Violating assumptions in twostep cluster analysis

[Martin Cernvall]

Hello,

 

I have some questions regarding twostep cluster analysis in PASW 18. It says in the manual that the Log-likelihood distance measure assumes that all variables are independent, but that it has been shown to be fairly robust even when this assumption is violated. Does anyone have suggestions on further reading regarding this issue in twostep log-likelhood cluster analysis?

 

I have longitudinal data with one main outcome variable assessed at 6 time-points and the different time-points are significantly correlated. Each time point is fairly normal distributed. I would like to tease apart different development trajectories over these 6 time-points, and when I run a two-step log-likelihood cluster analysis (and violate the assumption of independence) I get a solution that corresponds fairly well with theory and previous empirical work in this area. However, when I run a two-step Euclidian cluster analysis the solution is totally different.  Does anyone have suggestions based on this?

 

Thanks in advance

 

 

Martin Cernvall, M.Sc.

PhD candidate, clinical psychologist

Psychosocial Oncology and Supportive Care

Department of Public Health and Caring Sciences

Uppsala University

[hidden email]

[phone]

 

 

[Hector Maletta]

Euclidian distances certainly assume orthogonal dimensions (the sides of the rectangular triangle which Euclidian distance is the hypothenuse of), and therefore independence, although it is seldom specified in manuals. Just a thought in the middle of work: no time to analyze properly why the two methods give different outcomes, sorry.

Hector

----- Mensaje original -----

De: Martin Cernvall <[hidden email]>

Fecha: Lunes, Noviembre 30, 2009 7:17 pm

Asunto: Violating assumptions in twostep cluster analysis

> Hello,

>
>
>
> I have some questions regarding twostep cluster analysis in PASW
> 18. It says
> in the manual that the Log-likelihood distance measure assumes
> that all
> variables are independent, but that it has been shown to be fairly
> robusteven when this assumption is violated. Does anyone have
> suggestions on
> further reading regarding this issue in twostep log-likelhood cluster
> analysis?
>
>
>
> I have longitudinal data with one main outcome variable assessed
> at 6
> time-points and the different time-points are significantly
> correlated. Each
> time point is fairly normal distributed. I would like to tease apart
> different development trajectories over these 6 time-points, and
> when I run
> a two-step log-likelihood cluster analysis (and violate the
> assumption of

> independence) I get a solution that corresponds fairly well with

> theory and
> previous empirical work in this area. However, when I run a two-step
> Euclidian cluster analysis the solution is totally different. 
> Does anyone
> have suggestions based on this?
>
>
>
> Thanks in advance
>
>
>
>
>
> Martin Cernvall, M.Sc.
>
> PhD candidate, clinical psychologist
>
> Psychosocial Oncology and Supportive Care
>
> Department of Public Health and Caring Sciences
>
> Uppsala University
>
> [hidden email]
>
> [phone]
>
>
>
>
>
> ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD