Conference Agenda

Estimating network structures from cross-sectional data is most commonly achieved through regularization techniques (e.g., Friedman et al., 2008; Williams, 2020). An alternative approach relies on neighborhood selection via node-wise regressions, using the Bayesian Information Criterion (BIC) for model selection (Williams et al., 2020). This method has recently been extended to handle missing data (Nehler & Schultze, 2024), increasing its relevance for applied psychological research. However, despite these advances, methodological evaluations have largely focused on continuous, multivariate normally distributed variables. This is a notable gap, given that ordinal data—especially from questionnaire items—are widespread in psychology. Yet, dedicated neighborhood selection techniques tailored to ordinal variables remain underdeveloped (e.g., Isvoranu & Epskamp, 2023). In this talk, we address this gap by investigating the performance of neighborhood selection using BIC applied to ordinal data, both with and without missing values. For complete data, we compare a newly adapted estimation procedure for ordinal variables with an existing approach originally designed for continuous data (Williams et al., 2020). In the context of missing data, we evaluate two strategies, based on maximum likelihood estimation and multiple imputation. Both have shown strong performance with continuous data (Nehler & Schultze, 2024), and we test their robustness when applied to ordinal variables. The performance of the proposed methods is evaluated through a simulation study that varies network size, number of observations, and proportion of missing data. Evaluation criteria include the recovery of the network structure as well as potential biases in estimated edge weights and strength values.

Adequate handling of missing data on physical activity assessments is crucial in longitudinal accelerometer studies. This study aimed to evaluate the effectiveness of various imputation methods for handling missing data in an empirical application which utilizes wearable accelerometers. We employed a simulation approach to assess performance under different missing data scenarios, including Missing Completely at Random (MCAR), Missing at Random (MAR), and Missing Not at Random (MNAR). Our findings revealed that mean imputation and hot deck imputation applicated with a fine degree of matching criteria (participant and time of the day) outperformed discard-based methods under MCAR and MAR conditions. Notably, no imputation methods performed well under MNAR scenarios. We recommend conducting simulation studies tailored to specific study designs to compare imputation methods, implement strategies for improving data quality, gather information on non-wear periods, and ensure continuous monitoring and participant compliance, thereby reducing bias in activity level estimates.

Intensive longitudinal data from ecological momentary assessment (EMA) is widely used in clinical research but often suffers from dropout, leading to reduced statistical power, invalid results, and poor treatment outcomes. Predicting dropout could help with its prevention. While existing methods utilise baseline covariates, few studies account for the temporal dynamics of EMA data or identify the exact timing of dropout. Joint models (JM) enable simultaneous modelling of longitudinal processes and time-to-event data, offering dynamic predictions. However, conventional JMs assume limited measurement occasions and do not account for the autocorrelation inherent in EMA data. We extended the standard JM by incorporating an autoregressive submodel, capturing temporal dependencies in EMA measurements. We validated our approach through simulation studies, demonstrating good parameter recovery across different missingness mechanisms (MCAR, MAR, MNAR) and high dropout prediction accuracy. We applied the JM to an existing empirical EMA dataset, using baseline (e.g., depression) and time-varying (affect, intermittent missingness) predictors of dropout. The extended JM outperformed a baseline-only survival model in predicting dropout. The sensitivity analysis of the missingness mechanism revealed that fixed effect estimates remained stable across different missing data mechanisms, whereas random effect estimates for autocorrelation were sensitive to these assumptions. By integrating autoregressive components, the extended JM accommodates temporal dependencies and dynamically updates predictions of dropout risk. This approach improves dropout prediction in EMA studies and highlights the importance of utilising JMs for predicting clinically relevant outcomes while integrating EMA data.

Missing data are a common challenge in multilevel designs, and multiple imputation (MI) is one of the most commonly recommended methods for handling them. Past research has shown that multilevel MI can be extremely effective at handling missing data in multilevel designs, provided that the imputation model adequately takes the multilevel structure into account. This is particularly important in multilevel analyses that involve nonlinear effects or random slopes, and many specialized methods have been developed for these applications. However, multilevel MI can be difficult to apply in practice, where the multilevel structure is often not very pronounced or not of immediate interest in the intended analyses. In these applications, existing methods can become unstable and often struggle to provide reliable results. In the present talk, we present a fully conditional specification approach to multilevel MI that combines single-level imputation methods with group means (GM) or adjusted group means (AGM) to accommodate the multilevel structure. In addition, we present the results for a theoretical investigation and several simulation studies, in which we evaluated the statistical properties of these methods and their performance in different applications, including applications with balanced designs, unbalanced designs, and larger numbers of variables. Finally, we discuss the strengths and weaknesses of these methods and their implications for practice.