re FAQ How can I account for clustering when creating imputations with mi impute?

[Ian Dohoo]

I have a couple of questions:

1. How can I access the 3 data sets used in this FAQ

2.Do you have any insights / suggestions of approaches to use when it is a group-level variable which is missing? (I think the examples in the FAQ are all based on scenarios in which the missing values are in variables which vary within-group.)

Thank you
Ian Dohoo

[Carlo Lazzaro]

Ian, as far as question #1 is concerned, you may want to send a private message (see FAQ) to Yulia Marchenko, who co-authored the reply to the thread you quoted.
Unfortunately, I am not clear with your second question: do you mean a categorical variable like sex or race?
Kind regards,
Carlo

[Ian Dohoo]

Thanks Carlo
I will write to Yulia for the data.
I work in veterinary epidemiology where we very frequently have clustered data. I might be investigating factors that affect disease risk in dairy cows. There will be individual-level predictors such as cow age which will vary between cows within a herd, but there will also be group (herd)-level predictors such as a hard management practice which will be consistent for all cows in the herd. I think the strategies outlined in the FAQ are all appropriate for individual-level missing data, but I do not think they will work if there is a group-level predictor missing. Unfortunately, this is likely to be a common scenario for us (we collect lots of data on herd management practices and some of the responses will be missing). If you (or anyone else) has insights on imputing group-level variables, I would appreciate it.
Thanks
Ian

[Yulia Marchenko (StataCorp)]

You can access the datasets in Stata by typing

. use http://www.stata.com/support/faqs/data1
. use http://www.stata.com/support/faqs/data2
. use http://www.stata.com/support/faqs/data3

We will update the FAQ to include the location of the datasets.

[Carlo Lazzaro]

Ian wrote:

...(we collect lots of data on herd management practices and some of the responses will be missing)...

.
As a first step, I would try to investigate why these data are missing or, put differently, which is their underlying missing mechanism (ignorable or not).
In some instances, if missing is ignorable, listwise deletion may not be evil (http://www.statisticalhorizons.com/l...n-its-not-evil).
If data are missing not at random, I would introduce some research hypotheses about the missing group-level predictors and present different scenarios (sensitivity analysis) in order to test the robustness of your findings: what's the impact on dependent variable if the cows with missing level-group predictor were managed according to (say) type-1 herd management practice (and what if according to type-2 or a mix between type-1 and type-2)?

Kind regards,
Carlo

[Ian Dohoo]

Thanks Carlo ... Ian