Dynamic Inverse-Probability Weighting Estimator

• Dynamic IPW estimators adapt traditional inverse-probability weighting by using data-driven normalization to minimize finite-sample variance.
• They integrate control variate techniques and model-based propensity scores to deliver improved efficiency and robust inference across various data structures.
• Extensions to dynamic IPW include methods for partial adherence, dynamic treatment regimes, and online learning, ensuring consistent and optimal estimations.

The dynamic inverse-probability weighting (IPW) estimator is a methodological generalization of classical IPW, allowing for improved efficiency and robustness in statistical estimation for a range of modern data structures, particularly longitudinal and time-to-event data, partially labeled data, @@@@1@@@@, and streaming updates. Dynamic IPW aims to optimize finite-sample performance (principally variance) by adaptively modulating normalization, leveraging observed covariate structures, and borrowing ideas from control variate and prediction-powered inference frameworks without compromising fundamental large-sample properties.

1. Motivation and Fundamental Structure

Traditional IPW, such as the Horvitz–Thompson (HT) and Hajek estimators, targets population means or causal effects by reweighting each observation by the inverse of its sampling or treatment assignment probability, restoring unbiasedness under informative sampling or treatment assignment. However, strict reliance on fixed normalization (by sample size or sum of weights) is often suboptimal in finite samples—especially as weights become unstable or distributions are highly skewed.

Dynamic IPW estimators introduce adaptivity, defined as normalization or weighting schemes that are data-driven—often optimized to minimize asymptotic or finite-sample variance. A commonly studied one-parameter family interpolates between HT and Hajek by normalizing using an affine combination of sample size and total inverse weights:

$$\hat\mu(\alpha) = \frac{S}{\alpha n + (1-\alpha) \hat n}, \quad S = \sum_{i=1}^n \frac{Y_i I_i}{p_i},\, \hat n = \sum_{i=1}^n \frac{I_i}{p_i}, \quad 0 \le \alpha \le 1,$$

where $p_i$ is the known (or estimated) inclusion or treatment probability. The optimal $\alpha$ is chosen data-adaptively to minimize the large-sample variance $V(\alpha)$, yielding the adaptively normalized IPW estimator (Khan et al., 2021).

2. Asymptotic Properties, Variance Minimization, and Control-Variate Link

The asymptotic variance for this family is

$$V(\alpha) = E\left[ \frac{1-p}{p} (Y - (1-\alpha)\mu)^2 \right ],$$

and the minimizing value

$$\alpha^* = 1 - \frac{T}{\pi \mu}, \quad T = E\left [\frac{1-p}{p} Y \right ], \quad \pi = E\left [\frac{1-p}{p} \right ].$$

This estimator always attains variance never worse than HT or Hajek, with strict improvement except in degenerate cases. The result can be interpreted algebraically as an optimal control variate adjustment, regressing out the variability due to the weights themselves, yielding

$$\hat\mu_{AN} = \frac{S}{n} + \frac{T_n}{\pi_n} \left( 1 - \frac{\hat n}{n} \right ),$$

with sample analogues substituted for population quantities. Consistency, asymptotic normality, and variance estimation via plug-in or bootstrap are straightforward under standard boundedness and positivity conditions (Khan et al., 2021).

3. Extensions to Time-to-Event and Longitudinal Settings

Dynamic IPW estimators have been developed for time-to-event data, competing risks, and recurrent event models by incorporating time-dependent or history-dependent probability weighting, stable variance normalization, and control for partial adherence:

• In competing risk or time-to-event analysis, adaptively weighted Nelson–Aalen and cumulative incidence function (CIF) estimators use IPW where propensity scores may be estimated parametrically or non-parametrically, with influence function decomposition accounting for PS uncertainty. This introduces an additional variance term, generally minor relative to sampling variability (Deng et al., 2024).
• For recurrent event data with intercurrent events (such as treatment switching), dynamic IPW frameworks weight each risk period by the inverse survival probability with respect to the intercurrent event, modeled using Cox or discrete-time hazard models. Resulting estimators of hypothetical estimands retain consistency, asymptotic normality, and proper inference properties (Sun et al., 6 Jul 2025).
• Continuous-time extensions use stochastic calculus to derive a weight process as the Radon–Nikodym derivative between observed and hypothetical interventions, allowing for weighted Cox (hazard ratio) and Kaplan–Meier (RMST) estimation with guaranteed consistency and inference under local regularity and positivity (Chatton et al., 2020).

4. Flexible and Partial Adherence Weighting in Dynamic Treatment Regimes

Standard dynamic IPW for regime value estimation strictly requires perfect adherence: only individuals who exactly follow a prespecified treatment sequence are included. This produces data inefficiency and high variance, especially with long treatment sequences. Recent developments introduce compatibility-weighted estimators:

• Generalized Adherence Weighting (GAW): At each stage, nonadherers are down-weighted rather than excluded using a parameter $\gamma_{t,i} < 1$, bounded such that variance is minimized without high bias; stabilization preserves the expected value of the weights (Si et al., 10 Dec 2025).
• Bootstrap Adherence Windowing (BAW): Subjects near the decision boundary are included by tolerating small deviations; window widths are selected to optimize bias-variance tradeoff via bootstrap MSE minimization. These strategies deliver consistently reduced finite-sample variance and improved effective sample size, while retaining (asymptotic) unbiasedness and double-robustness under mild additional constraints.

5. Prediction-Assisted and Streaming Dynamic IPW

Dynamic IPW estimators extend naturally to semi-supervised inference and online learning:

• Prediction-Powered Inference (PPI): Bias corrections for large unlabeled data are adaptively weighted by inverse inclusion probabilities (HT or Hajek style) to allow for informative labeling and prediction model error; using estimated propensities, the IPW-PPI estimator remains unbiased and nearly retains the variance reduction of unweighted PPI, with nominal coverage in simulations (Datta et al., 13 Aug 2025).
• Streaming Data / Online GLMs: Updatable IPW (UIPW) estimators update parameter and propensity estimates recursively with each batch, producing parameter estimates whose asymptotic variance matches the oracle batch learner. The per-batch update uses past data summaries, avoiding repeated access to raw data and reducing computation, and is shown to be both consistent and asymptotically normal (Zhang et al., 2023).

6. Dynamic IPW in Longitudinal, Micro-Randomized, and Causal Excursion Analyses

Dynamic weighting addresses variance inflation in micro-randomized and longitudinal trials by adjusting the product-structure of standard IPWs:

• Per-Decision IPW: For windowed binary outcomes, standard IPW products amplify variance due to zero weights when any deviation occurs. Per-decision IPW truncates the product after the first relevant event, reusing partial data and substantially reducing variance. The per-decision estimator is shown to be consistent, asymptotically normal, and achieves 10–50% variance reduction in practical settings (Bao et al., 2023).

7. Implementation and Application Guidance

Implementation is straightforward, requiring only modest adaptations to standard IPW workflows:

• Algorithmic steps typically include model-based propensity estimation, computation of dynamic/compatibility/partial adherence weights, and weighted estimation using standard statistical software.
• In real applications (e.g., HIV dynamic regimes, epidemiological time-to-event studies), these estimators deliver sharper confidence intervals, higher effective sample sizes, and greater analytic stability, especially under complex adherence or missing data patterns (Si et al., 10 Dec 2025, Chatton et al., 2020).

The dynamic IPW framework provides a general, theoretically justified, and computationally efficient extension of IPW methods for modern causal and predictive inference, accommodating time-varying, partially observed, or regime-defined data, while provably improving variance and robustness in nearly all practical sample regimes.