Identification of linear panel data models when instruments are not available

One of the major virtues of panel data models is the possibility to control for unobserved and unobservable heterogeneity at the unit (individual, firm, sector...) level, even when this is correlated with the variables included on the right hand side of the equation. By assuming an additive error structure, identification of the model parameters spans from transformations of the data that wipe out the individual component. We propose an alternative identification strategy, where the equation of interest is embedded in a structural system that properly accounts for the endogeneity of the variables on the right hand side (without distinguishing correlation with the individual component or the idiosyncratic term). We show that, under certain conditions, the system is identified even in the case where no exogenous variable is available, due to the presence of cross-equation restrictions. Estimation of the model parameters can rely on an iterated Zellner-type estimator, with remarkable performance gains over traditional GMM approaches.