Here's a fun pet project I've been working on: udreamed.com/. It is a dream analytics app. Here is the RUclips channel where we post a new video almost three times per week: ruclips.net/channel/UCiujxblFduQz8V4xHjMzyzQ Also available on iOS: apps.apple.com/us/app/udreamed/id1054428074 And Android: play.google.com/store/apps/details?id=com.unconsciouscognitioninc.unconsciouscognition&hl=en Check it out! Thanks!
A good common method bias paper is "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies" by MacKenzie and Podsakoff. The 0.20 threshold is just something I use. No particular reference for it except this video... Although I suppose you could quote Cohen who talks about effect sizes, with 0.20 as a moderate effect.
Thanks for such a fast reply ! I am not using a 2nd order factor. There are 4 factors: Attitude, Social Norm Support, Self Eficacy, and Entrepreneurial Intention. The instruments inted use is to use three of the constructs as dependent to predict entrepreneurial intention (4th) . Thank you so much !!!!!!!!!
Most data I have made available on my wiki, but the data in this video is not mine to give away. so sorry. It belongs to a colleague who has already declined my request to let others have it.
Cheers mate. I've used this method in conjunction with Harman's test in an effort to triangulate, and it appears that I have no real issue with CMV in any case, so all's good. Thanks for your quick response.
Exactly. If you are imputing for composites then you will be accounting for the common variance already, so you should not include the imputed CLF composite in the structural model. But, if you are NOT imputing (instead using the full latent model), you may want to retain the CLF so that the common variance can be accounted for.
Oh, ok. This makes more sense now. I don't know of any precedent for doing it three at a time, but on principle, that is perfectly fine. I would say that is sufficient evidence for no method bias. It would be a cool contribution in your paper as well to show that method bias is deceiving and fickle.
If someone wants to cite this, we have used this approach in our paper published in Journal of Business Ethics, You can cite the following: Serrano Archimi, C., Reynaud, E., Yasin, H.M. and Bhatti, Z.A. (2018), “How perceived corporate social responsibility affects employee cynicism: the mediating role of organizational trust”, Journal of Business Ethics, Vol. 151 No. 4, pp. 907-921.
After running it, you should be able to still see some output, although it will be unfinished. Go to the regression weights table to see which regressions have exceptionally high standard errors (like anything above 1 or 2). These are the culprits. If there are no problems there, check the standardized regression weights to see if any factor has a set of items that are negative. If so, then move the parameter constraint onto the positive one. Hope this helps!
Same thing sort of. The bias is referring to the confounding effect. The variance is referring to the amount of variance that is common to the entire set of items.
This is probably adequate model fit. We use multiple measures as a means of triangulation. It is uncommon to meet the threshold for every single one. You may also check Hair et al for more relative thresholds based on sample size and model complexity. I think it is table 12-4 in the 7th version of the book. hope this helps!
Dr. Gaskin, Thank you so much for this series. I owe my entire dissertation methods section to you. Quite honestly, you're videos have been a God-send. Would you be able to present a video on scalar (strong factoral) and strict (invariant uniqueness) measurement testing? I ask because relatively new literature (Putnick & Bornstein, 2016) suggest using a "ladder" of measurement testing going from configural -> metric -> scalar -> strict to ensure more robust invariance testing. Thank You!
Here is a more recent video that covers this ladder of validation: ruclips.net/video/sbTl-icEyD8/видео.html (I've queued the link to start about 3/5ths of the way through).
If you are looking for a CMB video in the SEM series, it would be in the one about CFA or measurement models. It is fine to have an SE greater than 1. abnormally high SE can be detected by comparing the SE to the unstandardized regression weight. If the SE is greater than the regression weight, then it is probably a problem.
Barbossa: "There are a lot of long words in there, [Vahid]; we're naught but humble pirates. What is it that you want?" If I understand you correctly, you want to know what the regression weights on the CLF means for each indicator? It means that the portion of variance in that indicator explained by the method is the standardized regression weight. The other portions of variance are accounted by error and trait (with the trait remaining on the original latent factors' indicators).
Dear James Gaskin, I have appreciated you for uploading this video. I want to use your all videos for spreading out research areas in my country if permission me.
Well done and thanks! I have a question. I noticed that regression weights of indicators regressed on the Common latent variable are negative (or zero). What do the negative values indicate? Additionally, I suppose some of the items do not load significantly on the Common latent variable. What does this lack of significant loading tell us? I suppose it would indicate that common method factor affects merely a portion of items not all. Would it be okay to delete the regression paths with p> 0.05
Negative regressions just mean that the indicator moves in the opposite direction to the majority of other indicators. Lack of significant loadings means that they are not affected by CMB. Do not remove non-significant paths because you want to see what variance is COMMON to ALL indicators in the model.
I cannot think of a way to do it off hand, but there is probably a way. I recommend looking at the SmartPLS forum. They are very helpful. I have never been disappointed with what I find there.
Yikes! that's trouble. However, the Harman's test is the least revered of CMB tests. I recommend using the CLF test if you don't have a marker variable (like social desirability).
I'm not sure I understand. If you are using a 2nd order factor for EI, then you still need to do the cmb test on the lower order dimensions. I hope this helps. If not, feel free to clarify.
Dear Prof. Gaskin, I learned a lot. thank you. Let me ask you one question. This 0.2 is used as the standard, but is this standard quoted from some previous research? If there is any previous research on this subject, please let me know. In setting this reference value, if there is a clear precedent research, I would like to cite that literature and write a paper.
That's odd about the imputing, although imputing is somewhat sensitive. It doesn't like you to have variable names that have spaces. If that is not the problem, then you may want to just move forward, retain the CLF, and use the full SEM with latent constructs.
I have a question, Is there a way to conduct a common method bias analysis using common latent factor with PLS? I don´t have a marker variable, and I would like to perform a common method bias test. By the way, I am a big fan of your videos. Yo have taught me a lot !!!
I appreciate your advice. I look forward to further videos from you. I wonder if you could teach us how to do CFA/SEM of dichotomous data on AMOS. I think AMOS does not have an option for using tetrachoric or polychoric correlation matrices, which are commonly used in estimating CFA models of dichotomous data. Do you recommend any certain approach?
Dear James, I have tried to find out the reference for the way you check CMB using Common Latent Factor (CLF), but failed, could you please suggest me some published articles that have use either following ways to check for CMB? (1) Compare the standardized regression weights from the model with CLC to the standardized regression weights of a model without the CLF. If there is no large differences (like greater than 0.200), then the CMB is not serious. (2) compare the model fit of the two CFAs, if CFA without CLF is better than CFA with CLF, thus CMB is not serious. Thank you very much in advance Angelina
I don't know about the model fit approach, and I've never seen it done (and it doesn't make a lot of sense to me...), but as for the comparing of standardized regression weights, no article talks about this or about a threshold for their difference. Podsakoff talks about method bias as does MacKenzie. The most recent article in marketing by the both of them would be a good resource. It's called: "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies"
James Gaskin Dear James, thank you very much for your prompt reply. May I ask whether this method to check CMB is equivalent to the way Liang, Saraf, Hu, and Xue (2007) "Assimilation of Enterprise Systems: The Effect of Institutional Pressures and the Mediating Role of Top Management" has done using PLS model with a CMF?
Angelina Hanh It is similar, but not the same. Liang et al., explain that the average loadings are much stronger for the factors than for the CLF and that they are mostly non-significant on the CLF. This is fine too.
I have looked around for a suggested cutoff, and no one has put a stake in the ground as far as I can tell. So, the 0.2 is simply an educated degree of tolerance (i.e., I'm willing to accept small differences between the effects, but not big ones). 0.2 is sometimes also used to distinguish a small effect from a large effect. I think Wynne Chin has published this threshold in MISQ. Others have also used this cutoff for effect size, but I can't remember who off the top of my head.
Thank you so much for your video. If I use Harman’s single factor test and the single factor accounts for more than 50% of the variance in the model, then what can I do? How can I overcome the problem?
Hello James, thanks for the good explanation, but I still have some questions: First, if you compare the standardized regression weights with and without CLF, would you keep the Marker variable in both models or also delete it in the without check? Second, would you do the CMB-test before you start testing the reliability and validity? and if yes and assuming CMB is an issue, would you keep the CLF all the time in the model to check for convergent and discriminant validity? It is not to easy to obtain validity with the lowered loadings and I couldn´t find something about the right order of doing all the pretests.
