Anonymous posted on Tuesday, October 12, 2004 - 12:35 pm
I am running a CFA and I'm wondering if it is possible to avoid having the path for the first item on each factor equal to 1. I would like to get factor loadings similar to those from an EFA. The MPlus user's guide mentions that there is an option for removing this restriction, but I couldn't find it. Thanks!
Anonymous posted on Tuesday, October 12, 2004 - 1:33 pm
I guess if you use the option in the syntax: Output: Standardized. You get standardized loadings not equal but similar to those estimated in an EFA. But i haven't found yet the possibilty for a "Rotation" (Varimax...) in an CFA. With the @-option in the syntax it should be possible to equal other factor loadings besides the first one to 1.
Re: October 12, 12:35pm -- see the BY option of the MODEL command in the Mplus User's Guide. There is a description of how to free the factor loadings for the first item and fix the factor variances to one.
Re: October 12, 1:33 pm - CFA does not do a rotation because the model is already identified. You can do an EFA in a CFA framework.
Anonymous posted on Wednesday, October 13, 2004 - 8:45 am
I am estimating a 3-factor CFA, any suggestions on how to interpret the factor loadings? Secondly, what goodness of fit statistics should I be looking at?
bmuthen posted on Wednesday, October 13, 2004 - 4:57 pm
Those are broad questions that you would have to turn to the literature for - or come to Day 1 of our upcoming short course. The website has several references related to factor analysis that can help with these types of questions.
I am familiar with other SEM stats programs, but I am learning about Mplus in one of my classes this semester. When looking at factor loadings, I have been used to see them as being less than 1.0, similar to an R2, but in mplus I am getting factor loadings greater than 1.0 (e.g., 2.569), but also some that are definitely less than 1.0 (e.g., .844). How do I interpret these numbers? Thanks for your help.
A factor loading is a regression coefficient. If factor loadings are continuous, they are simple linear regression coefficients and are interpreted as such. They can be greater than one. There is a discussion of this on the LISREL website under Karl's Corner.
If the factor indicators are categorical, then the factor loadings are probit or logistic regression coefficients depending on the estimator used in Mplus.
I measured a factor at two points in time with four indicators, respectively. The factors loadings (but not the indicator intercepts) are invariant over time. Now I want to construct a standardized test for that factor based on that findings. How can I use the factor loadings to construct a regression formula to compute the factor loadings from the indicators (in later studies)? Specifically, how do I account for the different indicator intercepts when building that formula? Or can I ignore them?
Sorry, there's a mistake in my previous posting. Here's th correct version:
I measured a factor at two points in time with four indicators, respectively. The factors loadings (but not the indicator intercepts) are invariant over time. Now I want to construct a standardized test for that factor based on that findings. How can I use the factor loadings to construct a regression formula to compute the factor VALUES from the indicators (in later studies)? Specifically, how do I account for the different indicator intercepts when building that formula? Or can I ignore them?
Boliang Guo posted on Thursday, October 06, 2005 - 3:52 am
hello Michael Schneider, if you wanan compute factor score, why not use the MPlus to get it?fscoefficient in outpust
Hi Boliang, I need the factor scores in studies were SEM is not appropriate (experiments with small samples).
bmuthen posted on Saturday, October 08, 2005 - 11:59 am
The formulas for the regression method of computing factor scores are given in Appendix 11 of the Version 2 Technical Appendices. If you want to compare factor scores across time with respect to their level, I would work from a model with invariant intercepts as well.
Tom Hardie posted on Wednesday, October 12, 2005 - 2:37 pm
I am doing a EFA in a CFA have examined the fscore file. Is there a way to export the subject ID to this file. It appears that several cased were dropped but I can not identify were to insert missing for a factor score for futher analysis. Thanks in advance
I would like to Validate a Questionnaire, I have the data collected from 300 patients, so i want some suggets on statistical analysis i can do, i know i can use factor analysis, but what other test can i do to validate the questionnaire.
If the factor is in a different metric, the factor loadings will be different.
SH posted on Thursday, December 15, 2005 - 10:56 am
mcole posted on Saturday, January 28, 2006 - 7:52 am
In answer to the first question posted above: "I am running a CFA and I'm wondering if it is possible to avoid having the path for the first item on each factor equal to 1. I would like to get factor loadings similar to those from an EFA" add this output statement to your input OUTPUT: STAND;
You will then report the StdYX values in the Model results as your factor loadings.
