Likelihood ratio tests and Markov Switching models

[paldama]

Hi,
I am currently working on EV 8 Student (MacOS), and I try estimating Markov Switching models.
Beside t-test and Wald test, I would estimate a MS model with linear constraints in order to do a likelihood ratio test.
I see that there is a routine implementing a LR ratio test for omitted variables bias -- but not for linear restrictions: Is there any way to do it "manually" ?
Thanks for any answer!
Pierre A.

[EViews Glenn]

Quite possibly no, but what kind of restriction do you wish to impose?

[paldama]

Linear restrictions such that c(1) = a * c(2).
I know I can do it through Student and Wald tests, but I'd like to compare these results to a likelihood ratio test.
More generally I was asking if there is any procedure to estimate a constrained MS regression.
Thanks for your answer.

[startz]

While there isn't a general procedure, it's usually possible to impose linear restrictions manually by substitution. In your example, wherever c(1) appears substitute a*c(2).

[paldama]

While there isn't a general procedure, it's usually possible to impose linear restrictions manually by substitution. In your example, wherever c(1) appears substitute a*c(2).

I just can't see how to do it in a Markov-switching model framework.

That would be easy implementable if I had to declare manually each regime equation: Suppose the following two-state MS model. I would estimate this model:
Eq1: y(t) = c(1) + c(2)*x(t) + c(3)*z(t) + u(t)
Eq2: y(t) = c(4) + c(5)*x(t) + c(6)*z(t) + u(t)
And then estimate this one, applying a restriction: c(2)=a*c(5):
Eq1: y(t) = c(1) + a*c(5)*x(t) + c(3)*z(t) + u(t)
Eq2: y(t) = c(4) + c(5)*x(t) + c(6)*z(t) + u(t)
Finally, I would implement a LR test, to test this restriction.
..but as far I know, the MS procedure in Eviews does not allow it.

[EViews Glenn]

I can't think of any easy way to impose that restriction. The closest you'll be able to do is to use a Wald test.

[startz]

If a=1, you could make x a nonswitching regressor. I admit I don't see a general solution.

[paldama]

Thanks for your replies.
Marcelo Perlin's MS_Regress Matlab package (https://sites.google.com/site/marcelope ... -in-matlab) can estimate constrained MS regressions.
But I cannot connect my Eviews (Student version) to Matlab to use it. And Eviews is obviously easier to use, to estimate and to test various models, controlling for serial correlation, etc.
Best regards.

[trubador]

Start by searching the journal databases for articles that has studied a similar subject and/or methodology. Also, a(ny) textbook that devotes at least a chapter on Markov Switching (MS) Models would do fine. After building your model, you can come back and ask for the practical aspects of the estimation phase, if necessary. Keep in mind that there are various forms of switching dynamics. As of version 9, EViews will be covering most of the commonly used univariate switching models (STAR models in addition to MS models to be exact).

Before all that, however, you should really provide sufficient evidence in favor of switching behavior in interest rates. Once you reject the one-state structure, the problem will no longer be confined to finding appropriate variables and/or lags or any other traditional diagnostics. You will be deciding not only the number of regimes but also the correct form of switching dynamics. At this point, economic and econometric theory will be your guide.

[Econometrist]

I searched methodology and smiler subject for example Hamilton and Engel.
And according to them I can use Markov switching model in interest rate.
And I decided my data that is monthly interest rate for 25 years for 2 countries.
I decided models and my first model is ar(1) ar(2) ar(3) ar(4) and my second model is ar(1-4) ma(1)
Now my first question is can I use 'ma' model in markov switching model in eviews 8?
And my secon question is should I start which test and for what?
I didn't understand how can I choose number of regime?
Please help me
Thank you very much