Balancing Weights in Causal Inference

• Balancing weights are sample-specific numerical multipliers that adjust covariate distributions to mimic randomized experiments for unbiased effect estimation.
• They are computed via propensity score modeling or direct convex optimization methods to enforce balance on prespecified covariate functions.
• Their effective implementation requires careful tuning, regularization, and diagnostics to mitigate issues like extreme weights and poor overlap.

Balancing weights are sample-specific numerical multipliers applied to observations in statistical estimation procedures to achieve covariate balance between groups with different exposure, treatment, or target attributes. They serve as a foundational tool in causal inference, survey sampling, standardization, policy evaluation, missing data analysis, and modern representation learning frameworks. At their core, balancing weights are designed to construct synthetic samples in which the covariate distributions of different groups are aligned, thus emulating randomization and underpinning unbiased estimation of population-level or subgroup-specific estimands.

1. Formal Definitions and Core Principles

Balancing weights are constructed so that, after weighting, prespecified moments or functionals of the covariate distributions are equated between groups. In the canonical causal-inference setting, consider units $(X_i,W_i,Y_i)$ with $W_i$ a binary treatment, $X_i$ covariates, and $Y_i$ the observed outcome. For identification of the average treatment effect (ATE), assumptions of strong ignorability ($W \perp (Y(0),Y(1)) \mid X$) and overlap ($0 < e(x) < 1$ for $e(x) = P(W=1|X=x)$) are imposed (Ben-Michael et al., 2021). Covariate balance is enforced by weights $w_i$ satisfying

$$\frac{1}{n}\sum_{i=1}^n W_iw_i f_j(X_i) = \frac{1}{n}\sum_{i=1}^n f_j(X_i),\quad \forall\,j=1,\dots,p,$$

for a specified basis $\{f_j\}$.

A key special case is inverse-propensity score weighting (IPW), where $W_i$0 with $W_i$1 an estimate of the propensity score. More generally, balancing weights can be viewed as solutions to method-of-moments or constrained optimization problems where the objective includes dispersion penalties (e.g., variance, entropy) and balance constraints.

2. Methodological Families: Propensity-Score and Direct Balancing

Two principal schools define the literature:

• Propensity-score-based approach: Fit a model for the assignment mechanism $W_i$2, then compute weights as the inverse probability of treatment or exposure. This approach is conceptually simple and doubly robust if paired with appropriate augmentation (Ben-Michael et al., 2021). However, it can yield poor finite-sample balance and unstable, extreme weights when overlap is weak.
• Direct balancing approach: Estimate weights by directly solving balance equations, typically via convex optimization. For a chosen feature set $W_i$3, one selects $W_i$4 to minimize a dispersion measure subject to exact or approximate balance constraints:

$W_i$5

and possibly $W_i$6. Penalties such as $W_i$7 (stable-balance), KL-divergence (entropy balance) or $W_i$8 norm are used for regularization and variance control (Ben-Michael et al., 2021).

Both classes can be unified: as the function class $W_i$9 becomes sufficiently rich, balancing weights converge to the inverse-propensity weights, ensuring the same asymptotic efficiency properties (Ben-Michael et al., 2021).

3. Theoretical Properties: Semiparametric Efficiency, Robustness, and Minimax Duality

Balancing weights estimators, when constructed with rich enough balancing sets and regularization that ensures imbalances vanish faster than $X_i$0, are regular and achieve the semiparametric efficiency bound for the estimand of interest (Ben-Michael et al., 2021). The influence function for the average potential outcome is

$X_i$1

where $X_i$2.

Doubly robust estimators (augmentation with an outcome model) retain consistency if either the propensity or outcome model is consistent, achieving asymptotic normality and the efficiency bound (Ben-Michael et al., 2021).

Minimax duality results have clarified that balancing weight optimization—subject to outcome model constraints—admits a dual convex loss form. Specifically, for outcome function class $X_i$3, minimax weighted mean-square-error bounds can be recast as single convex optimization problems over $X_i$4 involving Integrated Probability Metrics or variance penalizations (Bruns-Smith et al., 2022). The "minimum worst-case bias" $X_i$5 acts as a quantitative replacement for classical overlap assumptions, precisely characterizing the irreducible bias given function class restrictions and sample supports.

4. Generalizations: Target Populations, Complex Designs, and Modern Algorithms

Targeted Populations and Generalized Estimands: The balancing-weights formalism encompasses not just the ATE but a wide class of linear functionals, including the ATT, policy evaluation, dose-response, and individualized effects. In these settings, weights are shaped by the Riesz representer of the target functional, with balance enforced on features that parameterize the estimand (Ben-Michael et al., 2021).

Clustered and Structured Designs: In clustered or multilevel observational studies, balancing weights are adapted to accommodate intra-cluster correlation, complex group allocation, and design-based variance structure. Penalties can be tailored to simultaneously minimize both cluster-level and unit-level variance, subject to moment constraints on both levels (Keele et al., 2023).

