effect size (f2) and predictive relevance (Q2 and q2)

Frequently Asked Questions about the SmartPLS software. Please refer to this section first, if you have problems with the software.
tamarin06
PLS Junior User
Posts: 6
Joined: Thu Feb 01, 2007 4:00 pm
Real name and title:

effect size (f2) and predictive relevance (Q2 and q2)

Hello,

I have three questions relative to smartpls.

1) I think i manage to obtain a Q2 thanks to the blindfolding produre. However, is Q2 equals to CV Red or CV Com? If it is equal to CV Red, then what is CV Com?

2) When runing the blindfolding procedure, how to determine the Omission Distance?

3) Is it possible to calcule q2 (and not Q2) with smartpls?

4) Is it possible to calculate the indicator of effect size f2?

Thanks a lot for your help

Laurent

cringle
SmartPLS Developer
Posts: 797
Joined: Tue Sep 20, 2005 9:13 am
Real name and title: Prof. Dr. Christian M. Ringle
Location: Hamburg (Germany)
Contact:
Hi!

Since the Stone-Geisser criterion (usually) refers to the structural model, Q² should be used representing the cv-redundancy (Chin 1998, 318).
viewtopic.php?t=336
viewtopic.php?t=292

Please check the literature for additional explanations on the difference between cv-communality and cv-redundancy.

The research sets this number - I usually use seven - make sure that the number of observations is not a multiple of the omission distance (blindfolding does not work in that case).

Yes, q² = [Q²(included ) - Q²(excluded)] / 1 - Q²(included) - this is not an automatic function because the user must decide what to do when not only a single latent variable but complete model structures that are related to this variable are excluded.

Yes, f² = [R²(included ) - R²(excluded)] / 1 - R²(included) - this is not an automatic function because the user must decide what to do when not only a single latent variable but complete model structures that are related to this variable are excluded.

Best
Christian

tamarin06
PLS Junior User
Posts: 6
Joined: Thu Feb 01, 2007 4:00 pm
Real name and title:
Christian,

Thanks a lot but why is your formula for F2 different from Cohen (1988) formula where F2 = R2 / 1-R2 ?

All the best

Laurent

cringle
SmartPLS Developer
Posts: 797
Joined: Tue Sep 20, 2005 9:13 am
Real name and title: Prof. Dr. Christian M. Ringle
Location: Hamburg (Germany)
Contact:
Hi,

the equation is different because we intend to compute the (relative) effect of a certain latent variables that is related to the latent variable under analysis.

Best
Christian

tamarin06
PLS Junior User
Posts: 6
Joined: Thu Feb 01, 2007 4:00 pm
Real name and title:
Hi Christian,

Thanks for your answer. Actually I found a good reference where the authors use the same formula:

Exploring the influence of gender on the web usage via partial least squares. By: Sánchez-Franco, Manuel J.. Behaviour & Information Technology, Jan/Feb2006, Vol. 25 Issue 1, p19-36, 18p

Laurent

powermaro
PLS Junior User
Posts: 1
Joined: Fri Jan 20, 2006 5:39 pm
Real name and title:
Dear Christian,

thanks a lot for your information. You stress an important point with you ad 4) - how to calculate f2. So far I couldn't find the answer to my following question in the forum, even though I excessively used the search function. Maybe you might be willing to give me a hint:

Let's say, I have a structural model
A -> C
B -> C
C -> D
C -> E

It is obvious for me how to calculate f2(A) and f2(B), as there is only one dependent variable. However, how can I calculate f2(C)? Having two dependent variables, the process is not clear for me. Just adding the R2 of the two dependent variables?

One more (even simpler) question: Is f2 to be calculated for both formative and reflective constructs?

Thanks to everybody for a short help, best regards
Juergen

Schnuffel84
PLS Expert User
Posts: 31
Joined: Mon Jun 14, 2010 7:35 pm
Real name and title:
Hi all,

I have a question about q^2 and f^2. I don´t really understand how to calculate it.
Yes, q² = [Q²(included ) - Q²(excluded)] / 1 - Q²(included) - this is not an automatic function because the user must decide what to do when not only a single latent variable but complete model structures that are related to this variable are excluded.

Yes, f² = [R²(included ) - R²(excluded)] / 1 - R²(included) - this is not an automatic function because the user must decide what to do when not only a single latent variable but complete model structures that are related to this variable are excluded.
Ad 3)I do blindfolding and then I see the Q^2. But what do you exactly mean with Q^2(included) and Q^2(excluded)? What do I have to do to get them?
When I do blindfolding I only take one variable which I want to analyse, for example customer satsifaction. Then I see in the last column 1-SSE/SSO that the value is 0,4014, which is good because it´s higher than 0.
What do I have to do in order to get Q^2(included) and Q^2(excluded) or where do I have to look at the output?

Ad 4) The same questions for R^2 included and excluded? But here I don´t have to run blindfolding?

I would be very happy and grateful if somebody could help me and give me some hints for my problems.

Thank you!

