Multiple Imputation Analysis with Survey of Consumer Finances

[LarsD]

Hi,

I am trying to run regressions on the Survey of Consumer Finances data, which consists of multiple imputed data (m=5), and estimate correct Standard Errors.

My problem: The data set does not contain the original un-imputed data row. Instead, m=0 in my data setis the first imputed data row (flong format). Unfortunately, when I register my variables, the MI command treats m=0 as if it was the originally un-imputed row of data, and as a result makes some small adjustments that change the data, because it believes that there are some inconsistencies with the data.

My questions:
1. Is it possible to somehow force the MI command to not make these adjustments and just analyze the data without the original un-imputed data row?
2. If Stata's MI command is unable to run regressions on multiply imputed data alone (without the original un-imputed data row), maybe there is another user-written command?

Thanks,
Lars

[LarsD]

I have actually found a way to suppress Stata's attempt to "correct" my data, with the option "esampvaryok". However, running an OLS or QREG model with and without "mi estimate, esampvaryok:" delivers slightly different coefficients. This is puzzling, because I thought the coefficients should be the same.

[Yulia Marchenko (StataCorp)]

The mi command requires that the original data are provided in addition to the imputed data when importing external multiply-imputed data. mi performs various checks to verify that your imported multiply-imputed data are consistent and it needs the original data to perform these checks. You may find the FAQ "How can I use multiply imputed data where original data is not included?" from the UCLA website useful:

http://www.ats.ucla.edu/stat/stata/faq/no_m_eq_0.htm