Missing imputed values still present after doing*multiple imputation (MICE)

[Niels Henrik Bruun]

Hi
How do one handle missing imputed values still present after doing multiple imputation (MICE)?
It worked when I used (regress) on some scale variables, but then the imputed values were sometimes were quite outside the possible range.
So I tried (intreg) and (truncreg) instead of (regress), but then I missing values after the imputation.

I get the error text:

missing imputed values produced
    This may occur when imputation variables are used as independent variables or when independent variables contain missing values.  You can specify option force if you wish
    to proceed anyway.

I do not have independent variables containing missing values.

With option force I get the MI to work, but then I have some few missing values left.

Any suggestions on how to report this?

Thank you very much

[Weiwen Ng]

In your other post, Maarten Buis did already recommend predictive mean matching, which will definitely keep the scale in its proper range. If the scale is an ordinal question, I suggested ordered logit. If it's actually the sum of ordinal questions, you have the individual questions, and a number of respondents have some questions missing (and thus their entire scale score missing), you can impute ordered logit at the question level, then create the missing scale score via -mi passive-. That would be preferable to imputing the entire scale score.

I'm not certain why truncated and interval regression are producing missing values while regress is not, but I would guess that you didn't set the truncation or censoring points correctly. An example of your actual code would probably help. If you could describe your scale variable in more detail, that might help also.

[Rich Goldstein]

if I remember correctly (possibly a dubious condition), I complained several years ago about -mi- not obeying the upper and lower limits in -intreg-; this was confirmed by tech support and the response from developers was to make any imputed value outside the stated limits a missing value - not what I wanted or expected, but ...

[Niels Henrik Bruun]

Thank you both.
The question is still what to do with missings after imputation.
The error message and the option force makes it clear that the developers at Stata Corp recognize the problem.
So this can happen! What do you do?
Do you drop the imputation?
Do you carry on? if so, do how do report the missings after imputation?
Or do I go back to (regress) getting imputed data which are impossible in real life?

I know, I have to try (pmm). And I will.
I can not easily present data to support my question.
And given the matter of the question data should not be necessary .

Looking forward to hear from you

[Weiwen Ng]

Thank you both.
The question is still what to do with missings after imputation.
The error message and the option force makes it clear that the developers at Stata Corp recognize the problem.
So this can happen! What do you do?
Do you drop the imputation?
Do you carry on? if so, do how do report the missings after imputation?
Or do I go back to (regress) getting imputed data which are impossible in real life?

I know, I have to try (pmm). And I will.
I can not easily present data to support my question.
And given the matter of the question data should not be necessary .

Looking forward to hear from you

I hope I am not coming across as obtuse, but I would do exactly as I said in your other thread and in post #2 here.

Unless you at least describe the scale, as I asked earlier, there is nothing else to say. I agree, we don't need an example of your data. But unless you can at least describe the scale and its possible values, there's nothing to say.

For example, say I were trying to describe the EuroQol-5D scale to a mixed audience. I'd say: the EQ-5D scale measures health-related quality of life. It's composed of 5 ordinal items, with 3 levels. The health states reflected by all combinations of those items are then preference weighted to reflect utility. In principle, the EQ-5D produces a continuous summary score that can take on values from -0.5 to 1 inclusive; 1 reflects full health, 0 reflects death, anything negative reflects health states worse than death.

In this fictitious example, I'd say that my problem is that a bunch of people are missing one or two of the items on the scale, so what do I do? I'd probably tell myself to go impute those items using ordinal logit, then use -mi passive- to construct the scale score.

[Carlo Lazzaro]

Those interested in the subject of Weiwen's clear example (I've experienced the same imputation-related issues that Weiwen describes with the newer EQ-5D-5L questionnaire) can take a look at: https://euroqol.org/

[daniel klein]

The error message and the option force makes it clear that the developers at Stata Corp recognize the problem.
So this can happen! What do you do?

The error message indicates two specific sources of the problem, of which at least the second source is easily fixed. Neither of these two sources seems to be relevant in your case (I would double check, as I have run into this issue before and did indeed have missing values in the independent variables). Anyway, I would take this error message to indicate that there is something wrong with the imputation model that I am using. Therefore, the answers to your question seem to be: Find the specific problem with your imputation model, then fix that problem.

Best
Daniel

[Niels Henrik Bruun]

Thank you all.
What I wanted to know was whether I could carry on with missings after multiple imputation.
The short answer is: I cannot!
I thank you for comments and suggestions.
I will certainly use them in the process of getting no missings after imputation.
I am really grateful for your comments

[Vince Vo]

Thank you all.
What I wanted to know was whether I could carry on with missings after multiple imputation.
The short answer is: I cannot!
I thank you for comments and suggestions.
I will certainly use them in the process of getting no missings after imputation.
I am really grateful for your comments

Though I didn't have the error text as you did, I used the command mdesc to check if there is any missing data after multiple imputation. However, missing data still remained in a larger set of observations. Did you just get rid of missing values after MI?

[Niels Henrik Bruun]

After a couple of starting problems I have no more troubles using multiple imputations.
The key point is carefully formulating the imputing regressions.
And then Maarten Buis 's advice regarding predictive mean matching has been really usefull