Keynote Talk 1
Kenneth A. Bollen,
University of North Carolina at Chapel Hill
Which Longitudinal Model Should I Choose?
With the growing availability of longitudinal data comes the question of what model to use? In an ideal world, theory and substantive arguments would be sufficiently clear to dictate one. But in practice, there is little guidance and academic fads or the practice in researchers’ fields typically affect model choice. We illustrate how a general longitudinal model (LV-ALT) can help researchers in their selection. The LV-ALT model can specialize to other popular models such as the classic random or fixed effects, growth curve models, autoregressive, latent difference scores, and a variety of other hybrid structures. The LV-ALT model can help to defend the choice of one of these traditional models or it can suggest new hybrid models to consider. We illustrate our results with Add Health NLYS79 data on self reported health and an analysis from a 2021 Demography paper.
Keynote Talk 2
Pascal R. Deboeck,
University of Utah
From Data to Causes: Perspectives on Causation from Psychology, Physics, and Dynamical Systems
Causality, by definition, implies change across time. The difficulty making inferences about constructs that change across time are perhaps clearest in the mediation literature. In the mediation literature it has been demonstrated that inferences about longitudinal processes based on cross-sectional data are typically a poor reflection of the underlying processes. The latter is particularly true when it is not possible to implement experimental designs to detangle causes and effects, as is the case in many areas of psychology. Longitudinal data thus becomes essential for understanding the relations in a system of inter-connected components, whether those components consist of differing constructs or people. This presentation will begin by introducing a common longitudinal model for mediation, the cross-lagged panel model, and build to introducing a continuous-time counterpart. Through study of aspects of these models it will be highlighted that when working with longitudinal data it becomes necessary for any inference to carefully consider how time should be treated, explore the relations between components of a system, and consider that causation could imply effects that are not typically expected.
Keynote Talk 3
Less Casual Causal Inference
Correlation does not imply causation—but correlations are often all we can get. In this talk, I will provide an introduction to graphical causal models, a powerful tool for researchers working with observational data, but also for experimentalists. A broad range of common inference problems such as mediation analysis, missing data, and generalization across populations and settings can be tackled with such models. Thus, being less casual about causal inference really pays off, as it clarifies many methodological matters simultaneously.