Confirmatory factor analysis in AMOS (Sept 2020)
HTML-код
- Опубликовано: 21 авг 2024
- This video presentation provides a general introduction to using AMOS to perform confirmatory factor analysis. In it I cover how to set up a diagram to test your model (including discussions of various drawing program options), how to run your analysis (after selection of options under the Analysis Properties box), where to find various outputs, and how to interpret your results.
In the video I recommend you download this Powerpoint to follow along: drive.google.c...
Download the data for the first example here: drive.google.c...
Download the data for the second example here:
drive.google.c...
Download a copy of the .amw file with the final diagram here: drive.google.c...
For other videos on AMOS, please visit: sites.google.c...
Watching your newest video and studying its PPT on CFA for the confirmatory factorial analysis of my online student engagement scale for my dissertation! You are the best, Dr. Crowson!
I'm glad you are finding them useful! Hang in there on the dissertation. Hoping the best for you and your research. Cheers!
Dear Mike, thank you so much for informative video presentations, highly appreciated. These are GREAT and HELPFUL!!!
Thanks a ton mike! Just cannot express enough gratitude! The ppt is just on-point!! and with references!!!! 🥺🥺🥺Thanks once again!
Hi Vrushali, thank you for your comment. I'm glad this was helpful to you. Best wishes!
I have been finding all your videos extremely helpful and have been sending them to all my fellow students!
Greetings from Germany :)
Thank you so much for your comment! I'm so glad you find my videos helpful. Best wishes!
Just started watching. Thank you! This will be extremely useful for my Dissertation
Hi Mike, Thanks so much for the great vidoes in EFA and CFA. I really enjoyed them and it started me off on utilizing Amos in my research. Keep up the great work! Much appreciated!!!! Having produced lots of introductory statistics videos for my students I know the effort you put for each of the videos....Well done!
Thank you for your kind words, Mitra! Also, I'm so glad you are finding my videos useful. Best wishes!
Thanks, Mike! This was a very helpful video to get me through my SEM PhD class!! Much appreciated!
Just one word perfect.
Thank you so much . very usfull
Very helpful presentation. Thanks.
Thanks a lot, Mike! You've solved a problem!
Thank you again for this very helpful video.
I think you should also show cases where the model fit is not very good and give advices on how to solve these problems. Most of the videos about CFA show loadings > .09, so it is not very helpful when your model has a poor fit.
Thanks !
Hi there, Maximie. I totally understand the issue you have raised. To be honest, there are a number of reasons why a model might not the data well. Unfortunately, it is difficult to efficiently go through every permutation of why a model appears to not fit the data and then solving those problems. Simply put, there's no 'one-size-fits-all' approach to addressing lack of fit because everyone's data and models are different. That's likely why most presentations on CFA don't go down those rabbit-holes.
Nevertheless, here are some thoughts I have for you that may help you think through issues that could produce a lack of fit in your data. The most obvious culprit for lack of fit is that the model you are testing is misspecified. For example, maybe you have specified a one-factor solution, when a two-factor solution would fit the data better. Maybe the covariances among some of the items should be estimated, etc. Often, when model fit is not obtained researchers will go back and reconsider their theory regarding the factor model specification and re-specify the model based on those considerations.
It is also the case that researchers may re-specify their model based on empirical criteria, such as reliance on modification indices or insights drawn from studying the matrix of standardized residuals. I believe I have some information in this presentation on those couple of topics. The biggest drawback of using these procedures is that you can end up overfitting your model to your sample data, so that the revised model fails to generalize in new samples.
Now, let's say that your model is theoretically defensible and it actually should fit your data. Another factor to consider that might reduce fit is whether you have a violation of assumed multivariate normality. One effect of this violation is to inflate your chi-square value, which will also indirectly affect many of the other fit statistics (since many incorporate the chi-square value in their computation). This violation can also have the effect of decreasing standard errors associated with your parameter estimates, and increase Type 1 error risk. To address this question, you would need to evaluate your variables for evidence of multivariate non-normality. That might entail using Mardia's test in the AMOS output, examining your data for outliers (perhaps using Mahalanobis distance-squared), etc.
If you conclude your problem is a violation of multiariate normality, then in AMOS your main option is to use bootstrap procedures (such as Bollen-Stine to generate an empirical p-value for use in judging overall model it) and bootstrapping the standard errors for your parameter estimates. These options are available in AMOS.
