| Message/Author |
|
|
|
|
Hi,
I have complex survey data that has clustering, weights and stratification. I'd also like to do multiple imputation but I keep getting the error:
TYPE=COMPLEX is not currently available with multiple imputation with the
DATA IMPUTATION command.
What should I do instead?
Thank you! |
|
|
|
|
I would recommend that you do not use multiple imputation. If you have missing data on covariates you can convert these to dependent variables (usually by mentioning their variances) so that the missing data is modeled through the model. This would work well in almost all cases.
If you have to do MI and the above alternative is not an option you can use 3-level imputation where the weight variable or a log of it is added as a variable (so information from that variable can be used) during the imputation process, in addition to any other variables used for constructing the weights. |
|
|
|
|
Thank you for this response.
Unfortunately, I think I do need to do MI since complete case analysis would reduce my sample by 30%.
I am not sure I understand your 3-level imputation option but will look into it - or if you know of any resources I should consult, please direct me.
Alternatively, I might consider doing MI in Stata and then importing to MPlus for: 1) confirmatory factor analysis; 2)logistic regression; and 3) a fully longitudinal mediation analysis with a binary IV, continuous M, and a binary DV. Is that something you think might work?
Thank you,
Allison |
|
|
|
|
The suggestion
"If you have missing data on covariates you can convert these to dependent variables (usually by mentioning their variances) so that the missing data is modeled through the model. "
does not imply using complete case analysis - quite the contrary - it uses all available information ("FIML" or ML under MAR).
With MI you have fewer options down the line (in subsequent analyses). |
|
|
|
|
Ok thank you. Do you please mind addressing the question below? I'm not sure if it's possible.
If I do MI in Stata and then import the values into MPlus for factor analysis and mediation with my complex survey data, would that be possible? Is it just that MPlus won't do the imputation part?
Thank you again! |
|
|
|
|
Q1: Yes. Use Data: Type = Imputation.
Q2: Mplus does imputation. But analysis with imputed data has fewer bells and whistles (in any software); that's why I propose "FIML". |
|
|
|
|
Thank you. I will try to move forward with FIML then. But I do have a question about FIML.
I have a variable that has missingness at the item level and at the scale level - and I'm not sure how to handle this with FIML.
My knowledge score is comprised of 15 dichotomous items. I would like to use the average score as a predictor variable in my logistic regression. There is some missingness on these items - and also some people who didn't answer any of the 15 items - so, missingness at the scale level.
I was thinking I'd follow the guidance set by Mazza, Enders and Ruehlman (2015) to set the scale as missing if any item is missing and then also add auxiliary variables. Is this what you'd recommend?
They provide some code in their article and I just wanted to be sure I understood it... Are the variables after the first semicolon the auxiliary variables? What does the word "with" invoke?
MODEL:
interf on txgrp female age severity depress;
sever2 sever3 dep2-dep6 interf2-interf6 with
txgrp female age interf severity depress
sever2 sever3 dep2-dep6 interf2-interf6; |
|
|
|
|
Why use the average score and not the factor with its indicators instead?
The WITH statements say that those variables are allowed to covary. This is done for x variables in order to bring them into the model to avoid deleting subjects with missing on one of the x's. |
|
|
| Back to top |
