A Critique of the Random Intercept Cross-Lagged Panel Model

Abstract

The random intercept cross-lagged panel model (RI-CLPM) is an extension of the traditional cross-lagged panel model (CLPM) that allows controlling for stable trait factors when estimating cross-lagged effects. It has been argued that the RI-CLPM more appropriately accounts for trait-like, time-invariant stability of many psychological constructs and that it should be preferred over the CLPM when at least three waves of measurement are available. The basic idea of the RI-CLPM is to decompose longitudinal associations between two constructs into stable between-person associations and temporal within-person dynamics. The present article critically examines the RI-CLPM from a causal inference perspective. Using formal analysis and simulated data, we show that the RI-CLPM has limited potential to control for unobserved stable confounder variables when estimating cross-lagged effects. The CLPM with additional lag-2 effects sufficiently controls for delayed effects, as long as all relevant covariates are measured. Furthermore, we clarify that, in general, the RI-CLPM targets a different causal estimand than the CLPM. Whereas the cross-lagged effect in the CLPM targets the effect of increasing the exposure by one unit, the within-person cross-lagged effect in the RI-CLPM provides an estimate of the effect of increasing the exposure by one unit around the person mean. We argue that this within-person causal effect is typically less relevant for testing causal hypotheses with longitudinal data because it only captures temporary fluctuations around the individual person means and ignores the potential effects of causes that explain differences between persons.

How causal is a reciprocal effect? Contrasting traditional and new methods to investigate the reciprocal effects model of self-concept and achievement

“…Second, is there evidence for reciprocal effects of academic self-concept and achievement when using these methods for causal inference and how do results from traditional and new methods compare to one another? In addressing these questions, we will investigate reciprocal relations between student self-concept and achievement using (a) traditional CLPMs, (b) FF-CLPMs (e.g., Lüdtke & Robitzsch, 2021), (c) RI-CLPMs (Hamaker et al, 2015), and (d) two weighting approaches (covariate balanced generalized propensity score weighting [CBGPS-weighting] and entropy balancing [EB]). The two weighting approaches were explicitly developed to study causal effects of continuous treatment variables in observational studies (Fong et al, 2018;Hainmueller, 2012;Tübbicke, 2021).…”

“…Rubin defined the difference between these two potential outcomes (Y(1)-Y(0)) as the individual causal effect and the average of all individual causal effects as the average causal effect (e.g., Shadish, 2010;West & Thoemmes, 2010). Lüdtke and Robitzsch (2021) applied this framework to define the cross-lagged causal effect. Translated to the REM, one would define the causal cross-lagged effect of self-concept on grades as the following linear function:…”

“…Here, β1 constitutes the causal effect of increasing self-concept at the second measurement occasion by one unit on grades at the third measurement occasion. Usami et al (2019) outlined three assumptions that need to be satisfied in order for CLPMs to identify causal effects: (a) consistency, (b) strong ignorability/no unobserved confounding, and (c) positivity (see also Lüdtke & Robitzsch, 2021). Referencing to Rubin's potential outcome framework (e.g., Holland, 1986;Rubin, 1974Rubin, , 2004, consistency implies that the observed outcome of a person is identical to the potential outcome of this person, given her/his observed exposure history (Rehkopf et al, 2016).…”

“…A conceptual look at the REM seems helpful to derive potentially reasonable structural causal models on the relation between self-concept and achievement from the literature. Structural causal models are models of reality, which display considerations and assumptions about causal relations between variables (Cunningham, 2021;Lüdtke & Robitzsch, 2021;Pearl et al, 2016). They consist of exogenous and endogenous variables, operators (arrows), which indicate the direction of the respective causal effects (see Figure 1), and any functional form to link the two types of variables (not displayed).…”

Transmission of Basic Psychological Need Satisfaction between Parents and Adolescents: The Critical Role of Parental Perceptions

“…However, extending our research to three or more time points would be important to examine how the development of BPNS happens over longer periods of time. Testing the models at different ages could also be highly informative to make a better causal inference -recent studies have shown the possibility that models with more than two time points can potentially control for various types of confounders (Lüdtke & Robitzsch, 2021;Usami et al, 2019).…”