Help with registering imputed variables

[Alex Stead]

Hi all,

I am having trouble with multiple imputation. I have a variable, z, which is an index made up of three components: z1, z2 and z3. I have complete data on the weights of these components (w1, w2 and w3), but not on the components themselves.

Rather than imputing z directly, I impute z1, z2 and z3. Note that when z1 is missing, z2 and z3 may be present, and so on. So I impute some values of the three components. So far, I have no problems.

The problem comes when I try to combine these imputations for imputations of the index z and include this variable in a regression. Here's what I do:

gen z = w1*z1+w2*z2+w3*z3
mi register imputed z
forval i = 1/10{
replace _`i'_z = w1*_`i'_z1+w2*_`i'_z2+w3*_`i'_z3
}

This works fine, too. I get 10 new copies of z with blanks replaced with imputed values. However, when I try this:

mi estimate: regress y x z

I get the following:

(20 values of imputed variable z in m>0 updated to match values in m=0)

I think the variable isn't registered properly in some way, but I can't figure out what to change.

Thanks,

Alex

[Rich Goldstein]

why are you trying to register this as "imputed" rather than, say, "passive"? in fact, why aren't you generating this via "mi passive"?

[Alex Stead]

Hi Rich,

Thanks! I hadn't really looked into passive variables. That's fixed it for me.
or me.

Alex

[Alex Stead]

Update: actually, this hasn't fixed the problem. I generate the passive variable, and now I'm just getting a message:

(passive variable z unregistered because not in m=0)

[daniel klein]

You might have created what Stata calls super-varying variables. Run

mi varying

the start with

help mi set

read under the heading mi register and mi unregister.

Stata sometimes makes things hard if they are usually a bad idea. This does not imply that what you are doing is a bad idea, but it should give you the opportunity to reflect in more detail.

Best
Daniel