Markov switching regression - different estimates each run

For technical questions regarding estimation of single equations, systems, VARs, Factor analysis and State Space Models in EViews. General econometric questions and advice should go in the Econometric Discussions forum.

Moderators: EViews Gareth, EViews Moderator

nesreen
Posts: 8
Joined: Fri Jun 20, 2014 3:38 pm

Markov switching regression - different estimates each run

Postby nesreen » Fri Mar 19, 2021 7:21 pm

Hello,

I am using Eviews 11 university edition on mac and estimating a Markov switching regression model using Switchreg-Switching regression method. I keep getting different estimates (and significance levels) every time I re-estimate the regression equation using same data, knowing that I am selecting "EViews supplied" in the starting values options. I read in the manual that choosing "User supplied" would mean using values stored in C coefficient vector at the time of estimation as starting values which would explain getting different estimates depending on which regressions where run first. But, if I am choosing "EViews supplied" how are the starting values determined? and why do I end up with different estimates each time? and what can i do to solve it?

Would really appreciate your help,
Thank you.

EViews Gareth
Fe ddaethom, fe welon, fe amcangyfrifon
Posts: 13401
Joined: Tue Sep 16, 2008 5:38 pm

Re: Markov switching regression - different estimates each run

Postby EViews Gareth » Fri Mar 19, 2021 10:49 pm

Could you provide the workfile?
Follow us on Twitter @IHSEViews

nesreen
Posts: 8
Joined: Fri Jun 20, 2014 3:38 pm

Re: Markov switching regression - different estimates each run

Postby nesreen » Tue Mar 23, 2021 8:10 am

Hello,

Thank you for your prompt reply. kindly find the attached workfile. Every-time I estimate the equation using Markov switching method, I end up with different coefficient estimates as well as different significance.

First, I estimated eq1 using OLS (eq1ols) then estimated it using Markov switching model (eq1markov).
Then, I estimated eq2 using OLS (eq2ols) then estimated it using Markov (eq2markov).
Then, I re-estimated eq1 using Markov (eq1markov_reestimation) and as you can see the coefficient estimates have changed and all the standard errors and p-values displayed "NA".
I did the same for eq2, and re-estimated eq2 using Markov (eq2markov_reestimation) and the estimates have changed and all the standard errors and p-values displayed "NA" as well.
Afterwards, I re-estimated eq1 using OLS then re-estimated it again using Markov (eq1markov_reestimation2) hoping to end up with the same initial results, however, he coefficients significance as well as the estimates were different. The same conclusion was drawn when I did the same for eq 2, re-estimated eq2 using OLS then re-estimated it again using Markov (eq2markov_reestimation2).

All Markov models are estimated after selecting "EViews supplied" in the starting values options.

Would really appreciate your help,
Thank you.

EViews Gareth
Fe ddaethom, fe welon, fe amcangyfrifon
Posts: 13401
Joined: Tue Sep 16, 2008 5:38 pm

Re: Markov switching regression - different estimates each run

Postby EViews Gareth » Tue Mar 23, 2021 9:20 am

EViews supplied starting values uses random numbers. That's why there's the random number settings there. If you use the same random seed across estimations, you'll obtain the same results.
Follow us on Twitter @IHSEViews

nesreen
Posts: 8
Joined: Fri Jun 20, 2014 3:38 pm

Re: Markov switching regression - different estimates each run

Postby nesreen » Tue Mar 23, 2021 6:28 pm

Thank you again for your prompt reply. I read the manual again and found the point you are referring to "if you attempt to replicate the estimation using the default settings, you may obtain different results due to the different set of random starting values. You may use the random number generator seed settings to obtain the same starting values". However, I still cant fully grasp how changing the seed drastically changes the results, not only changing the values of the estimates, but the sign and significance as well. And if this is the case, Is there any criteria to help choose among the different outputs generated using different seeds before sticking to a specific seed? or do we choose arbitrary?
Thank you again, really appreciate your help.

EViews Gareth
Fe ddaethom, fe welon, fe amcangyfrifon
Posts: 13401
Joined: Tue Sep 16, 2008 5:38 pm

Re: Markov switching regression - different estimates each run

Postby EViews Gareth » Tue Mar 23, 2021 6:47 pm

That’s more of a theoretical question.
Follow us on Twitter @IHSEViews

startz
Non-normality and collinearity are NOT problems!
Posts: 3779
Joined: Wed Sep 17, 2008 2:25 pm

Re: Markov switching regression - different estimates each run

Postby startz » Wed Mar 24, 2021 6:28 am

Since Markov switching is maximum likelihood, you want to pick the estimate with the highest log-likelihood.

nesreen
Posts: 8
Joined: Fri Jun 20, 2014 3:38 pm

Re: Markov switching regression - different estimates each run

Postby nesreen » Wed Mar 24, 2021 9:01 am

Thank you both so much for your help.


Return to “Estimation”

Who is online

Users browsing this forum: Google [Bot] and 2 guests