Presumably, if you use E to decide in Newcomb's soda, the decisions of agents not using E are screened off, so you should only calculate the relevant probabilities using data from agents using E.

Can you show where the screening off would apply (like A screens off B from C)?