Significances / Covariation / Weights /Quality Indicators
Significances / Covariation / Weights /Quality Indicators
Hi there
I am new to PLS - just switched over from AMOS - and I tried to read up on some of the basic differences. And indeed PLS sounds quite intriguing especially considering the formative nature of its approach. Nevertheless I have a couple pressing questions:
1. Can two LVs covariate or is that not possible due to the different calculation as opposed to SEM? In SmartPLS I tried to draw a connection from LV1 to LV2 and vice versa but the pogram painted the arrow red as soon as I put in the vice versa connection and refused the calculation.
2. Where are the signficances of the regressions weights? I cannot find them or doesn't PLS make any assumption on these? How would I interpret the weights then? Or even when would I interpret them at all?
3. How would I interpret the quality indicators in the output. What is "good" and what is "bad"?
4. What is considered a "good" weight for an MV onto an LV factor? Especially if I have an LV with 20 MVs on it.
Looking forward to some enlightenment.
Regards
Niels
I am new to PLS - just switched over from AMOS - and I tried to read up on some of the basic differences. And indeed PLS sounds quite intriguing especially considering the formative nature of its approach. Nevertheless I have a couple pressing questions:
1. Can two LVs covariate or is that not possible due to the different calculation as opposed to SEM? In SmartPLS I tried to draw a connection from LV1 to LV2 and vice versa but the pogram painted the arrow red as soon as I put in the vice versa connection and refused the calculation.
2. Where are the signficances of the regressions weights? I cannot find them or doesn't PLS make any assumption on these? How would I interpret the weights then? Or even when would I interpret them at all?
3. How would I interpret the quality indicators in the output. What is "good" and what is "bad"?
4. What is considered a "good" weight for an MV onto an LV factor? Especially if I have an LV with 20 MVs on it.
Looking forward to some enlightenment.
Regards
Niels
- Diogenes
- PLS Super-Expert
- Posts: 899
- Joined: Sat Oct 15, 2005 5:13 pm
- Real name and title:
- Location: São Paulo - BRAZIL
- Contact:
Hi all,
1) In PLS isn´t necessary to put a arrow meaning covariation between LV, in the output we´ll have a matrix of correlations between LVs.
2) In the output we have the t-values of each path, then we can see what is significant or not (usually t > 2 is considered significant).
3) QUALITY INDICATORS:
3.1 FOR MEASUREMENT MODEL
3.1.1 For outward indicators (reflective or mode A)
AVE > 50%
Cronbach´s Alpha > 0,7
Composite reliability > 0,7
Loads > 0,7 (to communality > 50%) and t>2
3.1.2 For inward indicators (formative or Mode B)
The theory is even more important (Rossiter says that content vality is all that matters).
3.2 FOR STRUCTURAL MODEL
Paths with t>2
LV with R2>50%
4) A good weight is a significant weight in terms of theory building, but in terms of prediction we should have high values of R2.
Reference:
ROSSITER, John R. The C-OAR-SE procedure for scale development in Marketing. International Journal of Research in Marketing. v.19, p.305-335, 2002.
I hope it helps...
Best Regards.
1) In PLS isn´t necessary to put a arrow meaning covariation between LV, in the output we´ll have a matrix of correlations between LVs.
2) In the output we have the t-values of each path, then we can see what is significant or not (usually t > 2 is considered significant).
3) QUALITY INDICATORS:
3.1 FOR MEASUREMENT MODEL
3.1.1 For outward indicators (reflective or mode A)
AVE > 50%
Cronbach´s Alpha > 0,7
Composite reliability > 0,7
Loads > 0,7 (to communality > 50%) and t>2
3.1.2 For inward indicators (formative or Mode B)
The theory is even more important (Rossiter says that content vality is all that matters).
3.2 FOR STRUCTURAL MODEL
Paths with t>2
LV with R2>50%
4) A good weight is a significant weight in terms of theory building, but in terms of prediction we should have high values of R2.
Reference:
ROSSITER, John R. The C-OAR-SE procedure for scale development in Marketing. International Journal of Research in Marketing. v.19, p.305-335, 2002.
I hope it helps...
Best Regards.
