Dummy variables and likert scale

So I have two related questions. I am performing a logistic regression model and want to know how to enter these variables.

1) For one survey I have questions and the responses are Yes, Don't know, and No. Can I just create a dummy variable for Yes (1) with Don't know and No coded as 0 or do I need to exclude the Don't know's (this significantly cuts down my sample size).

2) I have a likert scale for another survey (1 strongly agree, 3 neither, 5 strongly disagree) what is the best way to enter this into the model? Just as categorical data. Some people have argued to use as continuous or again to create dummy variables.

The answer I give is not specific for logistic regression models, but it is also relevant for other types of regression models.

Excluding the "Don't Know's" is a very bad idea. You discard some important information. You are in some sense creating a missing-data problem. This can have more or less impacts on your results. You can create dummies for two of the categories and use the third one as a reference category. Which one you choose as a reference is mainly a matter of taste and convenience.

Likert scales can be handled in the same way. In your case, with three categories, you can create dummies for two categories and use the third category as a reference. I would not introduce it as a continuous variable. With the dummy coding you are able to capture non-linearities. This way you will obtain a much more flexible specification. More generally speaking, if you have $k$ categories, create $k-1$ dummies and use the remaining category as reference.

Depending on the software you use, the creation of dummies may be superfluous. There are ways to introduce categorical variables directly in regression models. In Stata you can have a look at help fvvarlist and in R at ?factor.

Note that there are other coding schemes you could use. Dummy coding is quite straightforward to implement and to interpret.

• So you would suggest using dummy variables for both surveys? For the larger likert scale you are suggesting to collapse the 5 categories into 3 with one being the reference (Agree, Neither and Disagree)? – user10720 Apr 18 '12 at 16:46
• I have 2 surveys, one has 3 categories, the other has 5. What is the benefit of dummy variables versus using the 5 categories and a categorical variable? – user10720 Apr 18 '12 at 16:53
• Dummy variables is using categories and a categorical variable. At least, it is one way of doing so. With 5 levels, you have 4 dummy variables, with 3 levels you have 2 dummy variables. There are other coding methods as well, such as effect coding. I find dummy coding most intuitive to interpret. I would not delete the "don't know", but it is probably not good to lump them with the "no". I would create two dummy variables. Any good statistics package (R, SAS, SPSS and probably every other one) will do this for you. – Peter Flom Apr 18 '12 at 17:01
• Ok thanks. Do you know how SPSS enters the data when you merely classify your likert scale data (which is labelled 1-5) as categorical. I wonder how that differs from dummy variables. – user10720 Apr 18 '12 at 17:25
• SPSS enters it in a regression coded as dummies, if you ask it. e.g. logistic regression honcomp with read science ses /categorical ses will result in ses being dummy coded. example from here, where the output is visible – vinnief May 13 '14 at 15:50