acumberl posted on Friday, June 18, 2004 - 8:50 pm
Hi.I have a question. I'm trying to estimate a model that has data from the same time period(the past year) for different groups of people. For instance,we have four courses in the model, but not everyone was in each of those courses for the year because they are taken consecutively.I want to use the missing data option because we do have random missing data, but I also want to specify that people in course 1 should not have data in say course 4 . . . I don't want mplus to impute those values. I was trying to use the pattern command for this, but it seems to eliminate my other variables that are missing at random. Is there a way to have both data that's missing at random and data that is missing with a pattern. If so, how would I specify this? Also, is there a more detailed book on mplus that you would recommend buying?
If you have numeric missing value flag(s), you can use the PATTERN option without specifying any missing value flags and then saving the data. Then in a second analysis using the saved data, you can use the missing value flags. This should get around cases being deleted. There are no books devoted exclusively to Mplus that I know of. We have many examples on the website and many papers that use Mplus that include inputs. Also, our course handouts can be purchased.
acumberl posted on Monday, June 21, 2004 - 8:08 pm
Hi Dr. Muthen. Thanks for your quick response. When you say to save the data, what exactly do you mean by this? I'm confused about what the pattern option is doing if there isn't any missing data specified. Also, we would love to get the course handouts. Can these be purchased on this site?
The PATTERN option rearranges the data. If you use SAVEDATA to save the rearranged data, then you can analyze it in a second run. See the homepage for information about ordering the class handouts.
acumberl posted on Wednesday, June 23, 2004 - 7:41 pm
Great. That makes sense. I sent a fax on Monday to order the course handouts. I hope that it was received. Thanks again.
acumberl posted on Friday, July 02, 2004 - 10:35 pm
Hi Dr. Muthen. I received this error when running a model: ESTIMATED SAMPLE STATISTICS
NO CONVERGENCE IN THE MISSING DATA ESTIMATION OF THE SAMPLE STATISTICS.
THE MISSING DATA EM ALGORITHM FOR THE H1 MODEL HAS NOT CONVERGED WITH RESPECT TO THE LOGLIKELIHOOD FUNCTION. THIS COULD BE DUE TO LOW COVARIANCE COVERAGE OR A NOT SUFFICIENTLY STRICT EM PARAMETER CONVERGENCE CRITERION. CHECK THE COVARIANCE COVERAGE, OR SHARPEN THE EM PARAMETER CONVERGENCE CRITERION, OR RERUN WITHOUT H1 TO OBTAIN H0 PARAMETER ESTIMATES AND STANDARD ERRORS.
I have the coverage criterion set to .00. Could you please help me understand the error and my next step. Is there just too much missing data to use the missing command? Thank you!
bmuthen posted on Friday, July 02, 2004 - 11:02 pm
When the coverage is very low (less than 0.10) for one or more cells in the printed coverage matrix, the computations of the ML estimation can be problematic. See the book by Schafer on the Mplus web site for details. So this indicates that there is just too much missing data. Even if you got convergence, you may not want to rely this heavily on the results because model assumptions play in too much when having so much missing data. You can delete variables that are most offending.
acumberl posted on Saturday, July 03, 2004 - 5:11 pm
Thank you for your quick response. I have tried to delete the variables with the most missing data already. Is there a way to use the pattern command to take care of having too much missing data or is it hopeless. I do have different people expected to have data on some variables and not others. Thanks!
bmuthen posted on Saturday, July 03, 2004 - 5:19 pm
If you have planned missing data, it's a different story because then the coverage is exactly zero and there would be no attempt at estimating the corresponding covariance element. It is the small, non-zero coverage instances that are problematic.
I would like to react on the previous message. You say that with planned missing data there would be no attempt at estimating the corresponding covariance element. How can I specify this in the syntax? (my model does not run now because of too much planned missing data and I am desperately looking for a solution). Thanks in advance.
Hello, we are running a latent growth model defined as:
Missing = all(-99) ; usev = ASIALCD1 ASIALCD2 ASIALCD3 Dgroup1 Dgroup2 ; Analysis: Type = meanstructure missing H1 ; Model: iu su | ASIALCD1@0ASIALCD2@1ASIALCD3@2 ; iu su on Dgroup1 Dgroup2 ; Output: tech4 sampstat residual standardized ;
And we get this warning when we are using two of our four continous variables.
WARNING: THE LATENT VARIABLE COVARIANCE MATRIX (PSI) IS NOT POSITIVE DEFINITE. THIS COULD INDICATE A NEGATIVE /RESIDUAL VARIANCE FOR A LATENT VARIABLE, A CORRELATION GREATER OR EQUAL TO ONE BETWEEN TWO LATENT VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE THAN TWO LATENT VARIABLES.CHECK THE TECH4 OUTPUT FOR MORE INFORMATION. PROBLEM INVOLVING VARIABLE SU.
Our assumption is becuase of the high number of missing values in our dependent variables. Any advise on how to address this will be greatly appreciated.
You should check for a negative variance for su or a correlation of one between su and another growth factor. If you can't see the problem, send your input, data, output, and license number to email@example.com.