Prof. Dr. Diogenes de Souza Bido
-
- PLS Junior User
- Posts: 6
- Joined: Mon Apr 03, 2006 8:21 am
- Real name and title:
hi,
regarding your 2nd question it is appropriate using the bootstrap resampling procedure. i suggest a number of 1000 samples, the cases should equal your n. then you can easily calculate your significances in the output on the t values (for infinite n) (e.g., t>2.576 means p<0.01 for a two-sided test). for more detailed information see the other threads in the methodology section.
best,
steffen
regarding your 2nd question it is appropriate using the bootstrap resampling procedure. i suggest a number of 1000 samples, the cases should equal your n. then you can easily calculate your significances in the output on the t values (for infinite n) (e.g., t>2.576 means p<0.01 for a two-sided test). for more detailed information see the other threads in the methodology section.
best,
steffen
- cringle
- SmartPLS Developer
- Posts: 818
- Joined: Tue Sep 20, 2005 9:13 am
- Real name and title: Prof. Dr. Christian M. Ringle
- Location: Hamburg (Germany)
- Contact:
Hi,
ad bootstrapping: the "search" function of this forum might be helpful as well. Please check this link (one of the search results):
viewtopic.php?t=151&highlight=bootstrapping
Best
Christian
ad bootstrapping: the "search" function of this forum might be helpful as well. Please check this link (one of the search results):
viewtopic.php?t=151&highlight=bootstrapping
Best
Christian
Prof. Dr. Christian M. Ringle, Hamburg University of Technology (TUHH), SmartPLS
- Literature on PLS-SEM: https://www.smartpls.com/documentation
- Google Scholar: https://scholar.google.de/citations?use ... AAAJ&hl=de
- Literature on PLS-SEM: https://www.smartpls.com/documentation
- Google Scholar: https://scholar.google.de/citations?use ... AAAJ&hl=de
Stefen and Christian,
Thank you for your follow-ups.
After reading the related threads, I came up the following questions:
1. Stefen once posted (viewtopic.php?t=206) that the literature is not clear enough whether we should use one-sided or two-sided test for t-values. Accordingly, the comments in this Forum seem to be mixed (for example, your thread recommends one-sided: viewtopic.php?t=151, while others use two-sided: viewtopic.php?t=223). Then, how could we know which to use and when?
2. Going back to Niels's original question 1, in PLS, how can we examine the "measurement model"? I recall some literature recommend two-step procedures in PLS (measurement -> structural model), just like SEM (e.g., Maes et al., 2005, "Measuring user beliefs and attitudes towards conceptual models: a factor and structural equation model, Working Paper, Universiteit Gent; Sanchez-Franco, M.J., 2006, "Exploring the influence of gender on the web usage via partial least squares, Beahviour & Information Technology, 25 (1), 19-36)
Any comments will be appreciated. Thanks.
Kind regards,
Shintaro
Thank you for your follow-ups.
After reading the related threads, I came up the following questions:
1. Stefen once posted (viewtopic.php?t=206) that the literature is not clear enough whether we should use one-sided or two-sided test for t-values. Accordingly, the comments in this Forum seem to be mixed (for example, your thread recommends one-sided: viewtopic.php?t=151, while others use two-sided: viewtopic.php?t=223). Then, how could we know which to use and when?
2. Going back to Niels's original question 1, in PLS, how can we examine the "measurement model"? I recall some literature recommend two-step procedures in PLS (measurement -> structural model), just like SEM (e.g., Maes et al., 2005, "Measuring user beliefs and attitudes towards conceptual models: a factor and structural equation model, Working Paper, Universiteit Gent; Sanchez-Franco, M.J., 2006, "Exploring the influence of gender on the web usage via partial least squares, Beahviour & Information Technology, 25 (1), 19-36)
Any comments will be appreciated. Thanks.
Kind regards,
Shintaro
-
- PLS Expert User
- Posts: 54
- Joined: Wed Oct 19, 2005 5:53 pm
- Real name and title:
Hi all,
1) regarding the one- or two-sided significance tests
I would apply the same logic as in any test of the significance of correlation coefficients. If you already have an expectation regarding the sign of your coefficient (based on theoretical considerations), then use a one-sided test. If you don't, use a two-sided test. For indicator loadings and weights, positive values are typically expected. Hence, a one-sided test should also be appropriate here.
My earlier comment regarding the non-explicitness of the literature in this regard just refers to the fact that authors typically do not report the critical t-values they used in bootstrap significance testing. They tend to just present the significance levels (and sometimes also the t-values).
2) two-step approach
A two-step approach doesn't really make any sense in PLS since the algorithm requires both the measurement and the structural model for parameter estimation. You simply can't just estimate the measurement model first without specifying structural relationships.
However, some people may choose to first create LV-scores in PLS (using both the measurement AND the structural model for their estimation) and then do additional analyses on these LV-scores in a different statistical package (e.g., for interaction effects, etc.). In effect, working with LV-scores is "fixing" the measurement model. This approach comes closest to the two-step approach you described.
Obviously, you could also just simply report it as two separate steps to make it look more "methodological" (which is probably what these authors did).
