debias_hte_conditional: Customized debiasing for 'hte_test_conditional()'

In dScoreTest: Debiased Score Tests for Goodness of Fit and Model Comparison

Customized debiasing for hte_test_conditional()

Description

For a hunt \hat{h}(t,z), it can be rewritten as

\hat{h}(t,z) = \hat{h}_0(z) + t \cdot \hat{h}_{\Delta}(z),

where \hat{h}_{\Delta}(z) := \hat{h}(1,z) - \hat{h}(0,z). Let the projection of \hat{h}(t,z) onto the null space be

m_{\hat{h}}(t,z) = m_0(z) + t \cdot m_{\Delta}(z_{S^c}).

This function returns the debiased hunt function \hat{h} - \hat{m}_{\hat{h}}.

Usage

debias_hte_conditional(
  h.hat,
  X.debias,
  fit.debias,
  predict_fun,
  weight_fun,
  wls_method,
  arg.wls_method,
  randomized = FALSE
)

Arguments

h.hat	Object of class hunt produced by hunt_optimal(), hunt_wls() or hunt_vanilla(), where h.hat$h(X) is \hat{h}(t, Z) for X=[t, Z].
X.debias, fit.debias, predict_fun, weight_fun, wls_method, arg.wls_method	See dScoreTest() for details. Argument arg.wls_method must contain field S that defines the model space.
randomized	If FALSE (default), the propensity e(Z) = \mathbb{E}[T \mid Z] is estimated with grf::probability_forest. If TRUE, T is assumed randomized (independent of Z), so e(Z) is taken to be the constant mean(T) on the debiasing sample, fitted upfront without cross-fitting.

Details

Function \hat{m}_{\Delta} is fitted by minimizing

\sum_i (T_i - \hat{e}(Z_i))^2 \, (\hat{h}_{\Delta}(Z_i) - m_{\Delta}(Z_{i,S^c}))^2,

where e(z):= \mathbb{E}(T \mid Z=z). The debiased hunt function is given by

(\hat{h} - \hat{m}_{\hat{h}})(t,z) = (t - \hat{e}(z))\,(\hat{h}_{\Delta}(z) - \hat{m}_{\Delta}(z_{S^c})).

Value

A list with elements:

m.h.fit

The null model fitted (over all columns of X) to project and debias \hat{h}.

h

The debiased hunt function \hat{h} - \hat{m}_{\hat{h}} with signature h(X).

dScoreTest documentation built on Sept. 3, 2026, 1:06 a.m.