Unobserved heterogeneity - Interpretation

[muriloz]

Dear all,

Could you please assist me in understanding if my conclusion regarding the FIMIX-PLS analysis is accurate? I couldn't find an explanation for when the results indicate two segments and the second segment solution lacks a minimum sample. Can I conclude that the model does not exhibit unobserved heterogeneity at critical levels?

See the analysis and conclusion.

"To identify unobserved heterogeneity, we employed the FIMIX-PLS procedure (Hair, et al., 2016). A power analysis conducted in g*power software, assuming an effect size of 0.15 and a power level of 80%, indicated a minimum sample size requirement of 74, allowing for the extraction of up to three segments. In our analysis, AIC3 suggested a two-segment solution, while CAIC pointed to a one-segment solution. AIC4 and Bayesian Information Criteria (BIC) both indicated a two-segment solution, which appeared densely clustered according to the EN criterion (0.801). Additionally, AIC suggested a solution fewer than three, and the Minimum Description Length with a factor of 5 (MDL5) indicated a solution greater than 1. However, despite the good fit of the two-segment solution, it did not meet the sample size requirement (Segment 1 = 80.3%, 233 cases; Segment 2 = 19.7%, 57 cases). Therefore, considering the discrepancy between AIC3 and CAIC and the insufficient sample size for valid analysis with the two-segment solution, we concluded that unobserved heterogeneity is not at a critical level."

Thank you in advance,

Murilo

[jmbecker]

Well, I think you may need to be careful with the way you formulate this. I think the criteria relatively clearly point to a two segment solution and thus potential unobserved heterogeneity. However, your sample size (overall) and particularly for the second segment may not be large enough to further reliably explore this unobserved heterogeneity.

[YvonneA]

Dear Prof. Becker & the PLS community,
I have a similar issue where all the indicators point to a 4 segment solution in a reliable manner. Three of the segments have sufficient sample size based on the 10 times rule, however, the 4th segment does not meet the sample size criteria.
In light of this, are there any options I can take to move forward or perhaps this is the end of the road for this analysis given that I will not be able to run a hard clustering of the 4th segment?
Thank you so much! :-D