XI Conference – European Congress of Methodology

Advancing Dynamic Methods for Modeling Change Over Time

Not scheduled

1h 30m

EAM2025

Symposium Statistical analyses Symposium : "Advancing Dynamic Methods for Modeling Change Over Time"

Abstract

A popular cost-effective way of collecting longitudinal data is the accelerated longitudinal design (ALD). In ALDs, participants from different cohorts are measured repeatedly but the measures provided by each participant cover only a fraction of the time range of the study. It is then assumed that the common trajectory can be studied by aggregating the information provided by the different converging cohorts. ALDs are, therefore, characterized by a very high rate of planned data missingness. Additionally, it is very common that most longitudinal studies present unexpected participant attrition leading to unplanned missing data. A way for analyzing this data is the latent change score (LCS) model within a Continuous-Time State-Space Modeling framework (CT-SSM). This CT-SSM model allows computing Kalman scores, which can be used to estimate individual observed and unobserved scores. We simulated an accelerated longitudinal design where we manipulated different conditions such as the sample size, the unplanned missing data mechanism (MCAR, MAR, MNAR), and the severity of the unplanned missingness. Results showed that the Kalman scores were able to estimate both (1) data points that were expected but unobserved and (2) data points that were outside the age range observed for each case (i.e., to estimate the individual trajectories for the complete age range under study). These results have important implications for practitioners in psychology and education because they make it possible to accurately forecast individual longitudinal trajectories and to make individual-level decisions considering the model predictions. This presentation summarizes part of the results of a recent publication: https://doi.org/10.1037/met0000664

Abstract

Evaluating change over time is one of the most interesting problems in behavioral sciences. In this symposium, we present a set of cutting-edge advances in dynamic modeling, and provide several perspectives on how they can be used to answer relevant substantive questions in psychology.
Each of these contributions proposes interesting methodological innovations and provides new tools for studying change in different applied contexts. Recommendations are also given on how to design longitudinal studies, and R code is provided to apply very novel models to analyze this type of data.

Abstract

Introduction. Experience sampling methods (ESM) are an increasingly popular strategy for studying affective processes (i.e., mood and emotions). In these studies, the emotional state of one or more individuals is measured several times a day during multiple days or weeks. A unique feature of these studies is the spacing of observations: measurements are frequent during waking hours but separated by a much longer interval overnight while participants sleep. This uneven distribution poses challenges for dynamic models, where emotional states are represented as a function of previous states and dynamic noise. Importantly, the overnight gap may induce changes in emotional dynamics that cannot be explained solely by the length of the interval. For example, emotional states at bedtime may exert an influence on morning affect that differs from daytime patterns. Despite its potential impact, the role of overnight lags has been largely overlooked in the literature. Typical approaches either ignore these effects or exclude nighttime intervals entirely, which simplifies the data structure but may overlook meaningful dynamics in the transitions between days. In this study, we evaluate the efficacy of various modeling strategies to address overnight effects within the framework of state-space models. Specifically, we investigate how overnight lags can be incorporated to account for changes in emotion dynamics that occur between consecutive days.
Method. To evaluate the performance of the strategies compared, we conducted a Monte Carlo study under a range of conditions that are frequent in experience sampling studies. We also applied the proposed approaches to existing datasets on affect dynamics to illustrate their implementation and practical utility.
Results and discussion: We discuss the implications of modeling overnight dynamics, highlighting the importance of accurately capturing these effects for a more nuanced understanding of daily emotional processes. Strengths, limitations, and future directions for improving the handling of irregular time intervals in ESM research are also considered.

Abstract

One of the key questions in longitudinal research is when to take measurements of the variables of interest. Panel studies usually focus on the dynamics between two processes over time (e.g., depressive symptoms and self-esteem), and include few repeated measures (<10). This forces researchers to find the most efficient way to design their study and collect their data. Recently introduced in Psychology, continuous-time models are very convenient in this context, as they can accommodate irregularly spaced measurements, both between and within individuals.
Previous research on deciding the optimal time interval between measurements have proposed various criteria, such as using the time interval at which the overall cross-effects are largest, or the time interval leading to best estimation reliability. However, relatively less attention has been paid to the effect of stochastic innovations (i.e., dynamic error) on the sampling design, despite its key role in the stability of the system.
In a Monte Carlo simulation, we used state-space continuous-time models to characterize the dynamics of two variables measured longitudinally through panel designs. We compared various sampling schedules, including those suggested in recent research, to evaluate their effectiveness in recovering bivariate trajectories under various levels of stochasticity. We discuss the strengths and weaknesses of different sampling approaches and provide practical recommendations on using continuous-time modeling in panel data studies.

Abstract

State-space models (SSMs) provide a powerful framework for modeling dynamic systems, capturing both intra-individual and inter-individual variability in longitudinal data. In the context of cognitive development research, one interesting feature of SSMs is their ability to model deviations, or “shocks,” in individual trajectories. Such shocks may signal atypical changes that could be considered outliers within developmental processes. In this study, we adapt a semi-exploratory procedure proposed by You et al. (2020) to the context of cognitive development, using the dynr package (Ou et al., 2019) in R.
Our main objectives are to: a) propose a novel SSM designed to detect outliers in developmental trajectories; and b) evaluate its performance in terms of accuracy of outlier detection and recovery of the population parameters.
To evaluate these objectives, we performed an extensive Monte Carlo study. First, we generated data based on empirical trajectories. We manipulated several simulation conditions, including sample size, number of time points per participant, timing of shocks, and proportion of participants affected by shocks. Next, we examined the impact of these factors on group-level parameter bias, and the balanced accuracy of the individual outliers identification. Based on our findings, we discuss the utility of SSMs to detect abrupt environmental changes affecting cognitive development.

Presentation materials

Intro. Advancing Dynamic Methods for Modeling Change Over Time.pdf