Two questions. 1) I am running a TWOLEVEL EFA and the user's guide pgs 481-483 are useful in that regard. However, my first attempt took 13 hours. I know through attending prior workshops that for multilevel regressions, you have some suggestions for shortening run time by changing the estimator. My model has ordered categorical DVs and so I am using WLSMV. Do you have suggestions for how I might shorten the run time or do you have sample Twolevel EFA input?
2) For a TWOLEVEL CFA, the examples provided in the workshops and user's guide seem to hold the factor loadings at the two levels equivalent. Substantively, what does that mean and statistically, why do it?
1. Please send your input, data, and license number to email@example.com so I can see exactly what you are doing. Computational time will be affected by the number of factor indicators and the range of factor solutions you are requesting.
2. We don't typically hold loadings equal across the two levels - see Example 9.6 in the Version 5 User's Guide.
Anonymous posted on Thursday, October 09, 2008 - 6:24 am
Dear Prof. Muthen,
Hello, I runned twolevel EFA using 14 binary variables with 35088 samples(including missing cases). The estimation stopped during bivariate estimation, without any output. gANALYSISh part of Input I wrote is gANALYSIS: TYPE = TWOLEVEL EFA 1 5 UW 1 5 UB;h I tested the same model with smaller samples (1170 samples), the model terminated normally. So, I think the input program wasnft wrong. Is this problem of my PC spec? My PC spec is OS: Windows XP professional version2002 service pack2 CPU: Intel(R) pentium(R) M processor 1700MHz, 1.69GHz Memory: 1.5GB RAM Could you please give me any hint? What is most likely the cause of this problem and what I could do about it? I would be very thankful for any suggestions. P.S. Bottom part of optimization history(in MS-DOS window)before stopped is below;
BIVARIATE ESTIMATION FOR SOCIAL AND CULTURE Within-level dimensions of integration: 0 Between-level dimensions of integration: 2 Within-level number of integration points: 1 Between-level number of integration points: 49 Total number of integration points: 49
ITER LOGLIKELIHOOD ABS CHANGE REL CHANGE ALGO TIME TOTAL TIME 1 -0.20215129D+05 0.0000000 0.0000000 EM 0.84 0.8
Does anyone know of guidelines/literature on how to interpret a two-factor EFA solution with binary/ordinal dependent variables? That is, what to consider when trying to figure out how many factors are optimal at each level?
Thank you. I didn't realize that some of the traditional EFA criteria for deciding on the number of factors would actually be available in a multilevel context with categorical indicators. I have one more question about your first reply in this thread re: holding the loadings equal across the two levels. If one wanted to test whether the loadings were the same or not on the within and between levels, is it possible to impose this constraint and use a chi-square difference test? Just to clarify, that would be in a multilevel CFA with categorical variables.
You can do a chi-square or loglikelihood difference test if chi-square is not available. In most cases, the factors do not have the same meaning on both levels so testing for equality may not be meaningful.
Simon Denny posted on Wednesday, November 25, 2009 - 12:30 pm
I am trying to run a two-level exploratory factor analysis on 13 dichotomous variables with about 8500 participants with missing data (about 10%). All of the variables have proportions less than 0.16.
I have tried MLR and WLSM but get the warning message for all the combinations of number of factors: FACTOR DETERMINACIES COULD NOT BE COMPUTED ON THE BETWEEN LEVEL FOR EXPLORATORY FACTOR ANALYSIS WITH 1 WITHIN FACTOR(S) AND 2 BETWEEN FACTOR(S). THE MODEL COVARIANCE MATRIX IS NOT POSITIVE DEFINITE.
Is there anything I can do about this? I get an output with Eigenvalues - are these interpretable?
Perhaps you have Heywood cases - negative residual variances on one or both levels. Often, the between level needs only 1 factor. If this doesn't help, please send your input, output, data, and license number to firstname.lastname@example.org.
Mukadder posted on Sunday, March 20, 2011 - 1:19 am
I ran a TWOLEVEL EFA (5 within and 5 between)and determined the factor loadings at the within and between level. The MPlus output showed me that the solution with the same number of factors at each level (5 within and 5 between)is better. And the twolevel CFA fit well. So I designed my further TWOLEVEL structural model input regarding that.
If I am to clarify, I want to test the interactions among these 5 factors at the within level and the same 5 factors at the between level with different factor loadings.
However, in an earlier post it is suggested that sometimes the factors do not have the same meaning on both levels so testing for equality may not be meaningful. Is it impossible to test the same model with same number of latents at both the within and between parts although they have different factor loadings? Thanks in advance.
I am running a two-level EFA with both categorical and continuous variables. If I try to run any model larger than 1 2 within and between the computer runs for days before I stop it. If I run a smaller model, say 1 1 within and between, I get the following message:
STANDARD ERRORS COULD NOT BE COMPUTED. PROBLEM OCCURRED IN EXPLORATORY FACTOR ANALYSIS WITH 1 WITHIN FACTOR(S) AND UNRESTRICTED BETWEEN COVARIANCE.
THIS PROBLEM IS MOST LIKELY CAUSED BY THE RESIDUAL VARIANCE OF P4HOMECM ON THE WITHIN LEVEL CONVERGING TO ZERO.
This message is repeated for all within and all between attempted factors (including unrestricted) and for numerous variables - all dichotomous thus far.
The error message suggests that when you run a model with one within factor and an unrestricted model on between you get a Heywood case, that is, a zero or negative residual on within. This suggests that a factor model is not suitable for these data. It then most likely does not help to increase the number of within factors, or to add a factor structure on between.
