GARCH (1,1)-M

[Tinoosh16259]

I have applied GARCH-M to analyse significance of volatility spillover effects from some developed countries to developing countries. My question is when I plug in lagged squared residuals of e.g., U.S. in Iran volatility equation to analyse significance of volatility spillover effects from U.S. stock market to Iran stock market, in comparison to when I plug in lagged squared residuals of Germany to analyse significance of volatility spillover effects from Germany stock market to Iran stock market, why significance of the variables in Iran mean equation plus GARCH and ARCH effects in Iran volatility equation is changing quite significantly? In the other word, why changing one exogenous variable, lagged squared residuals of U.S. to Germany, in the volatility equation of Iran changes significance of variables in the mean and volatility equation of Iran quite significantly? Thank you for your help in advance?
Cheers
Tinoosh

[trubador]

In GARCH models, variance is assumed to be time-varying conditional on its past values, which makes them highly data dependent. And keep in mind that conditional variance is actually unobserved and has to be extracted. Parameter estimations of GARCH models are (usually) obtained via maximum likelihood techniques through optimization. Since GARCH models are nonlinear, they are also sensitive to initial conditions and may experience local optima problems. Therefore, each time you build a new model or change the specifications of a current one, it is quite possible that you get different results. Garch-in-mean models further complicate the problem by adding this unobserved variable into the mean equation.

As an aside, volatility spillover effects are better handled through Multivariate GARCH models. Please see the following discussion: http://forums.eviews.com/viewtopic.php?f=4&t=1298

[Cabinho]

Hi,

If I am working with cross sectional data with tobit (binomial dependent variable) and logit (categorical dependent variables) models and have an ARCH structure in any of those, should I run the regression with an ARCH method instead of tobit or logit?

Also, will the ARCH method fix serial autocorrelation as well?

Thank you

[trubador]

How do you detect the ARCH effect in these settings?

[Cabinho]

Running an OLS first and testing for ARCH.

I figured if I ran an OLS first with the model I want to use to test for ARCH and find problems, I would have those same problems in the same model when running the regression with tobit or logit even though I cannot test for those problems after running tobit and logit.

I also tested for White heteroskedasticity and serial autocorrelation.

I am obviously a beginner but what do you think?

[startz]

Serial correlation or ARCH with cross sectional data is kinda weird, although not impossible.

[Tinoosh16259]

Thank you for your reply trubador. Also I have got one more question, about fulfilling the stationary condition of the volatility specification, my question is that I have got some cases where ARCH+GARCH effects are more than one (1); and do not know how to solve that, I want to know:
1. how can I solve that, considering the fact that I have included deviations in the mean equation instead of variances; and
2. If there is no way to get red of this matter, how should I explain this? what are the effects of not fulfilling the stationary condition?

Thank you for your help!
Cheers
Tinoosh

[Tinoosh16259]

also I wish to know why one country's spillover effects to my selected dependent country is "not significant" when I am considering standard deviation in the mean equation (GARCH-M), but the same country's volatility spillover shows "significant" effects through GARCH, not considering deviation/variance in the mean equation?
Thank you again!
Cheers
Tinoosh

[Cabinho]

Serial correlation or ARCH with cross sectional data is kinda weird, although not impossible.

Thank you. Ok, so I have three questions:

1. I am trying to propose tobit and logit models but I have been running a OLS first with the same variables to check for ARCH, serial correlation and heteroskedasticity to see if I will have these problems in my tobit or logit regressions. Is this acceptable?

2. If I find that I do have ARCH or serial correlation? Should I run the regression with the ARCH method instead even though my models use binary and categorical variables? If not then how can I solve these problems in a tobit or logit model?

3. How can I solve serial autocorrelation in tobit and logit models?

Thank you for your help

[startz]

Serial correlation or ARCH with cross sectional data is kinda weird, although not impossible.

How can I solve ARCH in a Tobit or Logit model? Should I run the regression with the ARCH method instead even though my models use binary and categorical variables?

What about serial correlation?

Thank you for your help

Both serial correlation and ARCH require your data to be ordered sequentially, which is not common with cross-sectional data. You may want to give a more extended explanation of what you're doing.

[Cabinho]

Serial correlation or ARCH with cross sectional data is kinda weird, although not impossible.

How can I solve ARCH in a Tobit or Logit model? Should I run the regression with the ARCH method instead even though my models use binary and categorical variables?

What about serial correlation?

Thank you for your help

Both serial correlation and ARCH require your data to be ordered sequentially, which is not common with cross-sectional data. You may want to give a more extended explanation of what you're doing.

My data is not ordered sequentially since is cross sectional however when I run the OLS and then test for ARCH it appears that I have ARCH. But it is cross sectional with categorical dependent variables so I am trying to figure out how to solve this.

[startz]

The ARCH model assumes the data is sequential. You might think about whether your data has been inadvertently sorted in some interesting way.

[Cabinho]

The ARCH model assumes the data is sequential. You might think about whether your data has been inadvertently sorted in some interesting way.

Interesting. I will check that.

One last questions: If I add a independent lagged variable in this type of cross-sectional study such that is X=X(-1) but my data is not ordered sequentially and the variable appears significant in the regression. Could it be still interpreted as the past value of X related to the present value of Y?

[startz]

The ARCH model assumes the data is sequential. You might think about whether your data has been inadvertently sorted in some interesting way.

Interesting. I will check that.

One last questions: If I add a independent lagged variable in this type of cross-sectional study such that is X=X(-1) but my data is not ordered sequentially and the variable appears significant in the regression. Could it be still interpreted as the past value of X related to the present value of Y?

No. X(-1) is simply the previous observation in whatever order you stored the data points in the workfile. In a cross-section this has nothing to do with dates.

[Cabinho]

The ARCH model assumes the data is sequential. You might think about whether your data has been inadvertently sorted in some interesting way.

Interesting. I will check that.

One last questions: If I add a independent lagged variable in this type of cross-sectional study such that is X=X(-1) but my data is not ordered sequentially and the variable appears significant in the regression. Could it be still interpreted as the past value of X related to the present value of Y?

No. X(-1) is simply the previous observation in whatever order you stored the data points in the workfile. In a cross-section this has nothing to do with dates.

You have saved my life so many times already today.

So really if I check for autocorrelation and the correlogram shows my dependent variable is autocorrelated should I just ignored this because it would not make sense to include y(-1) as an independent variable. Should I just rearrange the way my data is ordered?