I need to impute the missing values of a dataset of medical data in which several variables only make sense if another variable has a specific value. In the questionnaire the data come from they were conditional questions.

Example: one variable is "Have you ever had a stroke?"; and only the observations that have "Yes" here are then asked "At what age did you have the stroke?". Of course, all people that said "No" to stroke have a "missing" at the age question.

I have this problem in 4 out of 32 variables. Those 4 variables have something like 70-80% missing, because most people have not had infarct or stroke and similar (while most other variables have maximally 10% missing). The dataset is large, with 13000 observations. I am not an expert, but it doesn't seem right to impute in those 4 variables all of those missing values that do not even make sense for most of the observations.

What would be a correct approach? I am going to use R, package "mice".


  • $\begingroup$ Wuldn't a model that would need you to impute their values make no sense? It's not clear to me how the problem could reasonably arise $\endgroup$
    – Glen_b
    Mar 1, 2015 at 3:00
  • $\begingroup$ Thanks @Glen_b, maybe I wasn't clear. I am talking about imputing the missing values in a dataset before running a model. It is a very common procedure in statistics, the main alternative to just removing the observations with missing values. My supervisor, an epidemiologist, ordered me to do this, and a professor of statistics told me multiple imputation using "mice" and Predictive Mean Matching was the way to go. They just never bumped into this particular subquestion so aren't 100% sure. $\endgroup$
    – torwart
    Mar 1, 2015 at 10:42
  • 1
    $\begingroup$ Unfortunately, @Glen_b was right: one condition for using multiple imputation is that the values are actually missing, i.e. there is a value but you just did not observe it. Your missing values are thus not missing; they just don't exist and there is thus nothing to impute. I don't have the book here so I cannot give you the pages, but I am pretty sure it was discussed in the beginning of Little, R. J., & Rubin, D. B. (2014). Statistical analysis with missing data. John Wiley & Sons. $\endgroup$ Mar 1, 2015 at 13:04
  • 1
    $\begingroup$ A more general way of stating my problem: how does multiple imputation work for those variables that by their intrinsic nature HAVE to be missing in some observations? For instance, number of pregnancies in male subjects; PSA (prostate marker) level in women; years since a subject stopped smoking in subject that have never smoked, and so on. I have articles and book chapters deeply discussing multiple imputation, but I expected this to be a basic problem; yet, I haven't found it methodologically addressed anywhere so far. $\endgroup$
    – torwart
    Mar 1, 2015 at 15:32
  • 1
    $\begingroup$ you don't impute in those cases, as they aren't missing. If your variable is the explained/dependent/right-hand-side/y-variable people sometimes use Heckman selection models. If they are explanatory/independent/left-hand-side/x-variables you can give them any value you like and add an indicator variable identifying those with systematic missings. This is one of rare situations where that dummy-variable correction is legitimate. See footnote 4 of Allison, P.D. (2002) Missing Data. Thousand Oaks, CA: Sage. $\endgroup$ Mar 1, 2015 at 16:19


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.