[Gretl-users] Re: IRF interpretation

Am 04.09.2019 um 14:29 schrieb lassesskola(a)gmail.com: Hi, you asked a similar question on April 27th, 2018. (The internet doesn't forget, hehe...) There Allin pointed out that you had actually asked the question before on Aug 17th, 2017: https://gretlml.univpm.it/hyperkitty/list/gretl-users@gretlml.univpm.it/t... So this is turning into a "same procedure as every year" thing. However, I do acknowledge that back in 2017 the mailing thread was a bit open-ended and possibly not clear enough. So let me try to summarize this here. - Be aware that there is a difference between built-in VAR (and VEC) analysis within core gretl on the one hand, and a more general setup in the so-called SVAR addon on the other hand. (I agree that the difference is not very transparent to the user, and we probably should improve the situation somehow... @Allin, @jack) - Built-in: This is what you get when you go to Model/Time series/Multivariate/Vector Autoregression (or ... VECM). After estimation you get a model window where one of the menus says "Plots" and below that "Impulse responses (combined)" and some "Response of..." entries. There you can also do a bootstrap, but the only method for identification is the old-school Cholesky approach, where you choose the causal ordering of the variables. - Still Built-in: Here the documentation of the native irf() function is relevant, which -as Allin cited before- says: "...the estimated response of the target variable to an impulse of one standard deviation in the shock variable." So that is clear (but it might be worthwhile to copy that information into section 29.3 of the user guide as well). - SVAR addon: To run this you either go to the function package list ("fx" button in the main window) or if it has already been added to the menus, it's under Model/Time series/Multivariate/Structural VARs. There you can impose other identification schemes, not just Cholesky, and you have more refined bootstrap options. - Still addon: As noted in the thread in 2017, the numerical IRF results are the same for the built-in and the addon variants when the models are the same (modulo the typical minor differences in how the variance is estimated, with or without a degree-of-freedom correction). So the shock size convention definition is the same, too: One standard deviation. (I agree that the explanation in the SVAR doc on p.10 is not super clear, either.) - Addon continued: As also explained on p.10 of the SVAR doc, there is a "normalize" option (which however you can only access via scripting, not in the graphical interface) by which the shocks are rescaled such that each k-th shock will imply a unit response on impact on the k-th variable. This is what typically is called a unit shock size. Sorry, this has become longer than intended. If anybody (especially Jack) spots any mistakes or more confusion, please correct me. cheers sven