How to test for Common Method Variance in PLS?
How to test for Common Method Variance in PLS?
Dear all,
I am currently trying to test my dataset in PLS for common method variance (CMV) (possible bias in the data because both independent as dependent variables were collected at the sime time from the same source). Podsakoff 2003 gives several excellent methodological suggestions how one should deal with CMV in research.
I found an interesting article on the internet in which PLS is used to test for CMV (I also found similar published articles).
http://www.hfncenter.org/cms/book/export/html/243
However, this article contains the following CMS tess which I do not understand how to perform in PLS:
1) "Second, a partial correlation method was used (Podsakoff and Organ 1986). The highest factor from the principal component factor analysis was added to the PLS model as a control variable on all dependent variables. According to Podsakoff and Organ, this factor is assumed to contain the best approximation of the common method variance if is a general factor on which all variables load (p. 536). This factor did not produce a significant change in variance explained in any of the three dependent variables, again suggesting no substantial common method bias."
How does one add a factor from Principal component analysis to PLS? I know how to conduct a principal component analysis in SPSS, but I don't understand how you add the highest factor into PLS??
Thank you very much for your feedback in advance.
Kind regards,
Maurice
literature:
P. Podsakoff, S. MacKenzie, J. Lee, 2003, "Common Method Biases in Behavioral Research: A Critical Review of the
Literature and Recommended Remedies", Journal of Applied Psychology, Vol. 88, No. 5, 879–903
I am currently trying to test my dataset in PLS for common method variance (CMV) (possible bias in the data because both independent as dependent variables were collected at the sime time from the same source). Podsakoff 2003 gives several excellent methodological suggestions how one should deal with CMV in research.
I found an interesting article on the internet in which PLS is used to test for CMV (I also found similar published articles).
http://www.hfncenter.org/cms/book/export/html/243
However, this article contains the following CMS tess which I do not understand how to perform in PLS:
1) "Second, a partial correlation method was used (Podsakoff and Organ 1986). The highest factor from the principal component factor analysis was added to the PLS model as a control variable on all dependent variables. According to Podsakoff and Organ, this factor is assumed to contain the best approximation of the common method variance if is a general factor on which all variables load (p. 536). This factor did not produce a significant change in variance explained in any of the three dependent variables, again suggesting no substantial common method bias."
How does one add a factor from Principal component analysis to PLS? I know how to conduct a principal component analysis in SPSS, but I don't understand how you add the highest factor into PLS??
Thank you very much for your feedback in advance.
Kind regards,
Maurice
literature:
P. Podsakoff, S. MacKenzie, J. Lee, 2003, "Common Method Biases in Behavioral Research: A Critical Review of the
Literature and Recommended Remedies", Journal of Applied Psychology, Vol. 88, No. 5, 879–903
With kind regards,
Maurice de Rochemont
Maurice de Rochemont
- Diogenes
- PLS Super-Expert
- Posts: 899
- Joined: Sat Oct 15, 2005 5:13 pm
- Real name and title:
- Location: São Paulo - BRAZIL
- Contact:
Hi Maurice,
The article by Podsakoff et al (2003) was the Best Paper in Industrial and Organizational Psychology in 2005 http://www.siop.org/tip/backissues/July ... 1to136.pdf
For using PLS, I recommend:
LIANG, H.; SARAF, N.; HU, Q.; XUE, Y. Assimilation of enterprise systems: the effect of institutional pressures and the mediating role of top management. MIS Quarterly. v.31, n.1, p.59-87, mar/2007.
This article has a detailed explanation in its appendix E on how to use PLS to assess CMB, see pages 71 and 85-87.
About your question, if you want follow what was said in your item 1, you should run Principal Components in SPSS to identify the "highest factor" = first principal component before rotation, after that you could create a new model with this LV included.
Best regards
Bido
The article by Podsakoff et al (2003) was the Best Paper in Industrial and Organizational Psychology in 2005 http://www.siop.org/tip/backissues/July ... 1to136.pdf
For using PLS, I recommend:
LIANG, H.; SARAF, N.; HU, Q.; XUE, Y. Assimilation of enterprise systems: the effect of institutional pressures and the mediating role of top management. MIS Quarterly. v.31, n.1, p.59-87, mar/2007.
This article has a detailed explanation in its appendix E on how to use PLS to assess CMB, see pages 71 and 85-87.
About your question, if you want follow what was said in your item 1, you should run Principal Components in SPSS to identify the "highest factor" = first principal component before rotation, after that you could create a new model with this LV included.
Best regards
Bido
Hello,
I have 3 independent variables and a dependent variable. As per theory, all IVs are +vely related (positive path coefficients) to the DV.
