Myth: Shapley value attribution properly accounts for channel interactions

Myth: Shapley value attribution properly accounts for channel interactions

The question: Shapley value is praised for handling synergy — it averages each channel's marginal contribution across all possible orderings. Doesn't that mean it correctly captures how channels work together?

What the method guarantees, and doesn't: Shapley does fairly distribute credit given a defined 'coalition value' function, and it satisfies clean axioms (efficiency, symmetry, dummy, additivity). Within those rules it is the unique fair allocation. That's real and worth respecting.

The nuance is where the value function comes from. In marketing implementations, the coalition value is the observed conversion rate of users exposed to a given subset of channels. That number is riddled with selection bias — users exposed to 'search + email + social' are not a random sample; they self-selected into that exposure set by being high-intent. Shapley then fairly divides a credit total that was itself confounded. Fair division of a biased pie yields biased shares. Worse, most implementations bin channels and ignore timing and frequency within a channel, so genuine interaction effects (the email that mattered only because it followed the demo) get collapsed.

So Shapley's mathematical rigor is real but orthogonal to causal validity: it answers 'how do we fairly split this observed credit,' not 'what did each channel cause.'

Bottom line for practitioners: use Shapley when you want a principled, order-robust credit split — it beats linear and last-click for that purpose. But do not present its outputs as interaction-corrected truth. The coalition values are observational; calibrate the channels it elevates against experimental lift before reallocating budget on its say-so.