Community Detection and Unsupervised Model Construction in Continuous-Time
Restricted (Penn State Only)
- Author:
- Park, Jonathan
- Graduate Program:
- Human Development and Family Studies
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- May 22, 2024
- Committee Members:
- Douglas Teti, Program Head/Chair
Chad Shenk, Major Field Member
Sy-Miin Chow, Chair & Dissertation Advisor
Michael Hunter, Major Field Member
Michael Russell, Outside Unit & Field Member
Peter Cm Molenaar, Special Member
Zachary Fisher, Major Field Member - Keywords:
- dynamic network analysis
dynamic modeling
continuous-time
discrete-time
community detection
state-space modeling
time-series analysis
network modeling
network analysis
depression
anxiety - Abstract:
- Quantitative methods in the social and behavioral sciences is characterized by constant change. Fluctuations in what is deemed appropriate, popular, or “best practices” will continue to shift as we gain a better understanding of the world and gain access to data of higher quality. As with the field of methods, individuals likewise, are also in a state of constant flux; that is, they are always changing in some way. The past several decades of work in quantitative methods has focused strongly on the notion that individuals change heterogeneously through time. This increased focus has drawn attempts to reconcile the group with the individual by pooling, averaging, constraining, or borrowing information from individuals to inform the group- or subgroup-level structures that may be present in the data. Ultimately with the sole purpose of identifying group-level commonalities from person-specific results. This dissertation unfolds in three parts. The first chapter provides a discussion on how we think about processes compared to how we actually go about modeling them. Contextually, we view many relations between psychological, behavioral, and emotional variables as evolving continuously through time; however, the manner in which these data are analyzed are typically through the lens of discrete-time models. Chapter 2 investigates the consequences of modeling continuous processes with subgrouping methods based in the discrete-time vector autoregression (VAR) framework by means of a Monte Carlo simulation study. The results of these simulations highlighted the nature of how continuous processes evolve in time with particular emphasis in where information in the drift matrices manifests in the discrete-time VAR model given larger disparities in the sampling rate of the data. We find that subgroup membership can be reliably recovered even in the presence of very slow sampling rates if the VAR-based subgrouping algorithms leverage information from the process noises to model–and subsequently cluster on–contemporaneous information. The third chapter of this dissertation proposed a novel extension to the popular, discrete-time method: the group iterative multiple model estimation (GIMME; Gates & Molenaar, 2012) procedure to the continuous-time framework. Dubbed ct-gimme, it provides researchers with an automated model fitting procedure based on modification indices (MIs). Broadly, continuous-time models exhibit myriad benefits over discrete-time models including the relaxation of several key assumptions in the discrete-time framework (e.g., equally spaced data, sensitivity to differing time-scales; Park et al., 2022). These– among other–benefits make modeling in continuous-time an attractive option for many researchers. However, the application of continuous-time models still lags significantly behind those in discrete-time. Through ct-gimme, I sought to make continuous-time modeling more accessible and grant researchers access to a powerful tool for identifying and modeling within-sample heterogeneity in dynamic processes. I design a Monte Carlo simulation study to assess the strengths of ct-gimme and validate the strengths of the continuous-time approach. Then, I compare its performance to alternative approaches in both continuous- and discrete-time. Finally, I provide an empirical illustration of individuals with major depressive disorder and generalized anxiety disorder and compare and contrast the results to those derived in discrete-time.