If so, can you share your experience/thoughts?

Sorry, I'm just curious, but why should anyone do that? You would omit data that could (and probably will) be very important for your predictions. Example: If x is a member in the target file making claims c_x in year 3, and y is a member not in the target file, but making exactly the same claims in year 1 or 2, staying y_d days in hospital the following year. Woudn't it be a good approach to predict about y_d days for x? (Of course, it will hardly happen that you'll find a exact match for the claims x made - that's where it gets tricky, combining/merging information etc.)