Interaction variables work differently when population is split Let's say I have two predictors to predict financial risk: Gender and shopping habits. Gender has levels of "Male" and "Female", while shopping habits has "quick shopper" and "slow shopper". 
I am quite surprised to see this: When I split the population into "male" and "female", shopping habits as a single variable work quite well to separate the risk on female population (not for males). This also intuitively makes sense. But when I create an interaction variable of "gender:shopping habits" and look at logistic regression, there is no risk separation whatsoever even on the dummy created "female:quick shopper".
Could there be an explanation of this?
Note: I should add that the initial working performance is cross-validated (not an overfitting result).
Note2: I edited the question for better understanding.
 A: Edit: might not apply because the OP's mention of "age" was a mistake. Would only still apply if another "third" variable (besides shopping habits and financial risk, not counting gender) is changing in the model...
Original answer:
If you're including age as a predictor in the model as well (not just the interaction of age and shopping habits...which is a regression faux pas), age could could be mediating the relationship between shopping habits and financial risk in women. When a third variable fully mediates the relationship between two others, entering it alongside the first variable as a regressor in multiple regression predicting the second variable will wipe out the relationship between the first and second variable, because it explains all the same variance in the second variable that the first does, and then some (probably). 
Full mediation is somewhat rare with such indirectly related constructs as age, shopping habits, and financial risk though, so I doubt this is really the case if you've represented your variables accurately. I'm mostly suggesting this to imply that you should make sure you include your main effects with the interaction, and consider the possibility of (at least partial) mediation or other consequences of including a stronger predictor in the model.
