I am using the mice package to impute some missing values, and it works nicely.

I am facing a tricky strategic question though.

Basically, I am working on predictors of myocardial infarction (at time 3), with all patients having baseline features (e.g. age, gender - at time 1), despite a few missing values.

Some patients have performed also a stress test (at time 2), with specific continous details (eg stress duration), but others haven't. Notably, those without the stress test are not missing at random, as typically a stress test is not performed in those who are very sick, or have electrocardiographic abnormalities). Yet, it would be very important to be able to capture all stress test details, as it generates several prognostically informing variables (eg rate pressure procudct, stress duration, maximum heart rate, and so forth).

What should I do to capture the information associated with stress test features?

A complete case analysis will of course exclude all those without a stress test (roughly 50%).

Is it reasonable to impute with mice the stress features among also those who did not undergo any stress test?

Or should I best create a factor variable such as stress_status (0- no stress, 1-stress with low tolerance, 2-stress with high tolerance, and so forth)?

  • 1
    $\begingroup$ Welcome to CV! Several questions to clarify yours: What do you mean by "baseline" features? Are you working with panel data with risk factors for myocardial infarction (MI) at time1 and information on MI status at time 2? How many variables do you have on the stress test the patients had? Finally--and more importantly--is information on stress test missing at random, or is it systematic? Meaning: Are some patients more likely to be offered a stress test than others? If so, you may want to look into Heckman-type corrections for selectivity rather than imputation methods. $\endgroup$ Mar 8 '16 at 1:04
  • $\begingroup$ Dear marquisdecarabas, I will look into more detail in the Heckman-type corrections. I am adding also in the edited version of my question the specifics you required. Briefly, I have baseline data (time 1), stress test data (time 2 - but can be considered equivalent to time 1), and then events (time 3). Of course missingness is not at random, as patients without the stress test are those who are too frail or have specific ECG abnormalities. Finally, the stress test is very informative, if performed, with 4 or 5 quantitative variables generated, which would really be a pity to lose. $\endgroup$ Mar 8 '16 at 10:01
  • 3
    $\begingroup$ If the goal is to use this in clinical practice then it is not going to help the clinician using it if s/he has to impute the stress test results for people who cannot do the test. It would seem more useful to him/her to have two models: one for frail people, one for the rest who can do the stress test. $\endgroup$
    – mdewey
    Mar 8 '16 at 12:15
  • $\begingroup$ I see your point. My goal would actually use these features for confounder adjustment, and I am afraid that splitting the analysis into two (or more) models might lead to loss of power. $\endgroup$ Mar 8 '16 at 13:02
  • 1
    $\begingroup$ How big is your sample size? What are the proportions of people with MI vs. no MI among those who had a stress test? Among those who did not have a stress test? $\endgroup$ Mar 8 '16 at 15:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.