1
$\begingroup$

I have a variable that has 8 ordinal categories. These categories are made up by discretizing an interval variable; that is the categories (levels in R) are '<7.5', '7.6 - 10', ..., '17.6-20' such that all values between 0 to 20 are included. Give that I don't have lots of data, I am not sure whether I use 7 dummy variables in linear regression. So, I am thinking to create this as a interval variable by replacing the categories by their means, say '7.5' by 3.75 so on.

Does it make sense to do this?

$\endgroup$

1 Answer 1

2
$\begingroup$

No, and kind of perhaps.

First, numerically, the two coding schemes are nearly the same thing. Calling something 1, 2, 3, 4, 5, and 1.5, 3.5, 5.5, 7.5, 9.5 is simply a rescaling, instead of a jump of 1, it became a jump of 2 between every level. In other words, if the ordinal variable is unfit as a continuous independent variable, it'd still be unfit even in disguise behind the "proper" unit. I said "nearly the same thing" because not all ordinal categories are of equal widths. In those cases, recoding sometimes may improve the situation (see below.)

Second, what determines if the said ordinal variable can be model as categorical or continuous is if it is linearly related to your dependent variable. So, keep it either in its orginal form or the recoded form, use scatter plot (or component plus residual plot if it's a multiple linear regression) with the aides of a regression line and a LOWESS curve to see if linearity can be safely assumed. If yes, then possibly fine to model it as a continuous variable. If not, then some more drastic recoding such as aggregation or even dichotomization may be considered.

$\endgroup$
0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.