Prev: Negative binomial linear regression model
Next: LEAVE THE LIST
From: Daniel Nordlund on 28 Dec 2009 00:09
> -----Original Message-----
> From: SAS(r) Discussion [mailto:SAS-L(a)LISTSERV.UGA.EDU] On Behalf Of Dan
> Abner
> Sent: Sunday, December 27, 2009 4:54 PM
> To: SAS-L(a)LISTSERV.UGA.EDU
> Subject: Discriminant Analysis in SAS XXXX
>
> Hello,
>
> I am attempting to reproduce results in SAS that Huberty & Olejnik (2006)
> discuss in section 8.5 (p. 143, for example) where these authors perform a
> descriptive discriminant analysis (what SAS calls a canonical discriminant
> analysis) in SPSS. I am using PROC DISCRIM with the CANONICAL option,
> however:
>
> This situation is a 2 factor MANOVA situation. Apparently SPSS performs
> the
> MANOVA and DDA all part of the same MANOVA program. In SAS, I am using
> separate calls to GLM and DISCRIM and am unable to reproduce the DDA
> results
> in DISCRIM for each of the main effects (the GLM results match the SPSS
> output, but not the DDA results). I assume this is b/c I am not able to
> specify the correct Error SSCP matrix in DISCRIM (but I'm not 100% sure
> about that...).
>
> Does anyone have any suggestions or expertise in this area?
>
> Thank you,
>
> Dan

It has been some time since I have worked with SPSS (and I don't have access to it now). I do have some questions.
1. Why are you using Proc DISCRIM rather than CANDISC?
2. Why are you using GLM as well?
3. I am not familiar with the Huberty paper or results and you haven't shown your SPSS or SAS code, so it is hard to know what is going on. What do you mean when you say you can't replicate the results? Is there nothing that is similar?

I published a paper some time ago which dealt with the differences between SAS CANDISC and SPSS MANOVA specifically with respect to the differences in standardized discriminant coefficients.

Nordlund, D.J. and Nagel, R. Standardized Discriminant Coefficients Revisited. Journal of Educational Statistics, Vol. 16, No. 2 (Summer, 1991), pp. 101-108

SPSS and SAS presented different output results, and SPSS defaulted to using the Within-groups SSCP matrix for standardization while SAS defaulted to using the Total SSCP matrix. This can produce different standardized coefficients, and even result in small changes in the rank ordering of the magnitude of coefficients. Differences in output can also be the result of differing methods of parameterizing class variables.

So, there can be lots of reasons for not being able to "match" results. I think you will need to provide more information before anyone can give you any specific advice.

Sorry I can't be of more help,

Dan

Daniel Nordlund
Bothell, WA USA
 | 
Pages: 1
Prev: Negative binomial linear regression model
Next: LEAVE THE LIST