Currently I am studying VAR methodologies in the hope of constructing a model for a future project, and have a fairly limited understanding of the necessary criteria to be met to generate robust results. My readings of recent literature make few mentions of residual diagnostics, specifically the joint normality of the residuals.
I have found through trial and error that a small number of exogenous spike ( blip ) dummy variables at key dates such as financial crises or policy changes, through visual inspection of residuals, have corrected the non-normality issue, I have found almost no evidence of similar studies doing the same, which leads me to wonder whether such measures are in fact misspecifications.
Tests for co-integration, in my case the Johansen test, give warnings (Eviews 12) against adding any exogenous variables, so as not to invalidate critical values. However, lag selection criteria for a VAR model differs when correcting residual non-normality with exogenous dummies. My current understanding of creating a VEC model is that, as a preliminary measure, both lag length selection and co-integration tests are performed on the model in levels, assuming all series are I(1) processes. Therefore, my question is whether one should:
1) Perform a cointegration test without dummies and select the lag length with dummies.
2) Abandon the dummies altogether, and subsequently violate the normality assumption.
3) Perform both Lag length selection and co-integration testing on the VAR in levels, then add dummies to the VECM.
My intention is to follow the common empirical approach in analysing both IFR and VDC of the VECM model (assuming there is cointegration), should this have any bearing on the matter. My understanding is that normality impacts only the validity of hypothesis testing however, in my reading, I have found no evidence to suggest that IRF and VDC standard errors are robust to non-normality.
many thanks,
Andrew Slaven
(Undergraduate Student at Aberystwyth University)
I don't think that there is an easy answer to your question. The problems you discuss are covered in Juselius (2006), The Cointegrated VAR Model, Oxford. When I last used Eviews their Johansen routines would not have covered all the routines in this book. That was some time ago and things may have changed. Cats in Rats (estima.com) or the equivalent in Oxmetrics were better. For a more modern survey, you might look at Killian and LUtkepohl, Structural Vector Autoregressive Analysis, Oxford. These are graduate-level tests and are not easygoing.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size