If you need to include the don't know category of response in order to get more data points in your estimation, then you could use an ordered probit model to estimate the determinants of the likert scale responses (excluding the don't know cases) then use the model to predict the likert scale response for the don't know cases . Alternatively, you could recode the don't know cases to 0 and all other responses to 1 then estimate  a selection equation (1,0) is the response 1=all others or 0=don't know and then a conditional equation for the likert scale responses. Hope this helps? Cheers, Marc 

Thank you so much Marc for this valuable rapid answer.

best

Abdul

Using 5-point scale, "Agree" should have two points; Strongly Agree and  Agree. Likewise Disagree. Then, neutral which can be interpreted as "I do not know" should have the least weight in the coding irrespective of method  of analysis as suggested by Marc.

In summary, "I do not know" is allowed in 5-point scale   

Here is an article by two experts in survey research that includes a discussion of the use of Don't Know.

FYI If you are planning on combining your LIkert scored items into a scale, then you can treat that as a linear rather than getting into something as complex as pro bit analysis.

Yes, for me you can add "don't know" to a five point Lickert scale.While analyzing  you can treat "don't know" as missing value. If you lose to high a percentage of cases, recode them to 0, If you have other variables check if the "don't know" group is not significantly different from those who answered your question. If there is no difference you are in a slightly better position.

Including "I don't know" leaves a back door for those who cannot make up their mind. I would say that it very much depends on the question content whether you should permit this back door or not. If you ask for opinions it should in most cases not be an issue of knowing/not knowing, unless specific knowledge is necessary for developing this opinion (then the question might not be asked to the respondent at all and a filter could be appropriate). In sum, I'd rather avoid DK in likert scales.

I think the answer is "it depends". On your 5-point scale, is "don't know" a legitimate mid-point? If it is, you can treat the item as linear, with "don't know" as the midpoint. If it isn't you'll need to code the response as missing or do something more complicated. Whether or not it is legitimate will I'm afraid depend on the content of the item and the nature of the scale.

Let's be careful with this. In the analysis, using weight 3 for "I Do Not Know" may result to error in the computation as overall average can fall within the mid-point thereby interpreting the average as I do not know but using the least weight for the item/scale gives better result and reduces human error.

 Example :

    SA     A      N(DK)    D      SD

    3        3           0        2      3

 using weight 5, 4, 3, 2, and 1.

total respondents is 11

Weighted-mean = 34/11 = 3.09(Approximately 3)

the mean response is 3 (undecided).

BUT

 SA     A        D      SD    N(DK)     

  3        3       3        2         0

 using weight 5, 4, 3, 2, and 1.

total respondents is 11

Weighted-mean = 40/11 = 3..6(Approximately 4)

the mean response is 4 (agreed).

Thank you all for you opinions and explanations. Professor Morgan thanks for the Ref.

Best.

Abdul

yes, neutral choice can be considered as "I do not know" 

I disagree with some of the responses here.

A response of 3 is "neutral". This is different from an opt-out response of "don't know".

Including a "don't know" is common, and is a good idea to get honest data. You don't want to force people who don't have an opinion to pick a number.

Usually you just count opt-out answers separately.

So, if your responses were 5, 5, 4, 4, 3, 2, Don'tKnow, Don'tKnow:

You have 8 total responses.

25% "Don't know",

and a median response of 4.

Most of the scales used in social science research are intended to be analyzed as unidimensional.  Consider, for example, the Gaziano-Meyer scales for newspaper credibility; the individual items are assumed to "scale" as a group.  Item non-response means a *survey* non-response; if someone gives a DK/NA to an item, the individual, not the item, is discarded.  Imputing a median score, or something like that, for item non-response is ill-advised.  I suppose the answer depends on the set of scales; do you intend to do the validation yourself, or have the scales already been validated?

See Chapter 5 "To Answer or Not to Answer: Decision Processes Related to Survey Item Nonresponse" by Paul Beatty and Douglass Herrmann in Robert M Groves and colleagues'  seminal work on nonresponse: "Survey Nonresponse" Wiley 2002. Chapter 6 is worth your read, too, "The Causes of No-Opinion Responses to Attitude Measures in Surveys: They Are Rarely What they Appear to Be," by Jon Krosnick. They basically say that unless you are asking a cognitive question that the respondent really may not factually know the answer to (e.g. what is the speed limit on X highway?"), you should NOT provide a nonresponse answer. In other words, if you are asking an opinion or behavioral question then you should force a response. 

