Ask for help about State Space model wiht big z-Statistic

[smilesym1982]

today, i use the following equations
lex = c(1)+c(2)*lex(-1)+sv1*lex_pri+[var=exp(c(3))]
@state sv1=sv1(-1)
the estimation result amazed me, the z-Statistic is more than one hundred thousand,what wrong with my result, who can help me?
the estimation result:

Sspace: S_PRI
Method: Maximum likelihood (BHHH)
Date: 06/14/09 Time: 15:08
Sample: 2004M01 2009M12
Included observations: 72
Valid observations: 58
Estimation settings: tol= 0.00010, derivs=accurate numeric
Initial Values: C(1)=14.4758, C(2)=0.72838, C(3)=0.48158
Convergence achieved after 1 iteration

Coefficient Std. Error z-Statistic Prob.

C(1) 14.47580 0.000839 17254.59 0.0000
C(2) 0.728377 3.47E-06 210001.5 0.0000
C(3) 0.481586 5.31E-06 90765.00 0.0000

Final State Root MSE z-Statistic Prob.

SV1 2.977182 0.001927 1545.109 0.0000

Log likelihood -1333230. Akaike info criterion 45973.56
Parameters 3 Schwarz criterion 45973.67
Diffuse priors 1 Hannan-Quinn criter. 45973.61

[trubador]

It is difficult to detect the error (if any) without seeing the actual workfile besides the estimation output. Below are the few points I've noticed so far:

1) Your equation can be estimated much more easily with OLS, since the state variable is time invariant. The following code will equally serve the purpose:

Code: Select all

equation eq1.ls lex c lex(-1) lex_pri

2) You have missing values in the sample. It means you force the estimation to fill in these missing values. You should appropriately set your sample period for the estimation, if your intention is not forecasting of these values.
3) The initial values are very close to actual estimation results (as also stated by:"Convergence achieved after 1 iteration"). Either you have run the model more than one in a row or you have supplied very good starting values. When you change the specification of your model and run it or re-run the same model, EViews uses the values in the coefficient vector (i.e. c in your workfile) as the starting points and overrides them after each estimation.
4) You selected BHHH method for the optimization algorithm. The covariance matrices can differ significantly with respect to chosen optimization method.

[smilesym1982]

HI， I have added the eviews workfile on the attachment, so everyone can check my problems on the workfile, thanks a lot!

And i have checked trubador'solutions, but none of them suit my problem, z-Statistic is too big, who can help me to solve it, thanks!

while i have anther problem: when i run the estimation each time, the estimation result will change, and how i choose which model is the best one!

By the way, i want to test the varibles' (dex_pri, dex_qua, dex_str) time Time-varying influence on the varible dex, so i choose the state space model.

Thanks a lot, everyone!

[trubador]

Actually your problem suits to the second point in my previous post. Your sample period should be between 2004M03 and 2009M01 (not 2004M01 and 2009M12). Otherwise, you are forcing the model to estimate the missing observations. If you adjust the sample that way you'll obtain more feasible results.

The change in estimation results is due to change in starting values each time, which corresponds to the third point in my previous post.

If you are interested in estimating the time varying parameters, then you have to define coefficients in a dynamic manner (e.g. random walk or AR process). In your case, state variable is time invariant and estimation results will be very close to those of OLS. And this corresponds to the first point in my previous post.

By the way, if you search the forum, you'll find plenty of discussions on this type of modeling in EViews.