Open Golden-Retriever-1 opened 4 years ago
@vandenman
As a reminder, can you please look at this issue and please respond?
Cheers Frans
A key difference between Wasserman & Faust (1994) and the networks currently implemented in the network analysis module is whether the edges are stochastic. In Wasserman & Faust (1994), edges are observed. In the network analysis module, edges are estimated from the data. The question is whether the statistics for observed networks also make sense for stochastic ones. Currently, there is evidence to the contrary (e.g., see Dablander & Hinne, 2019).
Nevertheless, it would be nice to augment the network analysis module with the network analyses where the edges are observed. I'll put this on the to-do list, but I'm afraid I cannot give a time estimate.
References:
Dablander, F., & Hinne, M. (2019). Node centrality measures are a poor substitute for causal inference. Scientific reports, 9(1), 1-13.
I appreciate your efforts. Your claim above does not match and cover all applications and heavily depends on application! No generalizations please. It would be a pleasure adding this to the manual/documentation of JASP to make the users aware of it. Some other software tools for network analysis do not even specify the details of calculations.
I thank you for adding the topic to the to-do list. I am looking forward to its implemenation.
Regarding Dablander & Hinne (2019, p. 8): I keep the descriptive nature of SNA to describe one setting of occurring interactions in condition A and observe interactions of a network of the same agents in another condition B. The causal influence on the network comes from the conditions of A and B on those agents while the descriptive nature of SNA is kept as a whole. Hence, I compare the measures of centrality between both different conditions and not(!) within one network. Hence, the causality is only available in terms of having two conditions that lead to two different networks that are described by SNA. Finally, the reference is limited in its relevance and depending on context. Of course, a smaller sample of agents leaves more possibilities for causal interdependencies between them also when influenced by external factors than having more agents in one network which additionally can be influenced by external factors (which are in both not included into the network model). Hence, the calculations of Dablander and Hinne (2019) are just a gimmick representing the nature of the specific networks and the conclusions are very narrow. The conclusions are not generalizable to all applications which is why I wonder! The authors make their claims regarding "directed acyclic graphs (DAGs)" within one network! Generalizing is beating around the bush for high attention since rationality tells us that in a bigger network causality is more difficult to find since (also not-included factors) can per se influence the mathematical description of a network. Smaller networks have less points of attack for causal influences, no matter if included into a model or not. Dablander and Hinne (2019) map this matter of fact. The conclusions drawn are not generalizable and depend heavily on context of use and application as well as on model specification. Of course, you are more likely to fail when you choose for causality in centrality measures in a big network than in a smaller network. The causal factors of unknown complex influence that are varied and/or influence agents in their behaviour are not included into the model of the social network and the social network only includes the observed dependent variables. Of course, correlations drop in some centrality measures when you increase the network size of observed settings!
Better leave the specific context to the user.
The alternative solution would be using another software for calculating prestige and missing centrality measures. Hence, the feature request relates to improving calculating network analysis.