Skip to main content

Application of Model Selection Principles to Distinguish Network and Latent Variable Models of Psychological Constructs

Final Report Summary - MODEL SELECTION (Application of Model Selection Principles to Distinguish Network and Latent Variable Models of Psychological Constructs)

A recent framework for modeling psychological constructs is to represent them as networks: webs of interacting components (e.g. symptoms or items) that affect each other directly. This approach is radically different from the dominant modeling framework within psychology, which is to represent constructs as latent, unobservable constructs which are responsible for the (co-)variation in observed items or symptoms. As the network modeling approach becomes more common, it is increasingly relevant to ask whether and how these two frameworks can be distinguished for a given psychological construct. Supported by the Marie Curie Career Integration Grant, “MODEL SELECTION”, this project has made substantial progress in developing and studying such methods.
The proposal was structured in terms of 3 main objectives. The first objective is to explore the theoretical implications of network vs. latent variable models that could be used to discriminate between them. Two empirical papers examined what network models imply and reveal in the context of two very different psychology constructs. First, Kossakowski et al. (2016) applied network analyses to the construct quality of life and found that this approach unveils important differences in the network structure of quality life in cancer patients vs. healthy adults. Second, Rhemtulla et al. (2016) applied network analysis to data on substance abuse and found interesting similarities and differences in the networks of interactions among substance abuse symptoms across substances. Third, we are currently writing a theoretical paper that aims to give an overview of the conceptual and theoretical stances that would support a network model or a latent variable model for any psychological variable. An extension of Objective 1 was to develop a hybrid framework that includes latent variables within a network structure; this extension has been worked out and submitted for publication (Epskamp et al., 2016).
The second objective is to develop and study empirical data-based tests for discovering whether the true underlying model is a network or a latent variable structure. This objective has been the focus of research for Riet van Bork, who is a PhD student on this project. Van Bork has developed two such tests and studied their performance via Monte Carlo simulations. This paper is close to ready for submission. Also toward this objective, I have done initial research on a second avenue within this research objective, which is to study whether existing latent variable model fit statistics can identify when the true population structure is a network.
The third objective is to use the relations between elements of a construct and external criteria (predictors and outcomes) to determine which framework is most appropriate. Toward this objective, we are conducting a series of simluations to explore what happens when data that are generated by a network model are modeled as a latent variable. In particular, this research investigates the consequences of interpreting predictive and explanatory effects of a construct in relation to other constructs when that construct is inappropriately modeled.
In sum, we are making progress toward all 3 objectives of this project. The project has been cut short, however, due to the researcher’s decision to leave the EU to move her career to the United States. As such, the final impact of this research is yet to be known.