I am trying to assign a monetary value to each variable in a logistic regression model in order to describe contribution "worth" in a unit the business can understand.
I have built my model table (one-hot encoded), and fitted a model against my outcome variable which in this case is a sales event. Each regressor is a form of contact strategy we have applied to a customer, and carries a cost. Some customers (cases) have been contacted by just one strategy, some customers many. It's a typical regression problem! :)
What I want to do is understand for each case is how much 'credit' to assign to each regressor.
I have attempted to do this by zero-ing each regressor column (one at a time), then running a prediction from the model on this dummy data set to see the net reduction in sales total. In effect, attempting to simulate what would happen without each regressor in order to understand its contribution. However when I do this the results are nonsensical. In particular, a null model (all columns zero) accounts for almost 90% of the sales; it's logodds coefficient is negative (-1.3) where regressors are positive, and in the context of my model should account for only a small fraction of the sales.
How should one go about this process in order to correctly enumerate contribution?
Many thanks in advance.