I have some missing data for a particular item on a 5-item measure, which is called Attitudes Towards Ageing. Several participants have declined to respond to the item because of the wording of the item, which is "As you get older, are things better or worse than you thought they would be?". Participants then have a forced-choice response of "better" or "worse". These participants have insisted their answer is in fact "neither better nor worse". So, this is an issue with the wording of this item. However, it seems to me that this missingness is not MCAR, MAR, nor MNAR. Q1: Would that be right? (Not responding to this item is unrelated to the actual attitude towards ageing, and is unrelated to any other measure in the study.)

Q2: Are there ways to handle this missing data with multiple imputation? Or does the measure itself need to be discarded from my analyses? At baseline, 8% of items are missing solely for this reason; at followup, 36.6% of items are missing solely for this reason. Thanks

  • 1
    $\begingroup$ It is not clear to me in what way the missing item would enter into your analysis. Missing count in a contingency table? missing component of an overall score? or some other way? // For the future, it seems unwise to design a question that tries to force subjects to express opinions they do not have. $\endgroup$
    – BruceET
    Sep 5 '18 at 4:35
  • $\begingroup$ It would be a missing component of an overall score. $\endgroup$
    – R Chau
    Sep 5 '18 at 10:12
  • $\begingroup$ How many other components in overall score, is that 5? Is this overall score your only measure for the study? Are you doing a paired test on baseline vs followup? How many subjects? // Seems something like 40% of subjects might be involved; that would entail massive (and in my view potentially dangerous) 'imputation'. Trying to get an idea of the total impact of this issue. Can't do that without more details of overall picture. // What if you split data into two parts, those who have missing answ vs others? For missing answ part: ignore item. For others: include item. Analyze separately? $\endgroup$
    – BruceET
    Sep 5 '18 at 16:59

I would say that this is Missing Not At Random because the reason why you do not record the response to this question is directly related to the unobserved answer, namely that the aging attitude has not changed.

Imputation does not seem meaningful because you would impute these missing values with the responses 'better' or 'worse', which you know are not the correct answers to the question.

You could instead do a sensitivity analysis in which you treat the missing data as this new response category, and see how this influences the results.

  • 2
    $\begingroup$ This is the right answer. Another possibility is to just drop the item & analyze the responses to the rest. $\endgroup$ Sep 5 '18 at 20:18

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.