Hello,
I am new to R and Amelia. I have managed to pull a small Stata
dataset into Amelia, generate the imputations, and then push the
resulting imputed datasets back into Stata for analysis using the
-xtreg- command. So far, so good.
My actual dataset comes from the World Bank. It consists of about 50
variables for 40 countries over a 20 year period. There is a lot of
missingness. There is also a lot of non-normality. However, the
Amelia program guide 1.7.1 suggests that non-normality does not
necessarily mean problems with the imputation model, so I have started
with the variables untransformed.
However, the EM algorithm is generating a lot of error messages
'error: inv(): matrix appears to be singular'. I am not sure how best
to respond to this.
(1) Should I be looking to drop highly correlated variables in the
model? But how high is 'high'? Is there an easy way to identify the
'problem' variable(s)?
(2) Should I be looking to drop variables with a high degree of missingness?
(3) Is the lack of invertibility related to non-normality?
If all three are relevant, which should I prioritize?
Any suggestions/guidance would be most welcome.
Thanks.
Steve
Show replies by date