FIMIX-negative fit indices

[Verena]

Dear all,
I am working on a FIMIX segmentation at the moment. To reduce complexity, my four latent variables are each measured with only one reflective indicator. Whereas EN is very high for k = 2, AIC, CAIC and BIC are negative; the lnL is positive. Does anybody have an explanation for that?
Thanks a lot for your help, Verena.

[cringle]

Hi,

just look at the equations for computing the information critria: they usually use -lnL.

Some good ;)) literature on FIMIX-PLS (or just visit http://www.ibl-unihh.de/tea_rin.htm):

Ringle, C. M. / Sarstedt, M. / Mooi, E. A.: Response-based segmentation using FIMIX-PLS: Theoretical foundations and an application to American customer satisfaction index data, in: Stahlbock, R. / Crone, S. F. / Lessmann, S. (eds.): Annals of Information Systems, Special Issue on Data Mining, Vol. 4, Berlin-Heidelberg 2009, in print (available upon request).

Sarstedt, M. / Ringle, C. M.: Heterogenität in varianzbasierter Strukturgleichungsmodellierung: Eine Analyseprozedur zur systematischen Anwendung von FIMIX-PLS, in: Marketing ZFP (Zeitschrift für Forschung und Praxis), 30. Jahrgang (2008), Heft 4, S. 241-257, http://vahlen.becksche.de/zneu/vahlen/z ... aehlt=2008.

Sarstedt, M. / Ringle, C. M.: Erfassung unbeobachtete Heterogenität in varianzbasierter Strukturgleichungsmodellierung mit FIMIX-PLS, in: Meyer, A./Schwaiger, M. (Hrsg.): Theorie und Methoden der Betriebswirtschaft, ISBN 978-3-8006-3613-6, München 2009, in print, http://www.beck-shop.de/productview.aspx?product=26287.

Ringle, C. M. / Wende, S. / Will, A.: The finite mixture partial least squares approach: Methodology and application, in: Esposito Vinzi, V. / Chin, W. W. / Henseler, J. / Wang, H. (eds.): Handbook of partial least squares: Concepts, methods and applications in marketing and related fields, ISBN: 978-3-540-32825-4, Berlin-Heidelberg 2009, in print, http://www.ibl-unihh.de/ringle_wende_wi ... ndbook.pdf.

Cheers
Christian

[stefanperras]

Dear all,

my FIMIX results are as follows

s___AIC______BIC_______CAIC______EN
2___257,15___307,33____307,78_____0,71342
3___199,08___275,36____276,04_____0,859811
4___151,59___253,96____254,88_____0,929657
5___4,3981___132,86____134,02_____0,993377
6___-928,5___-773,95____-772,5_____0,999997

If i use more than 6 segments in SmartPLS I get an error message.
My overall sample size is just 55.

Has anyone an idea how many segments to choose regarding this output?
I could not find any hint in all the papers dealing with unobserved heterogeneity.
Or might there be any other reason for these strange results?

Best
Stefan

[cringle]

Hi

S=1 is the solution.

Your sample size is simply too small for any reasonable segment specific PLS-SEM estimations.

In general, for selecting the proper number of segments, a lot of misapplications have been published that do not consider that criteria such as BIC have a strong underfitting tendency. This simulation study might be useful to gain additional insights:

Sarstedt, M./ Becker, J.-M./ Ringle, C.M./ Schwaiger, M.: Uncovering and treating unobserved heterogeneity with FIMIX-PLS: Which model selection criterion provides an appropriate number of segments?, in: Schmalenbach Business Review (sbr), Volume 63 (2011), Issue 1, pp. 34-62.
http://www.fachverlag.de/sbr/pdfarchive ... 34-062.pdf

Best
Christian Ringle

[stefanperras]

Many thanks Prof. Ringle!