Dear stata list members,
I try to use mi impute chained to impute missing values in my rather small data set (n=191). Listewise deletion would reduce the dataset to n=148.
I have mostly categorial/dummy variables, but also squared terms of continous variables (means centered) and dummy interactions.
I use the following syntax for imputation:
mi impute chained (regress) pf_alterf84_c pm_alterf84_c pf_alterf84_sq pm_alterf84_sq f_isei_re m_isei_re ///
(logit, augment) p_childf84 f_migbackzp_v2 m_migbackzp_v2 p_migback f_isced_2katv2 m_isced_2katv2 p_edubi m_versicherung_v2 p_versicherung ///
p_gemeindegr2 f_f28_v2 m_f28_v2 p_finanzkat ///
, add(5) rseed(4409) savetrace(extrace, replace) burnin(500) noisily
When I try to run the imputation it doesnt work, meaning that I receive the following error message (or similar ones) :
note: 0.f_migbackzp_v2 != 1 predicts success perfectly
0.f_migbackzp_v2 dropped and 48 obs not used
note: 0.p_childf84 != 1 predicts failure perfectly
0.p_childf84 dropped and 17 obs not used
outcome = m_migbackzp_v2 <= 0 predicts data perfectly
error occurred during imputation of pf_alterf84_c pm_alterf84_c pf_alterf84_sq pm_alterf84_sq f_isei_re
m_isei_re p_childf84 f_migbackzp_v2 m_migbackzp_v2 p_migback f_isced_2katv2 m_isced_2katv2
m_versicherung_v2 p_gemeindegr2 f_f28_v2 m_f28_v2 on m = 1
I used the option noisily to check the outputs and it appears the error messages always occurs when interaction variables are to be imputed (all variables that start with p_) It is normally recommended to include interactions in the imputation model if they are relevant for the analysis model (which they are) so I dont want to discard them, but the imputation model only "works" if I discard all interactions.
I wonder if it is a problem that my interaction terms are not strictly multiplicative, meaning that p_var=1 if (f_var=1 or m_var=1) and p_var=0 if (f_var=0 & m_var=0) and thereby cause multicolinearity? This is because there are rather few observations for which the variable is true and for my purposes it is sufficient that either one is true.
Looking forward to your suggestions.
Jasmin
