6 / 6 Causal Inference

What does Neyman orthogonality protect against?

Let ψ(W;θ,η)\psi(W; \theta, \eta) be a score function for a target parameter θ\theta and nuisance parameter η\eta. What does Neyman orthogonality at the true nuisance value η0\eta_0 require?

A

The nuisance parameter must converge at the same n\sqrt{n} rate as θ\theta

B

The score must be statistically independent of every nuisance estimate

C

The derivative of the expected score with respect to local perturbations of η\eta must vanish at η0\eta_0

D

The Hessian of the empirical loss with respect to θ\theta and η\eta must be block diagonal