If you are concerned that people are going to "guess" an answer and so don't want to force an answer, then be very specific in your stem about how certain you want them to be if they are going to answer the question vs select no response. For example, if you wanted to ask how many calories their breakfast was that day, you can include in the stem that "a round estimate to your best judgement" is ok. Alternatively if you want to be exact, you can say "only answer if you looked up the exact calories." This is discussed in the aforementioned chapters as well.

I would also agree strongly with the above replies that recommend AGAINST using 3 as a "don't know" response. don't know is off the scale and is the same as a skipped question. 

I concur totally with the above.  The Likert scale should be presented as 

SD - D - N -A -SA 

or whatever, with no DK/NA option, and an omitted answer is coded as missing. 

If the scale is valid and reliable in its orginal construction and its home setting then no you cant add additional step. The case when you can do this, is a modifiaction to the sclae with the need to get validated and test its pscychometrics. The original scale has its total score and its way of calculating its meanings.

"Don't know functions as a statistical mean, whereby all the answers will gravitate to the "don't know" center. However, depending on the questions, you can have them leave blank anything they do not know, though this may present other problems. 

Why not? It's time to be a little innovative. Purposive applications of Likert scale can be expanded up to the fringes of innovative ideas too. However, actual survey results, and confusions (if any) should also be correlated and extensively researched with.

www.joe.org/joe/2012april/pdf/JOE_v50_2tt2.pdf

There are three cases: 1, you are conducting a new questionnaire OR 2, you are planning to modify an existing one. In the former case I'd say it is possible, but not required (and quite useless after all), in the latter case I'd say no way.

You can ask the perticipants to mark the median of the scale (let's say 3 on a 5 Likert scale) OR leave it blank and you can later decide to replace missing values with mean (i.e. the mean of all the participants' answer on that question).

I don't see any point in adding an I don't know option, because it is not a part of the scale and it can be considered as missing value, becasue you cannot use it later on.

3RD CASE: However, if you want to ask one specific question and you use only one scale for that (i.e. Your mother's highest degree? 1, She doesn't have any ... 3, Bechelors, 4, Masters ... 6, I don't know) - than it is a considerable option, becasue it is possible that someone does not know the exact answer and excluding this data would be far more useful than using a possibly false one.

I agree with Austin on this question.

Isn't it an error to posit that the mean of responses will equal 3 (neither)?  Missing responses, IMHO, should always be coded as 'missing', not imputed as something they are not, because of the deleterious effects on covariance structures.  A good read on the (possibly) complex nature of Likert scales, and the ways of dealing with them,  is Muthén, B. (1984). A general structural equation model with dichotomous, ordered categorical, and continuous latent variable indicators. Psychometrika, 49(1), 115-132.  Another is Muthén, B., & Kaplan, D. (1985). A comparison of some methodologies for the factor analysis of non-normal Likert variables. British Journal of Mathematical and Statistical Psychology, 33, 171-189. 

Yes, you are right, you cannot replace missing values with the theoretical mean, which is 3 on a Likert scale of 1-5. However, leaving missing values means that in case of a questionnaire, you cannot add up a scale. For example if you are using a Questionnaire that consists of 20 items and you need the total score, which is all the answers added together and you leave out 1 answer, then you cannot calculate the total score for that person, which is not a big deal, because it's just one person, however, if there're more people missing only one value (and the total missing value count is not too hihg), you can still miss a lot of people, because of only one missing value. So in this case you can replace that missing value with the mean score (of all your participants) of that question. As a result, you will be able to add up the scale for everyone and you will have a bigger sample to work on and you didn't manipulate your data, becasue if you replace it with mean, than you (statistically speaking) didn't do anything with your dataset.

Abdulrahman

If it helps I think Andrew, Muayyad and Mark (particularly Mark) are correct in what they say. A DK is a missing but there are conditions and outcomes surrounding it's use. Personally I would steer away from use of a median when applying 5 point scales, particularly if undertaking factor analysis followed by SEM.

How you interpret "DK" in this situation depends on how you have phrased the question. In general, I treat "DK" as missing. If I am asking an "agree-disagree" question, I will generally have only 4 possible answers: "strongly agree, agree (somewhat), disagree (somewhat), strongly disagree." I don't offer a middle category. But I might offer "or do you not have an opinion on this question? -- in which case "no opinion" or "neither agree or disagree" are treated as missing (and I know that the respondent heard or read the initial question and had no opinion). Whether later in the data analysis you want to impute values for the "missing" responses is a separate issue.

Yes