Nixen85 If you're using a marker variable, then the better approach would be to constrain the regression weights from the CLF to be equal, and then square their unstandardized loadings. This will give you the percent of shared variance. I would test validity prior to including the CLF. The CLF can extract shared variance which is actually trait variance, and so it can artificially deflate the construct reliabilities.
Hello James Thank you so much for your videos. I have learnt alot from them. Just a quick question regarding this video on CMV. You mention that do not include control variable in your structural model after imputing the data. But then when you convert CFA to SEM in this video, you retained the control. Aren't we suppose to delete the control variable, as its affects are going to be present anyway since the data has already been imputed? Thanks in advance.
Hi James, many thanks for your great videos. These have been very helpful! I have few questions in the case using CFA to confirm EFA results and using CLF to address the CMB after CFA: 1. How to calculate AVE and MSV to report the discriminant and convergent validity adjusted by CLF with composites as AMOS does not produce Correlation table anymore? (Am I mistaken anywhere?) 2. If the model fitness of the models with and without CLF is not much different, is it necessary to use the more complicated model (with CLF)? 3. Can you please recommend me any reference about the extent of the CMB that affects the model? How much difference between the standardised regression weight between 2 models can be considered small, medium or large? What to say about a difference >.2 with most items? Many thanks James!
+Van Nguyen 1. The correlation table should show up if you check the box for standardized estimates. 2. You only keep the CLF if there are some method bias issues. Otherwise, just get rid of it. 3. The best way to do it is actually to run it unconstrained with the CLF, and then run it with all paths from the CLF constrained to zero. Then you can do a chi-square difference test to see if the CMB is different from zero. I don't have a video for this yet...
+James Gaskin Thanks James! I really appreciate you addressing the comments and questions! 1. I adjusted all factors in the models (deleting all co-variance paths, replacing by single headed arrows). It's probably because there are no covariance anymore and that AMOS cant produce correlation table? What should I do in this case? 3. When I unconstrained the CLF (deleting all the a constraints as in the previous video), AMOS doesnt run at all. It popped up with the message about the one of the estimated covariance failed to be positive - it did not show up the output window for me to trouble shot the problem). Apart from AMOS instruction manual, would you recommend me any book to learn & play around with AMOS - as AMOS is obviously very sensitive?
+Van Nguyen 1. correct. if no covariances, then no correlations. You don't need correlations to run the model if it is just a causal path model. 2. The Barbara Byrne book "Structural Equation Modeling with AMOS".
Thanks for your reply. Then, if the latent variable CMB goes up by one unit, the magnitude of observed indicators goes down by Standardized Coefficients' magnitude. This seems to be statistical indication. Could you explain the substantive (theoretical) indication of this finding, please?
Hi, James. I learn a lot from your video. Thank you very much. It seems very clear and explicable. But I have a silly question. 'Cause I'm a new learner. I only know how to run SPSS. What's the technique tool you use (I can see in the video) to draw this beautiful model and do CFA, SEM, And CLF method? Thank you very much!
Dear Prof. Gaskin, Thank you so much for this great tutorial. In the video, you mentioned that the threshold for differences of standardized weights is 0.2. Could you please cite any reference mentioning it should be like this. I need it to refer in my thesis. Many thanks.
This difference is a personal preference (it's just exploratory after all). Nevertheless, there are citations that suggest it can be even closer (0.100). I think the PLS article by Gefen and Straub (2008?) talks about this. I think the Hair book (2010) also talks about this.
I am not familiar with the instrumental variable technique (unless that is referring to the marker variable technique - if so, I have a video for this as well).
Thank you so much for the excellent videos. I have found them very helpful to suppliment my reading and advice from my supervisor. Could you perhaps point me in the direction of your source for the .2 cutoff for difference in standardised regression between the two models? I've followed the citation trail from Podsakoff et al 2003, but I'd just like to make sure.
The only published threshold that I'm aware of is 50%, but this is pretty high. Most modern approaches at CMB run a chi-square difference test to see if the unconstrained model is different from a model where all paths from the CLF are constrained to zero (implying zero common shared variance).
It should work in the path model or hybrid model, but won't work in the CFA/measurement model. If you are in the hybrid model, then you might have simply forgotten to include error terms for new endogenous variables.
Hi James, Thanks for this video. I did not understand one point. First you say, you don't take the common latent factor into the SEM model, because you partial out the variance before by building composites. However, at the end of the video, you still leave the common factor in there, when running the SEM model. Could you please explain? Would you get the same results, when constraining the factor loadings in the SEM model to the decreased values you got when including the common factor in the correlation analysis, but then taking out the common factor in the SEM model? Thanks so much, Fu
Fu Yin-Heinberg If you decide to retain the CLF when imputing composites, then the composites already account for the method bias which has been extracted by the CLF. Thus you don't need to include the CLF variable when drawing the path model using composites. However, if you decide to not create composite, but instead use a latent path model, then you should retain the CLF in the model (if CMB was an issue). If CMB is not an issue, then you can remove the CLF before creating composites. As for constraining vs. not constraining the paths from the CLF, you can try it and see, but the unstandardized results will differ.
Hi James, thanks for taking the time to put this video together. I tried your method but the model did not run for some reason. Would you have any idea why it did not run?
hishonline when looking at standardized estimates, 0.200 is a meaningful difference (in my opinion). Sort of like in an EFA when you are looking for cross-loadings within 0.200. I'm not sure if there is anything published about a threshold though.
Dear James I really like your videos. I have studied various articles about common method bias and don't found the 'common latent factor' technique. They have used other names and I am really very confused about it. I have read your suggested artile Sources of Method Bias in Social Science Research and Recommendations on How to Control It byPhilip M. Podsakoff,1 Scott B. MacKenzie,2 and Nathan P. Podsakoff3. But they have used different terms. Please guide which method reflects CLF?
Hello Professor, Thank you so much for this excellent video. I have two questions: 1) Can we perform common latent factor technique on path analysis model instead of measurement model? 2) Is common latent factor test also called as Harman's single factor test or both are different?
Hi James..Thanks for these wonderful videos..what reference/citation can we give for using this correct way? podsakoff orany other? also, whom do we quote when we say that the threshhold value of difference of standarized estimates (with and without common factor) is 0.2..Thanks
Thx, James. I watched the "inputing composites" video and my model works. but that one is without CLF. It still doese not work with CLF. My factor names do not have spaces and there is no missing data.
Dear James, Your videos are very informational, thanks a lot ! I am testing Linan & Chen (2009) entrepreneurial intention scale for CMB. This instrument has 4 dimensions with 20 itens. As expected the EI is highly correlated with attitude. So when I use a single factor, I get a factor loading on the com. fact, but when I do test 3 dimensions at time (3 and CFACT), do not. Do you think it is enough to provide evidence of no CMB ? Have seen any advice to test when variables are highly corr?
Many thanks for directing me to this video follwoing my earlier query. I have followed this and when I remove the CLF, my df increases by 1 as expected but my st. regression weights remain exactly the same. My delta scores are zero throughout. Is there any possible explanation for this?
Fantastic video series...thanks! Could you please clarify the process for handling CLF in a 2nd order model? I've added the CLF to my 2nd order model, it runs, and the fit is good. But when I impute I get an error that says "a sample of parameter values was inadmissable." Just for fun I've tried this on a 1st order version of my model and impute works...do I need to impute in my 2nd order model or can I just go straight to the structural model and keep the CLF in the picture? thanks....
This is quite helpful. I've been racking my brain trying to figure out how to do this in SAS PROC CALIS. I haven't found much in the documentation or at SUGI; has anyone had any luck with something like this in SAS?
Hi James. I'm wondering how this newer latent factor method relates to the marker variable analysis. Should I add in the marker variable and then compare the regression weights (with and without the common factor)? or Should I stick to the marker variable method presented in the other video?
Hi James, Thank you so much for sharing your wonderful videos. How to adjust common method method bias in 2nd order constructs as data imputation is not done in this. In SEM, can we show the Common latent factor with the 2nd order construct the way you have shown and used in your video?
the common method bias occurs at the item level, so whether it is second or first order makes no difference. Check it at the first order level, and if it is a problem, then simply retain the CLF when creating the 2nd order model.
