Dynamic and longitudinal models I (05)

Chair: Rung-Ching Tsai, Wednesday 22nd July, 11.45 - 13.05, Castlereagh Room, Fisher Building. 

Tom Lodewyckx, Francis Tuerlinckx, Peter Kuppens, Department of Psychology, University of Leuven, Belgium, Nicholas Allen, University of Melbourne, Australia, Lisa Sheeber, Oregon Research Institute, USA. Studying latent affective dynamics with a Bayesian state-space approach. (172) ♥

Wilfried Cools, Francis Tuerlinckx and Peter Kuppens, Research Group of Quantitative Psychology and Individual Differences, Katholieke Universiteit Leuven, Belgium. Modeling the individual specific oscillatory evolution of emotion in continuous time. (190)

Vassilis Vasdekis, Department of Statistics, Athens University of Economics and Business, Greece, Sylvia Cagnone, Department of Statistics, University of Bologna, Italy,Irini Moustaki, Department of Statistics, London School of Economics, UK, and . Full and limited information estimation methods for latent variable models for ordinal longitudinal data. (005)

Mark de-Rooij, Institute of Psychological Research, Leiden University, The Netherlands. Mixed effect ideal point models for the analysis of longitudinal multinomial data. (156)

 ABSTRACTS

Studying latent affective dynamics with a Bayesian state-space approach. (172)
Tom Lodewyckx, Francis Tuerlinckx, Peter Kuppens, Nicholas Allen and Lisa Sheeber
In the last years, emotion research has been focusing on the conceptualization of emotions as multicomponential, dynamical systems. This development created a new set of challenging research questions, concerning for instance autoregressive dependencies (related to concepts of emotional homeostasis) or cross-lagged relations (related to the mutual influence of emotion components). In a first part, we want to introduce a state-space approach for the dynamical modeling of emotion components. It will be shown how Markov chain Monte Carlo methods are used to estimate the model parameters. Various model extensions are discussed (e.g. external covariates, regime-switching). In a second part, we apply this framework to high resolution psychophysiological and behavioral data obtained during emotionally evocative adolescent-parent interactions and illustrate how it can be used to obtain new insights in the dynamical nature of emotions.  

Modeling the individual specific oscillatory evolution of emotion in continuous time. (190)
Wilfried Cools, Francis Tuerlinckx and Peter Kuppens
The change of emotion over time depends on both external influences and the internal dynamics of the emotion system, like self-regulating processes that are the focus of our paper. Self-regulating processes often elicit damped oscillations, which is also hypothesized to be the case for emotions. To investigate potential oscillations we propose a stochastic second-order differential equation, appropriately implying a continuous-time process. The link with discrete-time observations is made through a state-space representation with the model’s parameters estimated by optimizing a Kalman-filter based likelihood which flexibly incorporates irregularly spaced and individual-specific sampling schemes for panel data. The model allows for an hierarchical extension at the level of the regulatory process and is implemented in an existing R-package. The model is illustrated using panel data from an experience sampling study on the change in the two dimensions of core affect (valence and arousal) in which information about the impact and pleasantness of exogenous events is measured. 

Full and limited information estimation methods for latent variable models for ordinal longitudinal data. (005)
Vassilis Vasdekis, Sylvia Cagnone and Irini Moustaki
Analysis of longitudinal data allows studying variables across time. Often we are interested in studying the evolvement in time of latent constructs such as traits or attitudes. We develop here a latent variable model for longitudinal ordinal data. Associations within time are explained through time-dependent latent variables and covariates where associations across time are modelled through an AR(1) model for the latent variables. In addition, item speci¯c random e®ects are used for explaining the association of same items across time. The model has been estimated using full maximum likelihood estimation. Full maximum likelihood is very computationally intensive especially as the number of items increases. Alternative estimation methods that require only the use of univariate and bivariate information are being explored. Results from real applications
and a simulation study will be presented.

Mixed effect ideal point models for the analysis of longitudinal multinomial data. (156)
Mark de-Rooij
Maximum likelihood estimation of mixed effect models for multinomial longitudinal data can be prohibitive due to the integral dimension of the random effects distribution. We propose to use multidimensional scaling methodology to reduce the dimensionality of the response variable and thereby the dimensionality of the random effects. As a by product readily interpretable graphical displays representing change are obtained. For the simple case, where we have measurments on two time points and no other explanatory variables it will be shown that the model reduces to the slide vector model of Zielman and Heiser. Our new model thus provides a new interpretation to this model in terms of personal change. A more elaborate example will also be presented that shows further advantages of the modeling framework.