In logistic regression: What is the proper way to report the overall "odds ratios" for a non-linear continuous variable

36 Views Asked by At

I am fitting a logistic regression with multiple binary variables as well as a single continuous variable (AGE) for which I have a linear and a quadratic term.

$$\log\frac P{1-P} = \text{constant} + a \cdot \text{AGE} + b \cdot \text{AGE}^2 + c \cdot\text{GENDER} + d \cdot\text{PREGNANCY} + \cdots$$

For any of the binary variables, it is custom to report their "odds ratio" (OR). For example, the OR for GENDER is

OR_GENDER = exp(c)

However, for AGE (which is continuous and has a non-linear effect), it is less clear what is the proper way to report an overall odds ratio.

What I am currently doing is reporting the ratio between the maximum and the minimum of the odds as a function of AGE within the range of ages in our data (0

OR_AGE = exp( max(a * AGE + b * AGE^2) - min(a * AGE + b * AGE^2) )

where the min and max are evaluated within the range 0-100 and so may appear either at the extreme points 0, or 100, or in the middle if the function has extremum within this range.

My questions are:

Is this method of reporting OR for non-linear variable ok and proper to do?

If so, is there a name for this procedure / or a reference I can cite?

If not, what would be a proper way to report overall ORs for AGE that can be compared on equal footings with those of GENDER, PREGNANCY and others?