(), verbose = TRUE)

Do you have any idea about it?

“In that paper, the NCT was only validated for Pearson correlations. ”

I think you are right, I read that paper. I estimated networks using Spearman correlation and have no idea how to do the network comparisons. But I observed someone else do this using NCT, the author of the following paper estimated networks using partial Spearman correlations and did the NCT, which confused me further.

https://journals.sagepub.com/doi/abs/10.1177/10731911211050921

It’s been a while since I looked into the NCT literature. The original NCT paper by Claudia van Borkulo was available only as preprint for the longest time, so I’d check if that’s published. In that paper, the NCT was only validated for Pearson correlations. That means the sim studies and thresholds are for networks of normally distributed variables where the variance/covariance matrix which networks are constructed upon is estimated via Pearson correlations.

Anything else is easily possible to do, by exchanging the correlation function in NCT, but it’s important to note this isn’t validated.

“I am wondering whether the methods or correlations used in network estimating will change the results of NCT?”

That depends. If your data are normal, it probably doesn’t matter if you use Pearson or Spearman or polychorics (the latter only with sufficient observations, otherwise you can run into issues). If data are heavily skewed or something similar, Pearson is not appropriate, and changing to e.g. Spearman will change your correlation matrices, and, correspondingly, your networks, which can affect the NCT results of course.

]]>Thank you for sharing your excellent work! As a newbie, I have learned a lot from your blog. It inspired me a lot!

I am working on a network comparison using Network Comparison Test and I am confused at one point after reading this blog. In the “Complications” part, you mentioned that if polychoric correlations were used instead of CorMethod =” cor” when estimating networks, the connectivity for the networks will change to 8.30 vs. 8.38. But you didn’t suggest how to compare these two new networks in this situation. I observed that only original data sets were used in Network Comparison Test, i.e. NCT(data1, data2, it=1000, binary.data=FALSE).

I am wondering whether the methods or correlations used in network estimating will change the results of NCT? For example, whether there are differences in NCT if the networks were estimated using CorMethod=” cor” or CorMethod=” spearman”? Any advice would be appreciated!

Would it be possible, using these procedures, to group participants based on communities of items (i.e. something similar to what is done in latent class analysis or cluster analysis).

Cheers

]]>