Dynamic latent variable models and network models

• Applicant: Professor Dr. Augustin Kelava
• Subject Area: Personality Psychology, Clinical and Medical Psychology, Methodology
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 528244660

The collection and processing of so-called itensive longitudinal data (ILD) has increased substantially in psychology (and other social and behavioral sciences) in recent years (Trull & Ebner-Priemer, 2014). Dynamic latent variable models have been applied to ILD (e.g., dynamic structural equation models, DSEM, Asparouhov, Hamaker, & Muthén, 2018), which combine autoregressive time series models with confirmatory latent measurement models and multilevel modeling (for the purpose of accounting for inter-individual differences in trajectories). Although dynamic latent variable models are a rather general class of procedures, they have certain limitations with respect to the Bayesian estimation procedures applied and their confirmatory nature of the measurement models. A more explorative alternative to the analysis of multivariate relationships of observed variables is offered by so-called network models (e.g., Borsboom, 2008). These methods allow an unstructured examination of the relationship of variables for the purpose of detecting significant activation patterns. They have recently been further developed for ILD and inter-individual differences (cf. Epskamp, 2020), but have few discussed similarities and differences to dynamic latent variable models mentioned above in the literature. E.g., in the group of network models, the representation of heterogeneity (e.g., via latent classes) is not as flexible as in dynamic latent variable models. In this research project, four goals shall be addressed: 1. equivalent submodels from both classes of methods shall be described and a systematic, taxonomic discussion of the models shall take place. This will be done against the background of the identification of relevant model gaps and the transferability of estimation procedures. 2) Reanalyzes of multilevel ILDs from a study on dropout in mathematics will be conducted using network models that have already been investigated with dynamic latent variable models in order to identify new patterns of correlation in the high-dimensional data. 3) Taking into account the taxonomic discussion, we aim to transfer and develop frequentist estimators for the submodels of dynamic latent variable models (e.g., FIML and multi-step approaches), for which Bayesian estimation has dominated so far. 4. building on the publication of a general Forward Filtering Backward Sampling (FFBS) algorithm for dynamic latent variable models (Kelava, Kilian, Glaesser, Merk, & Brandt, in press), a forecasting algorithm for network models will be developed so that patterns of variables can be predicted (e.g., cognitions associated with dropout or clinical symptoms). The project will involve the external expertise of Prof. Stephen Paul Reise (UCLA).

• DFG Programme: Research Grants
• International Connection: USA
• Cooperation Partner: Professor Steven Paul Reise, Ph.D.