Poster #79 - Identification of Developmental Trajectory Classes: A Case Example of Studying Problem Behavior

• In Event: Poster Session 14

Abstract

Key theoretical models typically suggest development not only has aspects that are nomothetic (i.e., shared by all people) but more importantly, features that can be group differential (i.e., shared by only groups of people; Emmerich, 1968). Developmental researchers are often interested in mapping differentiated developmental trajectories of a particular construct (e.g., childhood mental health) which provide insights pertaining to variations in people’s change over time. For example, how many distinct groups (i.e., differential trajectories) can be identified based on children’s aggressive behaviors? Growth mixture modeling (GMM) has become a standard statistical approach to discovering heterogeneity in people’s development over time (Muthén, 2004; Masyn, 2013). This approach identifies subgroups of individuals with differentiated trajectories based on differences in each group’s intercept and slope parameters. Despite its popularity and usefulness to developmental researchers, one unresolved issue in applying GMM with longitudinal datasets is to determine the number, shape and size of classes since how the model parameters being specified can lead to different results. This is a critical issue since the number, shape, and size of classes are associated directly with how developmental researchers interpret results and make inferences about differentiated developmental patterns. Since mixture modeling is a flexible data-driven approach it can generate biased results when certain parameters are misspecified (McNeish & Harring, 2021; Sijbrandij et al., 2019; 2020). There are various frequently applied statistical methods available for developmental researchers to adopt, namely (a) latent class growth analysis (LCGA: Nagin, 1999); (b) growth mixture modeling (GMM: Muthén & Muthén, 2000); and (c) covariance pattern growth mixture model (CPGMM: McNeish & Harring, 2020). The LCGA (Figure 1A) has been criticized for its over-extracted classes (Sijbrandij et al., 2019) and GMM (Figure 1B) often encounters convergence issues (McNeish & Harring, 2020). The CPGMM has been recently advanced which is similar to the traditional GMM except for the adjustments in the covariance pattern, which facilitates estimation. Although studies have shown that different methods can result in different outcomes (Morin et al., 2021), these three methods have not yet been compared in one study using real (non-simulated) data. The goal of this study is to compare how these statistical methods can result in different outcomes and impact the final class enumeration results. We use a population-based sample of school-aged children (ECLS-K: 2011; N = 17,648) and examine the developmental trajectories of externalizing problems from kindergarten to Grade 5. Our results (Table 1, indicated that GMM and CPGMM had similar 2-class results (note the three-class solution did not converge for both methods). Though the trajectories had similar shapes, the class counts and proportions exhibited variations. Moreover, the LCGA indicated a 3-class solution (with the four-class solution not converging) which yielded class shapes and counts substantially different from those of the CPGMM and GMM. Our study highlights the possibility that different methods may result in diverging results in terms of class number, shape, and proportions. Although many of the issues discussed here have been noted elsewhere, we hope this proposal makes them more salient to developmental researchers.

Authors

• Qinxin Shi, University of Utah
  Presenting Author

• Idean Ettekal, Texas A&M University - College Station

• Yoonkyung Oh, University of Texas Health Science Center at Houston