8
$\begingroup$

As title, I am thinking of merging both into "missing data", which is to name it as NA in R. Since I don't see it will make much sense (or even any sense), to separate the "don't know" row out and to compare the information with other rows.

Is it OK for me to do so?

$\endgroup$

2 Answers 2

12
$\begingroup$

Well, you should also considered that "don't know" is at least some kind of answer, whereas non-response is a purely missing value. Now, we often allow for "don't know" response in survey just to avoid forcing people to provide a response anyway (which might bias the results). For example, in the National Health and Nutrition Examination Survey, they are coded differently but subsequently discarded from the analysis.

You could try analyzing the data both ways: (1) treating "don't know response" as specific response category and handling all responses set with some kind of multivariate data analysis (e.g. multiple correspondence analysis or multiple factor analysis for mixed data, see the FactoMineR package), and (2) if it doesn't bring any evidence of distortion on items distribution, just merge it with missing values.

For (2), I would also suggest you to check that "don't know" and MV are at least missing at random (MAR), or that they are not specific of one respondents group (e.g. male/female, age class, SES, etc.).

$\endgroup$
2
$\begingroup$

It depends on the type of question/response in your survey. If they are like "I like", "I dislike", "Don't know", chl answers partially to your question.

The first solution is chl's answer. You have to check if "Don't know" doesn't hide anything. You have to analyse separately these values to see if it highlights a specific profile of respondents.

I'm not about imputation but... "Frenchy" software do it for MCA, ... often considering MAR assumption. It supposes that these answers are randomly distributed (you pick randomly another modality of response).

You can also use a more sophisticated approach : if "Like" is at 30% and "Dislike" at 70% you pick an uniform random number distributed on (0,1) and choose "Like" if your number is at or below 0.3. If you pick a number between 0.3 and 1 you choose "Dislike".

A more modern approach is Multiplie imputation (see MI PROC in SAS and mice package in R). Imputation is very efficient... But it can't recreate atypic profiles...

If you're working in educational testing or if you need to compute a score, let me know I will complete this answer about scores estimation.

Ref:

Multiple Imputation for Nonreponse in survey, Rubin (1987). Wiley. mice package: http://cran.r-project.org/web/packages/mice/index.html Survey Methodology, Robert M. Groves, Floyd J. Fowler & al. Wiley.

$\endgroup$

Your Answer

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

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