Many different models exist for examining causal effects in panel data through cross-lagged associations. Choosing among these models depends in part on assumptions about patterns of stability in the variables being examined. In this paper, we first use simulations to examine how different data-generating models affect bias in commonly used models that include cross-lagged effects. We then examine patterns of stability in over 400 variables from a 22-year panel study from Australia to assess plausibility of the parameter values from the simulations and to help evaluate risk of bias. Empirical results show that patterns of stability in real data are often inconsistent with expectations from commonly used models that fail to incorporate a state component. Simulation results show that bias in estimates from these commonly used models will likely be high, given these patterns of stability. Alternative models that include state components are recommended to more accurately account for patterns of stability and to reduce bias in estimates of lagged effects.
Lucas et al. (Thu,) studied this question.