NLS: convergence / starting values

[tvonbrasch]

Hi,

i am estimating an Error Correction model using NLS (see attached workfile):

for %industry 15 25 30 45 50 55 63 64 71 74 81 85
equation ass3nls_{%industry}.ls(cov=hac, c=1e-10, m=10000) d(omegahigh{%industry})=c(1)+c(2)*d(wagepremium{%industry})+c(3)*d(logkint{%industry})+c(4)* ((omegahigh{%industry}(-1))-c(5)*(wagepremium{%industry}(-1))-c(6)*(logkint{%industry}(-1))-c(7)*@trend)
next

But, if you look at the equation Ass3NLS_15, it has not converged and the estimates do not make sense. If I increase the number of iterations (m) it does not help. I have also tried to change the convergence criterion and the initial parameter values.

If I however manually open the equation object, click the estimation button and press ok, the equation is reestimated and it converges. How should i change the code above so that the NLS estimation converge in one step?

Sincerely
Thomas von Brasch

[EViews Glenn]

Nonlinear estimation results can depend on starting values. Every time you go through the loop you are estimating beginning with the values from the previous industry. In this one case, those values do not work well. You might want to try estimating a restricted version of the model for the industry to get starting values prior to doing the nonlinear estimation

[tvonbrasch]

Thanks for you reply Glenn, i will try that.

But is it not strange that it converges when i estimate the model manually by clicking the estimate button, and that it will not converge if I estimate the model as a part of a program?

why is that?
Thomas

[EViews Glenn]

If there is anything in causes the starting values to differ in the two cases then it is not strange. When you go through the loop, the values that are in the starting values prior to estimation are those from the previous estimates. That, I presume, is not the case when you do it manually.