Recent debates about the appropriate way to model cross-lagged associations in panel data have primarily focused on whether to include a stable-trait component. For instance, the random-intercept cross-lagged panel model is a version of the simpler cross-lagged panel model that adds a stable-trait component, and simulations show that failing to consider stable traits can lead to spurious lagged effects. In this talk, I examine patterns of long-term stability for hundreds of variables assessed over multiple decades, focusing on how often state components are needed, either in addition to or instead of a stable-trait component. Results show that state components are often required, even after accounting for measurement error. Considering the role of reliable state variance in cross-lagged models has implications for estimation problems that have frequently been noted in regard to these models and for causal inference.

There are various statistical models that try to address the estimation of cross-lagged causal effects in panel data. In the present contribution, we describe how these statistical models can control for unobserved time-invariant confounders. Our aim is to help researchers to understand differences in these statistical models from a causal inference perspective. Assuming that the true data generation model (i.e., causal model) has time-invariant confounders that were not measured, we compared different statistical models (e.g., dynamic panel model and random-intercept cross-lagged panel model) in terms of the conditions under which they can provide a relatively accurate estimate of the target causal estimand. We provide practical suggestions for researchers to select a statistical model when they want to control for unmeasured time-invariant confounders and emphasize that the choice depends on what type of confounder effects are predominant in the researcher’s belief about the causal models.

In this talk we compare different approaches for estimating cross-lagged effects with panel data from a causal inference perspective. We clarify that cross-lagged panel models (CLPM) rely on the assumption of no unmeasured confounding (i.e., all relevant confounders are measured). By contrast, extensions of the CLPM such as the random intercept cross-lagged panel model (RI-CLPM) use additional latent variables (e.g., stable traits) to adjust for the effects of unmeasured time-invariant confounders. However, the possibility to adjust for unmeasured confounding by including additional latent variables comes with the price of restrictive assumptions about modeling the longitudinal process (e.g., no lag-2 effects). We also argue that latent variable-type models such as the RI-CLPM target a different quantity (i.e., within-person cross-lagged effect) than CLPMs (i.e., cross-lagged effects of the undecomposed scores) and highlight issues that need consideration when choosing between the different approaches in a specific application.

Regarding personality, a frequent interest consists in investigating the directionality underlying effects between variance in characteristics that are relatively stable over time. To this end, neither the traditional cross-lagged panel models, underestimating long-term stability, nor proposed alternatives, such as the random-intercept cross-lagged panel model, removing substantial, stable variance from those variables actually interesting to derive the relationships, constitute appropriate approaches. We suggest an alternative approach to disentangle whether variance in one characteristic predicts variance in another one or vice versa. The model accounts for stable variance over time by including a latent variable per characteristic. But it also considers stable variance when estimating the predictive effects – and does not separate the associations among time-independent latent variables from those among time-dependent residuals. We illustrate the consequences of different modelling strategies for the estimates of effects between variables using real-data examples and discuss how the models’ appropriateness depends on specific research questions.

Cross-lagged panel modeling approaches are well-established in psychological research for investigating causal relations between variables over time using panel data. However, critics of this practice in the causal inference literature state that SEM models depend heavily on parametric assumptions; since these are likely to be violated – at least to some degree – in practice, they argue that SEM models are prone to bias when used for causal inference. Obviously, this claim should raise concerns, but disciplinary differences hinder SEM users to appreciate the arguments, concerns, and alternative modeling approaches that are put forward by critics in fields like epidemiology and biostatistics. To address this issue, this presentation introduces the phases of the potential outcomes approach to causal inference, discusses the assumptions that are made throughout the causal inference process, and discusses reasons why cross-lagged panel modeling approaches might be best avoided when the goal is to investigate a causal research question.

The symposium closes with a moderated general discussion.