So long,
Stefan
1) regarding the one- or two-sided significance tests
I would apply the same logic as in any test of the significance of correlation coefficients. If you already have an expectation regarding the sign of your coefficient (based on theoretical considerations), then use a one-sided test. If you don't, use a two-sided test. For indicator loadings and weights, positive values are typically expected. Hence, a one-sided test should also be appropriate here.
My earlier comment regarding the non-explicitness of the literature in this regard just refers to the fact that authors typically do not report the critical t-values they used in bootstrap significance testing. They tend to just present the significance levels (and sometimes also the t-values).
2) two-step approach
A two-step approach doesn't really make any sense in PLS since the algorithm requires both the measurement and the structural model for parameter estimation. You simply can't just estimate the measurement model first without specifying structural relationships.
However, some people may choose to first create LV-scores in PLS (using both the measurement AND the structural model for their estimation) and then do additional analyses on these LV-scores in a different statistical package (e.g., for interaction effects, etc.). In effect, working with LV-scores is "fixing" the measurement model. This approach comes closest to the two-step approach you described.
Obviously, you could also just simply report it as two separate steps to make it look more "methodological" (which is probably what these authors did).
So long,
Stefan
Hi,
I have problem with significant and non-significant indicators of formatives. what does it really mean when we say "A good weight is a significant weight in terms of theory building, but in terms of prediction we should have high values of R2. " significant weight is somehow confusing. for loadings, we have 0.7 , but what about the formatives and significant weight?
I have problem with significant and non-significant indicators of formatives. what does it really mean when we say "A good weight is a significant weight in terms of theory building, but in terms of prediction we should have high values of R2. " significant weight is somehow confusing. for loadings, we have 0.7 , but what about the formatives and significant weight?
- Diogenes
- PLS Super-Expert
- Posts: 899
- Joined: Sat Oct 15, 2005 5:13 pm
- Real name and title:
- Location: São Paulo - BRAZIL
- Contact:
Hi Shima,
For reflective indicators we must have loadings > 0,7 (communality > 0,5), it´s the same logic as factor analysis. But, for formative indicator we use the same logic as multiple regression (the beta values mustn´t be greater than a specific value).
A good example of construction of a measurement model with formative indicators could be saw in:
HELM, Sabrina. Design a formative measure for corporate reputation. Corporate Reputation Review. v.8, n.2, p.95-109, 2005.
She begins with 59 possible indicators, and after a three-step pretest only 10 indicators remain in the final model (6 of them with significant weights).
Best Regards.
Bido
For reflective indicators we must have loadings > 0,7 (communality > 0,5), it´s the same logic as factor analysis. But, for formative indicator we use the same logic as multiple regression (the beta values mustn´t be greater than a specific value).
A good example of construction of a measurement model with formative indicators could be saw in:
HELM, Sabrina. Design a formative measure for corporate reputation. Corporate Reputation Review. v.8, n.2, p.95-109, 2005.
She begins with 59 possible indicators, and after a three-step pretest only 10 indicators remain in the final model (6 of them with significant weights).
Best Regards.
Bido
-
- PLS Junior User
- Posts: 1
- Joined: Fri Oct 19, 2012 7:11 pm
- Real name and title:
Further to Stefan's first point above. So if we have a model in which we have an expectation regarding the sign of SOME of the path coefficients (based on theoretical considerations) but not others, can we use a mix of one-tailed and two-tailed tests in determining the significance of the various path coefficients in our model?stefanbehrens wrote:Hi all,
1) regarding the one- or two-sided significance tests
I would apply the same logic as in any test of the significance of correlation coefficients. If you already have an expectation regarding the sign of your coefficient (based on theoretical considerations), then use a one-sided test. If you don't, use a two-sided test. For indicator loadings and weights, positive values are typically expected. Hence, a one-sided test should also be appropriate here.
My earlier comment regarding the non-explicitness of the literature in this regard just refers to the fact that authors typically do not report the critical t-values they used in bootstrap significance testing. They tend to just present the significance levels (and sometimes also the t-values).
2) two-step approach
A two-step approach doesn't really make any sense in PLS since the algorithm requires both the measurement and the structural model for parameter estimation. You simply can't just estimate the measurement model first without specifying structural relationships.
However, some people may choose to first create LV-scores in PLS (using both the measurement AND the structural model for their estimation) and then do additional analyses on these LV-scores in a different statistical package (e.g., for interaction effects, etc.). In effect, working with LV-scores is "fixing" the measurement model. This approach comes closest to the two-step approach you described.
Obviously, you could also just simply report it as two separate steps to make it look more "methodological" (which is probably what these authors did).
So long,
Stefan
Many thanks for a reply and good day!
Regards
Paz