Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: st: "Can Your Results be Replicated?" (Stata error?)
From
[email protected] (Jeff Pitblado, StataCorp LP)
To
[email protected]
Subject
Re: st: "Can Your Results be Replicated?" (Stata error?)
Date
Thu, 19 Sep 2013 13:56:33 -0500
Stas Kolenikov <[email protected]> mentions that -mlogit- too will have
problems with perfect predictors:
> of other models/estimators, besides those brought up by Joao, -mlogit-
> will also be affected by the perfect prediction problems. How does
> Stata treat -mlogit- in this respect? I vaguely remember jumping
> through the hoops of specifying a bunch of dummy variable "this
> category vs. all others combined" to deal with it a few years back.
In addition, Richard Williams <[email protected]> mentions -glm-
and his own user-written commands:
> Well, while we are at it, we can go after glm too:
> webuse repair, clear
> logit foreign b3.repair
> glm foreign b3.repair, link(logit) family(binomial)
> My own user-written programs, oglm and gologit2, also have trouble with
> perfect prediction. My experience has been that when there is a problem, you
> get monstrous standard errors and extremely small test statistics. I don't
> know if that is what always happens, but it is hopefully some sort of clue
> that there is a problem somewhere.
Currently -mlogit- has no special code for detecting perfect predictors.
We will look into this, but I make no short-term promises for this one.
As for -glm-, we agree with Richard and will look into adding the checks for
perfect predictors to family binomial, the bernoulli case.
As for Richard's experience with perfect predictors, we believe you can expect
one of the following outcomes when dealing with perfect predictors:
1. The model will fail to converge because of the lack of a concave Hessian.
2. The model "converged", but there are some extreme valued coefficients with
just as extreme standard errors. Sometimes, but not always, the fitted
intercept will be extreme too. For -mlogit- with the -rrr- option, these
extreme values tend to translate to zero relative-risk ratios with missing
standard errors.
--Jeff
[email protected]
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/faqs/resources/statalist-faq/
* http://www.ats.ucla.edu/stat/stata/