QuantDev Brownbag - Incorporating Dynamic Fit Indices in Unsupervised Time Series Model Construction

Google map

Incorporating Dynamic Fit Indices in Unsupervised Time Series Model Construction

Abstract: 

The recent reduction in the burden associated with collecting time series data in the social, behavioral, and health sciences has coincided with an increased interest in the study of dynamic processes in human physiology, emotions, and behaviors. One popular approach for analyzing multivariate time series data is the structural equation modeling (SEM) framework, which allows for both the estimation of parameters of interest as well as the computation of metrics assessing model fit. Recent work has used these fit indices in an exploratory, data-driven manner to construct models that adequately represent individual-level dynamic processes, such as in the procedure implemented by the Group Iterative Multiple Model Estimation (GIMME) algorithm. Notably, because these fit indices lack an inherent null hypothesis, it has become standard practice to rely on cutoff values proposed by Hu and Bentler (1999), despite warnings against generalizing these values to models not examined in the seminal study. Indeed, recent work by McNeish and Wolf (2023) has shown that these cutoff values can change depending on the characteristics of the data and model, leading the authors to propose a simulation-based method for deriving dynamic fit index (DFI) cutoff values. The current project seeks to extend the DFI approach to the context of multivariate time series analysis, with a specific focus on the procedure implemented in the individual stage of the GIMME algorithm. Results from a simulation study and empirical example will be presented and discussed.