Background: I'd like to use a 7-point Likert scale for an upcoming survey (focused primarily on opinions). For the most part, the questions will be centered around the "To what extent…?" stem. That is, the answers will range from, e.g., "To a very small extent" to "To a very great extent".
Using the 7-point Likert scale, it is most often recommended to include the "neutral" as the center/middle score. Based on my research thus far, however, I'd like to discourage people from not taking a stance on the subject. That is still debatable though. Now, to ensure I'm in no violation of "knowledge liability" (i.e., respondents must know the answer), it was recommended to include the "N/A" in the question/scale. In my view, "N/A" does not equal the neutral. So, I don't think the center position would be the right location on the spectrum.
My questions:
- If I were to include the neutral in the survey questions, what's its location in the scale?
- How should N/A answers be numerically coded?
If it would be the most-outer left point (code value = 1), it would take the place of the "To a very small extent"… which then has been shifted by 1 score to the right (code value = 2) and thereby isn't the total opposite of the "To a very great extent" (code value = 7).
Best Answer
If you give it a place in the ranking of your likert scale, this is just as arbitrary as making it equivalent to "neutral". Clearly, this is not the way to go.
The cleanest way to handle this type of situation is probably by way of missing data algorithms. Might I recommend that you read "Statistical Analysis with Missing Data" (by Little & Rubin)?
Bottom line, in your case (from what I read above - maybe I'm missing valuable information), I suggest using a multiple imputation method. Imputation amounts to replacing all missing values (N/A) by random (valid) values. This results in a completed dataset that you can use as normal, so you can perform any analysis you would. In multiple imputation, you repeat this a number of times, and then you average the results over your repetitions. In many circumstances, this correctly accounts for the missingness and can be quite efficient. In addition, a formula exists (see the book I mentioned) to correct the variances, so you continue to have valid inference?