Hi Mark,

I really can't tell the difference between the two techniques you refer to.
If both end up with the relationship each variable has with the dependant
(cluster)? If you have a variable which has a strong relationship to
cluster 1 but not 2 it will tell "One can then look at the distinguishing
characteristics of each cluster .."  which is essentially what the
relationship parameter in discriminant does (assuming you retrieve Fishers
coefficients).

Personally, I prefer other clustering techniques to K-means such as Two
step and there you have a very good way of identifying the driving params
to each segment.

Cheers,

F


On Thu, Sep 12, 2013 at 7:11 PM, Mark Levine <[email protected]> wrote:

> Hi folks,
>
> I have to respectfully disagree in part with Fredrik's comments.
>
> I totally agree with him that K-Means does not provide a unique solution.
> Repeated clustering to determine the stability of a solution is wise
> council.
>
> I think, however, that Frederik is off re the relationship between
> clustering and discriminant analysis. In discrim, you start with cases that
> have been assigned a priori to groups and the analysis seeks to find
> optimal (linear) combinations of predictor variables that properly assign
> cases to the groups. It's sort of like multiple regression with a nominal
> dependent variable. Well, sort of . . .
>
> What Jason has requested (as I have in a previous note) is a vector of
> scores that identifies which cases fall into which cluster. One can then
> look at the distinguishing characteristics of each cluster by
> cross-tabbing, AOVing, etc. the cluster assignment score against other
> variables.
>
> Without the cluster assignment scores for further analysis, doing K-Means
> -- or any kind of clustering procedure, e.g. Q-type factor analysis -- is
> somewhere between relatively and totally useless.
>
> At least in my not so humble estimation.
>
> Cheers,
>
> Mark
>
>
> On Sep 12, 2013, at 12:00 PM, [email protected] wrote:
>
> Send Pspp-users mailing list submissions to
>         [email protected]
>
> To subscribe or unsubscribe via the World Wide Web, visit
>         https://lists.gnu.org/mailman/listinfo/pspp-users
> or, via email, send a message with subject or body 'help' to
>         [email protected]
>
> You can reach the person managing the list at
>         [email protected]
>
> When replying, please edit your Subject line so it is more specific
> than "Re: Contents of Pspp-users digest..."
>
>
> Today's Topics:
>
>   1. PSPP K-means quick-cluster: assigning cases (Fredrik Clementz)
>
>
> ----------------------------------------------------------------------
>
> Message: 1
> Date: Thu, 12 Sep 2013 14:02:01 +0200
> From: Fredrik Clementz <[email protected]>
> To: [email protected], "[email protected]"
>         <[email protected]>
> Subject: PSPP K-means quick-cluster: assigning cases
> Message-ID:
>         <CA+y-Rt-WGxteonyFvLsJeyaeGtOzM8no22BiGLqv_Q=
> [email protected]>
> Content-Type: text/plain; charset="iso-8859-1"
>
> Hi Jason,
>
> The feature you're looking for is Discriminant analysis and is
> unfortunately not implemented in PSPP. I also have to mention that you
> should be vary about using K-means as a clustering technique. Please
> rerandomize data and run several times to see if results are similar as the
> technique is dependant on how the data i sorted.
>
> Cheers,
>
> --
> Fredrik
> -------------- next part --------------
> An HTML attachment was scrubbed...
> URL: <
> http://lists.gnu.org/archive/html/pspp-users/attachments/20130912/e0405bd1/attachment.html
> >
>
> ------------------------------
>
> _______________________________________________
> Pspp-users mailing list
> [email protected]
> https://lists.gnu.org/mailman/listinfo/pspp-users
>
>
> End of Pspp-users Digest, Vol 88, Issue 10
> ******************************************
>
>


-- 
Fredrik Clementz

Reflect AB
Fenixpalatset
Adolf Fredriks Kyrkogata 10, 4tr
111 37 Stockholm

Direkt: 08 - 412 26 94
Fax: 0708 - 27 79 37
Mobil: 0709 - 99 52 36

E-mail: [email protected]
Url: www.reflect.se

Om Reflect
Reflect är ett undersökningsföretag specialiserat på kvantitativa
undersökningar. Med undersökningar som verktyg lär vi företag att förstå
sina marknader, sin organisation och omvärld. Undersökningarna som
genomförs tillsammans med Reflect avser fungera som en språngbräda inför
framtiden. Vi levererar beslutsunderlag som snabbt skall kunna användas för
att öka utväxlingen i företag och organisationer
_______________________________________________
Pspp-users mailing list
[email protected]
https://lists.gnu.org/mailman/listinfo/pspp-users

Reply via email to