Multiple Imputation

[Rosie Chen]

Hi I haven't used the "ice" command for multiple imputation for a while. Today I tried to do an imputation, however, it gave me an error message "using required". Could anyone take a quick look and let me know what is wrong with my syntax? Many thanks!

ice varA varB varC ///
varD varE polstrimpt, ///
seed (8786998) m(5) saving(C:\Users\IMP, replace)

[Marcos Almeida]

Hello Rosie,

"Ice" is a user-written command (SSC).

Assuming you are using Stata 14, you have - mi - commands available for several kinds of multiple imputation.

[Richard Williams]

I would also encourage you to check out the built-in MI commands. But, if you want to stick with ice, make sure you have the most current version:

. which ice
c:\ado\plus\i\ice.ado
*! version 1.9.7 PR/IW 25oct2014. For history, see end of this file.

[Rosie Chen]

Marcos and Richard, thanks for your prompt responses. Very helpful. Yes, I checked my version of Ice and found it was version 1.0.4 PR 08jun2005. Just updated the version of the command, and it worked! Thanks so much!

If I use MI, what are the advantages compared to Ice?

[Richard Williams]

First off, here is a brief overview of MI:

http://www3.nd.edu/~rwilliam/stats3/MD02.pdf

I am also a big fan of

http://www.ssc.wisc.edu/sscc/pubs/stata_mi_intro.htm

I don't know ice well enough to outline what the advantages are. I do know Stata MI came later, and I don't think Stata Corp would have released something that wasn't at least as good as ice. Stata MI will also have official support and will probably continue to be enhanced.

[Marcos Almeida]

I couldn't agree less than a 100% with Richard's comment in #5.

Here (http://www.stata.com/support/faqs/st...s-ice-and-mim/), Rosie may find a Stata FAQ with the aptly-named title "What is the relation between the official multiple-imputation command, mi, and the user-written ice and mim commands?"

[Clyde Schechter]

Well, I will agree less than 100% with Richard's comment in #5. While in general Stata's official commands are better than the user-written commands they replace, in this particular instance I would say "most of the time, but not always." I have found that -mi impute chained- has a really hard time with -mlogit- imputations and frequently fails. While I always try -mi- commands first, when there is a difficult -mlogit- in the problem, -ice-, at least so far, has always succeeded where -mi- has failed.

[daniel klein]

I second Clyde's point here. Fitting mlogit models is almost always a pain and often not feasible at all. Unfortunately, ologit is not much better. I frequently find myself using pmm (a linear model!) to impute categorical data - it is kind of surprising that the results are usually still quite good when comparing the imputed values to the observed ones. I think that mi's augment option is supposed to solve some of the problems with these models, but there seems to be some undocumented(?) limit to how many observations are augmented.

Anyway, whenever I observe such behavior, I ask myself whether the developers at StataCorp could not get it done or the failure of mi has some substantial meaning. Maybe ice and other software give you answers where there better should not be one.

Best
Daniel

[Clyde Schechter]

I ask myself whether the developers at StataCorp could not get it done or the failure of mi has some substantial meaning. Maybe ice and other software give you answers where there better should not be one.

FWIW, I often entertain the same nagging doubt in this setting.

[Richard Williams]

As luck would have it, a student the other day said she was having trouble getting MI to work with mlogit. I haven't seen the code and output yet so I don't know why. Maybe I will tell her to try ice.

But yes, I have also seen instances where user-written programs quickly give incorrect answers while Stata struggles to find a correct one.

The article linked to by Marcos didn't seem to say that much about differences between MI and ice, other than MI had a few more methods. If, indeed, ice is better for mlogit, or ice gives wrong answers with mlogit, I wish the FAQ would say that.

[Rich Goldstein]

I have also had problems with MI and turned to -ice-; one of the "advantages" of -ice- is the "persist" option; I have spoken with a person at StataCorp (unnamed because I have not requested permission to quote them) and they did not like this option because the user does not know (is not told) how many tries were made; they may consider adding something like it but, if so, will provide info to the user on the number of tries

[Tim Morris]

Very interesting to read the above and Clyde's original remark rings true for me as well. I've several times experienced (or seen others experience) a failure of mi impute, then turned to ice and it has succeeded.

Although I think it's right that people generally use mi impute chained over ice these days, it's not always better. For example, the ice implementation of the matching algorithm in pmm (and lrd – not available in mi impute) is superior to that in mi impute, and PMM is a fairly popular method. See https://bmcmedresmethodol.biomedcent...471-2288-14-75.

I'm interested by the comment

Maybe ice and other software give you answers where there better should not be one.

Are there examples where this concern was justified? I.e. ice produced nonsense when mi impute refused to impute?

Tim

[daniel klein]

Are there examples where this concern was justified? I.e. ice produced nonsense when mi impute refused to impute?

I have never looked into this systematically. Perhaps everything is fine. It just seems strange to me that professional software developers would do a somehow worse job than amateurs - even if these amateurs know as much and probably more about the statistical theory behind.

Best
Daniel

[Rosie Chen]

Very insightful discussions. Thanks to all!