Multiple Imputation

[Victor Santawi]

Hi, I am using Stata 13 for analyzing clinical outcomes of some patients. I am trying to do multiple imputation, however I am missing several values from several variables (continuous data). I hope I can use some collected data from the incomplete variables to predict the missing data in those variables.

I hope I am making some sense. Your help is highly appreciated.

Thank you.

[Carlo Lazzaro]

Victor:
welcome to this forum.
If you are planning something like a so called hotdeck imputation (as it would seem from your post), please note that it has been superseeded by other methods, such as -mi-.
Start off with reading the -mi- related entries in Stata .pdf manual, which can also refer you to some substantive references.

[Richard Williams]

Here are some overviews/resources on multiple imputation.

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

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

https://stats.idre.ucla.edu/stata/se...stata_pt1_new/

https://www.stata.com/support/faqs/statistics/#mi

[John Eiler]

Victor:
If you are planning something like a so called hotdeck imputation (as it would seem from your post), please note that it has been superseeded by other methods, such as -mi-.

I used to assume this was the case but as best I can tell there is no direct analog to hotdecking included in mi, and I spent quite a bit of time reading the manual and trying to "trick" mi into doing a hotdeck. If you or someone else knows how to do that I'd love to hear about it.

I will note that mi does include the pmm command (predictive mean matching) which is similar in spirit to a hotdeck and I have found that it works well with the significant caveat that for most of my use cases the speed is excruciatingly slow, and forces me to use other methods.

[Carlo Lazzaro]

Is Victor or John the original poster? Please, clarify. Thanks.
That said, I was probably unclear in my previous reply. The advice was to go -mi- instead of -hotdeck-, as the latter is usually outperformed by -mi- for good theoretical reasons that are explained in https://onlinelibrary.wiley.com/doi/.../9781119013563 (pages 66-70) and https://www.guilford.com/books/Missi.../9781593853938 (page 182-184).
In sum, hot deck imputation works only under the (often hard to find) missing completely at random mechanism.

[John Eiler]

Hi Carlo,

Thanks. Victor was the original poster, I just happened to find this in search yesterday and was curious what you meant.

I tend to use hotdecking more for practical reasons than theoretical reasons (in particular to impute several variables from the same donor and to prevent obtaining predicted recipient values that are much higher or lower than donor values). In such cases I've about gone crazy trying to get plausible values out of mi while hotdecking worked nicely and quickly -- but I'm not trying to give a theoretical defense here to be clear, that's merely my practical experience. And purely practical here, but I've found many part of mi to just be excruciatingly slow, to the point of not being worthwhile.

At risk of starting a lengthy statistical discussion, the point of MCAR being necessary for hotdecking is well taken, but without MCAR it seems to me that you will always rely on some sort of (borderline) heroic assumptions no matter what method you use. But I say that as someone not very well versed in the more complicated statistical aspects of imputation.

-John

[Carlo Lazzaro]

John:
thanks for clarifying that you're not a Victor's alias!
That said, I see your point that, in my opinion, makes practical but not theoretical sense, as hotdeck, handy as it can be, while provides you with a random value, does not give you the within/between variations which are the core of -mi-.
That said, it may well be that in your research field hotdeck is defensible.
As an aside, I would recommend you this old but enlightening article: https://stefvanbuuren.name/publicati...Med%201999.pdf.

[John Eiler]

Thanks Carlo, I'll check that out along with the links in your earlier response. -j