Many thanks for your extremely helpful videos in RUclips. I have a quick question which I hope you could answer if at all possible. I followed your recommendations for analysis with a single common method factor, and found that common method bias is a problem in my data - however I’m neither able to impute the variables accounting for CMB nor to run a SEM keeping the CLF. The message I get is that 1 parameter sample is inadmissible. CFA analyses run normally, and error variances are not negative, but most covariances are. I read this may be due to my small sample size (N 170 for 34 observed variables) and that a solution could be using Bayesian estimation, but I am unsure about how to proceed. Is there a way of manually imputing the data based on the estimates I get from the CFA with CLF? What would be your advice?
+elisaalt It might be sample size, but probably not. 170 is probably fine. Check to see if there are weird loadings, or any parameters with really high standard errors. These will show you were the problems are.
Hi James, first, thank you for excellent Stat-wiki! Second, I really like the idea of using CLF in order to create "common method bias adjusted composites" but could you give me a reference on this? Thank you so much for the great work! Nico
Here is a paper I just had published at JAIS: Paul Benjamin Lowry, James Gaskin, Nathan Twyman, Bryan Hammer, and Tom L. Roberts (2013). “Taking “Fun and Games” Seriously: Proposing the Hedonic-Motivation System Adoption Model (HMSAM),” Journal of the Association for Information Systems (Volume 14, Issue 11, pp. 617-671). In this article we use CLF adjusted values (see page 630).
James Gaskin Thanks a lot, I will definitely look at the article once its published (I can't access it just yet... stuck at issue #10 :) ) and I'll be more than happy to use your article as a reference for the usage of the method if I get a chance. Thanks again!
Hi James, thank you for the great video's.When I include the CLF like you did, the model does not run. By this I mean that it looks like it is running, but there is no option to click on the output. I get no error whatsoever.
Does the model run when not trying to impute? Also, do any of the factor names have spaces in them (they can't...). I think I have a video about this. Search my channel for "imputing composites".
Hi Prof. James, Thanks a lot for the nice video. However, I have a QUESTION. I have followed all the steps in my model to test for the CMV using common latent factor having one second-order factor. But even after constraining the path coefficient to a letter 'a', I am not getting the same path value from the common latent factor to all the items. Could you please suggest the solution for the same?
It is probably because you have only one 2nd order factor. Usually a CFA will have multiple covaried latent factors. However, if your model has no covariances between latent factors, then AMOS sometimes won't work properly. If you have another latent factor you can include, that would be best.
Hi James -thanks for all the wonderful videos! Question - when I run the model with the CLF all latent variables in my measurement model remain intact except for 2 whose factor loading coefficients go way way down to almost 0. The other 4 latent constructs in the model remain almost unscathed. My IVs comprise a relatively new instrument in the literature with three latent sub-constructs. One is still perfect with high loadings and the other two are completely destroyed when you apply the CLF method. If it was a CMB problem would we not see evidence of this across all the constructs in the model instead of just those two?
This probably means that the variables being affected are common denominators among all the items. Here is a way to fiddle with this a bit: ruclips.net/video/YlwpiEnSSYQ/видео.html (check around the 5 minute mark)
@@Gaskination Thanks so much for the reply! I have literally learned SEM from you and also Mike Crowson to some extent. Where do you find the time?!! You are such a generous soul. I hope you are abundantly blessed!
@@chriskwaramba8607 Thanks! I just keep it to 30 minutes per day, every day. I do my best to respond to all comments and emails within 24 hours. Just part of my scholarly service.
Hi James, in CMB, if the difference of 'CLF' and 'Without CLF' (in excel sheet) is very small (such as 0.003) and several values are also in negative, is it alarming or acceptable?
James, what is the appropriate citation for this method? I have a paper in Revision where I used the method but did not offer a reference since I couldn't find it.
Hi Bridget. There are many, but the best is probably the article by MacKenzie and Podsakoff "Common Method Bias in Marketing: Causes, Mechanisms,and Procedural Remedies"
+James Gaskin +Bridget Nichols: Another reference: Sources of Method Bias in Social Science Research and Recommendations on How to Control It - Philip M. Podsakoff,1 Scott B. MacKenzie,2 and Nathan P. Podsakoff3. www.annualreviews.org/doi/pdf/10.1146/annurev-psych-120710-100452
Dear James, thanks very much for uploading the vedio. i cannot performe the inputation with my dataset (308 responds). the note is " No Compeleted Data Files were created".
Thank you for this. What are your thoughts on investigating CMB in a multi-country study? Should I look at one country at a time, or the pooled data? (sorry for the double post)
Do it however you will be testing the data later. If you are testing across countries (i.e., moderating by country) then you may want to test CMB for each country separately. Theoretically, if you achieve measurement invariance, then you could just do the CMB with all the data together.
Honestly I have no recommendation, as I do not often use dichotomous data. I played around with Bayesian analysis in AMOS the other day, which is supposed to be useful for non-continuous data, but I found it a bit clunky and difficult to interpret. Sorry to not be more help on this issue.
Hi James, Thank you very much for the valuable videos.when I added the common latent factor by constraining all the regression weights to be equal, the model didn't run and didn't show up any messages. When I removed these constraints, the model ran, but resulted regressions weights of the three of the indicators to the theoretical construct to be non-significant.Is its correct to perform this test without constraining the regression weights to be equal to the common latent factor? Shall I need to delete these non-significant items from the measurement model?when I deleted the non-significant indicators, the model reached the iteration limit.
+Sam Thara The best approach is to do a chi-square difference test between the unconstrained model and a model where the paths from the CLF are constrained to zero. If there is no significant difference, then the common variance is not significant and there is no common method bias.
Thanks for the video! Is the CLF's variance constrained to 1? I'm trying to figure out if that's the necessary step or if I should constrain one of the regression weights to 1. Also when I run the analysis constraining CLF variance to 1 I get the message: "Iteration limit reached. The results that follow are therefore incorrect." any ideas why? Many thanks
you probably have a negative error variance or high standard deviation. Watch my "iteration limit reached in amos" video as well as my SEM Series CFA video. These show how to deal with these things.
Hii James. Thanks for this wonderful video about common method bias. Pls tell me, how can I site that the difference of the estimates values of standardizing regression weight without and with CLF have been found below 0.2. Thus it confirms that this study is not influenced by CMB issue. I need papers or any other source to cite the above-mentioned threshold limit. Pls, suggest the citation. Thanks in advance.
Gefen and Straub wrote an article (Gefen, D., & Straub, D. (2005). A practical guide to factorial validity using PLS-Graph: Tutorial and annotated example. Communications of the Association for Information systems, 16(1), 5.). In this article they cite 0.100 (rather than 0.200).
Thank you so much Mr Gaskin I followed your direction and got a difference that was above 0.2. Can i just conclude that there is no common latent bias or do i have to impute the data? I conducted the data imputation but Amos seemed to work very slowly anytime i did it (then i had to close Amos). Anyway it created a new data file although i did not know if that was a finished file or not. I started the SEM procedure again by conducting EFA with the newly created data file, but it did not work. Please explain the reason why? Could you please guide me how to continue the Sem procedure after the data imputation? Is it true that we should restart the process with the new data file? Thank you!
If the regression weights differ by more than 0.200, then this is an indication that there might be some issues with method bias. Check to make sure the CLF didn't break any factors (e.g., very low, negative, or nonsignificant regression weights). If the CLF breaks the model (or any factor), then it may be better to assess method bias using a different approach. As for conducting SEM after imputing factor scores, you can use the new dataset to conduct a path analysis (with the newly imputed factor scores). The SEM speedrun video and the SEM Series playlists show how to go about this.
Quick question - when I used the previous method (constraining regression weights from CLF--items to be equal) and compared standardized regression weights, there were some differences in items I would have expected given low factor loadings w/out the CLF. However, I just tried to do it using this method (no constraints on CLF--item regressions) and now it appears that many of my items are showing differences of .5 and above - also the items that are showing these differences are in completely different dimensions... any reason why the CLF would identify items in one dimension using the first method and other dimensions using this method?
