You can convert them (no data police), but in most cases you should not. You lose information (in almost all cases).

It is a 40-year-old problem Ali..

Dichotomization of the quantitative variables comes with a huge cost. You loose valuable information (variability), forced to work with attenuated correlation coefficients, etc.. In principle, it is not a recommended practice.. Find the articles below for more info.

Unless your data is highly unreliable, that is, there is too much random error in responses to the questionnaire items, don't do it.

p.s. Even if you decide not to dichotomize, you have to recode "the not applicable" category as a missing data. You cannot keep it with the value 6, if you are going to treat your variables as quantitative..  

Article On the Practice of Dichotomizing Quantitative Variables

Article The cost of dichotomization

no

Yes, you can do with your data as per your choice and requirment. However, you will loose information further, as with likert scale you are already lost some information. 

no problem

Ali, you are going to lose vital information (affirmation with frequency) by doing what you proposed. kindly consider the purpose of that question when you were designing the instrument... if 'never' is 'no', how will you classify 'not applicable'.... Honestly, it is not advisable!

I think the consensus is that you can dichotomize the data, but that it would be unwise due to the loss of information. 

Can you tell us why you want to dichotomize? That might help you get better advice.

In my opinion, conversion of scale depends upon what exactly you want to extract from the data. If you want to have just frequency distribution against each behavior,  you may show FD against each scale point and then narrate how many "never" did it. Putting "not applicable" under "never" code won't be a right choice. 

I had similar type of data set and was making comparison of security practices between two type of respondents,  I opted chi square test on complete scale points rather than converting first to dichotomous variables. 

It can be done but depends on the purpose. It is done in Fuzzy setts as well by defuzzifying the data.

Sushil 

Quantitative theorists are going to argue that your data will lose veracity; however, the beauty of a methods and discussion section of a research manuscript is that it serves as a place to account for bias and limitations. On the flip side of the argument, you may find that - if you are utilizing a specific standardized scale that uses a Likert - a dichotomous scoring system may produce stronger (or weaker) correlations between study constructs. If this is the case, you will have to argue that the reliability of the original scale could have been dramatically changed; and, as I stated, quantitative theorists and statisticians will question your methods of data collection. The solution: test your new dichotomous scale over many, many trials.  

I think that is not the best thing to do, because you certainly will lost data. Try to create clusters with your data, and then, apply a statistic.

It depend on your objective. You will lose the information content and you will mix up the scale without direction i.e. Not applicable. 

I have seen researchers do that by collapsing scale with the same direction with the aim to simplify their discussion. They collapsing the scale for strongly agree with agree as one scale and strongly disagree with disagree as another scale. 

Thanks all, much appreciate your responds, 

for whom ask what is my objective or my question, my objective is want to know how many participant engage in this behavior ? 

If I cant convert, so what is the best way to analysis the likert scale and get the percentage of how many student engage and not engage in this behavior ? 

Since I am also working on the behaviors of the students (information security and privacy), I can say one thing for sure that behavior cannot be dichotomous. There is variation in the behavior for which each of the data point (never to very often) tells a different set of students. 

If you treat your data points as a ordinal scale (Likert Scale) you can simply take the mean to examine the data. You can then divide them into different groups, for example based upon their gender, age or any other IV you have in your research. You can make use of t-test, ANOVA, ANCOVA to check difference in behaviors of different student groups.

Another option is to treat them your data points as nominal (categorical) and for each data point calculate the frequencies. Thereafter, you can draw bar or pie chart to show %age of student exhibited a behavior very often, often, sometimes, seldom or never. In order to find if the students exhibiting one type of behavior are different from the exhibiting other type of behavior, compute Chi-square to see significant difference. Further Cramer V can be used to the effect size, if the difference is significant.

Hope it helps you out.

Much of the advice we have been giving you assumes that you want to use the items as either independent or dependent variables. If you simply want to describe the results, then basic graphics such as bar charts would be a useful format.

thank you so much, really appreciate your help... 

What would you all recommend for the following related scenario:

We are preparing to replicate a between-subjects experiment where participants were randomly assigned to read one of 3 versions of a story (1 experimental condition and 2 control conditions) and then asked whether the protagonist in the story "knows" or "only believes" the stated claim (binary response; 3x2 chi square contingency table). The original authors found no significant difference between the experimental condition and the "knowledge" control condition (p = .164) and a large significant difference (d ~= .5) between the experimental condition and the "ignorance" control condition, but they claim that a more recent study suggests that if people are asked the same question using a more scaled response type (e.g., VAS, Likert-type), there may actually be a small real difference (d~= .1) between the experimental condition and the knowledge control that went undetected in their first study (N = 135).

Therefore, we conducted a pretest using a visual analogue scale (VAS; where 0 represents "only believes" and 100 represents "knows") in lieu of the original binary response format. Based on the pretest data though, it appears that participants largely respond quite dichotomously (most responses are either grouped at 0 or 100 in each condition, but less so in the experimental condition).

If this were your study, would you stick to a binary response type (knows/only believes) or try the supposedly more sensitive measure (VAS) to appease the original study authors (and then dichotomize the responses in order to compare the scale/parametric results with the binary/non-parametric results)? Wouldn't highly dichotomized data violate the assumptions of most parametric analyses though (not normally distributed)?

My question is the opposite one. Can one post binary questions and build off the answers a likert scale ?

@Itzhack I don't think it's possible to transform binary data into scale data, given that the distribution is not what respondents think but what you think. It will

*sorry sent before completing the sentence *

Any such effort will create bias in the data.

I read somewhere that researchers calculate standardized scores, however, I am not sure if it will be helpful either.

Sometimes binary variables count things, so that the sum has an obvious meaning.