Combining Multiple Imputation with Hierarchical Regression

[Elias Kurta]

I have a data set with partially missing independent variables. Therefore, I would like to impute the data set firstly and then regress it afterwards.

(A) Imputation: I would like to use Multiple Imputation (Stata command: "mi")
(B) Analysis: I would like to use a Hierarchical Regression (Stata command: "nestreg")

Does anybody know, how I can use the "nestreg" command after the Multiple Imputation?

Many thanks!

[daniel klein]

What is the research question that you are trying to answer? nestreg appears to be most useful for model selection. Multiple imputation requires the imputation model and the analysis model to be congenial; in other words: you should know how your analyses model will look like before you impute missing values. From that perspective, the two approaches do not seem to work well together. The problem might not be as severe as long as the imputation model includes at least all variables that are used in the analyses model. However, MI was really designed to obtain unbiased (and consent) point estimates of population parameters of interest, such as regression coefficients, for inference not for model selection.

[Elias Kurta]

Hi Daniel,
many thanks for your quick answer.

My DV is the likelihood to sell the company; my independent variables are for example financial performance, innovation performance etc. Due to the fact, that each respondent answered multiple scenarios I need a hierarchical regression model (it was a vignette study). Do you think, that "hireg" would be a better fit after a Multiple Imputation?

Or would you recommend to use Full Information Maximum Likelihood (FIML) instead of Multiple Imputation?

[daniel klein]

I think you are confused by confusing terminology. When you say hierarchical regression model, you are referring to what some people call multilevel models, other call mixed-effects models, and yet others call random-effects models. This is completely different from how the term is used in nestreg* (and apparently hireg [from SSC?]). The latter refer to a sequence of models, each of which adds one or more predictor variables compared to the previous.

Look into

help mixed
help xt

for starters. Figure out the mode that you want to fit first, then get back to more complex tasks, such as handling missing values.


Edit: I have quickly reviewed the documentation and the term "hierarchical" does not appear anywhere in nestreg's documentation. The term seems to stem from hireg, which indeed appears to be from SSC.