| Message/Author |
|
|
| Jan Newman posted on Friday, March 01, 2013 - 10:47 am
|
|
|
For my dissertation, I am trying to run a logistic regression model - about 19 predictors on a single binary dependent variable. My missing data is not missing randomly and I cannot do multiple imputation. It was suggested that I use mplus to avoid listwise deletion.
I got an error message and need help running this analysis or knowing if it is possible.
I got the following warning:
THE MODEL ESTIMATION DID NOT TERMINATE NORMALLY DUE TO AN ILL-CONDITIONED
FISHER INFORMATION MATRIX. CHANGE YOUR MODEL AND/OR STARTING VALUES.
Here is the syntax I ran to get this error message based on recommendations from UCLA's site.
Title:
Diss LR Model - General Recidivism
Data:
FILE IS "C:\Users\jln111\Desktop\recjn.dat";
Variable:
NAMES ARE x1 x2 x3 d1;
USEVARIABLES ARE x1 x2 x3 d1;
categorical = d1;
MISSING ARE ALL (-99);
Analysis:
estimator = ml;
integration = montecarlo;
Model:
d1 on x1 x2 x3;
[x1 x2 x3];
Output:
sampstat standardized;
residual;
Our department's copy of mplus was purchased over a year ago, so I am not sure we can get customer support now. Please help!! |
|
|
|
|
| Try removing INTEGRATION = MONTECARLO; or use INTEGRATION = MONTECARLO (5000); |
|
|
| Jan Newman posted on Friday, March 01, 2013 - 10:11 pm
|
|
|
| Thank you so much, Dr. Muthen for your prompt reply. I will try this next. I read something about using MLR instead of ML as the estimation method. Would that address this issue or is it unrelated? |
|
|
|
|
| I don't think this will change anything but you can try. |
|
|
| Jan Newman posted on Saturday, March 02, 2013 - 1:44 pm
|
|
|
| Thank you, Dr. Muthen. I hope this is a final question. I rechecked my assumptions, and everything works for the other logistic regression assumptions, but there is non-normality that may be affecting my analysis. I have a large (n=1000) longitudinal dataset of a clinical sample, and I am looking at 19 predictors. These factors are accounting for much of the non-normality. Then the MNAR issue is due to different researchers phasing measures in and out and cases are missing for big segments of people. Most predictors are missing less than 5%, but 2 are missing around 18-23%. Does the MNAR and non-normality suggest a different estimation method or other step? I apologize for all of the questions - I am very new to mplus. |
|
|
| Jan Newman posted on Saturday, March 02, 2013 - 2:00 pm
|
|
|
| Also, all predictors are continuous... (I apologize for taking up 2 comment windows). |
|
|
|
|
| I don't think non-normality in the predictors with missing data on the predictors affect the results very much. Missingness that is at most 23% does not seem extreme. You say MNAR, but you may have MAR reasonably well approximated. Note that MAR allows selective missingness (despite its name). |
|
|
| Jan Newman posted on Saturday, March 09, 2013 - 8:15 pm
|
|
|
I was able to run my logistic regression model with MLR and monte carlo integration successfully after rescaling variables due to large variances.
Now, I would like to run the model and examining the two groups. Can I do this with mplus in logistic regression? I have two different groups of offenders. All of my predictors are continuous - only the grouping variable is categorical. When I tried to do it using GROUPING command it said that I needed a knownclass command, but since this isn't a latent structure, I wasn't sure if it would work? Is this in the user's guide and I am missing it? |
|
|
|
|
| MIXTURE KNOWNCLASS is the same as GROUPING. When the classes are based on an observed variable, there is not mixture modeling. See the CLASSES and KNOWNCLASS options. |
|
|
| Back to top |
