Multivariate normality for imputation by linear regression (PROC MI)

[grep2638]

Hello everyone, 

The SAS documentation about PROC MI (on this page) states that multivariate normality is required for imputation by linear regression (implemented in MONOTONE REG or FCS REG statement). 

Is it an assumption for residuals and/or variables ?

Where does this assumption is needed in the regression algorithm described here ?

Thanks in advance

[SAS_Rob]

The multivariate normality is related to the variables themselves--that is, those you include in the specific imputation model

[grep2638]

So, is it incorrect to include class variables in the model ?

[SAS_Rob · Accepted solution]

While it does assume multivariate normality, it is fairly robust to small departures from it such as when you have dummy variables.