BUG: Correct derivative for exponential transform. #1540

Merged
merged 1 commit into from Apr 2, 2014

Projects

None yet

2 participants

@jseabold
Member
jseabold commented Apr 2, 2014

See if this fixes #1539.

@jseabold
Member
jseabold commented Apr 2, 2014

It looks like it fixes the issue of score always matching numerical score, but this model is still bizarre. Might try comparing output to stata.

@jseabold
Member
jseabold commented Apr 2, 2014

See #1539. It appears to be sensitive to starting values. You need to make sure alpha >= ~1 or lnalpah >= ~0.

@jseabold
Member
jseabold commented Apr 2, 2014

Merge? Marking for 0.5.1.

@jseabold jseabold added this to the 0.5.1 milestone Apr 2, 2014
@josef-pkt
Member

wait for TravisCI to report back, then merge

BTW don't forget to add labels, it looks like we can add them directly now also to PRs

@jseabold
Member
jseabold commented Apr 2, 2014

Sure. Tests passed locally, which means this was untested. I think we only tested the default solves for which this wasn't such an issue.

@josef-pkt
Member

I think this is only "half" a bug, and I thought it was tested.

At the optimum, the score is approximately zero. multiplying it by a factor doesn't make any difference, (except for convergence tolerance in score which, I think, we don't check.)
postestimation results where based on untransformed parameters and were also correct.

The only problem was in the optimization itself, especially further away from the optimum or with a messy case like this.

@jseabold
Member
jseabold commented Apr 2, 2014

Makes sense. Still conceptually incorrect. Merging. Also marked to backport to 0.5.1. Need to go back and check those with milestone tags and no backport labels.

@jseabold jseabold merged commit 0fd2925 into statsmodels:master Apr 2, 2014

1 check passed

default The Travis CI build passed
Details
@jseabold jseabold deleted the jseabold:fix-1539 branch Apr 2, 2014
@josef-pkt josef-pkt added the PR label Apr 14, 2014
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment