Hi everyone
I’ve problems to confirm discriminant validity in my model. I use SmartPLS 2.0 software. Maybe some of you could help me.
First, my model is composed by the following variables:
- Service Quality (2nd order): “SQ”
- Satisfaction (1st order): “SA”
- Behavioral Intentions (1st order): “BI”
This is the structural path: SQ -> SA -> BI (common in marketing). After PLS algorithm, items loadings are higher than 0.707 and AVE is higher than 0.50. Then, convergent validity is confirmed.
Here is my problem with divergent validity. Following Fornell-Larcker criterion, correlation between “SQ” and “BI” is higher than AVE of “SQ”. Thus, it is not possible to confirm discriminant validity in my model. “SQ” and “BI” are conceptually different constructs. I have thought that maybe I have collinearity problems.
To deal with this situation I could delete some items from my “BI” construct and convert it in a single-item scale, but this would be more difficult to justify theoretically, and I think that single-item scales are less reliable than multi-item. If I'm right, for single-item scales is not necessary to assess discriminant validity. Would it be possible?
Thank you in advance.
Dear Michael
Thanks for your quick response.
What do you mean with "standards"? Do you think I should use other statistical software such as SPSS Amos or EQS? I chose PLS mainly due to the small sample size.
Otherwise, will my construct discriminate in, i.e., SPSS Amos if they don't in PLS? I think I should follow Fornell-Larcker criterion.
Thank you
Dear Manuel
Yes, I compared the Square root of AVE (SQRT AVE) against The Correlation of Latent Variables. I have two IV and one DV.
Following this criteria, I can guarantee discriminant validity of all the variables in my model (SQRT AVE > Corr.), except for one IV that does not discriminate with my DV.
According to the paper you suggest, this new criteria for assesing discriminant validity is available in SmartPLS 3.0, and I use 2.0 (free license).
In respect of deleting items, yes... I've the same opinion.
I'll have a look at the reference. Thank you.
Dear Michael
Thank you very much for your explanation. I'll consider what you tell me. I'll also read the references provided.
Regards
Here I am agree with Michael R. Frone . AMOS or LISREL is based on covariance matrix whereas SmartPls depends upon partial least squares. They can not be used interchangeably. First of all, you should know whether you should use CB-SEM or PLS-SEM. Here are some key points to be considered for both of them:
If the goal is predicting key target constructs or identifying key 'driver' constructs, select PLS-SEM.
If the goal is theory testing, theory confirmation, or comparison of alternative theories, select CB-SEM.
If the research is exploratory or an extension of an existing structural theory, select PLS-SEM.”
Dear John,
could you also state any reasons which justifiy the use of PLS(c)?
Best regards,
Flo
Hi everyone
Do you recommend me to run my analysis with other software, i.e., SPSS Amos?
Thank you
It depends on your theoretical model and how the constructs are modeled. Of course a comparison of PLS and MLE for a SEM with common factors is not a fair comparison since PLS does not estimate a factor model. In contrast PLSc does. And it is a consistent estimator, see Dijkstra Henseler (2015). Even though PLS does not seem to perform that bad in case of common factors (Sarstedt et al., 2016). I think there is more research needed to figure out in which conditions PLSc outperforms MLE ( and other estimators usually used in case of common factors) (misspecifications, sample sizes, etc..). Most of the studies investigates PLS and not PLSc. I'm also not sure if 2SLS is the only estimator which has the properties described by John, e.g., what about 3SLS, a mixture of GLS and 2SLS?
Best regards,
Flo
Dijkstra, T. K. & Henseler, Jö. Consistent and asymptotically normal PLS estimators for linear structural equations Computational Statistics & Data Analysis, Elsevier, 2015, 81, 10-23
Sarstedt, Marko, et al. "Estimation issues with PLS and CBSEM: Where the bias lies!." Journal of Business Research 69.10 (2016): 3998-4010.
Dear John,
thanks for clarification. I will read your work, to better understand your point mentioned about 'using your work '.
So thank you and best regards,
Florian
- Badges
- Science topic
- Similar topics
- Methodology
- Research Methodology