Validity Discriminant Problem on Second Order Construct

r00kie_researcher · Post by **r00kie_researcher** » Fri Feb 01, 2013 6:14 am

Hello, friends

I’m a student at postgraduate program in accounting and now working on my master thesis. This my first using PLS SEM then facing discriminant validity problem to my model which consist of eight latent construct. My model have two second-order constructs each include three first-order constructs, they are reflextive-formative type. I put dependency relationship between two second-order constructs. Before I relate those constructs I tested each second-order construct, each having problem with discriminant validity means second-order construct that endogenous have AVE (Average Variance Extracted) less than .5, I've compared AVE with Square Latent Variable Correlation and results AVE < Square Latent Variable Correlation. My question:

1. Is that ok for second-order construct to having AVE less than .5?
2. How can i repair my model to make AVE > Square Latent Variable Correlation?
3. How can i refine my model? Or solve this problem

Regards,
Triyoga

Hengkov · Post by **Hengkov** » Mon Feb 11, 2013 11:08 pm

Hi, Yoga

You have second-order Type II.
1. Is that ok for second-order construct to having AVE less than .5?
Actually No. Second-Order construct result low AVE (underestimate), please compute by hand for true value.

2. How can i repair my model to make AVE > Square Latent Variable Correlation?
Check outer loading > 0.7 and check significant construct dimension

3. How can i refine my model? Or solve this problem
Delete some items with outer loading low (first and second-order).

Best Regards,
Hengky

samaro · Post by **samaro** » Wed Feb 13, 2013 11:24 am

I Believe that if the second order construct is formative, you should not be looking at the AVE. You have to look at other criterions to assess formative constructs.

Hengkov · Post by **Hengkov** » Thu Feb 21, 2013 2:49 pm

Hi,

Second-order Type II (reflective-formative), If first-order mode A, second-order mode A too for repeated indicators. So, evaluation loading, AVE etc possible for first and second-order, because delete items first order (low outer loading), delete second-order too.

No evaluation AVE, CR, etc just for second-order construct Type III and IV.

Best Regards,
Hengky

r00kie_researcher · Post by **r00kie_researcher** » Mon Mar 18, 2013 10:58 pm

Thank you friends,

I’m still wondering. If I have three latent variable in my first order construct and a latent variable in my second order construct and use reflective-formative type and put my all observed variables (indicators) at first order to second order construct then I found my LVC (Latent Variable Correlation) for latent variable at second order having more value than my AVE Square Root.

Example, I have 203 respondents who replies and fully answer my questionnaire and my latent variable at my second order is “Managerial Accountability” while three other construct at first order are “Information”, “Personal Value”, and “Law Enforcement”. The result shows managerial accountability has 0.71 AVE (0.84 Square root AVE) which lower than LVC between managerial accountability to information (0.87), to personal value (0.92) and to law enforcement (0.95).

Since the rule of thumb of discriminant validity is Square root AVE more than LVC or correlation between latent variable so I think it’s difficult to make discriminant validity establish in second order.
Is my conclusion right? Please help me again.

Thank you very much

Regards,
Wahyu Triyoga

Hengkov · Post by **Hengkov** » Tue Mar 19, 2013 10:03 am

Hello,

AVE 2nd order underestimate, please compute by hand and also check collinearity (formative 2nd) because correlation very high.

Best Regards,
Hengky

Hengkov · Post by **Hengkov** » Tue Mar 19, 2013 10:04 am

Hello,

AVE 2nd order underestimate, please compute by hand and also check collinearity (formative 2nd) because correlation very high.

Best Regards,
Hengky

r00kie_researcher · Post by **r00kie_researcher** » Wed Mar 20, 2013 11:37 pm

Hi, Hengky

I’ve examined my collinearity problem in SPSS from my LV Scores, there is no problem, all three latent constructs which are part of managerial accountability such as information, personal value and law enforcement don’t have VIF more than 5. Issues come up with me when you said compute by hand for AVE 2nd order, since AVE is average sum square correlation indicator and latent construct so I calculate from outer loading table in managerial accountability column but the result is same 0.71.

Is AVE computation between reflective and formative using different way?

I put my indicators on my 2nd order construct from my 1st order construct using type II reflective. Is the AVE for 2nd order construct (formative) always compute by hand?

How to compute formative AVE by hand?

Thank you very much

Best Regards
Yoga (Wahyu Triyoga)

Hengkov · Post by **Hengkov** » Thu Mar 21, 2013 2:29 am

Hello,

I recommend used WarpPLS for check collinearity. In SPSS you just know vertical collinearity (among predictor), but in WarpPLS you know vertical and also lateral (among predictor and outcome) collinearity problem. Rencently, some paper PLS suggest VIF must less than 3.3 for stable estimate, and I aggree. Check out my paper in PLS Literature.

For, AVE 2nd order SmartPLS result 0.71, right? So, it's mean AVE 1st order very high, for each > 0.85?

Regards,
Hengky

r00kie_researcher · Post by **r00kie_researcher** » Thu Mar 21, 2013 2:51 am

Hi, Hengky

That’s right in SPSS we only know VIF but if VIF less than 3.3 it might my model having problem with collinearity.

You right AVE 1st order very high. Maybe I have to delete some indicator that make loading so high. Is that right way?

Thank you

Best Regards,
Yoga

Hengkov · Post by **Hengkov** » Sat Mar 23, 2013 2:39 am

Hi,
Yes, I think so. You have collinearity problem, because correlation among variable very high.

Collinearity have two type:
1. Vertical (among predictor).
2. Lateral (among predictor to criterion).
In SPSS, you just test vertical, don't lateral, So I recommend used WarpPLS for know all.

Best Regards,
Hengky