The RATS Software Forum

This demonstrates several of the LM tests from Eitrheim and Terasvirta(1996), ""Testing for the Adequacy of Smooth Transition Autoregressive Models", Journal of Econometrics, vol 74, pp 59-75. These tests require saving the derivatives of the STAR model (done with the DERIVES option on NLLS) and running auxiliary regressions on those derivatives plus some set of variables which are specific to each test and seeing the significance of the test variables. The test for serial correlation is simple (auxiliary regression requires merely lags of the residuals). A test for remaining (threshold) nonlinearity involves adding powers of the (assumed) omitted threshold variable interacted with the regressors. Note that the latter will almost always produce several redundant regressors (not really mentioned in their paper), but the RATS EXCLUDE instruction detects that and automatically adjusts the degrees of freedom.