Abstract
A key issue in spatial models is to specify the spatial effect. Robust LM tests have long been popular in spatial econometrics for discriminating between spatial lag and spatial error processes. Their application has recently been extended to spatial panels, a context where further issues arise the tests were not designed to address in the first place: individual heterogeneity and time persistence. Through Monte Carlo simulation, we show that RLM tests become virtually useless as a specification device under substantial individual or time heterogeneity, regardless whether correlated or not. If the heterogeneity is correlated (fixed effects) then the tests are theoretically inappropriate. If it is not (random effects), their properties severely deteriorate. The simple solution of eliminating individual effects by time-demeaning the data,and/or time effects by adding time dummies, greatly improves the situation, making the RLM tests a viable solution. Optimal (likelihood-based) conditional tests are available. We compare their performance to that of the RLM tests. Conditional tests are to be preferred if possible; otherwise the RLM on demeaned/augmented can be a computationally simpler second-best alternative to the optimal ML-based tests.
Collegamento Microsoft Teams
Organizzazione
Anna Gloria Billé