LISTSERV at the University of Georgia
Menubar Imagemap
Home Browse Manage Request Manuals Register
Previous messageNext messagePrevious in topicNext in topicPrevious by same authorNext by same authorPrevious page (June 2002, week 3)Back to main SAS-L pageJoin or leave SAS-L (or change settings)ReplyPost a new messageSearchProportional fontNon-proportional font
Date:         Tue, 18 Jun 2002 10:21:57 -0700
Reply-To:     Dale McLerran <stringplayer_2@YAHOO.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Dale McLerran <stringplayer_2@YAHOO.COM>
Subject:      Re: Untenable results from PROC GENMOD
In-Reply-To:  <OF156738AA.2ABE6D0D-ON80256BDC.004E1297@unauthorised.ici.com>
Content-Type: text/plain; charset=us-ascii

David,

The problem here is that you do not have sufficient information to fit the model you are trying to estimate. You state that you have n=31 observations. When you specify the SUBJECT effect, you use up all your degrees of freedom since there is only one observation per SUBJECT. On top of that, you want to fit a model for effects PHASE (1 df), PRODUCT (6 df), and PRODUCT*PHASE (6 df). Having used up all your df for the SUBJECT effect, you cannot fit the model which you are really interested in. You are overparameterized. It is not that PROC GENMOD needs tweaking. Your model needs to be overhauled. The common, LARGE standard error stated for every parameter is symptomatic of overfitting and not being able to estimate the effects in the model. If you look at your log or lst file, you should see a statement that the model did not converge.

Note that even if you remove the SUBJECT effect from the model, you still do not have sufficient information to estimate the remaining effects in the model. For a binomial model, you should have at least 5 events and 5 non-events for each parameter you wish to estimate. With 14 df (including the intercept), that would mean that you should have at least 140 observations to fit the interaction model. You just do not have enough data to estimate any model except a model for PHASE alone. You cannot estimate a model for PRODUCT alone, as there are less than 5 observations/product.

Sorry to have to be the bearer of bad tidings.

Dale

--- david.mcnulty@QUESTINTL.COM wrote: > Hi Folks, > > I am trying to estimate some proportions in a three way table using > Proc > GENMOD and am getting some very curious results. In brief after back > transforming the estimated LSMeans, the proportions have standard > errors > greater than 1. > > Example: LSmean Prod_1 = 2.1478, se = 1622.330 > proportion = Exp(mu)/(1+Exp(mu))= 0.895 > approx se = p(1-p).SE(LSMean) = 151.86 > > (Since n is 31, the se using the approximation sqrt(p(1-p)/n) is > closer to > 0.05) > > Is there something I can tweak in Proc GENMOD to get some sensible > answers? > > Thanks > > Dave. > > > > Additional Info > ================ > > In the following code the binomial response "correct" has the form 1 > or 0 > rather than the alternate events/trial syntax. Product "Prod_8" is > excluded > from the analysis since the empirical proportion for Product="Prod_8" > and > phase="P3" is one. Since log(1/(1-1)) is undefined SAS warns the > Hessian is > not positive definite and refuses to calculate the remaining LSMeans. > > Code > ==== > proc genmod data=master descending; > class subject phase product; > model correct=subject phase|product / type1 type3 dist=bin; > lsmeans phase|product ; > where product ne 'Prod_8'; > quit; > > Sample Output > ============= > The SAS System 10:57 > Tuesday, > June 18, 2002 126 > > The GENMOD Procedure > > LR Statistics For Type 3 Analysis > > Chi- > Source DF Square Pr > > ChiSq > > Subject 30 58.02 > 0.0016 > Phase 1 0.17 > 0.6775 > Product 6 26.30 > 0.0002 > Phase*Product 6 2.97 > 0.8124 > > > Least Squares Means > > Standard > Chi- > Effect Phase Product Estimate Error DF > Square Pr > ChiSq > > Phase P2 1.9958 1622.330 1 > 0.00 0.9990 > Phase P3 2.1003 1622.330 1 > 0.00 0.9990 > Product Prod_1 2.1478 1622.330 1 > 0.00 0.9989 > Product Prod_2 2.7500 1622.330 1 > 0.00 0.9986 > Product Prod_3 2.3726 1622.330 1 > 0.00 0.9988 > Product Prod_4 1.5691 1622.330 1 > 0.00 0.9992 > Product Prod_5 0.8329 1622.330 1 > 0.00 0.9996 > Product Prod_6 2.4147 1622.330 1 > 0.00 0.9988 > Product Prod_7 2.2489 1622.330 1 > 0.00 0.9989 > Phase*Product P2 Prod_1 2.3594 1622.330 1 > 0.00 0.9988 > Phase*Product P2 Prod_2 2.6068 1622.330 1 > 0.00 0.9987 > Phase*Product P2 Prod_3 2.6068 1622.330 1 > 0.00 0.9987 > Phase*Product P2 Prod_4 1.5691 1622.330 1 > 0.00 0.9992 > Phase*Product P2 Prod_5 0.7539 1622.330 1 > 0.00 0.9996 > Phase*Product P2 Prod_6 1.9362 1622.330 1 > 0.00 0.9990 > Phase*Product P2 Prod_7 2.1384 1622.330 1 > 0.00 0.9989 > Phase*Product P3 Prod_1 1.9362 1622.330 1 > 0.00 0.9990 > Phase*Product P3 Prod_2 2.8933 1622.330 1 > 0.00 0.9986 > > > ETC ... > > IMPORTANT NOTICE: > This email is confidential, may be legally privileged, and is for the > intended recipient only. Access, disclosure, copying, distribution, > or > reliance on any of it by anyone else is prohibited and may be a > criminal > offence. Please delete if obtained in error and email confirmation > to the > sender.

===== --------------------------------------- Dale McLerran Fred Hutchinson Cancer Research Center mailto: dmclerra@fhcrc.org Ph: (206) 667-2926 Fax: (206) 667-5977 ---------------------------------------

__________________________________________________ Do You Yahoo!? Yahoo! - Official partner of 2002 FIFA World Cup http://fifaworldcup.yahoo.com


Back to: Top of message | Previous page | Main SAS-L page