Symposium: Advancing Dynamic Models of Psychological Processes

Time series models are increasingly used to investigate fluctuations in psychological processes over time. Advanced time series models such as Dynamic Structural Equation Models have proven a powerful tool to capture, for example, mood fluctuations and to subsequently inform diagnosis and treatment. Arguably, temporal dynamics of mood are not gradual but involve states, especially in clinical populations. Therefore, we propose an extended time series model in which we implement states and allow temporal dynamics parameters to vary across these states. We show how these models outperform stateless models in capturing mood fluctuations and how model parameters can be used to inform clinical practice.

This is an in-person presentation on July 20, 2024 (11:20 ~ 11:40 CEST).

Interest in psychological dynamics can span multiple timescales, ranging from fast fluctuations at a resolution of seconds, to occasions (morning, afternoon, evening), days, and even months. However, currently available implementations of DSEM are restricted to two-levels forcing researchers to opt for suboptimal choices, either ignoring some timescales or building models using multi-step approaches with known psychometric pitfalls (e.g. loss of uncertainty in estimates). By leveraging the flexibility offered by the probabilistic programming language Stan we can extend DSEM to allow for the contemporaneous modeling of processes across multiple timescales. We showcase this modeling innovation in a rich dataset (VEKTOR) tracking children’s cognitive training progress. We show that children’s cognitive performance can fluctuate over the span of seconds and days using a 3-level DSEM. Researchers can easily extend this parameterization of DSEM to specify a hierarchical DSEM with their desired number of levels.

This is an in-person presentation on July 20, 2024 (10:20 ~ 10:40 CEST).

Commonly used time series models are able to capture temporal dynamics of psychological processes over time. However, these models assume that the temporal dynamics function similarly regardless of the current value of the time series. We argue that many psychological processes likely involve asymmetrical temporal dynamics, for example, that the effect of more-than-usual sleep (for a given person) on wellbeing is different from that of less-than-usual sleep. We illustrate how overlooking asymmetrical dynamics can adversely impact inferences one draws, for example, of the effect of sleep on wellbeing. We demonstrate how we can adapt commonly any time series model to allow for asymmetrical dynamics in, and how doing so improves model fit and inferences.

This is an in-person presentation on July 20, 2024 (11:00 ~ 11:20 CEST).

Although research investigating the dynamics of complex systems is a mainstay in fields such as epidemiology, economics, and certain sub-disciplines of psychology, such investigations are increasingly gaining traction in cognitive psychology. This surge in interest is largely attributed to the emergence of Dynamic Structural Equation Models (DSEMs). DSEM represents a powerful statistical tool for modeling longitudinal data, integrating elements of time-series analysis, multilevel modeling, and structural equation modeling. The availability of DSEM is increasingly broadened due to its translation into free software platforms like R and Stan. However, before embracing DSEM on a larger scale, it is crucial to thoroughly investigate its psychometric properties and parameter estimation capabilities. In this context, our study embarks on an exploration of a multiverse of methodological options related to DSEM application. Specifically, we examine the impacts of varying sample sizes, number of timepoints, priors, and sampler settings—including the number of iterations, seeds, and initial start values—on the robustness and reliability of the model outcomes. This comprehensive analysis aims to provide deeper insights into the optimal use of DSEM in cognitive psychological research, paving the way for their more informed and effective application in the field.

This is an in-person presentation on July 20, 2024 (10:00 ~ 10:20 CEST).

Night gaps are inherent to experience sampling method (ESM) data, yet they are often overlooked or approached solely as a dynamic modeling problem rather than an avenue of research. As a result, the theoretical implications of current solutions are typically not (fully) considered or discussed. These solutions involve (1) ignoring the night gap by considering the night interval as a regular interval, (2) removing the night gap by not regressing the first measurement of the day on the last measurement of the previous day, or (3) treating the night gap as a missing data problem. In this presentation, I propose an alternative method that estimates the relation between the first observation of the day and the last observation of the previous day freely, as a separate parameter. Additionally, I make explicit for each of these methods their theoretical implication for the process of interest within the context of the fundamental first-order autoregressive (AR[1]) model. Since the current methods are special cases of the proposed alternative, we can test which method best describes the process of interest. Through an empirical N=1 example of various ESM variables, I demonstrate that the best-fitting method differs per variable. This implies that some processes may exhibit different dynamics during the night than during the daytime, providing a stepping stone to understanding and modeling night gaps in ESM.

This is an in-person presentation on July 20, 2024 (10:40 ~ 11:00 CEST).