Apologies if this is more of a statistical question.
I am currently dealing with a multiple imputation problem that I am attempting to address in Stata. After imputing, I would like to estimate E(Y,X,Z|C=0) where Y and Z are complete (before imputation) and X and C have partially imputed values.
In Stata, my multiple imputation command is:
and my post-imputation estimation command is:
[CODE]
mi estimate, esampvaryok post: regress Y X Z if C==1
[\CODE]
I use esampvaryok to force estimation. By default Stata returns an error when estimation sample varies across multiply imputed datasets. This will happen here since the conditioning variable, C, is imputed, and the distribution of imputed values varies by imputation. Even with esampvaryok specified, Stata returns the following warning:
The main issue here is that the substantive research question requires C=0 for estimation. The following is an simplified version of my problem. Y denotes an asthma diagnosis (yes/no), X is a binary indicator of patient-reported respiratory difficulty, and X is obese status. The hypothesis is that breathing difficulty may be more apparent in obese children, even without asthma. Thus, in obese populations, respiratory problems may result in misclassification as asthmatic.
My primary concern is the warning statement that Stata gives. I am wondering if anyone can recommend papers that might tackle this specific question?
I am currently dealing with a multiple imputation problem that I am attempting to address in Stata. After imputing, I would like to estimate E(Y,X,Z|C=0) where Y and Z are complete (before imputation) and X and C have partially imputed values.
In Stata, my multiple imputation command is:
Code:
mi impute chained ///
(logit) C ///
(regress) X = Y X1 X2 X3 X4, ///
add(5) augment noisily rseed(123)
and my post-imputation estimation command is:
[CODE]
mi estimate, esampvaryok post: regress Y X Z if C==1
[\CODE]
I use esampvaryok to force estimation. By default Stata returns an error when estimation sample varies across multiply imputed datasets. This will happen here since the conditioning variable, C, is imputed, and the distribution of imputed values varies by imputation. Even with esampvaryok specified, Stata returns the following warning:
Warning: estimation sample varies across imputations; results may be biased. Sample sizes vary between 643 and 680.
My primary concern is the warning statement that Stata gives. I am wondering if anyone can recommend papers that might tackle this specific question?