To assess the common method bias, when I add the highest factor from the principal component factor analysis to the PLS model as a control variable and run the model, all path coefficients between IVs and DV are changing to -ve. I am not sure why this is happenning, and how to interpret it.
I'll apprecite your help and suggestions.
Thanks
George
I have 3 independent variables and a dependent variable. As per theory, all IVs are +vely related (positive path coefficients) to the DV.
To assess the common method bias, when I add the highest factor from the principal component factor analysis to the PLS model as a control variable and run the model, all path coefficients between IVs and DV are changing to -ve. I am not sure why this is happenning, and how to interpret it.
I'll apprecite your help and suggestions.
Thanks
George
-
- PLS Expert User
- Posts: 37
- Joined: Tue Mar 25, 2008 9:16 am
- Real name and title:
Dear Bido,Diogenes wrote:Hi Maurice,
For using PLS, I recommend:
LIANG, H.; SARAF, N.; HU, Q.; XUE, Y. Assimilation of enterprise systems: the effect of institutional pressures and the mediating role of top management. MIS Quarterly. v.31, n.1, p.59-87, mar/2007.
This article has a detailed explanation in its appendix E on how to use PLS to assess CMB, see pages 71 and 85-87.
I checked the article you referred to regarding CMB. If I understood their procedure correctly, they converted each indicator to a first-order-construct which is measured by the corresponding indicator. The first-order-constructs are then linked to a second-order-construct (= the first-order-construct before transformation).
Wouldn't it be problematic to model this approach in Smart-PLS because the 2nd-order-construct then lacks indicators for measurement? What would be a feasible approach there? Maybe using the repeated indicator approach?
Any comments would be highly appreciated.
Thanks,
Stephan
-
- PLS Junior User
- Posts: 8
- Joined: Wed Dec 09, 2009 10:39 am
- Real name and title:
- Diogenes
- PLS Super-Expert
- Posts: 899
- Joined: Sat Oct 15, 2005 5:13 pm
- Real name and title:
- Location: São Paulo - BRAZIL
- Contact:
Yes,
For the new 2nd order LV use the repeated indicator approach.
One way to do this faster is:
1) Copy the original model in the SmartPLS
2) For each 1st order LV:
- Hide its indicators (it will be a 2nd order LV in the final model)
- Create one new LV for each indicator (one LV / indicator)
- Connect the 2nd order LV with these new LV (that has just one indicator)
3) The “method” LV
- Create a new LV
- Use all indicators (repeated indicator approach)
- Connect this LV to all first order LV (that was created for each indicator)
4) Run and compute AVE
Best regards,
Bido
For the new 2nd order LV use the repeated indicator approach.
One way to do this faster is:
1) Copy the original model in the SmartPLS
2) For each 1st order LV:
- Hide its indicators (it will be a 2nd order LV in the final model)
- Create one new LV for each indicator (one LV / indicator)
- Connect the 2nd order LV with these new LV (that has just one indicator)
3) The “method” LV
- Create a new LV
- Use all indicators (repeated indicator approach)
- Connect this LV to all first order LV (that was created for each indicator)
4) Run and compute AVE
Best regards,
Bido
-
- PLS Expert User
- Posts: 37
- Joined: Tue Mar 25, 2008 9:16 am
- Real name and title:
Thanks a lot Bido! This is very helpful!
One last question: Do you know whether this CMB-approach is only feasible for pure reflective measurement models or is there a possibility to compute formativ and 'mixed' models (in a way that some constructs are measured reflective and some formative) as well? At least the 'mixed' mode would lead to problems when assigning the indicators to the method factor...Using LV-scores would not be a proper workaround, would it?
Best regards, Stephan
One last question: Do you know whether this CMB-approach is only feasible for pure reflective measurement models or is there a possibility to compute formativ and 'mixed' models (in a way that some constructs are measured reflective and some formative) as well? At least the 'mixed' mode would lead to problems when assigning the indicators to the method factor...Using LV-scores would not be a proper workaround, would it?
Best regards, Stephan
-
- PLS Junior User
- Posts: 8
- Joined: Wed Dec 09, 2009 10:39 am
- Real name and title:
Thank you for your reply! However, I now have another question to which I couldn't find an answer on previous threads. It would be great if someone could shed light on this issue:
Liang et al. state that to assess the extent of common method bias, "the squared values of the method factor loadings were interpreted as the percent of indicator variance caused by the method, whereas the squared loadings of substantive constructs were interpreted as the percent of indicator variance caused by substantive constructs" (p.87).
I am unsure about which numbers to compare. In the Liang et al. paper, I conclude from Figure E3 that what they actually used are the path coefficients (connecting the latent first-order variables to Method and Second order variables). However, using the repeated indicator (superblock) approach, there is a choice between
a) these path coefficients
b) loadings of the "superblock" indicators on the latent method and second order variables
Which is the appropriate number? One of these? Something else entirely?
