University of Minnesota
Model Identification and Selection for Nonlinear Mixed and Crossed Random Effects Models for Longitudinal Data
Abstract
dc:description.abstractIn education research, longitudinal data are used to describe patterns of change over time and determine factors that influence these patterns of change. These longitudinal trajectories can take on many shapes or functional forms (e.g., linear, nonlinear); however, for many educational and psychological datasets, they commonly follow a curvilinear shape. A popular statistical framework for modeling nonlinear growth is nonlinear mixed- (or random) effects modeling. Nonlinear mixed-effects models (NLMEMs) are advantageous because they are highly adaptable and can be modified to fit the data on hand to answer research questions. However, the flexibility of NLMEMs can become a double-edged sword as the complexity of the model can increase drastically. This leads to concerns about how to empirically identify and estimate these models and how to select the best model to represent the underlying theory/population, and not the current sample (a.k.a., “overfitting”). In the current dissertation, we aim to address the issues of model identification and selection for longitudinal NLMEMs and an extension of these models relevant to education research. First, in Chapter 2, we tackle the topic of model selection and the impact of functional form complexity using a Maximum Likelihood estimation framework. Various model selection criteria are evaluated to assess under what conditions these criteria can recover the nonlinear shape of the growth trajectory and the role functional form complexity has on model overfitting. The results highlight criteria that are more vulnerable to overfitting and provide recommendations to researchers about which methods should be implemented in practice. In Chapter 3, we extend the current NLMEMs to account for the influence of group membership on individual outcomes (e.g., the impact of schools or teachers on student achievement over time). These are known as crossed random effects models (CREMs). First, we introduce a Bayesian piecewise CREM that can both estimate the impact of group membership on individual growth and model nonlinear growth trajectories. Then, we address the issue of model identification for longitudinal data applications where researchers want to implement CREMs. Both linear and nonlinear longitudinal CREMs are assessed to determine what data conditions are necessary to empirically identify these models. Lastly, in Chapter 4, we once more tackle the issue of model selection. However, in this chapter, we expand upon the work done in Chapter 2 to address model selection for both NLMEMs and the CREM extension. Furthermore, we borrow from the work done in Chapter 3 to evaluate model selection methods that are used in a Bayesian inference framework—which provides a set of model selection methods not available in the Maximum Likelihood framework. Various Bayesian model selection methods are evaluated on their ability to recover the shape of the growth trajectory and their susceptibility to overfitting. As in Chapter 2, the results highlight which methods perform best for model selection, when the methods are more vulnerable to overfitting, and provide recommendations to researchers about which methods should be implemented in practice. We conclude with a summary of our findings and implications for education research.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Rohloff, Corissa
Subjects
dc:subject × 5Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Citation
- Rohloff, Corissa. (2024). Model Identification and Selection for Nonlinear Mixed and Crossed Random Effects Models for Longitudinal Data. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/281806.