Are those likert variables are item scores or a composite of some item scores? If you use an item score directly you should treat it as ordinal. But if you are taking a composite of multiple items then you could be using it as continuous.

Logistic regression can handle any kind of independent variables including ordinal and nominal. When the dependent variable is ordinal, then you should consider multinomial logistic regression.

I attached a similar question link.

https://www.researchgate.net/post/Is_a_Likert-type_scale_ordinal_or_interval_data

I agree with Mehmet 

I agree that you need to think about whether you have separate variables that use Likert scoring, or whether you have scales that would result from combining those variables.If you can combine variables to create scales, then that would allow you to treat those composites continuous variables.

Also, I disagree with the statement that you can use ordinal variables as predictors in standard logistic regression, which assumes interval-level predictors. 

David Logistic regression accepts any type of variables as predictors. What is wrong with ordinal variables?

You can enter any variable you want into any regression equation, but the interpretation will all be how the dependent variables responds to a one-unit change in the independent variable. When you have an interval level variable, that one-unit change means the same thing across all the scores on that variable, but when you have an ordinal variable, there is no way to assume that a one-unit change means the same thing at each level on the variable. 

OK. So your concerns are about interpretation.

David, I should add more. How will the interpretation differ for ordinal and nominal predictors? Will ever somebody interpret a unit increment in an ordinal variable or will she interpret it against a reference level? I guess nobody will interpret a unit increase in an ordinal variable.

Disregarding nominal and ordinal predictors because of interpretation dangers seems not appropriate to me.

With ordinal variables, the distance between scores of 1 & 2 is not the same as the distance between scores of 2 &3, etc. So, there is no meaningful way to interpret the meaning of a "one unit changes."

Putting a mis-specified variable into a model will not improve that model, so strictly ordinal variables should not be entered into a regression.

Binary predictors are another matter, since there is only one difference in levels, hence so-called "dummy variables" are quite common in regression.

I have no problem with interpretation danger in ordinal variables in general. My point is different. Potential interpretation danger should not refrain one from including an ordinal variable into a logistic regression.

Remember we have a logistic regression in consideration and we are going to have odds ratios.Those odds ratios we get for an ordinal variable would be ones with respect to a reference category of that ordinal variable, would not be like an odds ratio for a continuous variable and would be very meaningful, because they will compare odds with respect to that reference category. There would not be any danger of interpreting "one unit increment" because that ordinal variable is not treated as continuous. Nobody will say that when the ordinal variable is increased  "one unit"  then the dependent variable would change this much etc. When including an ordinal variable one should dummy-code it, not just use it as it is. If that ordinal variable was entered as a continuous variable and treated as continuous then there would have been an interpretation danger. 

The same is true for nominal variables.

No, the use of an odds ratio for an ordinal variable in a logistic regression does not make the results any more meaningful than they would be for Ordinary Least Square.

And this is not a matter of interpretation -- the linearity of the independent variable is a fundamental assumption for all forms of regression.

Consider Gender as a predictor with three categories: Male, Female, Other/No answer. Score it 1, 2, 3 and then ask: Is "Other" really three times "Male"? Of course not, and that is not just a matter of interpretation, it is how the variable is defined.

In that case, you need to break the three-category predictor into two dummy variables, but you cannot use it as a nominal variable.

Think of it this way: would you put the respondent's ID into a regression as a nominal variable? From your point of view, why not, since anything and everything might be a predictor, not matter how it is scored.

sir i have separate variables which which are formed using likert scale. now tell whether these should be treated as continuous variables or not? 

David, thanks for the insights. I very much appreciate. But I still have some questions as natural consequences.

I agree that linearity is an important assumption. So let's define it first of all. Could you give your words about linearity? How do you check linearity for nominal/ordinal variables or for continuous ones? As far as I understand  you argue that linearity assumption can not be checked (or is never true) for ordinal variables. I wonder how you check it for nominal variables.

For Gender and ID examples, I still have problems with your sentences. We do not include Gender as a variable whose only values are 1,2,3. Instead we introduce dummy variables. I am not defending including Gender (or ID) as a variable with values 1,2,3.

Sarita, if you have combined some item scores and found your variables in this way, then you could treat them as continuous. If your variables are just item scores, then you can not.

