This video takes you through three approaches for detecting common method bias: Harman's single factor test, common latent factor, and marker variable. I use SPSS and AMOS.
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!
You are really a superb trainer. thank you. Can i ask for evidence on single factor common method bias suggestions? for example, less than 50% variance. Thank you
@@Gaskination May i ask you something related to EFA? three constructs combined my DV. in that case, when i run EFA, loadings fall under same variable. is it okay to continue CFA ? or i need to get more data to get separate factor loadings for each construct?
@@MatinsVlog When it is a 2nd order factor, this is not completely unexpected. You can try to run an EFA with just the items from this factor, and then constrain it to extract three factors. Or you can move directly to the CFA. The trouble you will have if you skip the EFA is knowing where discriminant validity issues are originating.
Hi James, I just wanted to let you know that your videos were a God send for me over the past few days when trying to perform SEM for my thesis. Thank you so much, you are a fantastic presenter, love your work, keep it up!
WHAT YOU ARE DOING HERE, IT IS BETTER THANT THE KNOWLEDGE WE RECEIVE IN REPUTABLE UNIVERSITIES AT THE GRADUATE LEVEL....THANKS AND CONTINUE YOU ARE EDUCATING MANY PEOPLE
Yes. That is the best way to do it. If you didn't capture the marker during data collection, then you can try to add a marker variable from your existing dataset that is from a construct that is theoretically distant from the rest of the constructs in the model. If that is not an option, then it would be best to use the common latent factor method (see my new video: "Common method bias using common latent factor"). Hope this helps.
Remove all path constraints on the common latent factor, leave it in the model pointing to all the indicators. This will partial out the variance due to the common method, and will provide a structural model of unique variance unbiased by the common method. I will probably soon make a new video about how better to handle common method bias.
I have a couple other videos about common method bias as well as a video about iteration limit reached. I recommend watching those videos. Then, if you still have trouble, feel free to email me. My email is on the RUclips channel homepage.
definitely together. It is their shared variance that we are interested in. So, they must be together. There are many things you can do to reduce bias, and there are many papers that talk about it. I recommend reading the recent one by MacKenzie and Podsakoff: "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies"
I don't know if I have ever seen a specific published cuttoff for any CMB approach except perhaps Harman's test (50%). I'm going to make a new CMB video next week hopefully that is updated with the best info. I'll make sure to include references as well.
For the CLF test, I would recommend to add all latent factors in the dataset to the test. This way we only extract common variance, and are less likely to also extract shared trait variance. If your model is simple, including a CLF can corrupt the model by extracting shared trait variance (not just due to a common method or a common bias).
This is from Podsakoff et al. 2003 about common method bias. Most of the effects will drop because you are removing the method variance. Variance is three-part: Trait Variance, Method Variance, and Error Variance. Since you are removing one part, then you will naturally have lower effects. Large differences indicate method bias issues. You can sometimes fix these by fiddling with removing one or more of the indicators with the largest change.
I would not rely on the Harman's single factor test. It is the least accurate. I'm actually in the process of redoing the CMB videos. New insights have recently been published that recommend traditional tests like the ones in this video are actually not effective (I learned about this last night...). I'm not sure how soon I'll be able to create the new video though. In the meantime, you could read: "Sources of Method Bias in Social Science Research and Recommendations on How to Control It"
Hi James. Thank you very much for all of the videos about SEM. I have learnt a lot. Last week I had my PhD viva. The examiner did not like how I use the marker variable technique as shown in this video (he also asked if I watched James Gaskin) and asked me to revise by following Lindell and Whitney (2001). This is weird as I can also see some researchers used your method and cite your name in high quality paper. I am confused. Is the marker variable technique in this video different from Lindell and Whitney (2001)? I have tried to read Ljndell and Whitney (2001) but did not have a glue, (I did not understand in other words) I can also see some researchers convert age and occupation into number and correlate with other variables, and conclude that there is no common method bias if there is no or low correlation. Thanks again James.
Methods to test CMB are many and not all scholars are in agreement on the "correct" method. I have not used Lindell and Whitney in my work. Here are some cites I have used: statwiki.gaskination.com/index.php?title=References#Method_Bias.2C_Response_Bias.2C_Specific_Bias
@@cunghoctienganh Yes, it is reviewed in these four: -Podsakoff, P. M., MacKenzie, S. B., Lee, J. Y., & Podsakoff, N. P. (2003). Common method biases in behavioral research: a critical review of the literature and recommended remedies. Journal of applied psychology, 88(5), 879. -MacKenzie, S. B., & Podsakoff, P. M. (2012). Common method bias in marketing: causes, mechanisms, and procedural remedies. Journal of Retailing, 88(4), 542-555. -Williams, L. J., Hartman, N., & Cavazotte, F. (2010). Method variance and marker variables: A review and comprehensive CFA marker technique. Organizational Research Methods, 13(3), 477-514. -Podsakoff, P. M., MacKenzie, S. B., & Podsakoff, N. P. (2012). Sources of method bias in social science research and recommendations on how to control it. Annual review of psychology, 63, 539-569.
@@Gaskination thanks James, can I say that it is the CFA marker variable technique? As I read, I noticed that there are several types of marker variable technique, I am unsure if what I read was correct. Thanks again
Probably the best reference is: Podsakoff, P. M., MacKenzie, S. B., Lee, J. Y., and Podsakoff, N. P. "Common method biases in behavioral research: a critical review of the literature and recommended remedies," Journal of Applied Psychology (88:5) 2003, p 879.
Common method bias is tested at the indicator level, so it doesn't matter if you have first order or second order. As for the drastic change in regression weights, this might be due to multicollinearity among the items. You might check the modification indices to see which items are strongly correlated.
For anyone wanting to do a post hoc test to determine if there is common method variance biasing data, I highly recommend reading Richardson, Simmering, Sturman (2009). A Tale of Three Perspectives: Examining Post Hoc Statistical Techniques for Detection and Correction of Common Method Variance; and also Williams, Hartman, & Cavazotte (2010). Method Variance and Marker Variables: A Review and Comprehensive Marker Technique. Both in the journal Organizational Research Methods.
common method effects could be judged based on competing model with and without introducing common method factor. if the delta chi square improved significantly with CLF, then it is likely a common method effect exist. Howwever, based on many recommandation ( delta chi square is not a good indicator for fit improvement), it is therefore suggested that incremental in CFI above 0.01 is a better indication of model improvement. if delta CFI is improved by less than 0.01 then we should not worry about CMB, given that the items remained significantly loading on their constructs .Regards
I recommend reading the recent one by MacKenzie and Podsakoff: "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies". 10.89% is not anything to worry about.
The most recent article on all of this is "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies" by MacKenzie and Podsakoff in 2012 in the journal of retailing.
Dear James Gaskin, Thank you for your great tutorials! Your new tutorial only include CLF without Marker Variable. As I would like to use the Marker Variable method, I would like to check with you what the last steps of the Marker variable method are : 1) In order to judge wether there is a bias : a) Should we use the squared regression path? If yes, what is the threshold limit to judge if there is a bias? or b) Should we calculate the difference between the regression path with and without the marker variable? And then, what about the threshold: is it 0.2 ? 2) Then, what are you doing before doing the structural model ? Do calculate composites (just like the CLF method)? Do you keep the Marker variable in you model when you run the structural model in Amos? I am looking forward to your answer. Best regards
1. Use the squared regression path. There is no published threshold, except for in reference to the single factor approach, which is 50%... 2. Yes, same as with the CLF. If you have method bias, then keep the marker and common variable in the model while creating composites, or while running the hybrid model.
