0
$\begingroup$

I‘m calculating a simple regression with one predictor and one dependent variable. Missings treatment is done with full information maximum likelihood (FIML). Should I do outlier detection, i.e. Mahalanobis distance and leverage points, with complete cases only or is there a need (and a way) to do it with estimated values as well?

Improve this question
$\endgroup$
3
  • 1
    $\begingroup$ FIML does not impute missing values. It simply uses all available data points. Therefore, I'm not sure I understand your question. That is, there would not be any imputed values that could count as "outliers." $\endgroup$
    – Christian Geiser
    Commented Feb 28 at 0:14
  • $\begingroup$ Let‘s assume, my dataset has 100 cases, DV is complete in all cases, but in 10 cases IV is missing. Then I can calculate mahalanobis distances for 90 cases but how do I know that the 10 values of the DV that don‘t have a corresponding IV value behave „outlier-ish“? Just doing an univariate outlier detection? $\endgroup$
    – Madamadam
    Commented Feb 28 at 0:32
  • 1
    $\begingroup$ It doesn't really matter whether they do or not, since you aren't going to use them in the estimation procedure regardless. $\endgroup$
    – jbowman
    Commented Feb 28 at 2:48

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.