Sarita, the argument by David is the strict doctrinaire approach, playing by all the rules. The question is whether you can relax the rules. That depends on your data. Before you decide whether you can assume that the ordinal variables can be considered to be approximately linear, test this out. Plot out the responses your sample gave you and also subject them to empirical tests of linearity and/or normality. How did respondents, in fact, interpret the "ordinal" questions? That matters more than how they were supposed to interpret them! In some cultures, people tend to interpret ordinal questions as if they were on a straight line and in others they don't. Conduct empirical tests of the actual distribution of the responses (linearity, normality, skewness, kurtosis, etc...). Were they evenly distributed across all levels? Did they tend to clump into two or three groups? Let  the distribution of your data drive your practical decision about how to treat what where, in theory, ordinal questions. Be very transparent and report each of the steps, with results of tests of linearity etc... (Including p-values) listed. Your decision will then be seen as objective and empirical, not merely based on which school of thought you subscribe to. Hope this helps.

http://www.math.chalmers.se/~wermuth/pdfs/86-95/CoxWer94_Tests_of_linearity.pdf

http://www.statisticalassociates.com/assumptions.pdf

A normal distribution on an ordinal variables will not address the "unequal intervals" problem. A better approach would be to use a "test of linearity" across the set of categories. In SPSS, this test is easily accessible in ANOVA.

David, could you be more specific about the linearity test in SPSS?

It basically treats the values on a variable as a series of ordered dummies and attempts to fit a straight line through the means.

David do you mean contrasts in ANOVA? I do not get what you mean. What are ordered dummies?

Also in your previous post, you said "A normal distribution on an ordinal variables will not address the "unequal intervals" problem". I could not link this sentence to binary logistic regression/ANOVA.

With regard to the normal distribution, Anthony Waddimba recommended checking that for the ordinal variable, but I do not think that would add anything to the problem of using it as a predictor in a linear regression.

Contrasts in ANOVA are a way to order variables, but this test is designed to determine where you need the full set of predictor categories (non-linear effects) or whether they can be replaced by a single predictor variable (a linear effect),

I am still lacking the answers David, it is too cryptic.

Think of it as a form of "post-hoc" test in ANOVA, where you have a multi-category independent variable with an effect that is significant overall, so you want to examine the nature of that effect. For example, there might be just one category that is statistically different from all the others, or the (in this case) the category means might show a linear trend that could be captured in a single effect.

If this were a regression, each of your dummy variables would be significant and they would each increase in magnitude by about same effect size, indicating that each step you took from the omitted category would add about the same amount. If you got that pattern, then you would be justified in treating the original variable as linear.

There has been a discussion of this topic on ResarchGate:

https://www.researchgate.net/post/What_are_possible_post-hoc_tests_in_Kruskal-Wallis_and_Friedman_tests

 Thank you so much everyone for facilitate your valuable time. Definately your suggestions would help me a lot and i would keep them in mind in futher action. 

Sarita, what did you eventually use? I am trying to do a very similar analysis

Simply treat the likert scale variable as ordinal but ensure such order; and pay attention to the interpretation based on the referent on the scale.

As a statistician I completely agree with the all arguments of Mehmet Sinan Iyisoy. The argument is reasonable.

TREATING LIKERT SCALE IN LOGISTIC REGRESSION

No need to debate whether Likefrt scale is quantitative or qualitative. The researcher who employed the Likert scale should have defined or classified it prior to data collection during research panning stage. Therefore, when the data has been collected, just follow relevant procedure for testing.

1. QUANTITATIVE

If the Likert scale (1,2,3,4,5) is treated as quantitative, i.e. scale level in numerical format, then test it under continuous distribution.

2. ORDINAL

If the scale (1,2,3,4,5) is treated as ranking order (1sr, 2nd, 3rd, 4th, 5th), it is treated as ordinal data. In the ordinal data, it is necessary to identify the rank of interest, for instance, if we are looking for first choice or preference (1st) then 1st = 1 and else 0.

3. NOMINAL DATA

If the scale (1,2,3,4,5) has a category of interest then the category of interest is equivalent to yes = 1 and non-qualified ones are no = 0.

4. DATA CROSSING

The crossing or mixing o the X and Y may be in the following combinations row (X) and column (Y):

Q O N

Q qq qo qn

O qo oo no

N qn on nn

The non-parametric test for each combinations are:

QQ = Pearson correlation coefficient

QO = Biserial correlation coefficient

QN =Point biserial correlation coefficient

OO = Spearman Rho correlation coefficient

ON = Rank biserial correlation coefficient

QN = Point biserial correlation coefficient

NO = Rank biserial correlation coefficient

NN = Pearson C or phi

5. LOGISTIC MODEL

Let linear regression be Z = a + b1x1 = b2x2, one simple way of obtaining logistic regression is to formulate: Y = 1 / (1 + exp(-Z). See link below for explanations & illustration.

https://en.wikipedia.org/wiki/Logistic_regression

For test statistic for logistic regression, see link below:

https://digitalcommons.wayne.edu/cgi/viewcontent.cgi?referer=https://www.google.co.th/&httpsredir=1&article=1238&context=jmasm