Best regards,
Larissa

christian.nitzl
PLS Expert User
Posts: 248
Joined: Sat Jul 25, 2009 1:34 pm
Real name and title:
Hey Larissa,

I will try to explain it on an example:

Let us say we have a model with four variables (three independent and one depended):

A->D
B->D
C->D

If you want to know how big is the part of one of your variable for explanation R^2 or for prediction Q^2 in your model you should use the effect size formula.

For calculation you have to run your model one time with A (that mean ‘included’) and the other time without A (that mean ‘excluded’). Afterward you put the different R^2 or Q^2 for D in the effect size formula. In this way it possible to say how big is the influence of A on D.

I hope this will help,

Christian

Schnuffel84
PLS Expert User
Posts: 31
Joined: Mon Jun 14, 2010 7:35 pm
Real name and title:
Thank you very much! This helps a lot :)

Best regards,
Larissa

christian.nitzl
PLS Expert User
Posts: 248
Joined: Sat Jul 25, 2009 1:34 pm
Real name and title:

Christian

elianacarraca
PLS User
Posts: 13
Joined: Mon Oct 13, 2008 5:30 pm
Real name and title:
cringle wrote:Hi,

the equation is different because we intend to compute the (relative) effect of a certain latent variables that is related to the latent variable under analysis.

Best
Christian
Professor Christian,

By using and comparing f2, is it possible to state that on effect is relatively stronger than the other, assuming we have a model with 2 IV's and 1 DV? Is there a way to test one effect is significantly stronger than the other in smartpls?

Thanks!

Eliana

SMARB
PLS Junior User
Posts: 6
Joined: Thu Sep 16, 2010 4:51 pm
Real name and title:
Location: Hellevoetsluis, Tilburg, Arnhem - The Netherlands
Dear all,

Can f² be negative and if so, what does this conceptually means? = That the LV does make the structural model worse/has no effect on this structural model I use, even though the pad coefficient is significant and R² value is ok (0,475).

Should this LV be deleted from the structural model if this makes conceptually sense?

Bram

MarketingStudent
PLS Junior User
Posts: 5
Joined: Tue Aug 03, 2010 9:53 pm
Real name and title:

Bootstrapping after Blindfolding? See Tenenhaus et al.(2005)

Dear Professor Ringle,

given my complex model I cannot apply the q² formula to calcualte the effect size of Q² , since I have several dependent variables etc. When applying the formula and excluding the variable that is located in the middle of my model (being exogenous and endogenous at once) I cannot run the model. However I would like to at least evaulate the relevance of my Q² values, which are all above 0.

Is there a way to assess at least whether my Q² values (CV-redundancy) are significant?

Did I understand it the wrong way or does the follwoing quote by Tenenhaus et al. (2005; p. 174) suggest to run a bootstrapping test for the Stone-Geisser's Q²? But why? the bootstrapping results provide only t-values for path coefficients and loadings.

"Following Wold (1982, p. 30), the cross-validation test of Stone and Geisser fits soft modeling like hand in glove. In PLS path modeling statistics on each block and on each structural regression are available.The significance levels of the regression coefficients can be computed using the usual Student’s t statistic or using cross-validation methods like jack-knife or bootstrapping available in PLS-Graph. "[/size]

Furthermore, on page 182 it sais "We may notice that in the ECSI model only the block customer satisfaction has an acceptable cv-redundancy index F²" referring to the table. However, the constructs in the table are all above 0, which is why I don't understand why only the customer satisfaction construct has a relevant Q² (cv-redundancy index F²). (?)

Or is there any article stating acceptable levels of Q² (other than only Henseler et al. Q²>0)?

Thank you very much in advance for any input!

Best, michelle
cringle wrote:Hi!

Since the Stone-Geisser criterion (usually) refers to the structural model, Q² should be used representing the cv-redundancy (Chin 1998, 318).
viewtopic.php?t=336
viewtopic.php?t=292

Please check the literature for additional explanations on the difference between cv-communality and cv-redundancy.

The research sets this number - I usually use seven - make sure that the number of observations is not a multiple of the omission distance (blindfolding does not work in that case).

Yes, q² = [Q²(included ) - Q²(excluded)] / 1 - Q²(included) - this is not an automatic function because the user must decide what to do when not only a single latent variable but complete model structures that are related to this variable are excluded.

Best
Christian

vonbergh
PLS Expert User
Posts: 60
Joined: Tue Jan 26, 2010 10:06 pm
Real name and title:
Dear Christian,

cringle wrote:Ad 2): The research sets this number - I usually use seven - make sure that the number of observations is not a multiple of the omission distance (blindfolding does not work in that case).
I have three questions should this is the case (i.e. if my dataset contains 7*47=329 records):

(1) do I use a different omission distance (e.g. 6 or 8)?
(2) do I take different steps?
(3) how do I report these findings (i.e. do I include reporting the omission distance)?

All the best,
Dennis

cringle
SmartPLS Developer
Posts: 797
Joined: Tue Sep 20, 2005 9:13 am
Real name and title: Prof. Dr. Christian M. Ringle
Location: Hamburg (Germany)
Contact:
(1) do I use a different omission distance (e.g. 6 or 8)?
In that case: 8
(2) do I take different steps?
No
(3) how do I report these findings (i.e. do I include reporting the omission distance)?