I am trying to run a multilevel EFA with continuous and categorical variables measured at the individual level and at the cluster level. All cluster level variables are categorical (ordinal) or binary.
I specified all this in the CATEGORICAL, CLUSTER, and BETWEEN syntax. The individual variables are to be measured both in the between and the within, so they were not specified in the WITHIN.
I also have missing values specified as MISSING ARE all (-9999) ;
I wrote the following command: ANALYSIS: TYPE = TWOLEVEL EFA 1 4 UW 1 4 UB; An error appeared that says: *** ERROR in ANALYSIS command Analysis with between-level categorical variables is not allowed for TYPE=TWOLEVEL with estimators ULSMV, WLS, WLSM and WLSMV.
So I tried using ESTIMATOR = MLR ; and it appears that the input is read normally but I get no results. It also takes around two minutes to run.
Could you give me any advice on what may be going wrong?
I'm trying to do an EFA on a dataset with students nested in classes. All items were measured on a 5-point ordinal scale Because we are interested in 1 specific minority group and only include them in the analyses, the cluster size ranges from 1 to 14 (5.3 on average) level-1 N=164, level-2 N=31 ICCs for the 19 items range from .015 to .149
I'm not all that interested in the factor structure at the class level, just in accounting for the clustered nature of the data
When I do a 2-level EFA [Analysis: Type = TWOLEVEL EFA 2 4 UW 1 4 UB ;]
I get the error message:
STANDARD ERRORS COULD NOT BE COMPUTED. PROBLEM OCCURRED IN EXPLORATORY FACTOR ANALYSIS WITH 2 WITHIN FACTOR(S) AND 1 BETWEEN FACTOR(S).
and the same error message for chi-square. The error messages are repeated for each combination of within & between factors Is this because of the low number of level-1 units by cluster? Or is something else wrong?
Maybe it's good to mention that I haven't defined the items as categorical because I've read WLSMV doesn't perform well with N<200
I first thought I could just run [Type = COMPLEX EFA 2 4;]
But in the Mplus course video 7 I heard you say that the assumption behind this would be that the factor-structure is the same at the within and the between level and that this is rarely the case
Evelyn posted on Friday, October 28, 2011 - 4:21 am
Just to add: the 19 items refer to 'parenting style'
I try to run an EFA for 19 items (500 participants). There are two queries:
1) The observation group is 499 in stead of 500. I checked the file & it is normal.
TITLE: efa bull DATA: FILE IS C:\Users\kwass004\Desktop\efa.bul.dat; VARIABLE: NAMES ARE B1 - B19 w org; USEVARIABLES = B1-B19; MISSING = ALL (999); CLUSTER = org; ANALYSIS: TYPE = TWOLEVEL EFA 1 4 UW 1 4 UB;
*** WARNING Data set contains unknown or missing values for GROUPING,PATTERN, COHORT, CLUSTER and/or STRATIFICATION variables. Number of cases with unknown or missing values: 1 1 WARNING(S) FOUND IN THE INPUT INSTRUCTIONS
SUMMARY OF ANALYSIS Number of groups 1 Number of observations 499
2) Some of the result displayed below:
CHI-SQUARE TEST COULD NOT BE COMPUTED. PROBLEM OCCURRED IN EXPLORATORY FACTOR ANALYSIS WITH 4 WITHIN FACTOR(S) AND 1 BETWEEN FACTOR(S). STANDARD ERRORS COULD NOT BE COMPUTED. PROBLEM OCCURRED IN EXPLORATORY FACTOR ANALYSIS WITH UNRESTRICTED WITHIN COVARIANCE AND 1 BETWEEN FACTOR(S). NO CONVERGENCE. NUMBER OF ITERATIONS EXCEEDED.
Can you please assist me in sorting out these two queries? Thank you.
I am attempting a two-level EFA model with the Oblimin rotation and am wondering if the default loadings provided by MPlus are in an unstandardized or standardized metric?
A few of the loadings on various solutions (e.g., Item 5 on Factor 2) are slightly greater than 1.00 so it raised the question in my mind. I could not find the answer on the User's Guide. Can you please confirm whether the loadings are standardized by default?
VARIABLE: NAMES ARE item1-item30 clusterid; USEVAR ARE item1-item30; CATEGORICAL ARE item1-item30; CLUSTER IS clusterid; ANALYSIS: TYPE = TWOLEVEL EFA 1 5 UB;
I had a question. I've been attempting to estimate a multilevel EFA with 22 items and 1-5 factors at each level. I have found that when estimating an unrestricted structure at either the within or between, the performance of the EFA works well at the other level (i.e., the one that is not unrestricted). However, attempting to estimate a factor structure at each level results in Heywood cases or failure to converge for many of the models, even those that work well when estimated using an unrestricted structure at the other level. I was wondering whether this was a known issue or whether there was literature associated with why this would occur. I ask here because my assumption was that the structures were largely estimated separately at each level after the latent decomposition of the indicators. Thanks for any help you might offer.
Thanks for your reply, Bengt. I was able to find a little, at least that misspecification in one level can affect parameter estimates in the other (e.g., Yuan & Bentler, 2007), but although recognized as a potential issue it doesn't seem to be have been systematically investigated, or at least not widely. I've run a lot of these models and only recently has it been large enough of an issue that it caught my attention.
Yuan, K. H., & Bentler, P. M. (2007). 3. Multilevel Covariance Structure Analysis by Fitting Multiple Single-Level Models. Sociological methodology, 37(1), 53-82.