maximization issues including positive log likelihood

[William Lisowski]

I normally wouldn't "bump" another thread, but there are maximization issues in the thread linked to below that the title doesn't reveal. The user has forced convergence with a set of maximization tolerance options that I question, and in the final example in the post linked, the iteration log displays a positive log likelihood, which I also question, to say the least: it seems a fairly un-selfaware result for a statistical package to report.

I gave some initial advice on other aspects of the poster's problem, but lack experience in the details of maximization. If you have experience and expertise in maximization of difficult problems, please take a look.

https://www.statalist.org/forums/for...34#post1401134

[Nick Cox]

There is nothing impossible about positive log likelihood. http://blog.stata.com/2011/02/16/pos...values-happen/

However, I can't advise on use of xtfrontier.

[William Lisowski]

Thanks for the link on positive log likelihoods. I copied that to the referenced topic, where makes a certain amount of sense of where the positive log-likelihood was observed.

Let me clarify the intent of my posting this topic, which I titled "maximization issues including positive log likelihood" in hopes of distinguishing my concerns from concerns about xtfrontier.

While original topic in question is about the use of xtfrontier, that command uses Stata's common maximization code and accepts the standard maximization options common to all commands that do so. So I expect there may be more general expertise that applies to the underlying issue of maximization in Stata, even in the absence of specific experience or expertise with xtfrontier.

The user is having difficulty with convergence and has made a (likely ad hoc and unmotivated) choice of tolerance options that result in Stata declaring convergence, in particular

tol(1) ltol(1) nrtol(1)

where the defaults are on the order of 1e-5 or less, and then writes

I have no idea if this has an effect on the reliability of the obtained estimates..

As I wrote in post #1, I am soliciting anyone who has experience and expertise in maximization of difficult problems to comment, not on the interpretation of xtfrontier results, but on whether the maximization results are likely to represent actual convergence.