Bobby Bullock The CLF introduces so many additional parameters in the model, which can then create some instability. You can use either approach. Look at the unstandardized regression weights from the CLF if you are going to constrain them to be equal. Also, the CLF only extracts shared variance, which might not represent method variance. This can break several of the factors in the model...
Functionally, it is very similar if you are using the CLF approach. If using a specific bias factor or marker variable, this provides some difference with the bifactor model.
Hi James, I have tried using Common method bias in one of my paper. Once I'm using CLF in the measurement model, the standardized regression weights are exactly the same for constrained and unconstrained model (The model with CLF and the one without it ). Therefore, delta values cannot be calculated and no inferences regarding CMB can be made. The coefficients from CLF to the measurement variables are zero. Kindly revert on this.
Thanks again James, If I have an observed ordinal categorical variable (4 levels and endogenous in my path model) in there 1) how would I model it in AMOS (I can't find anything looking at categorical variables in your videos....but maybe I'm not looking hard enough!) and 2) can I still load the CMV latent factor onto it? Thanks in advance, Tom
Categorical variables have to be modeled as a set of dummy variables. I don't think I have any videos on this. I can't imagine sticking them in the measurement model, but I may be wrong.
James Gaskin Thanks James, is this the same when you have a binary response that is directly observed...I can get around my ordinal variable issue in AMOS by using MPlus, but I have a colleague using AMOS who is looking to use a binary endogenous variable in a structural model. Cheers, Tom
James, thank you for the video. Tom, do you happen to have the syntax for Mplus to perform such an analysis? I see the syntax for multiple imputation in the Mplus user guide, but I am uncertain if that's appropriate for this analysis. Will highly appreciate your assistance.
Hi James. Based on the CLF test, there is CMB in my data (I still can't run the marker variable test...not sure why). So, I tried doing exactly what you did here and include the CLF in my structural model because I didn't use the imputed variables. When I try to run the model, this is what I get: The model is probably unidentified. In order to achieve identifiability, it will probably be necessary to impose 1 additional constraint. However, I included a path constraint on all indicators of the two latent variables in my model and I constrained the variance of the CLF to be 1. What could this additional constraint possibly be? Thanks!
Dear Prof. James Gaskin, some reviewers ask me to use the CMB approach of Lindell and Whitney (2001): Lindell, M.K. and Whitney, D.J. (2001) Accounting for common method variance in cross-sectional research designs. Journal of Applied Psychology 86(1): 114-121. Do you know how to test it? Could you help to tutor me? Thank very much.
It is just the marker variable approach. I show how to do that here: ruclips.net/video/0GRob-VMPFM/видео.html (for Mplus) and here: ruclips.net/video/abzt5zTkCxk/видео.html (for AMOS)
HI, I have seen your answer to a user with a similar problem. I could watch the video "iteration limit reached in AMOS" but could not find the one in the SEM series, do you know how it is called please? Also, after following your guidelines in the iteration video I noticed that all the observed variables in two /different factors have S.E. bigger than 1. Do you know why does it happen? Surprisingly the issue is with the two factors with the best CR and AVE scores...Many many thanks
Dear James, There is a paper from 2010 [Williams, L. J., Hartman, N., & Cavazotte, F. (2010)] about the MARKER test. But your above CLF method is more recent seems to (humble) me more exact. do you think it makes sense to run CLF and MARKER test as well? and if so, how do you suggest I do the marker test with the new CMB -CLF method? Thank you, Gale
Gale Bren The marker variable approach is the more accurate test, if you have a theoretically selected marker variable. In this case, you would do as shown in my older video where you constrain the regression weights from the CLF to be equal, and then square the unstandardized loadings in order to get the percent of common variance.
Hi James Thanks for the very helpful videos. I have just run the common method bias test on some data and one latent variable for all its observed variables is displaying large differences between the model with the CLF and the model without. These rang from .68 to 1.1 hence the standardized regression weights on the latent factor are mostly negative or small when the CLF is included. Where as without the CLF they are all above .5 This only occurs for one latent variable with seven items the rest of the latent factors are unaffected by the CLF. Is this a problem because it does not look good? Thanks Alastair
Alastair Nightingale The CLF often creates instability. When this happens, I try moving the parameter constraint around on that factor. If that doesn't fix it, then I remove all error covariances on that factor (if any). If that doesn't fix it, then I resort to more primitive methods, such as the Harman's single factor test, and then just explain why I had to do that - i.e., because the CLF introduced too many new parameters, thus causing the model to be unstable.
James Gaskin Thanks James moving the constraint has solve the issue. What do you make of Conway and Lance's (2010) argument that post hoc control for common method bias using a common latent factor is "logically indefensible because it can easily remove trait variance when multiple traits have a common cause"?
Alastair Nightingale I completely agree. This is why I think CMB is stupid. You can totally mess up your model by including a CLF. The only validated way to control for CMB is through a theoretically selected marker variable.
"Paul Benjamin Lowry, James Gaskin, Nathan Twyman, Bryan Hammer, and Tom L. Roberts (2013). “Taking “Fun and Games” Seriously: Proposing the Hedonic-Motivation System Adoption Model (HMSAM),” Journal of the Association for Information Systems (Volume 14, Issue 11, pp. 617-671). In this article we use CLF adjusted values (see page 630)." I unfortunately have no access to this paper. So can you confirm that the threshold of 0.2 is named in the article? In that case I would cite it blindly. Or do you have any further reference for this? I just found one value out of 56 above .2 (=.21). Would you perform all subsequent analysis under inclusion of the CLF though?
I would just drop the CLF if that is the case. CMB is clearly not a big factor if only one item is having trouble. In the article you mention, we retained the CLF because there were some method bias issues. However, we do not explicitly state the use of a 0.200 cut off. No one does. I need to write a paper about this...
Hi mate, great video's, when I try this I can't do it with the number of iterations I ask it to run (no matter what I set the limit to), is there something I could be doing wrong? cheers, Tom
Hey James, I followed your steps, and found CMB problems with some paths (>0.2), do I need to do the Validity and reliability test since I had to do some adjustment to the model ( I had to delete some covariances that shows negative error " in the note for model"?? Thank you
Dear James.. I have done the peocess of checking std regression(with clf and without clf), and I got two variable, which is large than 0.2. unfirtunately, While I am doing the data imputation, I got message like "a sample of parameter value was inadmissible"... I had no space for variable name and no missing data... How should I do with this issue?
Here's a fun pet project I've been working on: udreamed.com/. It is a dream analytics app. Here is the RUclips channel where we post a new video almost three times per week: ruclips.net/channel/UCiujxblFduQz8V4xHjMzyzQ
Also available on iOS: apps.apple.com/us/app/udreamed/id1054428074
And Android: play.google.com/store/apps/details?id=com.unconsciouscognitioninc.unconsciouscognition&hl=en
Check it out! Thanks!
A good common method bias paper is "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies" by MacKenzie and Podsakoff. The 0.20 threshold is just something I use. No particular reference for it except this video... Although I suppose you could quote Cohen who talks about effect sizes, with 0.20 as a moderate effect.
Hi James, Thank you for the video. Your videos have helped me a lot. Moreover, I am citing your article in my PhD dissertation. Regards from Brazil.
Thanks for such a fast reply ! I am not using a 2nd order factor. There are 4 factors: Attitude, Social Norm Support, Self Eficacy, and Entrepreneurial Intention. The instruments inted use is to use three of the constructs as dependent to predict entrepreneurial intention (4th) . Thank you so much !!!!!!!!!
Most data I have made available on my wiki, but the data in this video is not mine to give away. so sorry. It belongs to a colleague who has already declined my request to let others have it.
Cheers mate. I've used this method in conjunction with Harman's test in an effort to triangulate, and it appears that I have no real issue with CMV in any case, so all's good. Thanks for your quick response.
You do such a nice job with these videos - thank you for taking the time to share your expertise!
It's a very helpful video for dealing with CMV. Thanks a lot.
Exactly. If you are imputing for composites then you will be accounting for the common variance already, so you should not include the imputed CLF composite in the structural model. But, if you are NOT imputing (instead using the full latent model), you may want to retain the CLF so that the common variance can be accounted for.
