Dear Feriel Dermechi

First, Akaike's information criteria must theoretically be significant and negative; otherwise, your result will not be correct. Regarding the lag selection criteria, I suggest you should leave it to the E-views software. In other words, do not use 2 lags or sect lag criteria by yourself. The software can appropriately select the lags. Just follow the function. Estimating VAR with the maximum likelihood method may eliminate the problem but is not sufficiently accepted. I hope this helps you.

Dear Asrat Araya,

Thank you for your answer.

I proceeded in the same way about the lag order but I realized that the residuals do not follow a normal distribution whatever the lag order, the residuals follow a Student's distribution. In this case, isn't maximum likelihood estimation most appropriate???

how to proceed with this estimation (ML) on eviews? otherwise, how to maximize LogL function on eviews command?

Thanks again for your help.

Dear Feriel Dermechi

You need to check whether the model has serial correlation or not, by checking the Residual Cross Section Dependency Test.

Thank you Sir.

I checked the serial correlation at first step, the series are independents...

Dear Feriel Dermechi

You are saying that errors in the model are serially independent. Right? If so, it is good. Because making sure that errors in the model are serially independent is one of the steps in analyzing data.

But another important step that you have to make sure of is that the model is dynamically stable. So, to ensure that the model is dynamically stable, you can exclude one of your variables from further testing and test the model's stability with the remaining variables. Finally, you can explain it in your result discussion as justification.

There are many relevant ECB working papers on the ECB website. You might start with WORKING PAPER NO. 6 THE DEMAND FOR M3 IN THE EURO AREA BY GÜNTER COENEN AND JUAN-LUIS VEGA (https://www.ecb.europa.eu/pub/pdf/scpwps/ecbwp006.pdf?14e04cc5febe813648551a4836f7ead1) or WORKING PAPER NO. 91 THE MONETARY TRANSMISSION MECHANISM IN THE EURO AREA: MORE EVIDENCE FROM VAR ANALYSIS BY GERT PEERSMAN T PEERSMAN AND FRANK SMETS (https://www.ecb.europa.eu/pub/pdf/scpwps/ecbwp091.pdf?1da206533d7d8e412a87eac433780af0). Working Paper 91 is from a conference on the monetary transmission mechanism and there are other papers from that conference on the list of papers.

In summary, you do need to do a bit more thinking about your model, data, and institutional details before you do your econometrics.

Asrat Araya You say

"Akaike's information criteria must theoretically be significant and negative;"

What do you mean by significant? Different software packages use different formulae for AIC and other information criteria. Even in a particular econometric package, different routines may use different formulae. Some may report a positive AIC which will be a positive monotone transformation of "your negative" value and give rise to the same selection. Have a look at exercise 7 on page 114 of Enders (2014), Applied Econometric Time Series, Wiley, that explains this in detail. The idea of blindly following what EVIEWS recommends is not the best idea.

You test model stability by some form of recursive estimation.

Hello John C Frain

Thank you very much for your advice regarding E-views vis other software. The link you provide is for STATA Users, from pages 113 to 114, which is very good for understanding the usefulness of STATA software. But this is for *augmented D-F regression of unit root test (four lags) */ on page 113. Thank you once again.

Asrat Araya

There is no mention of STATA on pages 113 or 114 of Ender's Textbook. Wiley has published separate instructors manuals for EVIEWS, STATA, RATS and SAS There is also an R manual available from https://sites.google.com/view/davidgabauer/econometric-code. The book itself does not depend on any software. The point being made is relevant to any calculation of AIC.

Are you looking at the STATA instructor's manual?

Feriel Dermechi How did you conclude that your residuals followed a t-distribution? How did you decide on degrees of freedom? A t-distribution with nu degrees of freedom has mean 0 and variance nu/(nu-2). Perhaps you intend some form of skewed generalized t-distribution (https://en.wikipedia.org/wiki/Skewed_generalized_t_distribution). I have not used EVIEWS for more than 10 years but I would think that there are better ways to implement maximum likelihood with this distribution for errors.

Remember that OLS will still give asymptotically optimum estimates of your coefficients even if your residuals follow a distribution such as this (provided the other requirements of OLS are satisfied).

John C Frain

Here you are

The link to the book is:

Enders, W. (2008). Applied econometric time series. John Wiley & Sons.

[CITATION] Applied econometric time series

W Enders - 2008 - John Wiley & Sons

Save Cite Cited by 16849 Related articles All 17 versions

Dear Feriel Dermechi

To assess and calibrate your theoretical model, you estimate a vector autoregression (VAR) model. To apply this method to E-views, you need to search and find out about the VAR model first and then use the function in E-views.

Asrat Araya The 2008 edition is the second edition. My reference is to the fourth edition. Exercise 13 on page 114 of the third edition is the same as exercise 7 on page 114 of the fourth edition. I do not have access to the second edition.

