Date:         Wed, 30 Apr 2008 19:04:27 -0400
Reply-To:     Sigurd Hermansen <HERMANS1@WESTAT.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Sigurd Hermansen <HERMANS1@WESTAT.COM>
Subject:      Re: Cost analyses using PROC GENMOD using link=log and dist=gamma
Comments: To: Amy Smith <asmith197401@yahoo.com>
In-Reply-To:  <865017.54462.qm@web45803.mail.sp1.yahoo.com>
Content-Type: text/plain; charset="iso-8859-1"

Amy:

See http://www.ats.ucla.edu/STAT/sas/output/sas_negbin_output.htm ,

especially,
>>
As assumed for a negative binomial model our response variable is a count variable, and each subject has the same length of observation time. Had the observation time for subjects varied, the model would need to be adjusted to account for the varying length of observation time per subject. This point is discussed later in the page. Also, the negative binomial model, as compared to other count models (i.e., poisson or zero-inflated models), is assumed to be the appropriate model. In other words, we assume that the dependent variable is ill-dispersed (either under- or over- dispersed) and does not have an excessive number of zeros.
>>

Overriding everything else, though, would be the risk of finding ostensibly significant (Type 2 error) but small differences in means or other statistics in two partitions of a large sample. The distributions of costs related to the two cohorts could very likely result from typical variations in samples. The confidence intervals of the means of costs for the two cohorts overlap. More complex models would have to include other covariates, not more complex assumptions about distributions, to explain a benefit to the intervention in some groups within the population.
S

-----Original Message-----
From: owner-sas-l@listserv.uga.edu [mailto:owner-sas-l@listserv.uga.edu] On Behalf Of Amy Smith
Sent: Tuesday, April 29, 2008 9:41 AM
To: SAS-L@listserv.uga.edu
Subject: Re: Cost analyses using PROC GENMOD using link=log and dist=gamma


Does anyone know a theoretical reason why it is right or wrong to use the negbin link and log distribution in PROC GENMOD for analysis of cost data if it seems to produce nice dev/df ratios?  Is there any other way to assess goodness of fit?  The gamma distribution is booting out my zero values. A


----- Original Message ----
From: Sigurd Hermansen <HERMANS1@WESTAT.COM>
To: SAS-L@LISTSERV.UGA.EDU
Sent: Friday, April 25, 2008 8:42:41 AM
Subject: Re: Cost analyses using PROC GENMOD using link=log and dist=gamma

Amy:
See http://www.listserv.uga.edu/cgi-bin/wa?A2=ind0601A&L=sas-l&P=R30742&D=1&
H=0&O=D&T=1&m=187776 as a starting point.
S

-----Original Message-----
From: owner-sas-l@listserv.uga.edu [mailto:owner-sas-l@listserv.uga.edu]
On Behalf Of Amy Smith
Sent: Thursday, April 24, 2008 11:43 PM
To: sas-l@listserv.uga.edu
Subject: Cost analyses using PROC GENMOD using link=log and dist=gamma


I am trying to analyze medical cost data that has about 25% zeroes and then a lot of low values and some extremely high values (extremely long tail on the high end).  My predictor is simply a cohort (control group vs intervention group).  I read that cost data can sometimes be analyzed with PROC GENMOD using link=log and dist=gamma, but the log tells me that it is removing my zero values and I do not want to delete these valuable observations.

proc genmod;
class cohort;
model cost = cohort / link=log dist=gamma;
lsmeans cohort / cl;

Can I add $1 to all my cost data so I don't lose my zero values or could I use a negative binomial link or what is recommended?  My dev/df ratio is quite close to 1 with the negative binomial but I don't know if it is wrong to use something that seems to be associated with count data.  I'm guessing that dollars don't really qualify as count data.

I read something about a zero-inflated gamma using a different procedure (NLMIXED perhaps) but I am totally lost about how to code something so complex.  NLMIXED baffles me.  If you recommend the zero-inflated gamma in a different procedure then please provide actual code that I can test.

Thank you very much, Amy



________________________________________________________________________
____________
Be a better friend, newshound, and
know-it-all with Yahoo! Mobile.  Try it now. http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ



      ____________________________________________________________________________________
Be a better friend, newshound, and
know-it-all with Yahoo! Mobile.  Try it now.  http://mobile.yahoo.com/;_ylt=Ahu06i62sR8HDtDypao8Wcj9tAcJ