Oh, ok. This makes more sense now. I don't know of any precedent for doing it three at a time, but on principle, that is perfectly fine. I would say that is sufficient evidence for no method bias. It would be a cool contribution in your paper as well to show that method bias is deceiving and fickle.
Dear James,
Thank you so much for uploading this, very helpful.
If someone wants to cite this, we have used this approach in our paper published in Journal of Business Ethics, You can cite the following:
Serrano Archimi, C., Reynaud, E., Yasin, H.M. and Bhatti, Z.A. (2018), “How perceived corporate social responsibility affects employee cynicism: the mediating role of organizational trust”, Journal of Business Ethics, Vol. 151 No. 4, pp. 907-921.
After running it, you should be able to still see some output, although it will be unfinished. Go to the regression weights table to see which regressions have exceptionally high standard errors (like anything above 1 or 2). These are the culprits. If there are no problems there, check the standardized regression weights to see if any factor has a set of items that are negative. If so, then move the parameter constraint onto the positive one. Hope this helps!
Same thing sort of. The bias is referring to the confounding effect. The variance is referring to the amount of variance that is common to the entire set of items.
This is probably adequate model fit. We use multiple measures as a means of triangulation. It is uncommon to meet the threshold for every single one. You may also check Hair et al for more relative thresholds based on sample size and model complexity. I think it is table 12-4 in the 7th version of the book. hope this helps!
Dr. Gaskin,
Thank you so much for this series. I owe my entire dissertation methods section to you. Quite honestly, you're videos have been a God-send.
Would you be able to present a video on scalar (strong factoral) and strict (invariant uniqueness) measurement testing? I ask because relatively new literature (Putnick & Bornstein, 2016) suggest using a "ladder" of measurement testing going from configural -> metric -> scalar -> strict to ensure more robust invariance testing.
Thank You!
Here is a more recent video that covers this ladder of validation: ruclips.net/video/sbTl-icEyD8/видео.html (I've queued the link to start about 3/5ths of the way through).
If you are looking for a CMB video in the SEM series, it would be in the one about CFA or measurement models. It is fine to have an SE greater than 1. abnormally high SE can be detected by comparing the SE to the unstandardized regression weight. If the SE is greater than the regression weight, then it is probably a problem.
Barbossa: "There are a lot of long words in there, [Vahid]; we're naught but humble pirates. What is it that you want?"
If I understand you correctly, you want to know what the regression weights on the CLF means for each indicator? It means that the portion of variance in that indicator explained by the method is the standardized regression weight. The other portions of variance are accounted by error and trait (with the trait remaining on the original latent factors' indicators).
Dear James Gaskin,
I have appreciated you for uploading this video. I want to use your all videos for spreading out research areas in my country if permission me.
You are welcome to use these videos in your lessons and teaching material. That is fine.
Hi James. I am a big fan of you in data analysis. Great explanations. Can you provide some literature references for this method, please?
Here you go: statwiki.gaskination.com/index.php?title=References#Method_Bias,_Response_Bias,_Specific_Bias
Well done and thanks! I have a question. I noticed that regression weights of indicators regressed on the Common latent variable are negative (or zero). What do the negative values indicate? Additionally, I suppose some of the items do not load significantly on the Common latent variable. What does this lack of significant loading tell us? I suppose it would indicate that common method factor affects merely a portion of items not all. Would it be okay to delete the regression paths with p> 0.05
Negative regressions just mean that the indicator moves in the opposite direction to the majority of other indicators. Lack of significant loadings means that they are not affected by CMB. Do not remove non-significant paths because you want to see what variance is COMMON to ALL indicators in the model.
Thanks for the information. I will look at the forum. Best regards, Ramón
I cannot think of a way to do it off hand, but there is probably a way. I recommend looking at the SmartPLS forum. They are very helpful. I have never been disappointed with what I find there.
Yikes! that's trouble. However, the Harman's test is the least revered of CMB tests. I recommend using the CLF test if you don't have a marker variable (like social desirability).
Yes. Usually scholars use these terms interchangeably.
I'm not sure I understand. If you are using a 2nd order factor for EI, then you still need to do the cmb test on the lower order dimensions. I hope this helps. If not, feel free to clarify.
Observed variables will not suffer from CMB since they are observed. Only self-reported perceptual items are vulnerable.
Dear Prof. Gaskin,
I learned a lot. thank you.
Let me ask you one question.
This 0.2 is used as the standard, but is this standard quoted from some previous research?
If there is any previous research on this subject, please let me know.
In setting this reference value, if there is a clear precedent research, I would like to cite that literature and write a paper.
No. The published threshold is actually 50% from the Harman's single factor test. I just prefer a more conservative value.
That's odd about the imputing, although imputing is somewhat sensitive. It doesn't like you to have variable names that have spaces. If that is not the problem, then you may want to just move forward, retain the CLF, and use the full SEM with latent constructs.
I have a question, Is there a way to conduct a common method bias analysis using common latent factor with PLS? I don´t have a marker variable, and I would like to perform a common method bias test. By the way, I am a big fan of your videos. Yo have taught me a lot !!!
I appreciate your advice. I look forward to further videos from you.
I wonder if you could teach us how to do CFA/SEM of dichotomous data on AMOS. I think AMOS does not have an option for using tetrachoric or polychoric correlation matrices, which are commonly used in estimating CFA models of dichotomous data. Do you recommend any certain approach?
Dear James,
I have tried to find out the reference for the way you check CMB using Common Latent Factor (CLF), but failed, could you please suggest me some published articles that have use either following ways to check for CMB?
(1) Compare the standardized regression weights from the model with CLC to the standardized regression weights of a model without the CLF. If there is no large differences (like greater than 0.200), then the CMB is not serious.
(2) compare the model fit of the two CFAs, if CFA without CLF is better than CFA with CLF, thus CMB is not serious.
Thank you very much in advance
Angelina
I don't know about the model fit approach, and I've never seen it done (and it doesn't make a lot of sense to me...), but as for the comparing of standardized regression weights, no article talks about this or about a threshold for their difference. Podsakoff talks about method bias as does MacKenzie. The most recent article in marketing by the both of them would be a good resource. It's called: "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies"
James Gaskin Dear James, thank you very much for your prompt reply. May I ask whether this method to check CMB is equivalent to the way Liang, Saraf, Hu, and Xue (2007) "Assimilation of Enterprise Systems: The Effect of Institutional Pressures and the Mediating Role of Top Management" has done using PLS model with a CMF?
Angelina Hanh It is similar, but not the same. Liang et al., explain that the average loadings are much stronger for the factors than for the CLF and that they are mostly non-significant on the CLF. This is fine too.
James Gaskin Thank you very much, James.
I have looked around for a suggested cutoff, and no one has put a stake in the ground as far as I can tell. So, the 0.2 is simply an educated degree of tolerance (i.e., I'm willing to accept small differences between the effects, but not big ones). 0.2 is sometimes also used to distinguish a small effect from a large effect. I think Wynne Chin has published this threshold in MISQ. Others have also used this cutoff for effect size, but I can't remember who off the top of my head.
thanks so much , pleas provide us a reference for this cutoff for effect size, 0.2
@@nonadad8431 For effect sizes, please refer to the discussion here: chat.openai.com/share/037cb9de-23db-4654-8bd1-dd6c6b027148
Thank you so much for your video.
If I use Harman’s single factor test and the single factor accounts for more than 50% of the variance in the model, then what can I do? How can I overcome the problem?
Hello James, thanks for the good explanation, but I still have some questions:
First, if you compare the standardized regression weights with and without CLF, would you keep the Marker variable in both models or also delete it in the without check?
Second, would you do the CMB-test before you start testing the reliability and validity? and if yes and assuming CMB is an issue, would you keep the CLF all the time in the model to check for convergent and discriminant validity? It is not to easy to obtain validity with the lowered loadings and I couldn´t find something about the right order of doing all the pretests.
Nixen85 If you're using a marker variable, then the better approach would be to constrain the regression weights from the CLF to be equal, and then square their unstandardized loadings. This will give you the percent of shared variance. I would test validity prior to including the CLF. The CLF can extract shared variance which is actually trait variance, and so it can artificially deflate the construct reliabilities.
