Different T-Statistics

[grkn7878]

I would like to ask why same rats programme gives different t-statistics. I run rats programme on my computer and my advicer computer it gives different t-statisticwhile our data and programme are same.


Best Regards.

[TomDoan]

The standard errors that are generated by using BFGS are only an approximation, and can change (sometimes noticeably) with minor changes to the guess values (which can change from version to version of RATS), convergence criterion, preliminary estimation method, etc.

The first of these is from the first GARCH instruction in the GARCHMV.RPF example, the second is for the same model with CVCRIT reduced to .00000001 and with more accurate numerical derivatives (NLPAR(DERIVES=FOURTH)).

Label     Coefficient  Standard Error  T-Stat    Signif
10. A(1,1)    0.105805805  0.006893563     15.34849  0.00000000
11. A(2,1)    0.093930096  0.005858416     16.03336  0.00000000
12. A(2,2)    0.128175482  0.007995778     16.03040  0.00000000
13. A(3,1)    0.088809390  0.005381811     16.50177  0.00000000
14. A(3,2)    0.113629169  0.006960113     16.32577  0.00000000
15. A(3,3)    0.111519241  0.006749644     16.52224  0.00000000



10. A(1,1)    0.105428635  0.007217789     14.60678  0.00000000
11. A(2,1)    0.093479002  0.005563105     16.80339  0.00000000
12. A(2,2)    0.127382082  0.007391909     17.23264  0.00000000
13. A(3,1)    0.088408633  0.005235857     16.88523  0.00000000
14. A(3,2)    0.112950912  0.006166255     18.31759  0.00000000
15. A(3,3)    0.110899113  0.006045944     18.34273  0.00000000

As you can see, the point estimates typically agree to about 3 significant digits (which is probably all you would ever show in publishing the results), but the standard errors are rarely the same beyond the first digit. And if you use BHHH rather than BFGS, you'll get (perhaps quite different) standard errors and the same if you add a ROBUSTERRORS option.