Substance use is one of the most commonly occurring health risk behaviors in adolescence and has been unambiguously linked to a variety of negative physical, biological, and psychological outcomes (e.g., Hingson &Kenkel, 2004;USDHHS, 2007). These serious public health issues have impelled substantial growth in the theoretical conceptualization of pathways to substance use, especially factors that exist in childhood and adolescence (e.g., Hussong et al., 2011;Zucker, Heitzeg, &Nigg, 2011). However, researchers currently disagree as to whether patterns of onset and escalation are best captured through the identification of discrete groups or "types" of individuals or through modeling individual variability across continuous trajectories of use. Growth mixture models are widely used to identify qualitatively different subgroups, yet key limitations of these models directly undermine the extent to which competing theories of substance use and abuse can be validly tested and compared. It has been extensively demonstrated that maximum likelihood (ML), the current gold-standard method for estimating growth mixture models, is unable to reliably reproduce the true population structure. Mixture models are highly sensitive to even slight model misspecifications or improper restrictions, and if data are non-normally distributed spurious classes will be detected (Bauer &Curran, 2003a;2004). If latent classes truly exist, it is difficult to correctly determine the number and form of latent trajectory classes using existing fit statistics (Tofighi &Enders, 2008). Compared to ML, there is the potential for more parameters and more complex models to be identified in a Bayesian analysis (Muthen &Asparouhov, in press), and Bayesian estimation is more stable and has more power in small samples (Asparouhov &Muthen, 2010). If a model with too many classes is estimated (Rousseau &Mengersen, 2011), Bayesian estimation will reliably "empty" the unneeded classes whereas ML estimation becomes unstable. In a fully Bayesian latent class analysis, the distribution of the number of classes can be estimated and examined as an unknown parameter (Richardson &Greene, 1997), whereas no information of this kind is available using ML estimation. However, Bayesian estimation still needs to be rigorously studied in the context of adolescent substance use. The three core aims of my project are to (1) compare Bayesian and ML estimation of growth mixture models when the population is truly homogenous;(2) compare Bayesian and ML estimation when multiple trajectory classes do truly exist;and (3) apply novel Bayesian methods to existing adolescent substance use data. My proposed project will fully integrate advanced quantitative methods with substantive theory so that researchers can reliably and validly test developmental theories of adolescent substance use.
The proposed project will fully integrate advanced quantitative methods with substantive theory by investigating how researchers can optimally model substance use trajectories. This project will evaluate existing and innovative estimation methods for modeling substance use trajectories. Secondary data analysis will show the utility and feasibility of these novel estimation methods for rigorously and reliably describing trajectories o adolescent substance use and testing theories of developmental pathways to disorder.
|Bainter, Sierra A; Bollen, Kenneth A (2014) Interpretational Confounding or Confounded Interpretations of Causal Indicators? Measurement ( Mahwah N J) 12:125-140|