Hello James
Thank you so much for your videos. I have learnt alot from them. Just a quick question regarding this video on CMV. You mention that do not include control variable in your structural model after imputing the data. But then when you convert CFA to SEM in this video, you retained the control. Aren't we suppose to delete the control variable, as its affects are going to be present anyway since the data has already been imputed?
Thanks in advance.
Hi James, many thanks for your great videos. These have been very helpful! I have few questions in the case using CFA to confirm EFA results and using CLF to address the CMB after CFA:
1. How to calculate AVE and MSV to report the discriminant and convergent validity adjusted by CLF with composites as AMOS does not produce Correlation table anymore? (Am I mistaken anywhere?)
2. If the model fitness of the models with and without CLF is not much different, is it necessary to use the more complicated model (with CLF)?
3. Can you please recommend me any reference about the extent of the CMB that affects the model? How much difference between the standardised regression weight between 2 models can be considered small, medium or large? What to say about a difference >.2 with most items?
Many thanks James!
+Van Nguyen
1. The correlation table should show up if you check the box for standardized estimates.
2. You only keep the CLF if there are some method bias issues. Otherwise, just get rid of it.
3. The best way to do it is actually to run it unconstrained with the CLF, and then run it with all paths from the CLF constrained to zero. Then you can do a chi-square difference test to see if the CMB is different from zero. I don't have a video for this yet...
+James Gaskin
Thanks James! I really appreciate you addressing the comments and questions!
1. I adjusted all factors in the models (deleting all co-variance paths, replacing by single headed arrows). It's probably because there are no covariance anymore and that AMOS cant produce correlation table? What should I do in this case?
3. When I unconstrained the CLF (deleting all the a constraints as in the previous video), AMOS doesnt run at all. It popped up with the message about the one of the estimated covariance failed to be positive - it did not show up the output window for me to trouble shot the problem). Apart from AMOS instruction manual, would you recommend me any book to learn & play around with AMOS - as AMOS is obviously very sensitive?
+Van Nguyen
1. correct. if no covariances, then no correlations. You don't need correlations to run the model if it is just a causal path model.
2. The Barbara Byrne book "Structural Equation Modeling with AMOS".
Thanks for your reply. Then, if the latent variable CMB goes up by one unit, the magnitude of observed indicators goes down by Standardized Coefficients' magnitude. This seems to be statistical indication. Could you explain the substantive (theoretical) indication of this finding, please?
But does the model run with the CLF. Forget imputing for now. Does the model run regularly with the CLF?
Hi, James. I learn a lot from your video. Thank you very much. It seems very clear and explicable. But I have a silly question. 'Cause I'm a new learner. I only know how to run SPSS. What's the technique tool you use (I can see in the video) to draw this beautiful model and do CFA, SEM, And CLF method? Thank you very much!
Dear Prof. Gaskin, Thank you so much for this great tutorial. In the video, you mentioned that the threshold for differences of standardized weights is 0.2. Could you please cite any reference mentioning it should be like this. I need it to refer in my thesis. Many thanks.
This difference is a personal preference (it's just exploratory after all). Nevertheless, there are citations that suggest it can be even closer (0.100). I think the PLS article by Gefen and Straub (2008?) talks about this. I think the Hair book (2010) also talks about this.
I am not familiar with the instrumental variable technique (unless that is referring to the marker variable technique - if so, I have a video for this as well).
It is called AMOS. It is made by the same people as SPSS (IBM).
Thank you so much for the excellent videos. I have found them very helpful to suppliment my reading and advice from my supervisor. Could you perhaps point me in the direction of your source for the .2 cutoff for difference in standardised regression between the two models? I've followed the citation trail from Podsakoff et al 2003, but I'd just like to make sure.
This is really a very helpful video, but I have a question, what is the threshold of the common latent factor.
The only published threshold that I'm aware of is 50%, but this is pretty high. Most modern approaches at CMB run a chi-square difference test to see if the unconstrained model is different from a model where all paths from the CLF are constrained to zero (implying zero common shared variance).
@@Gaskination noted and thank you, highly appreciated.
Also, is "estimate means and intercepts" checked? This might also throw it off.
It should work in the path model or hybrid model, but won't work in the CFA/measurement model. If you are in the hybrid model, then you might have simply forgotten to include error terms for new endogenous variables.
Hi James,
Thanks for this video. I did not understand one point. First you say, you don't take the common latent factor into the SEM model, because you partial out the variance before by building composites. However, at the end of the video, you still leave the common factor in there, when running the SEM model. Could you please explain?
Would you get the same results, when constraining the factor loadings in the SEM model to the decreased values you got when including the common factor in the correlation analysis, but then taking out the common factor in the SEM model?
Thanks so much, Fu
Fu Yin-Heinberg If you decide to retain the CLF when imputing composites, then the composites already account for the method bias which has been extracted by the CLF. Thus you don't need to include the CLF variable when drawing the path model using composites. However, if you decide to not create composite, but instead use a latent path model, then you should retain the CLF in the model (if CMB was an issue). If CMB is not an issue, then you can remove the CLF before creating composites.
As for constraining vs. not constraining the paths from the CLF, you can try it and see, but the unstandardized results will differ.
Hi James, thanks for taking the time to put this video together. I tried your method but the model did not run for some reason. Would you have any idea why it did not run?
+anaikeda So many possible reasons... Here is a video to help you troubleshoot it: ruclips.net/video/B7YOv7hSohY/видео.html
Hi James,
Why exactly did you choose a delta variance value of .2 as the cutoff point to decide whether there is a common method bias?
hishonline when looking at standardized estimates, 0.200 is a meaningful difference (in my opinion). Sort of like in an EFA when you are looking for cross-loadings within 0.200. I'm not sure if there is anything published about a threshold though.
Dear James I really like your videos. I have studied various articles about common method bias and don't found the 'common latent factor' technique. They have used other names and I am really very confused about it. I have read your suggested artile Sources of Method Bias in Social Science Research and Recommendations on How to Control It byPhilip M. Podsakoff,1 Scott B. MacKenzie,2 and Nathan P. Podsakoff3. But they have used different terms. Please guide which method reflects CLF?
I think they call it the unmeasured latent factor.
Hello Professor, Thank you so much for this excellent video. I have two questions: 1) Can we perform common latent factor technique on path analysis model instead of measurement model?
2) Is common latent factor test also called as Harman's single factor test or both are different?
1. No, only conduct CMB tests on measurement models.
2. Yes, it is the Harman's test, but for the CFA instead of EFA.
@@Gaskination Thank you so much Sir.
Hi James..Thanks for these wonderful videos..what reference/citation can we give for using this correct way? podsakoff orany other? also, whom do we quote when we say that the threshhold value of difference of standarized estimates (with and without common factor) is 0.2..Thanks
Thx, James. I watched the "inputing composites" video and my model works. but that one is without CLF. It still doese not work with CLF. My factor names do not have spaces and there is no missing data.
Dear James,
Your videos are very informational, thanks a lot ! I am testing Linan & Chen (2009) entrepreneurial intention scale for CMB. This instrument has 4 dimensions with 20 itens.
As expected the EI is highly correlated with attitude. So when I use a single factor, I get a factor loading on the com. fact, but when I do test 3 dimensions at time (3 and CFACT), do not. Do you think it is enough to provide evidence of no CMB ? Have seen any advice to test when variables are highly corr?
Many thanks for directing me to this video follwoing my earlier query. I have followed this and when I remove the CLF, my df increases by 1 as expected but my st. regression weights remain exactly the same. My delta scores are zero throughout. Is there any possible explanation for this?
Fantastic video series...thanks! Could you please clarify the process for handling CLF in a 2nd order model? I've added the CLF to my 2nd order model, it runs, and the fit is good. But when I impute I get an error that says "a sample of parameter values was inadmissable." Just for fun I've tried this on a 1st order version of my model and impute works...do I need to impute in my 2nd order model or can I just go straight to the structural model and keep the CLF in the picture? thanks....
This is quite helpful. I've been racking my brain trying to figure out how to do this in SAS PROC CALIS. I haven't found much in the documentation or at SUGI; has anyone had any luck with something like this in SAS?
Hi James. I'm wondering how this newer latent factor method relates to the marker variable analysis. Should I add in the marker variable and then compare the regression weights (with and without the common factor)? or Should I stick to the marker variable method presented in the other video?
