I have a silly little question that I wouldn't want to make a new post on CrossValidated about. Maybe this is a more appropriate place? Basically, I am wondering how to appropriately conduct a regression when it is structurally impossible for a given continuous IV to take non-missing (or non-"NA") values depending on the value of a discrete IV.

To give a contrived (and rather crude and absurd) example, suppose we were modeling the time it takes to complete a horse race, and the sex of the horse is one of the predictors, but we are also interested in whether the length of the horse's penis is associated with race time. Obviously, this variable does not make sense among female horses, so it would be "NA" whenever sex = "Female". How should this problem be approached?