FIMIX path coefficients larger one

[mengchuanjin]

everybady,
I have a question about FIMIX path coefficients,

I think it should be <1,why my output the path coeffucients >1?

[Förster]

Hello everyone,

I have the same problems with the FIMIX path coefficients. I have a couple of them, which are in all segments > 1. There wasn't any problem with the PLS- and with the Bootstrap-results.

What could be the reason? Thanks for answering

Manuel Förster

[schroer]

Hi,

could it be a consequence of small segments, i.e., a) few people and b) many path coefficients in the model? This easily leads to path coefficients > 1 due to multicollinearity. In that case one should probably be careful to have a look at indicators of multicollinearity (tolerance/VIF, intercorrelations, etc.).

Best,

Joachim

[cringle]

Hi,

we also recognized this issue quite a while ago. A very intensive SmartPLS-FIMIX-PLS evaluation by our team and some colleagues from other institutes - thank you, Christian Reinstrom (Institute of Marketing at the University of Hamburg) and Jan-Michael Becker (Institute of Marketing at the University of Cologne) shows that all SmartPLS-results are correct.

However, there is one exception: a value of one is added to all relationships between latent endogenous variables - a little nasty software implementation error which has been corrected in our experimental system and will also be corrected in the next SmartPLS release.

This issue does not harm the other outcomes but leads to the values above one. A simple rule of thumb at the moment: subtract one from the final SmartPLS-FIMIX-PLS relationships between latent endogenous variables in the inner model. We tested this with artificially generated data. The expected value of such a relationship was 0.36 and the SmartPLS-FIMIX-PLS group-specific result was 1.35. However, the above comment by Joachim remains an issue (especially when you observe inner relationships above two).

Cheers,
Christian

[danielherrmann80]

Hello,
By making FIMIX-PLS analysis in the context of my diploma thesis, an unexpected kind of value occured for the path coefficients:
Values which are not between -1 and 1.

For the ones >1 I found the following explanation in the forum:
"subtract one from the final SmartPLS-FIMIX-PLS relationships between latent endogenous variables in the inner model"

[1]
Could anybody confirm my guess that in terms of figures below -1 (e.g. -1,76) I'll simply have to add one?
[2]
What about the values which are unfortunally above 2 / below -2?
"However, the above comment by Joachim remains an issue (especially when you observe inner relationships above two)" Do you have any suggestions/rule of thumb in that case?
[3] Quality Criteria
I try to conduct a benchmark study comparing FIMIX-PLS outcomes with the ones of an alternative method (i.e. k-means segmantation using the likert-scale answer figures of my survey as segmentation variables and an added PLS analysis of each extracted segment, due to the lack of 'real' alternatives for segmentations which rely on formative constructs in the SEM).
Could the "FIMIX Variances" figures for the latent variables of each segment interpreted like the R Square of PLS results? This seems to me the only way to compare the results of each method (apart form the path coefficients... )
And furthermore: Is there a way to apply the Blindfolding and Bootstrapping procedures to the FIMIX-PLS analysis for each single segment or?

Thank you in advance for your support

Best regards

Daniel

[cringle]

Hi,

ad [1] no, only subtracting "1" is correct.

ad [2] there is a problem in your data / number of segments - you obtain inadmissible FIMIX-PLS solutions. This can happen when you successively increase the number of segments but there is no additional segment. FIMIX-PLS "forces" observations to belong to this extra segment (these segment sizes usually are relatively small).

ad [3] compare FIMIX-PLS to sequential segmentation strategies. The latter are not capable to identify heterogeneity in the structural model. Jedidi et al. 1997 (Marketing Sience) show this for FIMIX-SEM and I presneted this issue for the first time in the following working paper:

Ringle, C. M.: Segmentation for path models and unobserved heterogeneity: The finite mixture partial least squares approach, Research Papers on Marketing and Retailing No. 035, University of Hamburg, Hamburg 2006, ISSN 1618-8985, http://www.ibl-unihh.de/RP035.pdf.

You also find my response to [2] in this working paper (and my subsequent FIMIX-PLS handbook and journal articles).

You may also be interested in an analysis of alternative FIMIX-PLS evaluation criteria:

Sarstedt, M. / Schwaiger M. / Ringle, C. M.:Determining the number of segments in FIMIX-PLS, Proceedings of the Academy of Marketing Science (AMS) Annual Conference, Balitmore, USA 2009, http://www.ibl-unihh.de/ams2009_download.pdf.

Blindfolding and Bootstrapping has not been proposed for FIMIX-PLS, yet.

Best
Christian

[danielherrmann80]

Hi Christian (and anybody else too, of course),

First of all: thumbs up for the outstanding reaction time!
Nevertheless it's a rather unfortunate finding for me, that you already performed a FIMIX-PLS vs. K-Means&PLS comparison in the research paper you mentioned, because I was looking forward to accomplish exactly that in my thesis ;-(

Although I'll now have to consider doing something else, I am still somewhat curious about the segment specific quality criteria reported by SmartPLS for FIMIX-PLS results:

How could the "Fimix Variances" be interpreted? / How are they related to "R Square" figures (of a normal PLS analysis)? / How are they calculated?

Thank you for (hopefully) crushing my mental blocks & Best regards

Daniel

[lenfitze]

Hello,

as well for me it is not clear how to calculate the R2 for the FIMIX segments. Also how could the "Variance" be interpreted?

Best regards

Lena

[cringle]

Hi

while R² is the amount of explained variance, the variance represents th eunexplained variance. THus, one minus the FIMIX-PLS variance should approximately provide you with the explained amount of variance.

Best
Christian

[lenfitze]

Hello Christian,

Thank you for the fast reply. I have another question regarding fimix results, especially the sign of the log-likelihood (LnL).

The fimix result of my satisfaction pathmodel presents a positive LnL. All quality criteria (AIC, BIC, CAIC) are negative. All concepts, from two to six segments for all criteria, present the change of the sign. In all papers I know, the fimix results show only negative LnL values and therefore positive AIC, BIC or CAIC results. Do you have an explanation for the result?

Could be for example multi-collinearity a reason? Or does fimix work with positive and negative LnL values and I "just" have to interpret these results in the opposite way to define the “best” number of segments?

Best Regards

Lena

[egon1999]

Dear Pr. Dr Ringle,
I have the same problem :
AIC, BIC, CAIC, EN are negative .
Is there a problem and how to explain it ?

Thank's a lot

[theresiadominic]

I need help on how to interpret FIMIX PLS results,
starting with
initial partition and 'final partition reports

so how do you assign cases to each segment?

please

[admin]

Based on the probabilities of membership.

Why don't you check the numerous articles with FIMIX-PLS applications and interpretations in www.smartpls.de --> announcements --> literature?

Best
CR

[theresiadominic]

which LnL value do we take from 'LnL development report' ?

another question
where do you get the LnL, AIC, BIC...etc values when K=1 ? because the software can generate FIMIX -PLS reports when segments is higher than one

please help

Theresia

[cys0r]

LnL Value: Quality Criteria -> Fit Indices - LnL

K=1 Just choose 1 segment ... it works at least in Smart PLS 3