| Message/Author |
|
|
|
|
Dear Dr. Muthen,
I am self-learning Mplus. I would like to use path analysis with binary mediator and continuous outcome. The missing pattern for outcome variable is not missing at random. I'd like to use multiple imputation first and then do the path analysis.
However, I do not know how to combine path analysis with multiple imputation in syntax. Could you please give me some example? thank you.
Thank you!
This is a part of the command for path analysis without imputation.
ANALYSIS: BOOTSTRAP = 1000;
MODEL:
acrfreg2 ON d_wk2 d_wk3 d_wk4
age female married income_h bd001 ja001_2
chnsg cd_w1 srh_1 srh_2 acrfreg;
cd_w2 ON d_wk2 d_wk3 d_wk4
age female married income_h bd001 ja001_2
chnsg cd_w1 srh_1 srh_2 h1kcntf h1fcany
acrfreg2 acrfreg;
MODEL INDIRECT:
cd_w2 IND d_wk2;
cd_w2 IND d_wk3;
cd_w2 IND d_wk4;
OUTPUT:
standardized CINTERVAL (BCBOOTSTRAP); |
|
|
|
|
There is no need for multiple imputation - no advantage. Just use ML and its default MAR method for handling missing data.
Also, you need a counterfactual approach to estimate indirect effects with a binary mediator - see
http://www.statmodel.com/Mediation.shtml |
|
|
|
|
Thank you so much!
I still have another question.
I am now using WLSMV. I am not sure either WLSMV or ML is better for my situation. I think both of them could handle missing values of DV.
And my continuous DV is a little bit right skewness. So do you think which estimator is better? |
|
|
|
|
| I would use MLR. |
|
|
|
|
| Thank you again! |
|
|
|
|
Dear Dr. Muthen,
I found the MLR or ML with bootstrap doesn't give some model fit statistics as WLSMV. So,
1. how to convince the model is fit well?
2. Could I use MLR or ML with bootstrap to get coefficients and SEs, because of I have missing data for my outcome variable. Then I also use WLSMV to show there are good CFI and RMSEA.
3. I am not sure what is the difference between Type=Complex and Type=general, if I'd like do a path analysis with national survey data.
Thank you! |
|
|
|
|
1. You can use TECH10 or you can use "neighboring models", that is, a less restrictive model where you check if its extra parameters are significant.
2. No, you have to stick to one estimator.
3. Read the intro to UG chapter 9 and the references given there. |
|
|
|
|
Thank you for so much help!
1. TECH10 cannot be printed because of having covariates in my model.
2. My mediator is binary, but my outcome variable is continuous. So could I calculate other fit indices?
3. I know ML is better to handle missing values than WLSMV.
But if the results didn't change much. Could I use WLSMV instead in order to get fit indices.
I have 16% of missingness for the continuous outcome variable.
Best, |
|
|
|
|
2. Try "neighboring models" that are less restrictive as I suggested earlier.
3. You can mention the WLSMV fit but you can't fully rely on it.
You can also try the Bayes estimator which gives fit info (but you need to study our Bayes videos first). |
|
|
| Sara Babad posted on Monday, April 13, 2020 - 1:07 pm
|
|
|
Hi Drs. Muthen,
I am running a path analysis with a dichotomous dependent variable. The dependent variable is created using a define command. The variables used to create the dependent variable are missing some data so the dependent variable is missing daya. I know that mplus automatically imputes missing data for dependent variables.
When I look at the output, the number of observations is consistent with the full sample size. However, when I look at the frequencies in each category of the dependent variable, the sample size indicates that the missing data is still missing (1/3 smaller than full sample).
SUMMARY OF ANALYSIS
Number of groups 1
Number of observations 292
UNIVARIATE PROPORTIONS AND COUNTS FOR CATEGORICAL VARIABLES
RESIL
Category 1 0.645 111.000
Category 2 0.355 61.000
My question is - Does Mplus output only reflect the values of observed outcome variables prior to multiple imputation? In which case, this "discrepency" is to be expected?
Or, is there something wrong with how I am defining the outcome variable that is not allowing for imputation to occur?
Thank you in advance for your help! |
|
|
|
|
It's not clear if you are doing a Multiple Imputation analysis or just doing a regular analysis, hoping that "FIML" (ML using all available data)will help you. In the latter case, FIML doesn't help when you have a single DV. As an aside, FIML does not impute missing data.
If this doesn't help, we need to see your full output - send to Support along with your license number. |
|
|
| Sara Babad posted on Saturday, April 25, 2020 - 8:19 pm
|
|
|
| Thank you for clarifying, Dr. Muthen! This explanation helps a lot! |
|
|
| Back to top |
