multiple imputation for survey data

[Kun Wang]

Dear Statalist:

I'm working on a national survey data. I could use mi command to impute the data, and use jackknife replicate to weight, separately. But I couldn't use the mi svyset command to combine the imputation and weighting together. Here's what I did:
The first try: I svyset the data; use mi svyset flong command to declare; then it prompts error "data not mi set".

The second try: I directly impute the data first; then I use "mi svyset [pw = person_finwt0], jkrw(person_finwt1-person_finwt50, multiplier(.98))" to do the survey set; then it prompts "options not allowed"; then, I delete the "mi" premix, and try "svyset [pw = person_finwt0], jkrw(person_finwt1-person_finwt50, multiplier(.98))", it prompts "no; data are mi set, ......"

Is it because of the Jackknife replication, or did I do something wrong in the above procedure? I'm so confused...

Thanks!!!

[Marcos Almeida]

You may wish to take a look at this text.

[MarcusLBritton]

FYI, if I'm not mistaken, the mi estimate command is not compatible with using jackknife replication to estimate variance (?). In any case, I got the following error from an mi estimate command after successfully using mi svyset without the multiplier() suboption for the jkrw() option (see below):

vce(jackknife) previously set by mi svyset is not allowed with mi estimate

I'd don't really understand why jackknife replication is not allowed for variance estimation with mi estimate, but this post from the old version of Statalist points to some issues:

https://www.stata.com/statalist/arch.../msg00850.html

Judging from some data (i.e., the HINTS 5 Cycle 2 data) that one of my students is working with, the error message you got from mi svyset ("options not allowed") only arises when specifying the jackknife multiplier with the multiplier() suboption to the jkrw option. If I'm not mistaken, in the absence of the multiplier() suboption, Stata assigns the default value of the jackknife multiplier, which is equal to the number of replicate weights minus one divided by the number of replicate weights, i.e., (R-1)/R (see the Stata Survey Data Reference Manual, Release 15, p. 185, which uses the more formal notation: [n_subscript_h - 1]/n_subscript_h, where n_subscript_h is the number of PSUs in stratum h). If that's correct for your data (as it appears to be for my student's data), then you may be able to omit the multiplier() suboption from the mi svyset command and avoid the error message. I have no idea, however, why including the multiplier() suboption works with the non-mi version of svyset but not with mi svyset. Again, the more fundamental issue is apparently that mi estimate does not allow one to use jackknife replicates to estimate variance.