I am measuring 4 factors (five indicators each) at four time points in the same group of people. How do I evaluate whether there is measurement invariance over time? I assume that measurement invariance means that the factor loadings are invariant over time. Is this correct?
Whether you are assessing measurement invariance across time or groups, the steps are the same. See pages 345-347 of the Mplus User's Guide that is on the website for a brief description of the steps to use for testing measurement invariance.
Standardized factor loadings can be greater than one. There is a discussion of this on the LISREL website under Karl's Corner. If a raw coefficient is negative, it's standardized coefficient will also be negative.
Going back to Linda's posting on May 9, 2005 , quoted here "If factor loadings are continuous, they are simple linear regression coefficients and are interpreted as such....If the factor indicators are categorical, then the factor loadings are probit or logistic regression coefficients depending on the estimator used in Mplus."
Just to clarify - if factor indicators are both categorical and continuous in the same CFA model, does this mean that the obtained loadings are linear regression coefficients for the continuous indicators and probit regression coefficients for ordinal (1-3 scale) indicators using the WLSMV estimator?
I am running a CFA with 20274 observations and 5 factors, and getting acceptable model fit according to CFI and TLI, but RMSEA is little above acceptable 0.10 and WRMR is far from <0.90. Can this be explained? It's my first time I have to do CFA and I'm not sure if this result can be accepted or I have to try another model.
Her is par of the output:
TESTS OF MODEL FIT
Chi-Square Test of Model Fit Value 19748.320* Degrees of Freedom 79** P-Value 0.0000
Chi-Square Test of Model Fit for the Baseline Model Value 271312.327 Degrees of Freedom 10 P-Value 0.0000 CFI 0.928 TLI 0.991
Thank you for your answer. I now did EFA, wich shows, that there are probably no more than 2 factors, but if I run CFA on these 2 factors the fit statistics are even more contradictory. The same happens if I choose 1, 3 or 4 factors. I'm confused...
I suspect that you have large cross-loadings for some of your factor indicators so that when you go from EFA to CFA, this causes the misfit. If the factor indicators were developed to load on only one factor and they load on both, you might want to think about why this is happening.
I'm a bit confused...is there a difference between factor loadings and factor scores? I understand that the obtained loadings in a CFA are regression coefficients (probit or logit) in the case of categorical indicators. But is there a difference between scores and loadings?
If not, then I should just add the following to the output statement (OUTPUT: STAND;) to get loadings similar to what is obtained with an EFA? And use the loadings from the StdYX column?
I have three ordinal-level indicators for a single factor at two levels. At the between level, the stdYX for all three indicators is 1.000. The std and the unstandardized estimates are not all the same. Does this make sense?
TYPE=TWOLEVEL; ESTIMATOR=MLR; %Within% fw BY u1-u3; %Between% fb BY u1-u3;
StdYX for a factor loading is equal to (factor loading * s.d. factor) divided by the square root of (factor loading squared times the factor variance plus the residual variance). See below for a factor loading of .5, a factor variance of 4, and a residual variance of 0:
.5 * 2/ sqrt (.25*4 + 0) = 1
Tracy Witte posted on Thursday, June 05, 2008 - 12:30 pm
Going back to the question about factor loadings being greater than 1, I thought that they were only regression coefficients in models that are NOT congeneric. Otherwise, I thought that they should be less than 1. I have a CFA with 2 factors (some variables are categorical and some are continuous) in which in indicator only loads onto one factor (i.e., it is congeneric). Some of my factor loadings are greater than 1 - is this a heywood case? Are the factor loadings given in the mplus output standardized or unstandardized?
Factor loadings are always coefficients in linear regressions of each y on f's. With more than one factor that are correlated there is no need for loadings to be less than 1 even when the y variance is 1 (as when analyzing correlations); the residual variance can still be positive as required. With CFA, the y variance is typically not 1 and loadings can be big, certainly bigger than 1. Mplus reports unstandardized loadings unless you also request standardized ones. Heywood cases occur when residual variances are negative.
Michael M posted on Friday, January 27, 2012 - 8:25 am
First - thanks for having such a great resource available to us modelers.
I've run a CFA with categorical responses (dichotmomous) using the WLSMV estimator. I've got a standardized factor loading greater than 1 (and from what I understand, this is an indication of a negative residual variance). Is it possible to fix the residual variance to zero for this item?
