|
|
 |
| MNAR Data with Multiple Mediating Var... |
 
|
| Message/Author |
|
|
|
|
Hello, I am looking to estimate a MNAR longitudinal model with two mediating variables and multiple covariates (control variables). My exogenous variable is a categorical (ordinal) latent variable on multiple mediating (one count and one categorical) variables and a endogenous variable that is a average of multiple items into a single count variable (computed outside Mplus).
Variable:
Missing are . ;
AUXILIARY = (M) V4136 V4138 V4140 V4156 V4158 V4160;
ANALYSIS: ESTIMATOR = MLR;
Model: LatentIV BY V1032 V1033 V1034 V1035;
Latentmediator1 BY V3122 V3123 V3124 V3125;
COUNTMEDIATOR2 ON IV;
Latentmediator1 ON IV;
DV ON Latentmediator1;
DV ON COUNTMEDIATOR2;
I have included auxiliary variables to help with the data being MNAR which without covariates such as age, gender, etc. does give me an acceptable model (but could be better). When I run the model with the additional covariates the model fit is awful. I would like to know what I am doing wrong or if there is a much better way to deal with MNAR data then what I am doing.
Also, should the covariates (control variables) be ran only on the DV? I have tried this many different ways and when I add the additional covariates my model fits very poorly and even brings some associations to non-significance when they should be significant. Thank you very much. |
|
|
|
|
This is a difficult analysis to do well due to having a count mediator M as well as a count distal Y. When regressed on the IV, the count mediator M is treated as a count variable, so ok. But when Y is regressed on M, M is treated as a continuous variable which may not be what you intended - you end up with a mixed treatment of M. Unlike categorical and censored variables, there is no natural underlying continuous latent response variable for a count variable. Because of this, there is not a linear regression of M* on IV that is combined with a linear regression of Y on M* - hence there is not a simple indirect effect like a*b. Also, with a count DV you want to use counterfactually-defined effects which is further complicated by having multiple mediators.
I also wonder what fit measure you are referring to because there is no chi-square test of fit for count DVs. |
|
|
|
|
| Hello Dr. Muthen, thank you for your reply. The fit measures I am referring to are particularly CFI and TFI. These measures are reduced when adding any control variables. I do have a dichotomy mediating variable that I could use in place of the count mediator assuming the model would improve. I only have a count DV so that would need to stay the same. Is this an approach I should consider? Also, when adding the control variables (e.g., age, biological sex, race, etc.) I would regress these covariates on the DV only? I tried adding these in previously but not only did it lower the TFI and CFI but several relationships that should be significant were reduced to insignificance. My apologies if I am missing anything that should be obvious. |
|
|
|
|
| Sorry, referring to my last post the fit measures I mean are TLI and CFI. With the addition of my control variables, these are reduced to unacceptable levels and many relationships become insignificant when they should be significant. Without adding the control variables to the model, the TLI and CFI are acceptable. |
|
|
|
|
We need to see your full output to see why you get fit measures with count DVs - please send to Support along with your license number.
A dichotomous mediator is much easier to handle than a count mediator.
Yes, all DVs should typically be regressed on all control variables. |
|
|
| Back to top |
|
|