Here is the most current approach and video: ruclips.net/video/abzt5zTkCxk/видео.html
Hi James, Thank you so much for sharing your wonderful videos. How to adjust common method method bias in 2nd order constructs as data imputation is not done in this.
In SEM, can we show the Common latent factor with the 2nd order construct the way you have shown and used in your video?
the common method bias occurs at the item level, so whether it is second or first order makes no difference. Check it at the first order level, and if it is a problem, then simply retain the CLF when creating the 2nd order model.
Many thanks for your extremely helpful videos in RUclips. I have a quick question which I hope you could answer if at all possible.
I followed your recommendations for analysis with a single common method factor, and found that common method bias is a problem in my data - however I’m neither able to impute the variables accounting for CMB nor to run a SEM keeping the CLF. The message I get is that 1 parameter sample is inadmissible. CFA analyses run normally, and error variances are not negative, but most covariances are. I read this may be due to my small sample size (N 170 for 34 observed variables) and that a solution could be using Bayesian estimation, but I am unsure about how to proceed.
Is there a way of manually imputing the data based on the estimates I get from the CFA with CLF? What would be your advice?
+elisaalt It might be sample size, but probably not. 170 is probably fine. Check to see if there are weird loadings, or any parameters with really high standard errors. These will show you were the problems are.
Hi James, first, thank you for excellent Stat-wiki! Second, I really like the idea of using CLF in order to create "common method bias adjusted composites" but could you give me a reference on this? Thank you so much for the great work! Nico
Here is a paper I just had published at JAIS:
Paul Benjamin Lowry, James Gaskin, Nathan Twyman, Bryan Hammer, and Tom L. Roberts (2013). “Taking “Fun and Games” Seriously: Proposing the Hedonic-Motivation System Adoption Model (HMSAM),” Journal of the Association for Information Systems (Volume 14, Issue 11, pp. 617-671).
In this article we use CLF adjusted values (see page 630).
James Gaskin
Thanks a lot, I will definitely look at the article once its published (I can't access it just yet... stuck at issue #10 :) ) and I'll be more than happy to use your article as a reference for the usage of the method if I get a chance. Thanks again!
Hi James, thank you for the great video's.When I include the CLF like you did, the model does not run. By this I mean that it looks like it is running, but there is no option to click on the output. I get no error whatsoever.
‘Thank you very much’
Does the model run when not trying to impute? Also, do any of the factor names have spaces in them (they can't...). I think I have a video about this. Search my channel for "imputing composites".
Dear James,
Can I have the data that you have used in your videos?
I need them to practice.
Thanks
Hi Prof. James,
Thanks a lot for the nice video.
However, I have a QUESTION.
I have followed all the steps in my model to test for the CMV using common latent factor having one second-order factor. But even after constraining the path coefficient to a letter 'a', I am not getting the same path value from the common latent factor to all the items.
Could you please suggest the solution for the same?
It is probably because you have only one 2nd order factor. Usually a CFA will have multiple covaried latent factors. However, if your model has no covariances between latent factors, then AMOS sometimes won't work properly. If you have another latent factor you can include, that would be best.
Hi James -thanks for all the wonderful videos! Question - when I run the model with the CLF all latent variables in my measurement model remain intact except for 2 whose factor loading coefficients go way way down to almost 0. The other 4 latent constructs in the model remain almost unscathed. My IVs comprise a relatively new instrument in the literature with three latent sub-constructs. One is still perfect with high loadings and the other two are completely destroyed when you apply the CLF method. If it was a CMB problem would we not see evidence of this across all the constructs in the model instead of just those two?
This probably means that the variables being affected are common denominators among all the items. Here is a way to fiddle with this a bit: ruclips.net/video/YlwpiEnSSYQ/видео.html (check around the 5 minute mark)
@@Gaskination Thanks so much for the reply! I have literally learned SEM from you and also Mike Crowson to some extent. Where do you find the time?!! You are such a generous soul. I hope you are abundantly blessed!
@@chriskwaramba8607 Thanks! I just keep it to 30 minutes per day, every day. I do my best to respond to all comments and emails within 24 hours. Just part of my scholarly service.
Hi James, in CMB, if the difference of 'CLF' and 'Without CLF' (in excel sheet) is very small (such as 0.003) and several values are also in negative, is it alarming or acceptable?
Thanks - do you know if controlling for CMV is similar to controlling for CMB as you've just shown in this video?
James, what is the appropriate citation for this method? I have a paper in Revision where I used the method but did not offer a reference since I couldn't find it.
Hi Bridget. There are many, but the best is probably the article by MacKenzie and Podsakoff "Common Method Bias in Marketing: Causes, Mechanisms,and Procedural Remedies"
+Bridget Nichols www.sciencedirect.com/science/article/pii/S0022435912000681
+James Gaskin +Bridget Nichols: Another reference: Sources of Method Bias
in Social Science Research and Recommendations on How to Control It - Philip M. Podsakoff,1 Scott B. MacKenzie,2 and Nathan P. Podsakoff3.
www.annualreviews.org/doi/pdf/10.1146/annurev-psych-120710-100452
Dear James, thanks very much for uploading the vedio. i cannot performe the inputation with my dataset (308 responds). the note is " No Compeleted Data Files were created".
Thank you for this. What are your thoughts on investigating CMB in a multi-country study? Should I look at one country at a time, or the pooled data? (sorry for the double post)
Do it however you will be testing the data later. If you are testing across countries (i.e., moderating by country) then you may want to test CMB for each country separately. Theoretically, if you achieve measurement invariance, then you could just do the CMB with all the data together.
Honestly I have no recommendation, as I do not often use dichotomous data. I played around with Bayesian analysis in AMOS the other day, which is supposed to be useful for non-continuous data, but I found it a bit clunky and difficult to interpret. Sorry to not be more help on this issue.
Hi James, Thank you very much for the valuable videos.when I added the common latent factor by constraining all the regression weights to be equal, the model didn't run and didn't show up any messages. When I removed these constraints, the model ran, but resulted regressions weights of the three of the indicators to the theoretical construct to be non-significant.Is its correct to perform this test without constraining the regression weights to be equal to the common latent factor? Shall I need to delete these non-significant items from the measurement model?when I deleted the non-significant indicators, the model reached the iteration limit.
+Sam Thara The best approach is to do a chi-square difference test between the unconstrained model and a model where the paths from the CLF are constrained to zero. If there is no significant difference, then the common variance is not significant and there is no common method bias.
Thanks for the video! Is the CLF's variance constrained to 1? I'm trying to figure out if that's the necessary step or if I should constrain one of the regression weights to 1. Also when I run the analysis constraining CLF variance to 1 I get the message: "Iteration limit reached. The results that follow are therefore incorrect." any ideas why? Many thanks
you probably have a negative error variance or high standard deviation. Watch my "iteration limit reached in amos" video as well as my SEM Series CFA video. These show how to deal with these things.
Thanks!
Hii James. Thanks for this wonderful video about common method bias. Pls tell me, how can I site that the difference of the estimates values of standardizing regression weight without and with CLF have been found below 0.2. Thus it confirms that this study is not influenced by CMB issue. I need papers or any other source to cite the above-mentioned threshold limit. Pls, suggest the citation. Thanks in advance.
Gefen and Straub wrote an article (Gefen, D., & Straub, D. (2005). A practical guide to factorial validity using PLS-Graph: Tutorial and annotated example. Communications of the Association for Information systems, 16(1), 5.). In this article they cite 0.100 (rather than 0.200).
Thank you so much James.
Thank you so much Mr Gaskin
I followed your direction and got a difference that was above 0.2. Can i just conclude that there is no common latent bias or do i have to impute the data?
I conducted the data imputation but Amos seemed to work very slowly anytime i did it (then i had to close Amos). Anyway it created a new data file although i did not know if that was a finished file or not. I started the SEM procedure again by conducting EFA with the newly created data file, but it did not work. Please explain the reason why?
Could you please guide me how to continue the Sem procedure after the data imputation? Is it true that we should restart the process with the new data file?
Thank you!
If the regression weights differ by more than 0.200, then this is an indication that there might be some issues with method bias. Check to make sure the CLF didn't break any factors (e.g., very low, negative, or nonsignificant regression weights). If the CLF breaks the model (or any factor), then it may be better to assess method bias using a different approach. As for conducting SEM after imputing factor scores, you can use the new dataset to conduct a path analysis (with the newly imputed factor scores). The SEM speedrun video and the SEM Series playlists show how to go about this.
Quick question - when I used the previous method (constraining regression weights from CLF--items to be equal) and compared standardized regression weights, there were some differences in items I would have expected given low factor loadings w/out the CLF. However, I just tried to do it using this method (no constraints on CLF--item regressions) and now it appears that many of my items are showing differences of .5 and above - also the items that are showing these differences are in completely different dimensions... any reason why the CLF would identify items in one dimension using the first method and other dimensions using this method?
Bobby Bullock The CLF introduces so many additional parameters in the model, which can then create some instability. You can use either approach. Look at the unstandardized regression weights from the CLF if you are going to constrain them to be equal. Also, the CLF only extracts shared variance, which might not represent method variance. This can break several of the factors in the model...
Dr Gaskin, in your opinion, what is the difference between CMB procedure and bifactor analysis?
Functionally, it is very similar if you are using the CLF approach. If using a specific bias factor or marker variable, this provides some difference with the bifactor model.
Hi James, I have tried using Common method bias in one of my paper. Once I'm using CLF in the measurement model, the standardized regression weights are exactly the same for constrained and unconstrained model (The model with CLF and the one without it ). Therefore, delta values cannot be calculated and no inferences regarding CMB can be made. The coefficients from CLF to the measurement variables are zero. Kindly revert on this.
Try this new approach: ruclips.net/video/CFBUECZgUuo/видео.html
Thanks again James, If I have an observed ordinal categorical variable (4 levels and endogenous in my path model) in there 1) how would I model it in AMOS (I can't find anything looking at categorical variables in your videos....but maybe I'm not looking hard enough!) and 2) can I still load the CMV latent factor onto it?
Thanks in advance, Tom
Categorical variables have to be modeled as a set of dummy variables. I don't think I have any videos on this. I can't imagine sticking them in the measurement model, but I may be wrong.
James Gaskin Thanks James, is this the same when you have a binary response that is directly observed...I can get around my ordinal variable issue in AMOS by using MPlus, but I have a colleague using AMOS who is looking to use a binary endogenous variable in a structural model. Cheers, Tom
James, thank you for the video.
Tom, do you happen to have the syntax for Mplus to perform such an analysis? I see the syntax for multiple imputation in the Mplus user guide, but I am uncertain if that's appropriate for this analysis. Will highly appreciate your assistance.
Dear james
Please explain how will we interprete the results in our paper? Which value we will plot and in what way?
Here are some excellent references for that: statwiki.gaskination.com/index.php?title=References#Method_Bias.2C_Response_Bias.2C_Specific_Bias
Hi James.
Based on the CLF test, there is CMB in my data (I still can't run the marker variable test...not sure why). So, I tried doing exactly what you did here and include the CLF in my structural model because I didn't use the imputed variables. When I try to run the model, this is what I get:
The model is probably unidentified. In order to achieve identifiability, it will probably be necessary to impose 1 additional constraint.
However, I included a path constraint on all indicators of the two latent variables in my model and I constrained the variance of the CLF to be 1. What could this additional constraint possibly be?
Thanks!
You are welcome to send it to me to troubleshoot it. Most likely there is just a path constraint missing. My email is "james.gaskin@byu.edu"
I don't use SAS very often, so I'm not sure.
Dear Prof. James Gaskin, some reviewers ask me to use the CMB approach of Lindell and Whitney (2001): Lindell, M.K. and Whitney, D.J. (2001) Accounting for common method variance in cross-sectional research designs. Journal of Applied Psychology 86(1): 114-121. Do you know how to test it? Could you help to tutor me? Thank very much.
It is just the marker variable approach. I show how to do that here: ruclips.net/video/0GRob-VMPFM/видео.html (for Mplus) and here: ruclips.net/video/abzt5zTkCxk/видео.html (for AMOS)
HI, I have seen your answer to a user with a similar problem. I could watch the video "iteration limit reached in AMOS" but could not find the one in the SEM series, do you know how it is called please?
Also, after following your guidelines in the iteration video I noticed that all the observed variables in two /different factors have S.E. bigger than 1. Do you know why does it happen? Surprisingly the issue is with the two factors with the best CR and AVE scores...Many many thanks
Dear James,
There is a paper from 2010 [Williams, L. J., Hartman, N., & Cavazotte, F. (2010)] about the MARKER test. But your above CLF method is more recent seems to (humble) me more exact. do you think it makes sense to run CLF and MARKER test as well? and if so, how do you suggest I do the marker test with the new CMB -CLF method?
Thank you, Gale
Gale Bren The marker variable approach is the more accurate test, if you have a theoretically selected marker variable. In this case, you would do as shown in my older video where you constrain the regression weights from the CLF to be equal, and then square the unstandardized loadings in order to get the percent of common variance.
James Gaskin
wow! that is highly unlikely, but not impossible. If it is not an error, then this means CMB is not an issue.
Hi Gaskin, If there is only 1 value higher than 0.2, how can we interpret it, can we consider it as no CMB problem.?
These are just targets and guidelines. If it is a little over, it is fine.
Hi James Thanks for the very helpful videos. I have just run the common method bias test on some data and one latent variable for all its observed variables is displaying large differences between the model with the CLF and the model without. These rang from .68 to 1.1 hence the standardized regression weights on the latent factor are mostly negative or small when the CLF is included. Where as without the CLF they are all above .5 This only occurs for one latent variable with seven items the rest of the latent factors are unaffected by the CLF.
Is this a problem because it does not look good?
Thanks Alastair
Alastair Nightingale The CLF often creates instability. When this happens, I try moving the parameter constraint around on that factor. If that doesn't fix it, then I remove all error covariances on that factor (if any). If that doesn't fix it, then I resort to more primitive methods, such as the Harman's single factor test, and then just explain why I had to do that - i.e., because the CLF introduced too many new parameters, thus causing the model to be unstable.
James Gaskin Thanks James moving the constraint has solve the issue. What do you make of Conway and Lance's (2010) argument that post hoc control for common method bias using a common latent factor is "logically indefensible because it can easily remove trait variance when multiple traits have a common cause"?
Alastair Nightingale I completely agree. This is why I think CMB is stupid. You can totally mess up your model by including a CLF. The only validated way to control for CMB is through a theoretically selected marker variable.
"Paul Benjamin Lowry, James Gaskin, Nathan Twyman, Bryan Hammer, and Tom L. Roberts (2013). “Taking “Fun and Games” Seriously: Proposing the Hedonic-Motivation System Adoption Model (HMSAM),” Journal of the Association for Information Systems (Volume 14, Issue 11, pp. 617-671).
In this article we use CLF adjusted values (see page 630)."
I unfortunately have no access to this paper. So can you confirm that the threshold of 0.2 is named in the article? In that case I would cite it blindly. Or do you have any further reference for this?
I just found one value out of 56 above .2 (=.21). Would you perform all subsequent analysis under inclusion of the CLF though?
I would just drop the CLF if that is the case. CMB is clearly not a big factor if only one item is having trouble. In the article you mention, we retained the CLF because there were some method bias issues. However, we do not explicitly state the use of a 0.200 cut off. No one does. I need to write a paper about this...
Thanks for your prompt answer. And yes, you should definitely write this article!
Hi mate, great video's, when I try this I can't do it with the number of iterations I ask it to run (no matter what I set the limit to), is there something I could be doing wrong? cheers, Tom
Hey James, I followed your steps, and found CMB problems with some paths (>0.2), do I need to do the Validity and reliability test since I had to do some adjustment to the model ( I had to delete some covariances that shows negative error " in the note for model"??
Thank you
+Riad Cheikh If you have a changed "final" model for the CFA, then you should redo the validity and reliability for the final model.
Sorry , the three cosntructs are Independent to predict EI
Dear James..
I have done the peocess of checking std regression(with clf and without clf), and I got two variable, which is large than 0.2. unfirtunately, While I am doing the data imputation, I got message like "a sample of parameter value was inadmissible"... I had no space for variable name and no missing data... How should I do with this issue?
Hi James, is common method bias similar to common method variance?