For some reason, I have it in my head that I can not do this with the WLSMV estimation.
You may have too many factors. Check your factor correlations in TECH4. Some may be very high. The residual variance is not a parameter in the model so you can't fix it. You could do this with the Theta parametrization but I'm betting the problem is with your model.
Michael M posted on Wednesday, February 01, 2012 - 2:25 pm
Thanks Linda - again this is a great resource you provide to all of us.
Lin Gu posted on Wednesday, February 08, 2012 - 9:15 pm
Does the product of two factor loadings in the same column of Lambda provides an estimate of the correlation between the two observed measures involved?
In my CFA model with two latent variables (parent's and children's religiosity) measured by 4 items each, the (unstandardized) loadings of these items are 1.5 to 2.5 times higher in the full model (including moderations of the effect of one latent variable on the other), compared to the model in which only the CFA is specified. Is this problematic? If so, what can I do about it? All items are binary and the loadings are constrained to be the same for parents and children. The specification of the CFA is as follows:
The size of a factor loadings depends on the scale of the factor and the variable. You should consider its significance. You don't omit a factor when a loading is small - you may want to omit the variable.
Linh Nguyen posted on Friday, August 01, 2014 - 12:02 am
Acceptable loadings in a second-order CFA model?
I am running CFA for a second-order construct with 4 first-order factors (measured by 14 observed variables). I am wondering what is an acceptable loading for first-order factors in a second-order CFA model?
The loadings of 4 first-order factors on the second-order factor are: .70, .34, .46, .82 ( all significant at p<.001). Model fit: CMIN/df = 2.31, CFI =.93,RMSEA =.08, SRMR =.08. Should I keep the .34 loading? This first-order factor is measured by 3 observed variables, so I cannot delete any observed variable.
No. Standardized loadings are computed after model estimation using standard deviations from each group. Unless all groups have the same standard deviations, the standardized loadings will not be the same.
Dear MPLUS support, I performed a CFA with 6 indicators to measure one latent variable. I obtained 4 positive loadings, and 2 negative loadings. A good chi-square and good RMSEA. Do you think it's a problem to present a CFA with 2 negative loadings? should I reverse the 2 variables with the negative loadings to have positive loadings in the model? Because I knew that with some negative loadings we have a reliability analysis with terrible indexes. Is it true? I can not understand why negative loadings should be a problem.
I am running structural equation models with a measurement model that contains categorical indicators. I have noticed that when I change which item is the reference variable with the loading set to 1 for each factor, the value of other estimated parameters changes. For example, there are differences in the values for structural path coefficients between the factors and outcome variable, factor variances, and factor correlations when I use different items as the reference variable. Why would this be the case? I would expect the other estimates besides the factor loadings to remain the same regardless of which item is used to scale the factor. If not, is there a recommendation on which item should be fixed to 1? For instance, is using the item that loads highest on the factor as the reference an appropriate choice? The model fit statistics do not differ. Because I plan to do multi-group analysis, I do not intend to use standardized values since the group variances may differ, so I'm trying to determine which output I should be relying on.
The first factor loading for each factor is set to one as the default to set the metric of the factor.
Rich Mohn posted on Wednesday, November 29, 2017 - 7:20 am
I have run an EFA in SPSS (shame on me) and gotten factors where all item loadings are negative for three of four factors - not a problem I don't believe, just need to name them right. When I use the same data set in a CFA (I am actually using two samples, but was having this same issue using the second sample, so decided to make sure it wasn't a sample issue), all items load positively on all four factors. The factor correlation matrices make sense in each analysis, but of course there are negative correlations in the EFA and positive correlation in the CFA, given the reverse item loadings on three of the factors.
This is a known indeterminacy in factor analysis: All loadings on a factor can change signs and the model will fit the data the same (the corresponding factor correlations will also change signs). So just ignore the issue. If you obtain all negative loadings, feel free to change them to positive (with corresponding factor correlation sign change).
Rich Mohn posted on Wednesday, November 29, 2017 - 12:53 pm
I am using indicators that involve different likert scoring ranges (1-5, 1-7, 1-9); is it appropriate to use the Zscores for these indicators rather than the raw scores given these scoring differences?
No need to go to z-scores - just keep them as they are.
A problem would arise only if you want to test measurement invariance across variables with different scoring ranges - but I don't know that z-scores is the right answer in that case. Also, try SEMNET.