If I understand correctly, you want to know why the estimates differ so greatly after including the common factor, yes? This is because a fair portion of the variance in those variables is explained by the common factor (supposedly method variance). To decrease the effect of the common factor, try including additional variables that were collected at the same time as the ones currently in the model. You don't have to use these additional variables any other time, just during CMB analysis.
I have looked near and far for a long time for a reference for the recommended threshold for common methods bias. I have never found one, except when referring to the Harmon's Single factor test (50%). Less is better. I would say that anything less than 20% is nothing to worry about, less than 30% is probably fine as well, more than 30% is troubling, more than 40% is a strong indication that you have a problem. More than 50% is not admissible.
@alexanderalexiev That is a good question! I'm not sure really. I guess it might be because we want to know what the common shared variance is (on average). I just know that this is the accepted method for estimating common method bias.
Harman's text doesn't actually teach anything beyond factor analysis. That is, it doesn't propose a test for common method bias. Rather, the name "Harman's One Factor Test" seems to first appear in print in Podsakoff & Organ (1986), although it was first used by Schriesheim (1979). So, one looking for a test of CMB in Harman's text will come up short (as I found out after much, much reading).
@vahidaryadoust As I understand it, you do not need to conduct a common method bias if you used more than one method. This test is also sometimes called "single method bias". If you used more than one method, then you can justify excluding this kind of a test.
You can make the amos background white, or you can remove the background for the picture by using the Word Picture functionality. Just google "how to remove color in word"
The problem is the incredibly high standard error values. To fix this, do NOT constrain all regression weights to be equal. Simply run the test with them unconstrained. Then compare the estimates from this model to the estimates of the model without the common factor. Notice the differences in standardized estimates. If the difference is greater than 0.200 for any estimate, then you will want to retain the common factor as you impute composite variables or as you move on to the structural model.
I have made a new video about this. I'm sorry this one is a bit outdated. I'll update it now. Anyway, please see my new video about using the common latent factor. Your professor is correct that we should not constrain the paths.
Thank you. Yes, Podsakoff et al do seem to indicate that would be the right approach, but they do not seem to discuss how to do it (other than what you show very nicely in your video). Would you be able to explain how to create a composite variable in AMOS that could then be used as a control variable in SPSS? Many thanks for your generosity.
Many thanks for your useful videos, which have taught me so much. I have run these a few times and don't generally run into problems but today my regression weights came out as .00 and remained at zero when I included the marker varaible. Does this sound possible? I should mention that I have 73 items across seven latent factors. Thanks for being incredibly patient with all of the queries with which you are inundated!
@sallgoo2009 This is probably because you are missing a constraint somewhere (like on a regression line). Feel free to email me directly and we can work it out. Make sure you have one constraint per factor. Also, you may need to impose a constraint to the variance of the CMB factor.
thank you dear James as usual all of your clip is very useful... thank's a lot James, I need an article about Harman’s method that described good and complete you told if variance more than fifty explained, it's trouble. but I have to write reference. thank you again
Last week I realized that the approach I recommend in this video is not the most appropriate. Instead of constraining all the paths from the common factor, they should be left unconstrained. Then examine the difference in regression weights for the latent factors --> items. If the regression weights change a lot after adding the common latent factor, then you likely have CMB issues. To then control for it, simply retain that latent factor as you move on to testing a structural model.
I am not positive, but I think you could create a composite variable in AMOS for the CLF and then use that as a control variable in SPSS. You might want to check Podsakoff et al 2003 to see if they say something about this.
Dear James: In this video, you said that the first factor should explain less than 50% variance. So, I guess that you are using Harman's test. m I correct? How about the second and third method? What might be the less and acceptable percentage for 2nd test? 20%? Would you please let me know which article I can have those?
Hi Prof. James, Thank you for these wonderful lectures. I have learned everything of my data analysis from your these lectures. I was wondering is there any plugin of model biasness test (CMB) for AMOS 21 version? I tried from the statwiki, but couldn't find it.
Hello James, I tried the second test in LISREL instead of Amos and, all things considered, it works. My question for you is: What is the threshold of acceptable common variance for items and the common latent factor? In your example, they share about 2% of variance. How much shared variance would be too much? Follow-up question: Does the fact that we also have other exogenous factors (here, your UB, FB, Aut, and PP) in the model decrease the amount of shared variance with the common method factor? Is that the reason the second number (2%) is so much lower than the number in the first test (33%)? Thank you in advance for your attention!
1. There is no published threshold, except for Harman's test (50%). Here is the most recent info on method bias including a video: statwiki.kolobkreations.com/index.php?title=Confirmatory_Factor_Analysis#Zero_and_Equal_Constraints 2. Correct. The more items we can include in the test, the more accurately it will filter out shared trait variance and focus in on just the shared method variance.
Thanks for the great video. I'm sure I'll be watching much of your playlist in the future. Is there an established threshold for common variance? I see that you've referred several folks to MacKenzie & Podsakoff 2012. Perhaps there's a threshold or guideline in there that I missed? Also, can you reference a recommended way to report common variance info? Thanks again!
There is not a good published threshold except with Harman's single factor test, we would hope that less than 50% of the variance is explained by a common factor. For the method in this particular video, I like to see less than 30%, but that's just me. As for how to report it, you can refer to my wiki: statwiki. kolobkreations. com and go to the general guidelines section for an Example Analysis (toward the top). It shows examples of how to report many of these SEM analyses.
Dear James, you mention that the apporach by Podsakoff is not correct (including an unmeasured latent method factor, direct paths to all indicators, constraining all loadings to be equal). However, Podsakoff et al. (2003) state on page 894 that you can do so. To overcome possible disadvantages of this approach he explicitly state "some reasearchers have constrained the measurement factor loadings to be equal"... With regard to Liang et a. (2007), who adapted this approach to mainly PLS studies, the constraining of the paths to be equal should be no problem. I know there are more accurate ways to check for common method variance. However, i think your approach is definitely not wrong ;) What is your opinion about it? Did i unterstand you correctly? Thanks in advance!
+Fabian Abed When no marker variable is provided, the common latent factor approach will often introduce errors in AMOS, resulting in an apparent 0% method variance. My current preferred approach is to use a marker variable (and CLF) and constrain the paths from the CLF to zero, and then compare (using chi-square difference test) this model to an unconstrained model. This will essentially tell you if a model with zero method variance is different from the unconstrained model.
Re: Harman's single factor test for CMB. Let's say that Principle Components reveals the variables do not load on one factor, but load on the expected factors from prior CFA validated factors. However, the percent of variance of first factor is 40 and the total accumulated variance of five factors is 70 (over 50 % on the first factor). Now I interpret this as CMV. What can I do with the findings from my MANOVA and Chi Sq Test results that show significance
Thanks for this video. Very useful! I have a question and wish you could help me. What if there are 40 test items loading on two factors but theoretically been influenced by, say, four test methods. In your example, you have only one test method. I appreciate your kind assistance.
Hi James: Thank you for this! I am wondering, do you include ALL observed variables in the model? All of my data comes from the same survey and I am wondering if I need to include all of the observed variables (including demographics such as income, internet use, etc.) when testing for bias. Thank you again -- this was very helpful! Laura
Try searching google scholar for "common method bias author:podsakoff" The top hit will be your best reference. As for a threshold, I have never found one, except when referring to the Harmon's Single factor test (50%). Less is better. I would say that anything less than 20% is nothing to worry about, less than 30% is probably fine as well, more than 30% is troubling, more than 40% is a strong indication that you have a problem. More than 50% is not admissible.
many thanks James for sharing your useful information. i have a question. in both method (SPSS and AMOS) should we calculate indicators in IVs and DVs separately or should calculate together? and how we can reduce bias?