Thank you so much in advance!!
Liang et al. state that to assess the extent of common method bias, "the squared values of the method factor loadings were interpreted as the percent of indicator variance caused by the method, whereas the squared loadings of substantive constructs were interpreted as the percent of indicator variance caused by substantive constructs" (p.87).
I am unsure about which numbers to compare. In the Liang et al. paper, I conclude from Figure E3 that what they actually used are the path coefficients (connecting the latent first-order variables to Method and Second order variables). However, using the repeated indicator (superblock) approach, there is a choice between
a) these path coefficients
b) loadings of the "superblock" indicators on the latent method and second order variables
Which is the appropriate number? One of these? Something else entirely?
Thank you so much in advance!!
-
- PLS Junior User
- Posts: 8
- Joined: Wed Dec 09, 2009 10:39 am
- Real name and title:
OK, thinking about it again, I decided to go with a)- the path coefficients of the "latent indicator variables".These seem appropriate to me as they are really the indicators of the second-order variables.I am unsure about which numbers to compare. In the Liang et al. paper, I conclude from Figure E3 that what they actually used are the path coefficients (connecting the latent first-order variables to Method and Second order variables). However, using the repeated indicator (superblock) approach, there is a choice between
a) these path coefficients
b) loadings of the "superblock" indicators on the latent method and second order variables
Which is the appropriate number? One of these? Something else entirely?
-
- PLS Junior User
- Posts: 6
- Joined: Sun Apr 11, 2010 8:40 pm
- Real name and title:
-
- PLS Expert User
- Posts: 248
- Joined: Sat Jul 25, 2009 1:34 pm
- Real name and title:
Hey Criston,
It is easy to produce a LV with "repeated indicator”. You have only to take a second time the indicators which you already used (drag and drop).
How it works you can see here: http://www.youtube.com/watch?v=7XDDTofpboM
Furthermore there is an working paper that could be interesting for the discussion her:
Ylitalo (2009): Controlling for Common Method Variance with Partial Least
Squares Path modeling: A Monte Carlo Study; http://www.sal.tkk.fi/Opinnot/Mat-2.108 ... yli09b.pdf
Best regards
Christian
It is easy to produce a LV with "repeated indicator”. You have only to take a second time the indicators which you already used (drag and drop).
How it works you can see here: http://www.youtube.com/watch?v=7XDDTofpboM
Furthermore there is an working paper that could be interesting for the discussion her:
Ylitalo (2009): Controlling for Common Method Variance with Partial Least
Squares Path modeling: A Monte Carlo Study; http://www.sal.tkk.fi/Opinnot/Mat-2.108 ... yli09b.pdf
Best regards
Christian
Last edited by christian.nitzl on Sat Jun 19, 2010 9:40 am, edited 1 time in total.
-
- PLS Junior User
- Posts: 6
- Joined: Sun Apr 11, 2010 8:40 pm
- Real name and title:
-
- PLS Junior User
- Posts: 6
- Joined: Sun Apr 11, 2010 8:40 pm
- Real name and title:
-
- PLS Expert User
- Posts: 248
- Joined: Sat Jul 25, 2009 1:34 pm
- Real name and title:
- Diogenes
- PLS Super-Expert
- Posts: 899
- Joined: Sat Oct 15, 2005 5:13 pm
- Real name and title:
- Location: São Paulo - BRAZIL
- Contact:
Hi,
1) To compute the variances we used the path coefficients, as Kristina had said.
See viewtopic.php?t=640
2) Common method variance (one assumption is that there is a common cause, what is connected with the reflective way of modeling).
This view was used in Ylitalo (2009):
p.8) Classical test theory
p.9) Harman´s single factor test
p.11) In these papers, a proxy for common method variance was formed by conducting an explanatory factor analysis on all items in the model, and using the first emerging unrotated factor as a control variable in the inner model.
p.15) reflective indicators
For these reasons, formative indicators were not used in the CMV calculations.
Best regards,
Bido
1) To compute the variances we used the path coefficients, as Kristina had said.
See viewtopic.php?t=640
2) Common method variance (one assumption is that there is a common cause, what is connected with the reflective way of modeling).
This view was used in Ylitalo (2009):
p.8) Classical test theory
p.9) Harman´s single factor test
p.11) In these papers, a proxy for common method variance was formed by conducting an explanatory factor analysis on all items in the model, and using the first emerging unrotated factor as a control variable in the inner model.
p.15) reflective indicators
For these reasons, formative indicators were not used in the CMV calculations.
Best regards,
Bido