If you want to step outside AMOS, there are other possibilities for dealing with non-normality, such as the use of robust fit indices (e.g., Satorra-Bentler) and use of robust standard errors. I would LOVE IT if one day AMOS would incorporate this into the program.
Ok, well that's all I have for you right now. Again, thanks for visiting and your comment. Hopefully this wasn't too long-winded :) Cheers!
Thank you so much for the insightful video!!
Thanks for all the time and efforts that you put on these videos.
Do you have a video that shows how to conduct a CFA using AMOS when you have a variable (continuous carriable) with a single item?
Do you have a video that shows how to conduct a CFA using AMOS when you have a variable (categorical carriable) with a single item?
As far as I know, they don't recommend having a factor with less than 4 items. It comes from the EFA.
You may want to do a regression probably?
Thank you very much for this video. You mentioned around the 31st minute that the RMSEA model fit with 0.10 would be fit for a different model. What model is that? Because mine is actually 0.10. I hope you see my message and reply. Thank you once more for your time and effort to create this video.
Hi, Thank you very much!
Thanks for visiting! Best wishes!
A fantastic tutorial, Mike!
Is there any workaround for missing values when running CFA in AMOS? When I've checked 'Estimate means and intercepts', modification indices, SRMR, and GFI cannot be generated.
Hello,
Great and informative video!
I may have missed this, but why did you include correlated errors?
For this demonstration, I basically added the correlated errors to show you how this can be accomplished. Substantively, one reason a person might include one or more correlated errors is due to the belief that part of the covariation between items is explained by factors that are not included in the model. For example, a common method artifact that might increase the correlation between two items might account for part of the association apart from anything a common factor might produce. Or that same method artifact could help to account for the association between items associated with different factors. Another common reason why folks add correlated errors (which I'm not a huge fan of) is that they look at modification indices and then use them as a basis for adding in parameters that would improve the fit of the model. If the additions make sense that's one thing. But sometimes folks get so wrapped up in trying to improve fit that the additions they make improve fit but may not be defensible from a conceptual or theoretical point-of-view.
cheers!
thank you very much
Just one quick question about your comment about toggling the "estimate means and intercepts". If my training were correct, this option tells Amos that you are going to use non-centered data and that the measurement & structural equations all contain intercepts and that you have several extra parameter matrices to estimate (mean of the latent var, intercepts of x and y). Since this configuration will likely complicate the estimation, I am not sure why it matters for the existence of missingness. The handling of missing is determined by the estimator and your choice of list-wise or pair-wise in the other place of AMOS. Just my humble two cents. :)
Hi there Yanchen, and thanks for your comment. Basically, you do not have the option in AMOS for listwise deletion or pairwise deletion as you have in other programs such as LISREL and MPLUS. If you wanted to use listwise deletion, you'd have to manually delete cases with missing values on your variables. And I cannot think of any other approach for pairwise deletion with AMOS. My understanding is when you have missing data, then a full-information maximum likelihood approach in AMOS is used to estimate model parameters - with that approach taking into account missingness on your variables. You'll see this option available in LISREL, MPLUS, Stata, etc. I'll be honest, I haven't spent a whole lot of time unpacking the theory on how this is all done. However, I do know that the AMOS program will not allow you to run your model if you do not request Estimate Means and Intercepts. Here's a statement from the AMOS user's guide: 'When confronted with missing data, Amos performs estimation by full information maximum likelihood instead of relying on ad-hoc methods like listwise or pairwise deletion, or mean imputation." You might check out the user's guide for more details (the guide also describes imputation-based approaches too): secure.ulster.ac.uk/isd/downloads/media/IBM_SPSS_Amos_User_Guide_v26.pdf
I hope you find this helpful! Best wishes!
@@mikecrowson2462 Thank you so much for your consideration and explanation, Mike!! The key issue I am confused about is that checking the estimate means and intercepts seems irrelevant to how AMOS behave to handle the missingness. You already selected ML as the estimator on the left panel. AMOS will do ML anyway. Here is the AMOS guide "Menu: View→Analysis Properties→Estimation→Estimate means and intercepts
:
When Estimate means and intercepts is checked, means and intercepts are estimated, and you can constrain them. (See To constrain parameters.) If you leave Estimate means and intercepts unchecked, means and intercepts will not be estimated.
"
So the question at hand is model configuration vs missingness handline.
