STata

[Jukka Ollgren]

Stata 16 mi mlogit CI bugs?

[Jukka Ollgren]

Stata 16 mi mlogit nonlinear parameter estimates calculates variances wrong. Point estimates are ok but their CI:s are wrong. For example,

uusi_outcome | exp(b) Std. Err. t P>|t| [95% Conf. Interval]
---------------------------+----------------------------------------------------------------
Successful | (base outcome)
---------------------------+----------------------------------------------------------------
Died |
1.male | 8.575755 7.665708 2.40 0.016 1.487302 49.44765


but using nonlinear estimate command: the CI is completely wrong! The point estimates match.


male: exp([Died]1.male)

(Within VCE adjusted for 21 clusters in shp)
------------------------------------------------------------------------------
uusi_outcome | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
male | 8.57648 7.667113 1.12 0.263 -6.450786 23.60375
------------------------------------------------------------------------------
When this is gonna be fix it?

[Jukka Ollgren]

Sorry: the variances are not wrong but the CI:s.

[Richard Williams]

Jukka, your output is very hard to read. You should use code tags. See pt 12 of the list FAQ on asking Qs effectively.

Also read this FAQ and see if it addresses your concerns.

https://www.stata.com/support/faqs/s...cs/delta-rule/

[Jukka Ollgren]

Hi, Thanks.

This exp is so nonlinear that it can cause troubles in roughs approximations like this delta method is sometimes. Actually, the problem was another thing but I thought that the problems in calculating variances could be the cause.

When square of the age should be in the model and one use also linear term age there and the linear age is highly nonsignficant in one outcome class, it could be relevant in other outcome classes and in the interactions (and maybe the hierarchical model reason it should be there age if age squared etc?). If I calculate the total age effect in different ages in the outcome class in mlogit, the confidence intervals are very large so that it seems that there is "no" age effect at all in different individual ages but there is very strong age effect which can be seen using only age as a covariate or from predictive margins (when the quadratic term in in the model).

[Richard Williams]

Jukka, without seeing your commands and output (in legible form) it is very hard to comment on your concerns. You are now talking about age and age^2, and neither was in the brief example you did post. If you really think Stata is doing something wrong, you should contact tech support.

Just from what you say, it is not clear to me that there is an error. These large CIs may just reflect statistical reality.

I also wonder if bootstrapping would be a better way for you to get the CIs if you don't like what Stata is giving you. Or, maybe you should use lincom to provide some sort of joint test involoving age and age^2.

I'm not saying you are wrong. I am just saying you are providing too little information for us (or at least me) to really understand your concerns.