Friday, March 17, 2006

Re: st: GLMs that fail - an effect (Hauck-Donner)

Allan Reese (Cefas) <r.a.reese@cefas.co.uk> wrote:

> I've been fitting a logistic glm to some count data: six treatments, each > duplicated, response is r/n: > > treat r n > 1 8 13 > 1 9 15 > ... > Since the effect is due, in this and similar examples, to groups having zero > variance, is it not possible to modify glm (and other estimating commands?) to > detect this and either issue a warning or switch automatically to robust > estimators?

What Allan observes in these data is what we refer to as "perfect failure predictors" in logit models. The categories 4 and 6 of treat (or the corresponding indicator variables defining these categories) predict failure perfectly, i.e. if we are to expand these data to the equivalent binary representation

treat y 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 0 1 0 1 0 1 0 2 1 ...

then Pr(y = 0 | treat==4) = 0 and Pr(y = 0 | treat==6) = 0. Stata's -logit- and -logistic- commands report the appropriate warning message in such situation and drop perfect predictors from the model. A more detailed description of this may be found in [R] logit p.96.

Since Alan has grouped data he can use Sata's -blogit- command to fit logit model:

. xi: blogit r n i.treat i.treat _Itreat_1-6 (naturally coded; _Itreat_1 omitted) note: _Itreat_4 != 0 predicts success perfectly _Itreat_4 dropped and 2 obs not used

note: _Itreat_6 != 0 predicts success perfectly _Itreat_6 dropped and 2 obs not used

Logistic regression for grouped data Number of obs = 118 LR chi2(3) = 13.50 Prob > chi2 = 0.0037 Log likelihood = -54.186571 Pseudo R2 = 0.1108

------------------------------------------------------------------------------ _outcome | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- _Itreat_2 | 2.203739 .8279165 2.66 0.008 .5810527 3.826426 _Itreat_3 | .4119798 .5553943 0.74 0.458 -.6765729 1.500533 _Itreat_5 | 1.761907 .7211817 2.44 0.015 .3484164 3.175397 _cons | .4353181 .386953 1.12 0.261 -.3230959 1.193732 ------------------------------------------------------------------------------

Note that -blogit- identified the "perfect predictor problem" and issued the corresponding warning messages.

You can fit the logit model equivalently using -glm- with family(binomial) and logit(link). However, -glm- does not replicate the behavior of the corresponding logit model since it is thought of a facility to fit a set of generalized linear models. Therefore, we tend not to specialize it for the particular members of this family as, in this case, logit models.

-- Yulia ymarchenko@stata.com * * For searches and help try: * http://www.stata.com/support/faqs/res/findit.html * http://www.stata.com/support/statalist/faq * http://www.ats.ucla.edu/stat/stata/


Tag:


Links to this post:

Create a Link



<< Home

This page is powered by Blogger. Isn't yours?