Mixed-models and Multiple Imputation

[Ahinoa]

Hello, 

I am running into a problem that I cannot seem to figure out. I am fitting a mixed-models logistic regression using PROC GLIMMIX using complex survey data from the Behavioral Risk Factor Surveillance Survey. Because I have variables that are missing more than 10% of data, I am using PROC MI to imput these missing values and will be using PROC MIANALYZE to pool the estimates from PROC GLIMMIX. 

The problem that I am having is that I do not know who to pool the LR test estimates as well as the AIC to assess model fit. Thus, I am wondering if there is a macro that can be used to pool these estimates or whether there is some code that I can add to my syntax to have this performed?

I would really appreciate any comments/suggestions you have

[ballardw]

Are any of the "missing" values you are imputing for the results of questions that were intentionally skipped for respondents? The BRFSS has moderately complex flow through the survey and intentionally does not ask all respondents all questions. Examples: Men are not asked questions related pregnancy, Non-diabetics are not asked questions about diabetes management, people who have never smoked are not asked questions about current smoking frequency and amounts.

If you "impute" a value for any of these intentionally skipped items then you will need to be prepared to defend that approach.

I am also not sure how well Glimmix works with complex weights such as the stratified cluster weighting scheme in BRFSS though a quick search will find lots of examples where weights are modified to use BRFSS data.

The weights may well have an effect on the Imputation.

[Ahinoa]

Hi, thank you for your quick reply!

No the missing values are not coming from the skipped pattern of BRFSS. These are questions that for their nature (e.g., income) people tend not to answer during a survey.

As for the survey weights, I scaled the weights following the approach outlined by Asparouhov (2006) to make sure they are appropriate for their use if MLM.

The issue that I am having is that PROC MIANALYZE does not pool LR statistics nor AIC values, which are needed to determine overall model fit. Thus, I am wondering what the best approach to assess model fit is when working with imputed data.