Making an ordinal response variable binary


No cake for spunky
I don't really understand why making an ordinal response variable binary is a bad thing. In fact having read logistic regression books I came to the conclusion that making an ordinal response variable binary was safer in terms of interpretation or violation of assumptions.

"When the dependent variable is ordinal or continuous, classification through forced up-front dichotomization in an attempt to simplify the problem results in arbitrariness and major information loss even when the optimum cut point (the median) is used. Dichotomizing the outcome at a different point may require a many-fold increase in sample size to make up for the lost information187. In the area of medical diagnosis, it is often the case that the disease is really on a continuum, and predicting the severity of disease (rather than just its presence or absence) will greatly increase power and precision, not to mention making the result less arbitrary."


Less is more. Stay pure. Stay poor.
Probably lots of issues beyond those. How do you select the threshold, loss of information, what if there is a non-linear increase in associations, generalizing results from a variable that you created. So depends on the context. If there is a general association the signal should remain and get you in the right direction. Your estimate might just be off some, likely biased toward the null, given there is an underlying effect.


Active Member
Most of us grew up with school test questions that involved picking one of four possible answers.
Which can be modelled in modern Item Response Theory as an Item Response Function

It does binary dichotomies, and in some ways it's better than the Likert scale representation.
Last edited:


No cake for spunky
Item Response Theory often uses regression such as Rauch models.

I found this kind of amazing from the same author since most analysis treats regression as linear.

Relationships among variables are seldom linear, except in special cases such as when one variable is compared with itself measured at a different time. It is a common belief among practitioners who do not study bias and