3 - (SYM 37) Novel Computational Approaches to Psychiatric States-trait Phenotyping

Background: Disentangling heterogeneity in psychological constructs remains vital for identifying homogeneous subgroups that hold practical utility and theoretical importance. The increased popularity of intensive longitudinal designs enables the use of novel statistical methods for identifying classes that incorporate considerations of temporal dynamics. Despite advances in latent state-trait theories and methods, traditional methods for phenotyping have rarely differentiated between classes due to situational versus personal influences.

Methods: The current study describes novel latent state-trait modeling approaches with discrete states and a discrete trait with a Bayesian estimation. We demonstrate models with independent states and trait, interactive states and trait, as well as computational models with shared and with trait class-specific growth parameters. We use artificial data examples to illustrate the model compositions and real data examples to show model interpretation. 

Results: The proposed computational approach can simultaneously account for heterogeneity in stable traits reflecting interindividual phenotypes as well as transitory states representing intraindividual phenotypes. Given a continuously measured observed variable across multiple persons and multiple occasions, the proposed computational approach is able to model states at each measurement occasion for each individual (and their dynamics), individual traits, as well as trend either shared across or specific to identified trait classes.

Discussion: Theoretically, differentiating between time-varying and time-invariant components of class membership can harness the rich information in data such as from EMA designs to disaggregate situation- and person-specific heterogeneity. Methodologically, our computational approach complements existing methods for identifying mixture structure in latent states and traits.