Also, do you mean that AMOS will stop iteration if (missingness is present) AND (the Estimate means and intercepts is NOT checked) = true? If it is, that will make total sense & please excuse my hustle & bustle questions :) thank you for this great video!
@@yanchenzhang5623 Basically, AMOS requires that if there is missing data on any of your variables in the model you MUST click estimate means and intercepts for FIML to be enacted. If you have missing data and only have ML checked but do not have Estimate Means and Intercepts checked, then you will get a message saying the model cannot be estimated until you check that box to estimate means and intercepts. Hope this clears things up. It's weird I know. Seems like they could've programmed AMOS to automatically recognize missing data and then to make the adjustment on its own.
Best wishes and thanks for visiting and your questions :)
@@mikecrowson2462 Perfect! Now that clear things up! Indeed, Mplus and Lisrel is far better option for complex needs, although the UI of AMOS is better. Thank you so much for your time and help on this question, Mike!
Why to draw correlated errors while forming the model? Ain't we supposed to draw them after analyzing Modification Indices?
Hello Mr Mike, I've actually read that we have to do modification in scale of each factors. I mean aff.labilit've been modified among afflability scales. However you did not like that. Which one is true? I'm confused.
Dr. Crowson - Your videos are excellent! My thanks and my compliments. After many hours today unsuccessfully trying to find a solution to the following, I am hoping that you might be able to point me in the right direction. I am running AMOS 28 and am attempting to run a CFA in preparation for running a full SEM. In particular, I want to generate modification indices to try to improve the model fit. When I select Modification indices, however, I receive the error message “Modification indices cannot be computed with incomplete data”. I know that the solution proposed is to run a FIML to estimate means and intercepts, and this is what I want to do, but selecting that option in the Estimation tab has no impact on the above error. I seem to be caught in a pernicious recursive loop :-) Any suggestions you might be able to offer would be most appreciated. Thank you.
Nice
Hi Mike, really thanks for your tutorial, but a simple question. Why you draw correlation between the error of "r" (reversed) variable?
Hi Xin, correlated errors are sometimes used in CFA/measurement models to account for the possibility that the errors are correlated due to some factor (other than those laid out in the model) that accounts for part of the association between two measured variables. Think of it this way, the error/uniqueness for each item is comprised of: variation attributable to the factor, unreliable variation (random factors), and specific variation (which is reliable variation not attributable to the factor the item is loaded onto). One factor that may contribute to the correlations among items (apart from a common factor from your model) is the method by which they are measured. In other words, a common method artifact (such as direction of item wording) could explain part of the association between the two items - apart from them sharing the same common factor. That was basically my reason to include it in the example, although I was mainly focused on showing you it could be done :) One other thing: even if two items do not load onto the same factor, you could still have the association between them accounted for by method factors, such that you might specify correlated errors involving similar methods across scales. In general, it is most ideal to keep your model as parsimonious as possible, but this is a general reason why you might consider including correlated errors. Hope this helps!
@@mikecrowson2462 Get it, really thanks~
Connection of covariance between e12 and e13 is permissible. But connection of covariance among e12 , e13 and e7 are not permissible, I think.
Do we need fix to 1 if the there is single item in the factor?
Thanks for the great video. I have a question for running CFA. The indicators assigned a weight of 1s do not get a significance test (CR values), and the standardized regression weights show no such test as well. Is there a way to get the t-values for the standardized weights? Thank you for the help.
Hi Steven, AMOS does not incorporate the ability to generate z- or t-values for your standardized weights. However, if you are seeking to test whether those weights are significantly different from say 0 (as the null), you can use the bootstrap feature in the program to generate bootstrap standard errors and confidence intervals to perform your tests. Under Analysis properties, you'll go to the Options tab and click Standardized Estimates (this will get you the standardized weights). Then under the Bootstrap tab, click on Perform bootstrap (and select how many resamplings you wish - I generally go for about 2000), and then also click on the confidence interval approach (I often demonstrate the formation of confidence intervals using Bias-Corrected intervals) and set the %CI you are seeking. When you go under Estimates in the output, click on Standardized estimates and then go below and click on Estimates/Bootstrap --> bias corrected percentile method, and then under there you should find the bias-corrected (if that's what you go with) confidence intervals. If you are using a null of 0, then if 0 falls inside the interval (i.e., between the lower and upper bound) you maintain the null, whereas if it falls outside the interval you can reject the null. I give kind of a demonstration of this feature in AMOS in another video and powerpoint presentation on path analysis: ruclips.net/video/9tm4YqTSM6M/видео.html
Hope this helps!
