How to best code the N/A response of the Likert-type rating scale? Say I have a dataset of people's opinion/"rating" on something, and they have to choose 1 out of 5 possible answers for each question - Very happy, Happy, Neither happy nor unhappy, Unhappy and Very unhappy. In addition, some people don't answer anything - so let's say we mark these as N/A. We can map Very happy to 5, Happy to 4, and so on until Very unhappy is 1. But, how should N/A be mapped? 


*

*Should I regularize it and make it the mean of the possible options (i.e., 3)? What are the disadvantages of this method? 

*Should I make it 0, in which case it might affect some prediction results later? 

*Should I convert the features into one-hot encoding, in which case if there are n columns with these same possible answers, we'll have increased the number of features to 6n? 

*Should I make an extra column for only the N/A as one-hot encoding, so that there are 2 columns for rating - one for "non-N/A" (or valid scores), and one for N/A?

*Should I randomly assign a number from 1 to 5 to the N/As?

 A: Some of the answers here seem more complicated or hi-falutin' than may be needed or indeed justified. For example, in many projects short of say Ph.D. level, getting into imputation may be beyond the time available or the skill level expected. 
Also, the title says "N/A" which often means "not applicable". In the question itself the OP treats N/A as the researcher's coding for blank answers, in effect "no answer". Some of my answer covers N/A as a deliberate and allowed reply to a question. There is a lot of turbulent water between these interpretations. 
It seems to me that the simplest possibilities have not yet been mentioned at all. 


*

*N/A is just another category. A side-effect of that: the scale is no longer ordinal. 

*Leave out the N/As from the data analysis. That is ethically sensitive as well as statistically sensible. Seriously, if I fill in a questionnaire and am given N/A as an option and then use it, that's my right within the survey I undertake. It should be taken that I meant what I said. Suppose that I really don't play golf or use Twitter or whatever it is, and that is why a question on preference for golf clubs or my Twitter attitudes really doesn't apply. I don't want some fool of a researcher analysing the data to presume or assume that I "really" meant something else and that the rest of my answers somehow are informative on what that might be. This also applies if I leave a question unanswered, given scope to do that. 

*There is one alternative that occasionally may make sense, which is to guess that N/A is in some sense equivalent to the neutral category. This has to be argued carefully, but the spirit is, perhaps, that for some questions they may both be flavours of "don't care". 

*Whatever the research strategy, if N/As are at all common, then a really careful study will include some kind of sensitivity analysis and explore different ways of analysing them, and explain how much difference it makes to the results of the study. Researchers often prefer not to do this, partly because it tends to underline how lousy the data are. 
A: Filling in the missing values and then running your analyses would be called "single imputation," regardless of how you did so (e.g., setting them to the mean or setting them to some static or random numbers). Although this would be very simple to implement, it would unfortunately bias your results. If you used the mean, for instance, there would then be more data points near the mean and thus less variability in your data. As a result, your standard errors would be too small and you would become more likely to commit a type I error (false positive). Similarly, excluding people with missing data from the analysis would be called "listwise deletion" and would also bias results.
Instead, you should treat the missing values as missing values; modern statistical programs have a way to denote this (e.g., NA in R or NaN in MATLAB). Then use a missing data analysis technique to run your analysis in the presence of missingness. The two techniques to consider would be "multiple imputation" (MI) and "maximum likelihood" (ML). Both techniques can be combined with most common statistical analyses and will yield unbiased results assuming your data is "missing at random" (MAR) or "missing completely at random" (MCAR). These are complex concepts that you should read more about to fully understand, but know that a red flag (i.e., a sign that your data may not be MAR or MCAR) would be if you think that participants were more likely to not answer your question based on what their answer to that question would have been. For instance, if people who are extremely unhappy were more likely to leave this question blank than to answer it with a low score, this could violate the assumptions of MI and ML.
A: The safest way to manage this issue is Method 4 - create an indicator variable for answered/did_not_answer. 
The advantage of Method 4 is that you retain all the information from your survey respondents. Marking N/A may be qualitatively different from filling out a score from 1 to 5. By treating N/A as an indicator variable, you give yourself the ability use it as a predictor variable without making risky assumptions or imputations beforehand (or, worse, analysing the data and making post hoc 'predictions' about what N/A means). 
Related to the above paragraph, Methods 1, 2, and 5 decrease the informational value in your data. Method 3 would be better for the same information-preserving reasons that Method 4 makes sense, but the results will be harder to interpret, and you will lose the informational value that the 1-5 values are ordinal. 
So, as an example with some Stata pseudo-code, let's say you have a rating scale from 1 to 5, where 'Not applicable' is coded as 6. 
Then:
gen ratingapplicable =.
replace ratingapplicable = 1 if rating != 6
replace ratingapplicable = 0 if rating == 6
label define ratingapplicablelab 1 "Rating applicable" 0 "Rating not applicable" 
label values ratingapplicable ratingapplicablelab

// Then, you can regress some new variable (let's say it's 'bonus')
// on rating and ratingapplicable
regress bonus rating ratingapplicable
// Note that you'll get a coefficient for rating, let's say it's 3045. 
replace rating = 600 if rating ==6
regress bonus rating ratingapplicable
// Note that the coefficient for rating will still be 3045. Hoorah! 
// Note that you'll still *also* have a coefficient for ratingapplicable, 
// which represents the effect (on bonus) of the rating being applicable at all. 

Edit: Method 3 may make a lot of sense if you're not sure that you should treat the scores as ordinal. This would be equivalent in Stata to using i.rating as a covariate, which will treat rating as a factor variable with no specific ordering between Very happy, Happy, Neither happy nor unhappy, Unhappy, Very unhappy, and N/A. 