You are looking at the Instructor's manual for the second edition and not at the second edition of the actual textbook.

Asrat Araya The 2008 edition is the second edition. My reference is to the fourth edition. Exercise 13 on page 114 of the third edition is the same as exercise 7 on page 114 of the fourth edition. I do not have access to the second edition.

You are looking at the Instructor's manual for the second edition and not at the second edition of the actual textbook.

Dear Asrat Araya ,

Thank you very much for your reply and clarification.

I say that the variables series are not dependant , but the residuals series are dependants. It's the reason that i would to estimate by Likelihood Maximum method..... I don't know how to do :( :( :(

Dear Feriel Dermechi

Here are a few steps to perform the Likelihood Maximum

step 1: Go to the e-views function and Select VIEW,

step 2: Find and click on RESIDUAL DIAGNOSTICS,

step 3: Click on SERIAL CORRELATION Likelihood Maximum (LM) TEST,

Then you get results for Likelihood Maximum and the p-value

but make sure that the p-value is less than 0.05. Good luck.

Dear Asrat Araya,

The residuals do not follow a normal distribution so, VAR estimation under OLS is not consistent in this present case. The steps that you mentioned represent estimation of Gaussian VAR.

Thank you Sir.

In that case, go to the menu and click on quick and click on estimate equation and type on, in the dialogue box, y and your dependent variables by excluding the constant term as follows:

Y inflation exchange rate GDP interest rate depending on how you arrange them and click ok. Then you will find that all the factors as significant.

Feriel Dermechi

As far as I know, there is no direct way, using EVIEWS, to do maximum likelihood estimation of your VAR with some form of scaled t-distribution, In EVIEWS there are two possibilities

1. Use the Log Likelihood (LogL object). This is explained in Chapter 41 of the manual for Version 12 of EVIEWS and it would be very difficult to simplify that chapter here. You need to read it or seek local advice.

2. If this does not work you can use the EVIEWS command and programming reference.

The first is obviously preferred Documentation should be available on your EVIEWs Installation. I have seen scaled t-distributions used in Finance (risk analysis) but not in Economics. I have used maximum likelihood in R to estimate the parameters of GARCH systems with errors following a modified t-distribution. Details of the distribution are on page 126 of http://www.tara.tcd.ie/handle/2262/77570

Both of these approaches involve some programming of the likelihood function. If the optimization fails in method 1 you can try method 2. I have not used EVIEWS for over 10 years and do not have access to It at the moment. Thus I regret that I am not able to be specific about how you should do the estimation. the procedure that I have set out would be applicable in many econometric programs. Real econometrics is not easy.

I would suspect that after all that your estimates will not change much. While you may have learned some computational econometrics you may not have found any answers to your queries. As I explained in my previous answer I think that your model is misspecified and I have suggested several points that you might consider in revising your model. It may be easier to reconsider your model. If you are doing causal analysis the initial model and an understanding of your data is very important. Failure of normality and serial correlation tests are usually indications of model misspecification. In your case, you may be missing a money variable and possibly also "exogenous" external effects.

I find some of the other points on this quorum confusing,

1. "You need to check whether the model has serial correlation or not, by checking the Residual Cross Section Dependency Test." Residual cross-section dependency has nothing to do with serial correlation. I would expect residuals in a VAR to be contemporaneously correlated.

2. Dynamic stability is tested using a modified Chow test or using recursive estimation

Others I don't understand.

I hope that the noise here has not caused confusion. There are recommendations here that I don't understand. Perhaps I have misunderstood.

Dear John C Frain

Thank you very much for commenting on my answers. Please check the following video about serial correlation and cross-section dependency tests. Thank you once again.

Second Generation Unit Root Test|| EViews - YouTube

https://youtu.be/IkhRNTxewt8

Dear John C Frain

First, I have to apologize for the spelling mistakes, I'm very bad at English.

Second, I don't know how to thank you, thank you for your explanations, I really appreciate it.

Indeed, the maximum likelihood estimation requires a programming code on eviews. I applied the same code of a simple linear regression estimation , where the residuals follow a student distribution but I did not manage to do it correctly because the VAR contains the lags order and several equations at the same time,.........I ask myself a lot of questions at the same time....what the lags order that i should use that allows me the ML estimation since the estimate with OLS is biased in this case and the lag order choised by this method is also false .................... :(

Thank you for all your time and attention Sir.

Dear Asrat Araya

Thank you so much for all your time, effort and attention.

Asrat Araya Sorry but I did not see your answer addressed to me until Quora referred me to that question. The question is about Vector Autoregression (VAR). In general, the residual variance matrix of a VAR is not diagonal. One expects to see correlations between the residuals in VAR. Perhaps you have added the wrong links to your comment but the video you reference is about panel data estimation/testing and not about VAR - the subject of the question.

It does explain well how to do various panel data tests in Eviews. It adds very little as to why one would want to do such tests.