Some methodological issues

[agalvez]

Dear mates

I would be grateful if you could help me again.

Question 1

I think I have a problematic construct in my model. It is composed of 5 items (5 point Likert Scale). Original scale shows these results:

- AVE = 0.121; CR = 0.055; Alpha = 0.719

After deleting three items with < 0.40 outer loadings, AVE and CR improves (both > 0.70) but Alpha gets worse. After this, final report shows:

- AVE = 0.579; CR = 0.734; Alpha = 0.276.

This is not the only construct in which I see Alpha goes down when AVE improves. I think this last result allows me to report that AVE > 0.50 and CR > 0.50, and use this scale in my study. However, I don't know how to interpret this Alpha coefficient.

What I sould care about? This construct comes from a validated scale that have been used in other context with good results.

EDIT: I just read this post and I think I understand now: viewtopic.php?f=5&t=4003&p=13024&hilit=composite+reliability+alpha#p13024

Question 2

I have a model with four constructs of 3 items each. After running PLS Algorithm, results suggest that I have to delete at least two items, resulting in construct of two items. Final report is OK (AVE and CR). I have been told that construct needs at least three items to be considered. However, I don't know if this affirmation is only when you want to develop a scale, so that you keep at least three items in the case they are poor and need to drop them.

In any case, I wanted to ask you if there is any problem when using constructs of two items, being originally four.

Question 3

When it is advisable to remove an item with outer loading between 0.40 and 0.70? I have read that it depends on content validity. However, when I delete an item with an outer loading of 0.50, AVE changes from 0.55 to 0.65. Is this enough to consider remove the item? I ask this becase the significance of path loading changes depending on whether or not that item is included.

I hope you understand my two questions. Your answers are very important for my research.

Thank you in advance.

[jmbecker]

Question 1:
The original scale results seems to be quite strange. It is unusual that AVE and CR are so much smaller than Cronbach’s alpha. There might be negative loadings (for example, due to reverse coded items). I would check this first, before deleting items.

Question 2:
Usually, the more items the better. However, it is not necessary in PLS that constructs have at least 3 items. It is more a convention (that stems from CBSEM) and it makes sense in terms of the rule: the more the better.
I would always be careful with deleting items. You should not directly delete them, just because their loading is bad. If AVE and CR are ok, you can also keep them.

Question 3:
See above. I would not delete the item if your AVE is above the threshold and the same applies to Cronbach’s alpha and composite reliability.

[agalvez]

Question 1:
The original scale results seems to be quite strange. It is unusual that AVE and CR are so much smaller than Cronbach’s alpha. There might be negative loadings (for example, due to reverse coded items). I would check this first, before deleting items.

Question 2:
Usually, the more items the better. However, it is not necessary in PLS that constructs have at least 3 items. It is more a convention (that stems from CBSEM) and it makes sense in terms of the rule: the more the better.
I would always be careful with deleting items. You should not directly delete them, just because their loading is bad. If AVE and CR are ok, you can also keep them.

Question 3:
See above. I would not delete the item if your AVE is above the threshold and the same applies to Cronbach’s alpha and composite reliability.

Dear jmbecker

First of all I would like to thank your help.

Regarding Q1, we checked the existence of reverse items and this is not the cause. To be honest, the results provided are from the worst contruct of my scale (according to the report reusults). That's the reason I wanted to ask you. My other constructs show better results.

In respect of Q2 and Q3, what I understand is that that the main criteria is not "the more AVE, the better construct". On the contrary, if AVE > 0.50, it is not worth removing an additional item to get a 10% improvement in AVE score. Please, correct me if I am wrong.

Thank you again, jmbecker

[jmbecker]

Ad Q1: I would need to see the data if there are any other problems. Incorrect missing value treatment could also be a problem.

Ad Q2&3: Yes. If you have an established scale, I would not remove indicators to increase AVE if you are above the thresholds. The more AVE the better...yes... to some degree, but you would per definition have an AVE of 1 for single item constructs. These are, however, not generally better than multiple-item scales. Hence, keep your scale if it fulfills the basic quality criteria.

[agalvez]

Ad Q1: I would need to see the data if there are any other problems. Incorrect missing value treatment could also be a problem.

Ad Q2&3: Yes. If you have an established scale, I would not remove indicators to increase AVE if you are above the thresholds. The more AVE the better...yes... to some degree, but you would per definition have an AVE of 1 for single item constructs. These are, however, not generally better than multiple-item scales. Hence, keep your scale if it fulfills the basic quality criteria.

Thank you again for your explanation, jmbecker.

[VSVSVS27]

Hello,
I have a question about the methodology itself. I understand the difference between PLS-SEM AND CB-SEM, which is expressed in several academic articles.
But what is the difference to multiple regressions?
Do multiple regressions belong to CBSEM?

Plus what logic is applied in SPSS? They follow CBSEM right?

I want to differentiate and explain it in my master thesis, but I need a clarification for that issue at hand.

Thanks for helping out!