Discretizing continuous mediators in integrals simplifies computation but changes the target functional and naive discretization induces first-order bias in bin width.
We then propose a simple bias-reduced functional that evaluates outcome regressions at within-bin conditional means.
