Affine-Normalized IPW Family

• Affine-normalized IPW is defined by replacing the standard normalization factor with an affine combination of the sample size and inverse probability weights, unifying the Horvitz–Thompson and Hajek estimators.
• It enables adaptive selection of the combination parameter to minimize asymptotic variance while preserving efficiency, double robustness, and optimal regret rates.
• Extensions to augmented IPW and policy learning demonstrate finite-sample improvements and variance stabilization, especially under challenging overlap or model specifications.

The affine-normalized inverse probability weighting (IPW) family comprises a class of estimators in which the normalization factor in IPW is replaced by an affine combination of the sample size and the sum of inverse probability weights. This framework unifies the Horvitz–Thompson and Hajek (self-normalized) estimators as special cases and enables data-adaptive selection of the combination parameter to minimize asymptotic variance. The method extends naturally to augmented IPW (AIPW) and policy learning, providing finite-sample improvements and preserving key asymptotic properties such as efficiency, double robustness, and optimal regret rates (Khan et al., 2021, Rostami et al., 2021).

1. Definition and Structure of the Affine-Normalized IPW Family

Let $Y_i\in\mathbb{R}$ represent observed outcomes, $p_i \in (0,1)$ be known (sampling or treatment) probabilities, and $I_i \sim \mathrm{Ber}(p_i)$ indicate when unit $i$ is observed or treated. The standard IPW weights are $w_i = I_i / p_i$, with $\hat S = \sum_{i=1}^n Y_i w_i$ and $\hat n = \sum_{i=1}^n w_i$. The affine-normalized IPW estimator for parameter $\mu$ is

$$\hat\mu_\lambda = \frac{\hat S}{(1-\lambda)n + \lambda\hat n}$$

where $\lambda \in \mathbb{R}$. Notable special cases:

• $p_i \in (0,1)$0: Horvitz–Thompson estimator $p_i \in (0,1)$1
• $p_i \in (0,1)$2: Hajek estimator $p_i \in (0,1)$3

This family was originally proposed by Trotter and Tukey in Monte Carlo sampling contexts and provides a continuous interpolation between unnormalized and self-normalized approaches (Khan et al., 2021).

2. Asymptotic Variance and Statistical Properties

Assuming $p_i \in (0,1)$4 are i.i.d. with $p_i \in (0,1)$5 and $p_i \in (0,1)$6, Theorem 1 establishes asymptotic normality:

$p_i \in (0,1)$7

The variance function is minimized at an optimal $p_i \in (0,1)$8 dependent on the population moments, allowing adaptively normalized estimators to improve over both the Horvitz–Thompson and Hajek forms, except in pathological edge cases. This structure provides a theoretical foundation for variance reduction by data-driven normalization (Khan et al., 2021).

3. Data-Adaptive Calibration and Convergence

Direct plug-in estimation of the optimal $p_i \in (0,1)$9 can increase variance; instead, an iterative adaptive procedure is recommended:

$I_i \sim \mathrm{Ber}(p_i)$0

This sequence converges with high probability to a unique fixed point $I_i \sim \mathrm{Ber}(p_i)$1, characterized by

$I_i \sim \mathrm{Ber}(p_i)$2

This estimator, labeled "adaptively normalized" (ANIPW, Editor's term), is algebraically equivalent to the control-variates (regression control) estimator using weights as controls. The iterative mapping constitutes repeated minimization of estimated asymptotic variance (Khan et al., 2021).

4. Connections to Control-Variate Techniques

The affine-normalized estimator connects to classical regression controls. For control variate $I_i \sim \mathrm{Ber}(p_i)$3 applied to the pure-IPW estimate $I_i \sim \mathrm{Ber}(p_i)$4:

$I_i \sim \mathrm{Ber}(p_i)$5

The minimum-variance solution $I_i \sim \mathrm{Ber}(p_i)$6 matches the optimal affine combination parameter, and plugging its sample estimate yields the same form as $I_i \sim \mathrm{Ber}(p_i)$7. Thus, affine normalization generalizes and algebraically coincides with classical variance-reducing control-variate strategies (Khan et al., 2021).

5. Extensions: Augmented IPW and Policy Learning

Augmented IPW (AIPW)

With covariates $I_i \sim \mathrm{Ber}(p_i)$8:

$I_i \sim \mathrm{Ber}(p_i)$9

The affine-normalized AIPW substitutes the second term with its adaptively-normalized analog:

$i$0

Theoretical guarantees show $i$1 in probability under standard double machine learning/double robustness assumptions, i.e., adaptive normalization achieves semi-parametric efficiency. Simulations confirm systematic finite-sample reductions in mean-squared error (MSE), especially under model misspecification (Khan et al., 2021).

Policy Learning

Given candidate policies $i$2, the standard IPW-value estimator is

$i$3

The affine-normalized version introduces a control-variate correction, and selecting $i$4 to maximize $i$5 over a VC-class $i$6 achieves the same $i$7 regret rate as classical IPW. Empirical evidence from low- and higher-dimensional policy classes demonstrates strictly reduced policy regret using adaptively normalized value estimators (Khan et al., 2021).

6. Practical Considerations and Recommendations

Adaptive normalization provides a robust response to practical issues including overfitting in nuisance estimation and failures of positivity (overlap) assumptions. In neural network–based AIPW/AIPW-normalized (nAIPW) settings, normalizing the weights stabilizes variance under poorly-regularized or near-positivity-violating models. Empirically, nAIPW exhibits uniformly better bias, variance, and RMSE than unnormalized AIPW in neural settings, and is suggested in scenarios of weak overlap or model complexity (Rostami et al., 2021).

Key recommendations for practitioners:

• Always impose at least mild $i$8 (or $i$9) regularization in neural nuisance models.
• Prefer normalized or affine-normalized variants when fitted propensities approach 0 or 1.
• Use cross-fitting and monitor effective propensity ranges as part of model tuning.
• Estimate asymptotic variances using influence-function–based formulas, or consider subsampling-based estimators in high-complexity regimes (Rostami et al., 2021).

7. Summary of Statistical Efficiency and Empirical Behavior

Across mean estimation, ATE estimation, and policy learning, affine normalization never incurs asymptotic efficiency loss and yields finite-sample improvements over classical baselines. All core guarantees of unbiasedness, double robustness, Neyman orthogonality, and minimax regret in policy learning are preserved. Finite-sample simulations and theoretical analysis consistently demonstrate lower MSE, bias, and variance under adaptive affine normalization, especially under challenging overlap or model specifications (Khan et al., 2021, Rostami et al., 2021).