ml maximize didn't converge but the log likelihood and ptol don't improve

[Zhaohui Li]

I am running a NLLS regression.
After several hours, the bfgs technique gives flat region so I turned to the nr and bhhh technique, now the log likelihood stop to improve.
the Log file shows like this for hundreds of iteration and improve little for the whole night,

Iteration 70: log likelihood = 20955.101 (not concave)
vtol = 1.673e-10
Iteration 71: log likelihood = 20955.101 (not concave)
vtol = 2.844e-10
Iteration 72: log likelihood = 20955.101 (not concave)
vtol = 4.072e-11
Iteration 73: log likelihood = 20955.101 (not concave)
vtol = 1.363e-10
Iteration 74: log likelihood = 20955.101 (not concave)
vtol = 9.474e-11
Iteration 75: log likelihood = 20955.101 (not concave)
vtol = 5.717e-11
Iteration 76: log likelihood = 20955.101 (not concave)
vtol = 2.980e-11
Iteration 77: log likelihood = 20955.101 (not concave)
vtol = 6.942e-11
Iteration 78: log likelihood = 20955.101 (not concave)
vtol = 3.047e-12
Iteration 79: log likelihood = 20955.101 (not concave)
vtol = 9.194e-11
Iteration 80: log likelihood = 20955.101 (not concave)
vtol = 9.146e-12
Iteration 81: log likelihood = 20955.101 (not concave)
vtol = 5.839e-11
Iteration 82: log likelihood = 20955.101 (not concave)
vtol = 1.425e-11
Iteration 83: log likelihood = 20955.101 (not concave)
vtol = 3.658e-11
Iteration 84: log likelihood = 20955.101 (not concave)
vtol = 1.072e-11
Iteration 85: log likelihood = 20955.101 (not concave)
vtol = 3.636e-11
Iteration 86: log likelihood = 20955.101 (not concave)
vtol = 2.069e-11


I tried the"ml maximize, difficult " option, but the result is still the same.
What should I do now??

[Joseph Coveney]

I am running a NLLS regression.
After several hours, the bfgs technique gives flat region so I turned to the nr and bhhh technique, now the log likelihood stop to improve.
the Log file shows like this for hundreds of iteration and improve little for the whole night, . . . I tried the"ml maximize, difficult " option, but the result is still the same.
What should I do now??

I recommend re-assessing your model with the thought of simplifying it.

If you manage to get it to converge with the simplified model, then you can try again using the fitted coefficients of the simpler model as starting values for your current one.

But if that doesn't work, then you might eventually have to accept that your current set of data won't accommodate your current model.

[Zhaohui Li]

I recommend re-assessing your model with the thought of simplifying it.

If you manage to get it to converge with the simplified model, then you can try again using the fitted coefficients of the simpler model as starting values for your current one.

But if that doesn't work, then you might eventually have to accept that your current set of data won't accommodate your current model.

Thanks for your advice. Here, we actually use 7th order septic equation with 200+ dummy variables. The reason why we use 7th equation is because it fits the theory better and the semipar has already test a significance of 7th septic equation.
Do you mean that I need to lower down to a 5th quintic equation?

[Joao Santos Silva]

In addition to Joseph's sound advice, I suggest that you change the scale of your dependent variable in such a way that the objective function has values of around 1000. Surprisingly, Stata is sensitive to the scale of the objective function.

Joao