Conference Agenda

As intensive longitudinal data become more prevalent in psychology, there is a growing need for user-friendly software enabling applied researchers to investigate dynamic relations over time. Vector autoregressive (VAR) models offer a means to examine (cross-)lagged relations and inter-individual differences therein as random effects in multilevel model extensions. Furthermore, latent variable extensions allow for the estimation of the dynamics on the level of latent within-person factors. Multilevel latent VAR models were introduced within the Dynamic Structural Equation modeling (DSEM) framework in Mplus, however, open-source and free software for the estimation of these models using Bayesian MCMC sampling require substantive coding skills, limiting their accessibility for applied researchers. Addressing this gap, we introduce the mlts-package in R, designed to simplify the application of multilevel latent time series models (e.g., VAR[1]) by utilizing precompiled models specified in the probabilistic programming language Stan. These models incorporate latent-mean centering, accommodate within- and between-person covariates, individual (random) effects for dynamic parameters, and support multiple indicator models to address measurement error at different levels. We demonstrate the package features, discuss current limitations, and outline planned extensions for future releases.

Latent variable models are central to psychological science research but current R tools either lack breadth or suffer from convergence problems. The latent R package addresses these limitations by offering a unified, efficient, and user-friendly platform for estimating a wide range of latent variable models—including Structural Equation Modeling (SEM), Item Response Theory (IRT), and Latent Class Analysis (LCA)—with syntax compatible with the lavaan ecosystem. At its core, latent uses C++ implementations and advanced optimization strategies to provide a fast and flexible model estimation. For intance, the package solves common estimation issues such as Ultra-Heywood cases and non-positive-definite latent covariance matrices by estimating the parameters over the partially oblique manifold, a non-linear constraint that guarantees positive semi-definiteness while allowing for sparsity. Furthermore, latent supports parallelization to efficiently handling multiple starting values in models prone to local minima. For example, it can compute polychoric correlations among hundreds of variables in seconds and fit rapidly LCA models with many variables. These innovations make latent especially suitable for large-scale simulation studies and the estimation of complex models with large data. Overall, latent combines speed and generality in one open-source framework, significantly advancing the toolkit available for psychometric and structural modeling in R.

We present StructuralEquationModels.jl, a novel Structural Equation Modeling (SEM) software designed with extensibility as its core principle. Since the introduction of SEM, the range of problems it is applied to is rapidly expanding. However, SEM software implementations struggle to keep pace, so software flexibility has become one major factor limiting the adoption of innovations in the field of SEM. StructuralEquationModels.jl is designed to be modular and extensible to remove that bottleneck. This approach not only speeds up innovation in SEM but also ensures the integrity and reliability of existing functionality. We leverage the Julia programming language's ecosystem to achieve efficient implementations of loss functions and gradients and support modern optimization backends. We demonstrate the versatility of StructuralEquationModels.jl with examples such as adding support for regularized SEM with only a few lines of code and simulations comparing its efficiency to that of established SEM software. Even though our example implementations are simple enough to serve as illustrations, they surpass existing implementations regarding stability, computational efficiency, and available features.

Welcome to StatsQuest: The Inferential Statistics Adventure! This educational game is designed to enhance undergraduate students' understanding of the fundamentals of inference statistics in a fun and engaging way. In this talk, I will present a live demonstration of an early prototype of this serious game, which is designed to complement the statistics and research methods training of undergraduate students. Set in a classic top-down RPG-style university, students embark on an interactive journey through the MSB Medical School Berlin. Their mission is to gather data from various NPCs, download them and analyze their findings using their favorite statistical program like JASP or R. By choosing between different scenarios (i.e., populations), students can learn about various concepts of inferential statistics such as sampling distributions, p values, or confidence intervals, as well as about practical concerns such as sampling bias through convenience samples or non-response bias. It is my hope that through this experience, students gain motivation and hands-on experience that reinforces classroom learning.