PLS-POS Result interpretation

[olearus]

Dear SmartPLS users and developers,
I am trying to discover the occurrence of unobserved heterogeneity in a model which contains composites and common factors. I have started directly with PLS-POS due to the complex nature of the model and the predominance of composite over factors. Thus, when I specified the first run for two segment solution I obtained the following result:
Segment Sizes (Total)
Group1 Group2
Number 666.000 14.000
Segment Sizes (Relative)
Group1 Group2
Percentage: 97.941 2.059

From my point of view these results are pointed out to non-presence of unobserved heterogeneity since the second group accounts for only a 2% of the sample, and the model depicted by 14 individuals is unsuitable and non-reliable (taking the 10 times rule I would need at least 60 cases to set up a group).
Then, I continued running the PLS-POS algorithm with three segments and here I got stuck because I am retrieving different error messages:
1. Singular Matrix (it makes sense for me since the groups will be so small that there must be a lack of variance in some variables)
2. The sample size is too small. there must be at least 6 cases/observations (it makes sense as well since whether there is no unobserved heterogeneity it must be almos imposible to make the groups due to their tiny sizes).

So, could you please enlighten me about my reasoning and how to proceed to demonstrate reliably the lack of unobserved heterogeneity?

Best regards,
MAC

[olearus]

Dear SmartPLS users and developers,
I would like to get some suggestions regarding the issue. For me it makes sense my reasoning but I would like to check it... The scholars' opinion would be very welcomed.
Thank you so much in advance
Kind regards,
MAC

[MrcJnk]

Dear olearus,

for me your interpretation sounds legit. The very small group could be interpreted as outliers. If you increase the number of segments in FIMIX-PLS or PLS-POS you force each observation into the number of pre defined segments. As noted in 2. the sample size could become too small for estimating your outlier group or finally no observation are assigned to the least segment.

Best regards

Marc Janka

[jmbecker]

Generally, your interpretations are correct. If the algorithm does not find stable segments of sufficient size then it is likely that heterogeneity is not a problem. However, it is always hard to prove that there is no heterogeneity. Particularly, the algorithm could end in local optima. You should try sufficient starts (at least about 20) to ensure that local optima are not a problem in your case.