Interpretation of lasso (probit) output

[Tor Haug Anonsen]

Hello

I was wondering if someone could help me with interpreting what it means when stata gives me the message "convergence for the lasso penalty = x.xx not reached after 100000 iterations; solutions for larger penalty values returned. I did a similar try with linear lasso regression (with a different dependent variable of course), and it did not give the same "error" message. I know the methods are different as were the dependent variables, but the dataset itself was practically the same. If someone could give me any input to why i get this message, it would be much appreciated.

[Tor Haug Anonsen]

Also, in the lasso linear model the 5th column display out of sample R-squared, which is easy to interpret. With lasso probit, it shows out of sample dev.ratio. Does anyone know the difference?

[Ryan Mather]

Hi Tor,

I received the same error message when doing a linear lasso regression. After looking into it a bit, I think the issue has something to do with how at lower lambda values the lists of variables that Stata is trying to decide whether or not to include in the model is larger, and so the problem of picking a final list of variables is more difficult. Does anyone know how to make the maximum number of iterations higher than 100000?

Using the option "gridminok" will allow the code to keep going after this error message, and just pick the best lambda that it was able to calculate. I think pages 212 and 226 of the Stata reference for Lasso (https://www.stata.com/manuals/lasso.pdf ) talk about the sample dev.ratio. It is a way of measuring the additional "predictive ability" that you get by moving to the next lambda value.