Masterarbeitsthemen

Beschreibung: Structural equation models (SEMs) are foundational tools in psychological and social sciences for modeling complex relationships between observed and latent variables. While SEMs typically assume linear associations, real-world psychological processes frequently involve nonlinear interactions. The failure to account for nonlinearity in these models can lead to biased parameter estimates, misinterpretation of relationships, and misleading conclusions. Despite these potential pitfalls, the extent to which unmodeled nonlinearity impacts common model fit indices and parameter uncertainty has received limited systematic investigation.

This master?s thesis will use simulations to explore how unmodeled nonlinear relationships manifest in SEMs. Specifically, we will examine the sensitivity of model fit indices to the presence of unmodeled nonlinearity. Additionally, we will evaluate how ignoring nonlinearity influences parameter estimates and their associated uncertainty. By simulating data under controlled conditions, this research aims to provide insights into the robustness of SEMs to nonlinear effects, offer practical guidance for researchers, and highlight areas where existing model diagnostics may fall short.
Kontakt: Charles Driver, E-Mail

Beschreibung: The frequency of observations in longitudinal studies can play a critical role in shaping the conclusions drawn from statistical models. Classic longitudinal approaches, such as growth curve models and cross-lagged panel models, use the temporal structure of the data to infer developmental trajectories and dynamic relationships. However, the choice of observation frequency is often driven by practical constraints rather than theoretical considerations, potentially leading to biases or misinterpretations in model results, particularly when the models are (as is usual) not perfect.

Using both simulated and real-world datasets, we will evaluate the effects of different temporal sampling schemes on growth curve and cross-lagged panel model estimates. By systematically varying observation frequency in simulations and comparing these results to empirical data analyses, this research aims to identify the conditions under which observation frequency affects model performance and validity. The findings will provide practical recommendations for designing longitudinal studies and interpreting results in the context of varying temporal resolutions.
Kontakt: Charles Driver, E-Mail

Beschreibung: The use of machine learning (ML) methods, such as random forests and support vector machines, is becoming more common in psychology for predicting outcomes like mental health, personality traits, or behavior. These methods are often seen as more flexible than traditional statistical approaches like regression, as they can handle complex relationships and large datasets. However, questions remain about whether these advanced methods truly offer better predictions in practical settings and how they compare in terms of accuracy and ease of understanding.

This thesis will re-analyze real-world datasets from published psychological studies to compare the performance of machine learning and deep learning models with the statistical methods originally used. The focus will be on examining how well each approach predicts key outcomes, as well as considering differences in interpretability and practical usefulness. By directly comparing these methods, this research will help clarify when and how machine learning can complement or improve upon traditional techniques in psychology.
Kontakt: Charles Driver, E-Mail