The Cross-Lagged Panel Model Almost Always Provides Evidence for Causal Effects

The cross-lagged panel model is a widely used tool for testing causal effects in longitudinal data. Critiques of this model have focused on the fact that it fails to account for unobserved confounders, which can lead to spurious evidence for causal effects. Prior simulation studies suggest that the risk of spurious effects is high. In this paper, randomly selected pairs of variables that showed cross-sectional correlations between .20 and .70 in a large longitudinal study were analyzed using the cross-lagged-panel-model. Results showed that in 98% of these randomly selected pairs, evidence for significant causal effects emerged. The median effect size of these randomly selected effects was larger than the median effect from the literature. Alternative models that account for stable-trait or state variance led to fewer significant effects. These empirical results confirm recent theoretical and simulation-based critiques of the cross-lagged panel model and suggest that the model is not well-suited for testing causal effects in panel data.