Continuous variable significant in logistic regression, but not when categorized?

Question

I've got a variable (age) that is continuous. It's significant for my binary outcome in a logistic regression. However, when I try to categorize the age variable for easier interpretation, by splitting it into 10 year intervals, the variable is no longer significant. Or that's to say the variable is significant but if I run it with an i. (in stata) in front, no group is significant by itself.

I could achieve the same by simply dividing the variable age in 10 and then I would have the effect for every 10 years and this time around it would of course be significant as nothing had been fundamentally changed. How come my variable stops being significant when categorized?

Answer 1

Without seeing the details, I can think of a couple of reasons.

First, for each category of age you add beyond 2 you lose a degree of freedom in the analysis. That makes it progressively harder to find truly significant differences.

Second, depending on how you coded them, you might have lost the natural ordering of the age categories. Coding them as a set of ordinal predictors would maintain the natural ordering and in principle would allow you to specify that later age categories should have progressively larger effects on outcome.

In general, it’s best not to categorize continuous predictors for analysis. For display you can show model predictions, with confidence limits, at example values based on the continuous model.