Continuous vs categorical variable in regression

[Jameswei]

Dear all,

In doing regression analysis, I am sometimes confused with whether I should use a continuous predictor variable (e.g., age in years) or re-group it into a categorical variable (e.g., 5 age groups: 18 to 24, 25 to 34, 35 to 44, 45 to 54, and 55+). What is the best way to decide?

Any comments or references are welcome. Many thanks.

James

 

[Karabiner]

The grouping is artifical and it will cost you 4 degrees of freedom, instead of only one degree of freedom.
So it seems wise not to group. What would you achieve by it?

With kind regards

Karabiner

 

[hlsmith]

I agree with @Karabiner - splits are generally arbitrary. If there really is a biological, policy, or social phenomenon at play; you should look at those effects using splines (general additive models [GAM]) or polynomials. Splitting variable in piecewise regression could be acceptable in some scenarios, but likely after visualizing them via GAMs and validating them in a holdout dataset. Piecewise splits typically also assume linear relationships before and after inflection points if polynomials or GAM not used - so that may misrepresent your data as well if not validated based on theory or bare minimum - empirically.

 

[noetsi]

I have always hear you lose information, I assume degrees of freedom, when you categorize interval variables. Having said that the federal agency I am doing an analysis for split age into 7 variables which left me very puzzled.

 

[hlsmith]

One way you are losing information is because you are saying the top value in one group is different than the bottom value in the next group - when these values could be figuratively identical on any given day.