Thank you very much, as always. Do you happen to have a citation for this method? And do you have any suggestions if some of my formerly significant effects disappear? Does this mean my results were not valid? Or is there another approach that might provide another look... I do not have a marker variable, unfortunately. Many thanks.
Thank you, as always, for your wonderful videos and assistance. If we are using AMOS only to test for CMB, and intend to use regressions in SPSS to test the hypotheses (due to small sample size), is there a way to assess the significance of the regression weight changes in AMOS due to the addition of a common latent factor? A chi square difference test suggests including the CLF is a significantly better fit, although using CFA in SPSS suggests variance due to a single factor is only 28%. TX!
Often the CLF won't run because of negative error variances or really high standard errors. Check my "iteration limit reached" video or my SEM Series CFA video for remedies.
Dear James, Thank you so much for your videos. When we squared the common latent factor to obtain the common variance, how we use this value in our further analysis, i.e. to correct the CMB? We substract this value from coefficients to reduce the bais? Thank you. Mike
Hi James, Thank you very much for such an informative video. Just one quick question. When I follow your CMB test with a marker variable, it seems like that my model is not working properly- the AMOS output doesn't seem to be normal (I can't get standardized loadings and model fit stats are all 1.00). What do I have to do in this case? Also, do you have any reference for common latent factor CMB test threshold? You mentioned that you'd be worried if it goes over 30% but do you have a specific reference for that ?
+Tiffany S 1. This often happens when adding so many parameters. You can try to troubleshoot it with this video: ruclips.net/video/B7YOv7hSohY/видео.html or this video: ruclips.net/video/Vx24KFf-rAo/видео.html 2. No published threshold exists for method bias. One way to determine if it is too much is to do a chi-square difference test between the model you just ran, and one where all the paths from the CLF are constrained to zero. This would tell you if the CMB was statistically different from zero.
Many thanks for your awesome tutorials :) I have 6 IVs, 8 Moderating Variables and 5 DVs ... Total items exceed 100. Inserting all of these at a time in AMOS for detecting common method bias is very cumbersome. Hope you can understand? Can we test common method bias i.e. Harman's single factor test etc separately for our constructs i.e. separately for DVs, seperately for IVs and separately for Moderating variables. For example while convergent and discriminant validity we do it separately for each construct.
Actually, it's not. The only test that has demonstrated any empirical efficacy to detect CMB is the CFA Marker Technique, described by Williams et al. (2010) in Organizational Research Methods. While Podsakoff et al. (2003) recommend the common latent factor test, the only study to truly test its validity is Richardson, Simmering, & Sturman (2009), in Organizational Research Methods, which demonstrate that it is highly flawed.
Help! I used your video (thank you!) to do a common latent method to test CMV for my dissertation. My professor commented "I'm not clear why you set the regression loadings to equality?how realistic is is they will all have the same regression weights?it seems you would want the coefficients to be free to vary for a one factor model, and then compare to your a priori model so please explain why you constrained the loadings to equality?" Please provide info and references about why all are equal?
One question for a structural model: does one add only the independent variables into a Harman's one factor test or are the dependent and independent variables included together?
If you end up retaining the latent factor, how do you treat it in the structural model? It has a bunch of indicators; do you keep its variance set to 1? Do you have it actually load on any of the outcome variables? Or does it just sit there pointing to all the various indicators in your model? Thank you very much!!!!
Hello James. Thank you for your great wiki! I used the CLF (with and without marker) in order to generate my composites for the SEM. However, I both case, the data fit of the SEM suffers greatly (in comparison to the composites made without the CLF). Then should I stick to the composite without CLF?
It just depends on whether you are likely to have a method bias problem. If the common method bias test indicates there is no real bias, then just drop the CLF. Watch my newer video on using the CLF and I also have one in the SEM series that demonstrates how to adjust things to correct for issues during the CMB test. Hopefully these will help.
Hello James, Just a small question of how to report with the CLF - I reported the measures such as GFI, RMSEA, etc. before inserting the CLF and the model fit was good enough. Then, your next step is to insert the CLF if we have CMB, which I do. But after I insert it, all those measures change again (GFI, RMSEA, etc.) and then I have Validity concerns, which didn't happen before... - which ones should I report? And the figure, should be the one with the CLF or just the normal CFA without it? It looks really confuse in the image if I present the CLF as well. Thank you so much!
Hello .. Two more queries please : 1- Is there any threshold limit for regression weights as in your case it is -.13 and squaring it gives 0.0160 which you mentioned approximately 2% common variance. I have more than 100 items and it difficult to run it simultaneously on all so I ran it only on my Moderating Variables and I have .31, (squaring it = 0.0961) i.e approximately 10 percent common variance/shared variance.... Shoud I run additional test by insertin Market Value? 2. For marker value, on what basis you added APATHY? Why not anything else from your SPSS file? or APATHY is something special you have placed in your SPSS which has no concern with your constructs but only meant as "Marker Variable" ... Summing up, what may I use as Marker Variable? Any logic behind it?
1. No published threshold is available, except in reference to the Harman's single factor test (50%). If I did the CLF test and observed more than 30%, I'd be worried. However, I have a new video on this approach which is more reliable: Common method bias using common latent factor Also, you may benefit from this brand new video: ruclips.net/video/etPciNEgWGk/видео.html 2. The most recent school of thought on marker variables is that they are not helpful unless they are theoretically derived. And, in cases when they are theoretically derived, the only one that has demonstrated consistent reliable effects is social desirability bias. So, unless you collected data on social desirability bias, you should stick with the CLF, if anything...
I really appreciate for your valuable videos, Just I wanna know which one is the best way of common method bias testing? and just using Harman's method would be sufficient and acceptable?
The marker variable approach is most rigorous. The Common Latent Factor approach is next best. Then, in last place, the Harman's test. Most journals would probably request additional tests if you only report Harman's.
@oogunfow1 Nothing comes immediately to mind. You would have to send me your model and I could take a look. Feel free to contact me via email: james. eric. gaskin at gmail. com
Using common latent factor and marker variable, we test common variance bias or variance, and then the results value is less than .50, respectively in these cases, is it possible to conclude that cmb issue is not serious? Look forward to your response asap. Thanks,
Dear James. Please advise on the minimum number of indicators that a construct can have and if you have any reference. One of my construct has only two and am worried with discriminate validity issues. Thanking you
You can run it with two. The stated ideal is four (I can't remember the reference... sorry). Two just sometimes becomes unstable and creates weird error variances and standard error.
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.
Thank you. I do have metric for the measurement model. Thus, based on what you said, pooled test would be acceptable? Individual countries don't turnout so hot.
Hello Dr. Gaskin, as I did't have the items of Marker variable, can I only use the ULMC (Unmeasured Latent Method Construct) to analyse CMV? Could you advice how can I convince the reviewers with the ULMC? Thanks a lot for the valuable videos!
Correct. ULMC is the next most appropriate approach. There are many articles on these methods: statwiki.gaskination.com/index.php?title=References#Method_Bias.2C_Response_Bias.2C_Specific_Bias
Hi, thanks a lot for this video! for the common factor, I get 1.14 :( . Did I do something wrong? (my Harman's single factor analysis gives me one factor explains 47% of the variance, which is high but acceptable). Thank you for your assistance, and great job with the videos!
Hi James, Thanks for this video explaining the common latent method. My query is - when i run the model, i get the all the regression weights (a) = .00. Am I doing it correct? It is possible?
This probably means that there is some other error in the model. Check the regression weights for the latent factors. See if any of them are super weak or became inverted. If so, then maybe move around the parameter constraint like in this video: ruclips.net/video/B7YOv7hSohY/видео.html
Thank you for the demonstrating. However, I was wondering if my third test (marker variable) have higher figure than second test ( Common Method Bias) what should I do? Does it mean I have common bias issues?
Hi James, thanks for your helpful video on common latent factor. Could you refer to any reference on this? And is it necessary to constrain the path to be equal?
Please watch my new video on this (which clarifies that it is not necessary to constrain the paths). Here is a reference: MacKenzie, S. B. and Podsakoff, P. M. (2012), "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies," Journal of Retailing 88(4), pp. 542-555.
James Gaskin Thanks a lot. Is there also a reference for the test with the constrains of the paths being equal and calculating the varianc explained by the method factor?
And another question: Is it possible to calculate the variance explained by the method factor if one does not restrict the path to be equal? In your other video, you show how to do this method, but I would like to know if one can get an overall evaluation of explained variance by the method factor with such a method? Sorry for all these questions - but your videos have been very helpful so far for me.
Steffen Giessner If there is a reference for constraining, it would be the podsakof article, but I'm not positive. i would have to go back and read through it. The overall level of shared variance is difficult to calculate without constraining them, however, often constraining them produces erratic results...
Dear James, thanks for your incredible and valuable videos. I wondered what is the difference between CMB models and biofactor models, can you explain? and if any difference is it possible to estimate a bifactor model in amos?
The bifactor model includes theoretically related secondary factors. The CMB includes theoretically unrelated secondary factors. Yes, a bifactor model should be possible in AMOS.
Thanks for this video. Do you have any article where the common latent factor method was and reported. I got 10.89% which i assume is not very large. I need to know how to describe and report this. Thanks
Dear james gaskin, I am Rahul Raoniar from India. I have watched the videos and read the statwiki site which I found very interesting. The SEM models & the interpretation of the results I found quit interesting. I am doing my M.Tech thesis on Public transport performance evaluation of Metro and Bus services in Delhi. By seeing your effort on SEM based modeling I also started learning SEM software’s. I have collected about 3000 sample with 25 different questionnaires. I have installed the SEM software AMOS and LISREL, when I make the structure diagram and run the model it sometimes says unidentified model, some time says missing data, sometimes runs but the path diagram shows regression coefficient more than 1, some time it shows turn off the modification indices as it not run with missing data .I was facing lot of problems in my data analysis and I have some confusions i) I have used Different scale for different questionnaire, is it necessary to take constant scale for all questionnaire like other variable like income you have used that consist of three parts less than 5000, 5000-10000, >10000 may be coded as 1, 2, 3 , could we able to use different scale and if yes then how? ii) I have analysed my data using spss, it shows no missing data, so why during run process of model it shows missing data. iii) some time it runs but it shows estimates as unidentified and it does not show regression coefficient iv) How to check the data whether it fit for analysis or not. Could you help me to address those problems that I am facing during my analysis? I will be very thankful to you if you will help me to overcome those problems. Rahul Raoniar Central road research institute New Delhi , India Email id------rahul.raoniar@hotmail.com Facebook-- Rahul Raoniar
1. No. not required to have same scale. 2. Sometimes this happens when the last row (or set of rows) in your SPSS file are active but empty. You need to delete the entire blank row. 3. This happens when it can't minimize (find a good solution) due to too much error. This rarely happens when data screening and a good EFA have been conducted first. 4. Do Data screening (see my video for this in the SEM Series playlist) and do a good EFA and CFA. If you pass these analyses and demonstrate good convergent and discriminant validity and reliability, then you are fit for analysis. After that, you'll also want to check model fit (I have videos for all these things). Hope this helps.
hello james, nice effort indeed, my cmb is 52% in spss, but when i want check in amos through common factor then error shows (THE VARIABLE NAME 'COMMON FACTOR' IS INVALID, BECAUSE IT CONTAINS AN INVALID CHARACTER, so uploaded spss data file, so should i incorporate variable names common factor, its not in my questionnaire, thats may be amos give error, i hope you understand, please guide
Hi James,I saw your video Common method bias using common latent factor. If our measurement model include one second order factor and three first order we shoud keep the second order factor?
Hi Jeff! I have a quick question regarding the marker variable CMB test. I added a marker variable and a common factor, as shown in the RUclips video. I did exactly what you did, but when I ran the model, the results say that the model is probably unidenitifed and I need one additional constraint. So, I am not sure what I need to do. Thanks. Boryana
That's a new one. I've never been called Jeff yet :) My guess is that you either did not include a path constraint on one of the indicators for the marker variable, or you did not include a variance constraint on the common factor. I have a new video for this by the way. It is called something like "common method bias with a common latent factor in amos".
James Gaskin I am so sorry for calling you Jeff....Thank you for the hints. I just checked my model again and have all constraints, just like you do in this video...Not sure what is wrong. I watched the other RUclips video on CMB as well....
Boryana Dimitrova You can send it to me and I'll look at it when I have a moment. I'm traveling today, but I might be able to get to it in between meetings. My email is "james.gaskin@byu.edu"
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!
You are really a superb trainer. thank you.
Can i ask for evidence on single factor common method bias suggestions? for example, less than 50% variance. Thank you
@@MatinsVlog Some of these should help: statwiki.gaskination.com/index.php?title=References#Method_Bias.2C_Response_Bias.2C_Specific_Bias
@@Gaskination Thank you so much
@@Gaskination May i ask you something related to EFA?
three constructs combined my DV. in that case, when i run EFA, loadings fall under same variable. is it okay to continue CFA ? or i need to get more data to get separate factor loadings for each construct?
@@MatinsVlog When it is a 2nd order factor, this is not completely unexpected. You can try to run an EFA with just the items from this factor, and then constrain it to extract three factors. Or you can move directly to the CFA. The trouble you will have if you skip the EFA is knowing where discriminant validity issues are originating.
Hi James, I just wanted to let you know that your videos were a God send for me over the past few days when trying to perform SEM for my thesis. Thank you so much, you are a fantastic presenter, love your work, keep it up!
WHAT YOU ARE DOING HERE, IT IS BETTER THANT THE KNOWLEDGE WE RECEIVE IN REPUTABLE UNIVERSITIES AT THE GRADUATE LEVEL....THANKS AND CONTINUE YOU ARE EDUCATING MANY PEOPLE
Yes. That is the best way to do it. If you didn't capture the marker during data collection, then you can try to add a marker variable from your existing dataset that is from a construct that is theoretically distant from the rest of the constructs in the model. If that is not an option, then it would be best to use the common latent factor method (see my new video: "Common method bias using common latent factor"). Hope this helps.
Remove all path constraints on the common latent factor, leave it in the model pointing to all the indicators. This will partial out the variance due to the common method, and will provide a structural model of unique variance unbiased by the common method. I will probably soon make a new video about how better to handle common method bias.
I have a couple other videos about common method bias as well as a video about iteration limit reached. I recommend watching those videos. Then, if you still have trouble, feel free to email me. My email is on the RUclips channel homepage.
definitely together. It is their shared variance that we are interested in. So, they must be together. There are many things you can do to reduce bias, and there are many papers that talk about it. I recommend reading the recent one by MacKenzie and Podsakoff: "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies"
I don't know if I have ever seen a specific published cuttoff for any CMB approach except perhaps Harman's test (50%). I'm going to make a new CMB video next week hopefully that is updated with the best info. I'll make sure to include references as well.
Hi James, I have passed the Harman test but I obtained 77.44% (0.88*0.88) in the Common latent factor test. What should I do? I am really stressed :/
For the CLF test, I would recommend to add all latent factors in the dataset to the test. This way we only extract common variance, and are less likely to also extract shared trait variance. If your model is simple, including a CLF can corrupt the model by extracting shared trait variance (not just due to a common method or a common bias).
This is from Podsakoff et al. 2003 about common method bias. Most of the effects will drop because you are removing the method variance. Variance is three-part: Trait Variance, Method Variance, and Error Variance. Since you are removing one part, then you will naturally have lower effects. Large differences indicate method bias issues. You can sometimes fix these by fiddling with removing one or more of the indicators with the largest change.
I would not rely on the Harman's single factor test. It is the least accurate. I'm actually in the process of redoing the CMB videos. New insights have recently been published that recommend traditional tests like the ones in this video are actually not effective (I learned about this last night...). I'm not sure how soon I'll be able to create the new video though. In the meantime, you could read: "Sources of Method Bias in Social Science Research and Recommendations on How to Control It"
I love to watch your lectures. your videos are great source of learning for us
Hi James. Thank you very much for all of the videos about SEM. I have learnt a lot.
Last week I had my PhD viva. The examiner did not like how I use the marker variable technique as shown in this video (he also asked if I watched James Gaskin) and asked me to revise by following Lindell and Whitney (2001). This is weird as I can also see some researchers used your method and cite your name in high quality paper.
I am confused. Is the marker variable technique in this video different from Lindell and Whitney (2001)? I have tried to read Ljndell and Whitney (2001) but did not have a glue, (I did not understand in other words)
I can also see some researchers convert age and occupation into number and correlate with other variables, and conclude that there is no common method bias if there is no or low correlation.
Thanks again James.
Methods to test CMB are many and not all scholars are in agreement on the "correct" method. I have not used Lindell and Whitney in my work. Here are some cites I have used: statwiki.gaskination.com/index.php?title=References#Method_Bias.2C_Response_Bias.2C_Specific_Bias
@@Gaskination hi James, thanks for your reply. Can I say that the method you used in the video is from William, Hartmand and Cavazotte?
Sorry, I have read other comments, the reference for this is Podsakoff. Thank you
@@cunghoctienganh Yes, it is reviewed in these four:
-Podsakoff, P. M., MacKenzie, S. B., Lee, J. Y., & Podsakoff, N. P. (2003). Common method biases in behavioral research: a critical review of the literature and recommended remedies. Journal of applied psychology, 88(5), 879.
-MacKenzie, S. B., & Podsakoff, P. M. (2012). Common method bias in marketing: causes, mechanisms, and procedural remedies. Journal of Retailing, 88(4), 542-555.
-Williams, L. J., Hartman, N., & Cavazotte, F. (2010). Method variance and marker variables: A review and comprehensive CFA marker technique. Organizational Research Methods, 13(3), 477-514.
-Podsakoff, P. M., MacKenzie, S. B., & Podsakoff, N. P. (2012). Sources of method bias in social science research and recommendations on how to control it. Annual review of psychology, 63, 539-569.
@@Gaskination thanks James, can I say that it is the CFA marker variable technique? As I read, I noticed that there are several types of marker variable technique, I am unsure if what I read was correct. Thanks again
Probably the best reference is:
Podsakoff, P. M., MacKenzie, S. B., Lee, J. Y., and Podsakoff, N. P. "Common method biases in behavioral research: a critical review of the literature and recommended remedies," Journal of Applied Psychology (88:5) 2003, p 879.
Common method bias is tested at the indicator level, so it doesn't matter if you have first order or second order. As for the drastic change in regression weights, this might be due to multicollinearity among the items. You might check the modification indices to see which items are strongly correlated.
For anyone wanting to do a post hoc test to determine if there is common method variance biasing data, I highly recommend reading Richardson, Simmering, Sturman (2009). A Tale of Three Perspectives: Examining Post Hoc Statistical Techniques for Detection and Correction of Common Method Variance; and also Williams, Hartman, & Cavazotte (2010). Method Variance and Marker Variables: A Review and Comprehensive Marker Technique. Both in the journal Organizational Research Methods.
common method effects could be judged based on competing model with and without introducing common method factor. if the delta chi square improved significantly with CLF, then it is likely a common method effect exist. Howwever, based on many recommandation ( delta chi square is not a good indicator for fit improvement), it is therefore suggested that incremental in CFI above 0.01 is a better indication of model improvement. if delta CFI is improved by less than 0.01 then we should not worry about CMB, given that the items remained significantly loading on their constructs .Regards
+Basit Queiri Do you have any reference regarding this thing?
I recommend reading the recent one by MacKenzie and Podsakoff: "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies". 10.89% is not anything to worry about.
The most recent article on all of this is "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies" by MacKenzie and Podsakoff in 2012 in the journal of retailing.
Dear James Gaskin,
Thank you for your great tutorials!
Your new tutorial only include CLF without Marker Variable. As I would like to use the Marker Variable method, I would like to check with you what the last steps of the Marker variable method are :
1) In order to judge wether there is a bias :
a) Should we use the squared regression path? If yes, what is the threshold limit to judge if there is a bias?
or
b) Should we calculate the difference between the regression path with and without the marker variable? And then, what about the threshold: is it 0.2 ?
2) Then, what are you doing before doing the structural model ? Do calculate composites (just like the CLF method)? Do you keep the Marker variable in you model when you run the structural model in Amos?
I am looking forward to your answer.
Best regards
1. Use the squared regression path. There is no published threshold, except for in reference to the single factor approach, which is 50%...
2. Yes, same as with the CLF. If you have method bias, then keep the marker and common variable in the model while creating composites, or while running the hybrid model.
If I understand correctly, you want to know why the estimates differ so greatly after including the common factor, yes? This is because a fair portion of the variance in those variables is explained by the common factor (supposedly method variance). To decrease the effect of the common factor, try including additional variables that were collected at the same time as the ones currently in the model. You don't have to use these additional variables any other time, just during CMB analysis.
Yes, the apathy data was part of the dataset. I added it as the marker variable because it was not very much alike the other factors in the model.
Very educative and simplified. Thank you.
You can create the composites following the instruction I give in my video called: "Imputing Composite Variables in AMOS"
I have looked near and far for a long time for a reference for the recommended threshold for common methods bias. I have never found one, except when referring to the Harmon's Single factor test (50%). Less is better. I would say that anything less than 20% is nothing to worry about, less than 30% is probably fine as well, more than 30% is troubling, more than 40% is a strong indication that you have a problem. More than 50% is not admissible.
Hello James, sorry to bother u here i'm getting 56% of the variance since u are saying it is not admissible how should I fix it?
@alexanderalexiev
That is a good question! I'm not sure really. I guess it might be because we want to know what the common shared variance is (on average). I just know that this is the accepted method for estimating common method bias.
Harman's text doesn't actually teach anything beyond factor analysis. That is, it doesn't propose a test for common method bias. Rather, the name "Harman's One Factor Test" seems to first appear in print in Podsakoff & Organ (1986), although it was first used by Schriesheim (1979). So, one looking for a test of CMB in Harman's text will come up short (as I found out after much, much reading).
I definitely agree. The more appropriate method is demonstrated in my new video on how to use a common latent factor to test for CMB.
@vahidaryadoust As I understand it, you do not need to conduct a common method bias if you used more than one method. This test is also sometimes called "single method bias". If you used more than one method, then you can justify excluding this kind of a test.
You can make the amos background white, or you can remove the background for the picture by using the Word Picture functionality. Just google "how to remove color in word"
The problem is the incredibly high standard error values. To fix this, do NOT constrain all regression weights to be equal. Simply run the test with them unconstrained. Then compare the estimates from this model to the estimates of the model without the common factor. Notice the differences in standardized estimates. If the difference is greater than 0.200 for any estimate, then you will want to retain the common factor as you impute composite variables or as you move on to the structural model.
Wonderful. Thank you very much, as always!
ผมมีความสุขที่จะได้รับประโยชน์ ฉันขอให้คุณโชคดีที่สุด
I have made a new video about this. I'm sorry this one is a bit outdated. I'll update it now. Anyway, please see my new video about using the common latent factor. Your professor is correct that we should not constrain the paths.
I'll have to look into it. I have not yet read Williams et al 2010. I'll check it out. Thanks for the lead!
Thank you. Yes, Podsakoff et al do seem to indicate that would be the right approach, but they do not seem to discuss how to do it (other than what you show very nicely in your video). Would you be able to explain how to create a composite variable in AMOS that could then be used as a control variable in SPSS?
Many thanks for your generosity.
Many thanks for your useful videos, which have taught me so much. I have run these a few times and don't generally run into problems but today my regression weights came out as .00 and remained at zero when I included the marker varaible. Does this sound possible? I should mention that I have 73 items across seven latent factors. Thanks for being incredibly patient with all of the queries with which you are inundated!
@sallgoo2009 This is probably because you are missing a constraint somewhere (like on a regression line). Feel free to email me directly and we can work it out. Make sure you have one constraint per factor. Also, you may need to impose a constraint to the variance of the CMB factor.
thank you dear James as usual
all of your clip is very useful... thank's a lot
James, I need an article about Harman’s method that described good and complete
you told if variance more than fifty explained, it's trouble.
but I have to write reference.
thank you again
Harman's is considered an outdated method. I recommend this approach: ruclips.net/video/CFBUECZgUuo/видео.html
Yes. Should be fine. I usually am happy with anything less than about 30% (0.55).
Thank you James! This is so helpful..
Last week I realized that the approach I recommend in this video is not the most appropriate. Instead of constraining all the paths from the common factor, they should be left unconstrained. Then examine the difference in regression weights for the latent factors --> items. If the regression weights change a lot after adding the common latent factor, then you likely have CMB issues. To then control for it, simply retain that latent factor as you move on to testing a structural model.
I am not positive, but I think you could create a composite variable in AMOS for the CLF and then use that as a control variable in SPSS. You might want to check Podsakoff et al 2003 to see if they say something about this.
Dear James: In this video, you said that the first factor should explain less than 50% variance. So, I guess that you are using Harman's test. m I correct? How about the second and third method? What might be the less and acceptable percentage for 2nd test? 20%? Would you please let me know which article I can have those?
Hi Prof. James, Thank you for these wonderful lectures. I have learned everything of my data analysis from your these lectures. I was wondering is there any plugin of model biasness test (CMB) for AMOS 21 version? I tried from the statwiki, but couldn't find it.
I don't know if it works on versions that old, but you can try the one in the '23 or lower' folder called "CLF23.dll". Hope it works!
Hello James, I tried the second test in LISREL instead of Amos and, all things considered, it works.
My question for you is: What is the threshold of acceptable common variance for items and the common latent factor? In your example, they share about 2% of variance. How much shared variance would be too much?
Follow-up question: Does the fact that we also have other exogenous factors (here, your UB, FB, Aut, and PP) in the model decrease the amount of shared variance with the common method factor? Is that the reason the second number (2%) is so much lower than the number in the first test (33%)?
Thank you in advance for your attention!
1. There is no published threshold, except for Harman's test (50%). Here is the most recent info on method bias including a video: statwiki.kolobkreations.com/index.php?title=Confirmatory_Factor_Analysis#Zero_and_Equal_Constraints
2. Correct. The more items we can include in the test, the more accurately it will filter out shared trait variance and focus in on just the shared method variance.
Thanks for the great video. I'm sure I'll be watching much of your playlist in the future. Is there an established threshold for common variance? I see that you've referred several folks to MacKenzie & Podsakoff 2012. Perhaps there's a threshold or guideline in there that I missed? Also, can you reference a recommended way to report common variance info? Thanks again!
Specific to my current project: I'm running a common latent factor test w/.19 common variance (.44^2).
There is not a good published threshold except with Harman's single factor test, we would hope that less than 50% of the variance is explained by a common factor. For the method in this particular video, I like to see less than 30%, but that's just me. As for how to report it, you can refer to my wiki: statwiki. kolobkreations. com and go to the general guidelines section for an Example Analysis (toward the top). It shows examples of how to report many of these SEM analyses.
Dear James,
you mention that the apporach by Podsakoff is not correct (including an unmeasured latent method factor, direct paths to all indicators, constraining all loadings to be equal). However, Podsakoff et al. (2003) state on page 894 that you can do so. To overcome possible disadvantages of this approach he explicitly state "some reasearchers have constrained the measurement factor loadings to be equal"...
With regard to Liang et a. (2007), who adapted this approach to mainly PLS studies, the constraining of the paths to be equal should be no problem.
I know there are more accurate ways to check for common method variance. However, i think your approach is definitely not wrong ;)
What is your opinion about it? Did i unterstand you correctly?
Thanks in advance!
+Fabian Abed When no marker variable is provided, the common latent factor approach will often introduce errors in AMOS, resulting in an apparent 0% method variance. My current preferred approach is to use a marker variable (and CLF) and constrain the paths from the CLF to zero, and then compare (using chi-square difference test) this model to an unconstrained model. This will essentially tell you if a model with zero method variance is different from the unconstrained model.
Re: Harman's single factor test for CMB. Let's say that Principle Components reveals the variables do not load on one factor, but load on the expected factors from prior CFA validated factors. However, the percent of variance of first factor is 40 and the total accumulated variance of five factors is 70 (over 50 % on the first factor). Now I interpret this as CMV. What can I do with the findings from my MANOVA and Chi Sq Test results that show significance
@academicthings
Thanks! I'm glad you find it useful. I hope you will also visit my wiki: statwiki. kolobkreations. com
Thanks for this video. Very useful! I have a question and wish you could help me. What if there are 40 test items loading on two factors but theoretically been influenced by, say, four test methods. In your example, you have only one test method. I appreciate your kind assistance.
@randomfleming Both together if they were collected using the from the same set of people using a single survey (or single method).
I'm not aware of another one in SPSS.
Hi James:
Thank you for this!
I am wondering, do you include ALL observed variables in the model?
All of my data comes from the same survey and I am wondering if I need to include all of the observed variables (including demographics such as income, internet use, etc.) when testing for bias.
Thank you again -- this was very helpful!
Laura
No. Just the latent variables that can actually be affected by some sort of social or temporal bias.
Try searching google scholar for "common method bias author:podsakoff" The top hit will be your best reference. As for a threshold, I have never found one, except when referring to the Harmon's Single factor test (50%). Less is better. I would say that anything less than 20% is nothing to worry about, less than 30% is probably fine as well, more than 30% is troubling, more than 40% is a strong indication that you have a problem. More than 50% is not admissible.
many thanks James for sharing your useful information. i have a question.
in both method (SPSS and AMOS) should we calculate indicators in IVs and DVs separately or should calculate together?
and how we can reduce bias?
The procedure shouldn't be any different for versions 20 or 21.
Thank you very much, as always. Do you happen to have a citation for this method? And do you have any suggestions if some of my formerly significant effects disappear? Does this mean my results were not valid? Or is there another approach that might provide another look... I do not have a marker variable, unfortunately. Many thanks.
Thank you, as always, for your wonderful videos and assistance. If we are using AMOS only to test for CMB, and intend to use regressions in SPSS to test the hypotheses (due to small sample size), is there a way to assess the significance of the regression weight changes in AMOS due to the addition of a common latent factor?
A chi square difference test suggests including the CLF is a significantly better fit, although using CFA in SPSS suggests variance due to a single factor is only 28%.
TX!
Dear james
Please explain the interpretation for all these methods. Which values we will plot in our paper and in what way?
Often the CLF won't run because of negative error variances or really high standard errors. Check my "iteration limit reached" video or my SEM Series CFA video for remedies.
Dear James,
Thank you so much for your videos. When we squared the common latent factor to obtain the common variance, how we use this value in our further analysis, i.e. to correct the CMB? We substract this value from coefficients to reduce the bais? Thank you. Mike
Hi James,
Thank you very much for such an informative video. Just one quick question. When I follow your CMB test with a marker variable, it seems like that my model is not working properly- the AMOS output doesn't seem to be normal (I can't get standardized loadings and model fit stats are all 1.00). What do I have to do in this case? Also, do you have any reference for common latent factor CMB test threshold? You mentioned that you'd be worried if it goes over 30% but do you have a specific reference for that ?
+Tiffany S
1. This often happens when adding so many parameters. You can try to troubleshoot it with this video: ruclips.net/video/B7YOv7hSohY/видео.html or this video: ruclips.net/video/Vx24KFf-rAo/видео.html
2. No published threshold exists for method bias. One way to determine if it is too much is to do a chi-square difference test between the model you just ran, and one where all the paths from the CLF are constrained to zero. This would tell you if the CMB was statistically different from zero.
I've never used lisrel, but I assume the procedure would not change.
Thank you very much for your help.
Many thanks for your awesome tutorials :)
I have 6 IVs, 8 Moderating Variables and 5 DVs ... Total items exceed 100. Inserting all of these at a time in AMOS for detecting common method bias is very cumbersome. Hope you can understand?
Can we test common method bias i.e. Harman's single factor test etc separately for our constructs i.e. separately for DVs, seperately for IVs and separately for Moderating variables. For example while convergent and discriminant validity we do it separately for each construct.
The whole point of CMB is to identify a common latent factor across ALL indicators. So, splitting up the indicators would prove less effective.
Actually, it's not. The only test that has demonstrated any empirical efficacy to detect CMB is the CFA Marker Technique, described by Williams et al. (2010) in Organizational Research Methods. While Podsakoff et al. (2003) recommend the common latent factor test, the only study to truly test its validity is Richardson, Simmering, & Sturman (2009), in Organizational Research Methods, which demonstrate that it is highly flawed.
thanks for the quick response, I not sure how to code it in LISREL because it is just one item. thanks for the video though!
Help! I used your video (thank you!) to do a common latent method to test CMV for my dissertation. My professor commented "I'm not clear why you set the regression loadings to equality?how realistic is is they will all have the same regression weights?it seems you would want the coefficients to be free to vary for a one factor model, and then compare to your a priori model so please explain why you constrained the loadings to equality?" Please provide info and references about why all are equal?
One question for a structural model: does one add only the independent variables into a Harman's one factor test or are the dependent and independent variables included together?
If you end up retaining the latent factor, how do you treat it in the structural model? It has a bunch of indicators; do you keep its variance set to 1? Do you have it actually load on any of the outcome variables? Or does it just sit there pointing to all the various indicators in your model?
Thank you very much!!!!
Hello James. Thank you for your great wiki! I used the CLF (with and without marker) in order to generate my composites for the SEM. However, I both case, the data fit of the SEM suffers greatly (in comparison to the composites made without the CLF). Then should I stick to the composite without CLF?
It just depends on whether you are likely to have a method bias problem. If the common method bias test indicates there is no real bias, then just drop the CLF. Watch my newer video on using the CLF and I also have one in the SEM series that demonstrates how to adjust things to correct for issues during the CMB test. Hopefully these will help.
Hello James,
Just a small question of how to report with the CLF - I reported the measures such as GFI, RMSEA, etc. before inserting the CLF and the model fit was good enough. Then, your next step is to insert the CLF if we have CMB, which I do. But after I insert it, all those measures change again (GFI, RMSEA, etc.) and then I have Validity concerns, which didn't happen before... - which ones should I report? And the figure, should be the one with the CLF or just the normal CFA without it? It looks really confuse in the image if I present the CLF as well. Thank you so much!
both without CLF
James Gaskin Thank you so so much, you should be included as author in my papers :D
Hello .. Two more queries please :
1- Is there any threshold limit for regression weights as in your case it is -.13 and squaring it gives 0.0160 which you mentioned approximately 2% common variance. I have more than 100 items and it difficult to run it simultaneously on all so I ran it only on my Moderating Variables and I have .31, (squaring it = 0.0961) i.e approximately 10 percent common variance/shared variance.... Shoud I run additional test by insertin Market Value?
2. For marker value, on what basis you added APATHY? Why not anything else from your SPSS file? or APATHY is something special you have placed in your SPSS which has no concern with your constructs but only meant as "Marker Variable" ... Summing up, what may I use as Marker Variable? Any logic behind it?
1. No published threshold is available, except in reference to the Harman's single factor test (50%). If I did the CLF test and observed more than 30%, I'd be worried. However, I have a new video on this approach which is more reliable: Common method bias using common latent factor Also, you may benefit from this brand new video: ruclips.net/video/etPciNEgWGk/видео.html
2. The most recent school of thought on marker variables is that they are not helpful unless they are theoretically derived. And, in cases when they are theoretically derived, the only one that has demonstrated consistent reliable effects is social desirability bias. So, unless you collected data on social desirability bias, you should stick with the CLF, if anything...
I really appreciate for your valuable videos, Just I wanna know which one is the best way of common method bias testing? and just using Harman's method would be sufficient and acceptable?
The marker variable approach is most rigorous. The Common Latent Factor approach is next best. Then, in last place, the Harman's test. Most journals would probably request additional tests if you only report Harman's.
thanks for ur fast and good response
ขอบคุณมากค่ะ มีประโยชน์มาก
@oogunfow1
Nothing comes immediately to mind. You would have to send me your model and I could take a look. Feel free to contact me via email: james. eric. gaskin at gmail. com
Hi James, how about making a video on the implementation of Williams, Hartman, & Cavazotte's (2010) Comprehensive CFA Marker Technique?
Using common latent factor and marker variable, we test common variance bias or variance, and then the results value is less than .50, respectively in these cases, is it possible to conclude that cmb issue is not serious? Look forward to your response asap. Thanks,
Dear James. Please advise on the minimum number of indicators that a construct can have and if you have any reference. One of my construct has only two and am worried with discriminate validity issues. Thanking you
You can run it with two. The stated ideal is four (I can't remember the reference... sorry). Two just sometimes becomes unstable and creates weird error variances and standard error.
Thanks. I came across Little et al 1999 and they say its good have optimal numbers of items rather than quantities
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?
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.
Thank you. I do have metric for the measurement model. Thus, based on what you said, pooled test would be acceptable? Individual countries don't turnout so hot.
Thanks for the quick reply. Is there a different test in SPSS that is now utilized?
Hello Dr. Gaskin, as I did't have the items of Marker variable, can I only use the ULMC (Unmeasured Latent Method Construct) to analyse CMV? Could you advice how can I convince the reviewers with the ULMC? Thanks a lot for the valuable videos!
Correct. ULMC is the next most appropriate approach. There are many articles on these methods: statwiki.gaskination.com/index.php?title=References#Method_Bias.2C_Response_Bias.2C_Specific_Bias
Wow! good to know! Thanks!
Hi, thanks a lot for this video! for the common factor, I get 1.14 :( . Did I do something wrong? (my Harman's single factor analysis gives me one factor explains 47% of the variance, which is high but acceptable). Thank you for your assistance, and great job with the videos!
Hi James, Thanks for this video explaining the common latent method. My query is - when i run the model, i get the all the regression weights (a) = .00. Am I doing it correct? It is possible?
This probably means that there is some other error in the model. Check the regression weights for the latent factors. See if any of them are super weak or became inverted. If so, then maybe move around the parameter constraint like in this video: ruclips.net/video/B7YOv7hSohY/видео.html
Many thanks!
Thank you for the demonstrating. However, I was wondering if my third test (marker variable) have higher figure than second test ( Common Method Bias) what should I do? Does it mean I have common bias issues?
If the clf analysis results indicate that the clf accounts for a variance of over .2 should i use the clf as a control variable ?
Use my newer video about Common Latent Factor. This fixes the .00 issue.
please share the link of the video
@@mantashafiroz4431 Here is the newer video: ruclips.net/video/Y7Le5Vb7_jg/видео.html
Hi James, thanks for your helpful video on common latent factor. Could you refer to any reference on this? And is it necessary to constrain the path to be equal?
Please watch my new video on this (which clarifies that it is not necessary to constrain the paths). Here is a reference:
MacKenzie, S. B. and Podsakoff, P. M. (2012), "Common Method Bias in Marketing: Causes, Mechanisms, and Procedural Remedies," Journal of Retailing 88(4), pp. 542-555.
James Gaskin
Thanks a lot. Is there also a reference for the test with the constrains of the paths being equal and calculating the varianc explained by the method factor?
And another question: Is it possible to calculate the variance explained by the method factor if one does not restrict the path to be equal? In your other video, you show how to do this method, but I would like to know if one can get an overall evaluation of explained variance by the method factor with such a method? Sorry for all these questions - but your videos have been very helpful so far for me.
Steffen Giessner If there is a reference for constraining, it would be the podsakof article, but I'm not positive. i would have to go back and read through it. The overall level of shared variance is difficult to calculate without constraining them, however, often constraining them produces erratic results...
@Gaskination tnks you very much for your prompt answer, and for teach me to fish ;)
Dear James, thanks for your incredible and valuable videos. I wondered what is the difference between CMB models and biofactor models, can you explain? and if any difference is it possible to estimate a bifactor model in amos?
The bifactor model includes theoretically related secondary factors. The CMB includes theoretically unrelated secondary factors. Yes, a bifactor model should be possible in AMOS.
@@Gaskination Thanks a lot! I hope you consider uploading a video of a bifactor model example and its indices!
great thanks, would you please recommend some scholar articles about the latent common method bias?
Thanks for this video. Do you have any article where the common latent factor method was and reported. I got 10.89% which i assume is not very large. I need to know how to describe and report this. Thanks
Dear james gaskin,
I am Rahul Raoniar from India. I have watched the videos and read the statwiki site which I found very interesting. The SEM models & the interpretation of the results I found quit interesting. I am doing my M.Tech thesis on Public transport performance evaluation of Metro and Bus services in Delhi. By seeing your effort on SEM based modeling I also started learning SEM software’s. I have collected about 3000 sample with 25 different questionnaires. I have installed the SEM software AMOS and LISREL, when I make the structure diagram and run the model it sometimes says unidentified model, some time says missing data, sometimes runs but the path diagram shows regression coefficient more than 1, some time it shows turn off the modification indices as it not run with missing data .I was facing lot of problems in my data analysis and I have some confusions
i) I have used Different scale for different questionnaire, is it necessary to take constant scale for all questionnaire like other variable like income you have used that consist of three parts less than 5000, 5000-10000, >10000 may be coded as 1, 2, 3 , could we able to use different scale and if yes then how?
ii) I have analysed my data using spss, it shows no missing data, so why during run process of model it shows missing data.
iii) some time it runs but it shows estimates as unidentified and it does not show regression coefficient
iv) How to check the data whether it fit for analysis or not.
Could you help me to address those problems that I am facing during my analysis? I will be very thankful to you if you will help me to overcome those problems.
Rahul Raoniar
Central road research institute
New Delhi , India
Email id------rahul.raoniar@hotmail.com
Facebook-- Rahul Raoniar
1. No. not required to have same scale.
2. Sometimes this happens when the last row (or set of rows) in your SPSS file are active but empty. You need to delete the entire blank row.
3. This happens when it can't minimize (find a good solution) due to too much error. This rarely happens when data screening and a good EFA have been conducted first.
4. Do Data screening (see my video for this in the SEM Series playlist) and do a good EFA and CFA. If you pass these analyses and demonstrate good convergent and discriminant validity and reliability, then you are fit for analysis. After that, you'll also want to check model fit (I have videos for all these things).
Hope this helps.
Many many thanks to you for your feed back. All of my queries are now solved and amos works well ........Thanks
hello james, nice effort indeed, my cmb is 52% in spss, but when i want check in amos through common factor then error shows (THE VARIABLE NAME 'COMMON FACTOR' IS INVALID, BECAUSE IT CONTAINS AN INVALID CHARACTER, so uploaded spss data file, so should i incorporate variable names common factor, its not in my questionnaire, thats may be amos give error, i hope you understand, please guide
Hi James,I saw your video Common method bias using common latent factor. If our measurement model include one second order factor and three first order we shoud keep the second order factor?
Yes. For CMB testing, the more variables included the better.
so good..Thank you
Hi Jeff! I have a quick question regarding the marker variable CMB test. I added a marker variable and a common factor, as shown in the RUclips video. I did exactly what you did, but when I ran the model, the results say that the model is probably unidenitifed and I need one additional constraint. So, I am not sure what I need to do.
Thanks.
Boryana
That's a new one. I've never been called Jeff yet :)
My guess is that you either did not include a path constraint on one of the indicators for the marker variable, or you did not include a variance constraint on the common factor. I have a new video for this by the way. It is called something like "common method bias with a common latent factor in amos".
James Gaskin I am so sorry for calling you Jeff....Thank you for the hints. I just checked my model again and have all constraints, just like you do in this video...Not sure what is wrong. I watched the other RUclips video on CMB as well....
Boryana Dimitrova You can send it to me and I'll look at it when I have a moment. I'm traveling today, but I might be able to get to it in between meetings. My email is "james.gaskin@byu.edu"