On Fri, Nov 11, 2022 at 5:28 AM Agustín Alonso Rodríguez <aalonso(a)rcumariacristina.com&gt; wrote: > > How to refine or simplify a VAR model suppressing the insignificant estimated coefs? You have a reply from Sven, noting that gretl does not currently have a built-in procedure for what you describe. That's true, but I'll offer a somewhat different perspective. When you estimate a VAR with p lags of each variable, you should not be too concerned with the "statistical significance" of each coefficient -- and not too quick to eliminate terms that are "insignificant". If the VAR is in levels, in particular, it's likely that the several lags of a given variable will be strongly correlated. This then gives you the classic "multicollinearity" effect, whereby each individual coefficient may be "insignificant" yet a joint test strongly rejects the null hypothesis that the "collinear" variables all have coefficients equal to zero. Gretl shows you the joint (F) tests after estimating a VAR -- that is, tests for Granger causality -- and you might be better off using these to eliminate one or more variables from your specification (if warranted) rather than worrying about individual lagged terms.