@@mikecrowson2462 Mike, Thanks for your detailed explanation! There seems to be many people having the same problem but no one actually gives explanations as explicit and detailed than yours. You saved my day. Thanks again!
Hi, Thank you for detail explanation! My question is how we calculate SRMR in Amos ?
Hi there. Thanks for visiting and your comment! You know, that was something I'd assumed was no longer available in AMOS. However, with a bit of searching I found this link:
www.ibm.com/support/pages/node/420693
Basically, you need to open the plug-in first. A blank box will open up and when you run the model, the SRMR will show up. You learn something new every day! Cheers!
@@mikecrowson2462 thank you so much for your kind and helpful response!!
@@nuran_EK_PsycofReligion Hi there. I decided to put a video together on this topic, if you are interested.
ruclips.net/video/nYnzI31ohjU/видео.html
Cheers!
Bruv, set up one of those "Buy me a coffee" things so I can buy you a coffe as thanks
Hello. What´s the name of the program did you use to make the graph?
Hi. I just drew it out in Amos. Cheers!
Hi. Tqvm for your tutorial on CFA. May I know where I can find the PPT?
Hi there. You can find it at a link under the video description for the video. Best wishes!
Hi, I have question regarding response format.. Is it a problem if one of the subscales is composed out of binary items?
Hi there. Yes, this can be a problem if you are doing CFA with binary indicators - even if on a single factor. The default estimation approach in AMOS (and most SEM programs) is Maximum likelihood estimation. It assumes multivariate normality of your endogenous variables (which observed indicators in a factor analysis are endogenous in that type of model). A violation of multivariate normality can lead to inflation of the chi-square value, which is used in computation of various fit indices. Thus, you might be at greater risk of rejecting your model. Additionally, the violation can lead to smaller standard errors, leading to greater type 1 probability with respect to your tests of factor loadings. Technically, AMOS does allow for modeling binary endogenous variables, but only through the use of the Bayesian estimation option. Bayesian estimation is discussed in the user's guide (www.csun.edu/itr/downloads/docs/IBM_SPSS_Amos_User_GuideV24.pdf). But it's not terribly intuitive to use when you are used to working with non-Bayesian approaches to SEM. You might consider a different program that will allow a mixture of indicators that are categorical and continuous that collects the correlations into a polyserial correlation matrix which is then used for the analysis along with something like Diagonally Weighted Least Squares (DWLS) estimation. I believe MPLUS, Lavaan (in R), and LISREL have capabilities along those lines. These are more standard approaches for dealing with the issue. But as you can see, this is one of the limitations of AMOS - one that I personally find rather frustrating.
I hope this is helpful to you!
Is it okay if cfa and sem model Indices value are same
When i try to run the model a message saying " the model is probably unidentified. In order to achieve identifibilty, it will probably be neccesary to impose 2 additional constraints"
🥺🥺🥺🥺🥺
Any clue doctor ?
The most probable reason is that you have not set the measurement scale for 2 of your latent factors. Did you fix one loading from each factor to a reference variable to 1? In a three factor model you will need three loadings (one per factor) fixed to 1
@@mikecrowson2462 oh my god,
You saved my life doctor
ThAnk you so much 🙏🙏🙏🙏
How Dr. can you do a video on Iteration limit
Doctor, when we draw a corelation arrow between two error terms like the one you draw , what does that mean?
Hi Ahmed. The correlation between two error terms is based on the assumption that part of the correlation between two items is accounted for by factors that are outside of those specified in the model. A correlation between errors might be included to capture potential method factors (e.g., common scaling, direction of item wording effects, such as items that are worded, in negative direction acquiescence, etc) or other unmodeled factors that may account for why two variables are related. Ordinarily one has some rationale (such as the above) for including correlated errors; although modification indices might suggest adding as well. If the latter is the case, my suggestion is to use those indices sparingly and in a manner that has some logic to it (such as based on the above considerations). I hope this helps!
@@mikecrowson2462 well, thank you doctor for your help.
Yes, my model does not fit due to the RMEA is more than .08 (other GOF indces are fine ) so , the modifcation indces suggested by amos is to correlate some error terms ( 5 correlations made) that made the model fitness. but i do not know how to explain it ( i do not know, shall i just mention that it is as per the suggestion of the sofware?! 🙈)
Anyway thank you again for your help