Continuous and Regression Covariates: Continuous Weight Balancing (CWB) approaches rebalance a continuous trait (regression label, feature) by constructing weights as the ratio between target and kernel-estimated source densities, avoiding arbitrary binning or discretization (Wu et al., 2021).

High-dimensional and Neural Schemes: In high-dimensional or nonparametric regimes, regularization (e.g. $X_i$6, entropy) becomes essential for variance control and feasibility (Wang et al., 2017). Recent neural network-based approaches parameterize weights via density ratio estimation using $X_i$7-divergence variational objectives, notably $X_i$8-divergences, with early stopping and diagnostic procedures to mitigate generalization and curse-of-dimensionality issues (Kitazawa, 2022).

5. Practical Considerations: Tuning, Diagnostics, and Implementation

Implementation of balancing weights entails several decisions:

• Choice of basis/functions: The function class for balance (linear terms, polynomials, kernel features) should be sufficiently rich to capture confounding; underspecification incurs bias, overspecification inflates variance or renders the problem infeasible (Ben-Michael et al., 2021, Ben-Michael et al., 2022).
• Regularization parameters: Penalty strength (e.g. $X_i$9 in $Y_i$0 or entropy balancing) must be tuned to optimize the bias–variance trade-off. Data-driven heuristics such as the bootstrap-balancing criterion or cross-validation can be used (Wang et al., 2017).
• Effective sample size (ESS): The ESS, $Y_i$1, diagnoses over-dispersion; too small an ESS signals instability. Weight clipping and additional dispersion penalties may be necessary.
• Diagnostics: Covariate balance should be inspected via standardized mean differences or balance diagnostics post-weighting. For settings with poor overlap, both estimability and uniqueness of the estimand should be verified, often through parallel calculation of overlap weights or trimming (Ben-Michael et al., 2022).
• Algorithmic solvers: Modern implementations utilize convex QP solvers (e.g. MOSEK, Gurobi), coordinate descent, or neural-network-based SGD for high-dimensional regimes (Kitazawa, 2022, Keele et al., 2023).

6. Application Domains and Extensions

Balancing weights are deployed across a spectrum of scientific and statistical contexts:

• Causal Inference: Standard for treatment effect estimation and robust policy evaluation. Overlap weights minimize the asymptotic variance among balancing weights under homoskedasticity, and provide finite-sample mean-balance under logistic models (Li et al., 2016).
• Missing Data: Non-monotone or not-at-random missing data patterns are addressed by balancing-weighted estimating equations, leveraging pattern graphs and sequential balancing across observed response patterns (Dong et al., 2024, Dong et al., 18 Apr 2025).
• Mediation and Complex Estimands: Two-step minimal weights or similar frameworks are used in causal mediation analysis, enforcing approximate balance on both covariates and intermediate variables, resulting in semiparametric efficiency and robustness to finite-sample imbalances common to EIF and IPW estimators (Kawato, 10 Dec 2025).
• Machine Learning and Representation Learning: Joint learning of features and balancing weights underpins counterfactual representation learning; balancing weights modulate loss functions or form the basis for doubly robust ITE estimation, facilitating improved generalization in high-dimensional or flexible nonparametric regimes (Assaad et al., 2020, Kitazawa, 2022).
• Sensitivity Analysis: The framework naturally accommodates interpretable sensitivity analysis by parameterizing bias due to unmeasured confounding as a function of imbalance and outcome-association, yielding valid interval estimates for the causal effect even under finite levels of confounding (Soriano et al., 2021).

7. Limitations, Open Challenges, and Recommendations

Despite their flexibility and strong theoretical foundations, balancing weights approaches face limitations:

• Feasibility and Overlap: When overlap between groups is poor, exact balance may be infeasible or result in extreme weights, inflating variance and finite-sample bias. Approximate balance and entropy or quadratic regularization ameliorate but do not wholly resolve this (Ben-Michael et al., 2021, Ben-Michael et al., 2022).
• Curse of Dimensionality: For high-dimensional or nonparametric bases, variance grows and strong regularization or dimension reduction is essential (Wang et al., 2017, Kitazawa, 2022).
• Population Targeting: Specification of the target population via the tilting function $Y_i$2 is critical; misalignment can render the causal estimand scientifically uninterpretable (Li et al., 2016, Assaad et al., 2020).
• Functional Misspecification: If important confounding structures are omitted from the balancing set, residual bias persists even with perfect balance on observed features (Ben-Michael et al., 2022).

Best practice entails joint consideration of covariate selection, overlap diagnostics, regularization tuning, validation of effective sample size, and reporting of both population-targeting choices and diagnostic tables post-weighting. In applications with positivity violations, partially retargeted balancing weights—modifying only the necessary subspace of covariates—offer improved trade-offs (Barnard et al., 24 Oct 2025).

Balancing weights continue to be a central analytic device, underpinning semiparametric efficient estimation, robust inference in complex designs, and novel algorithmic strategies in the age of machine learning (Ben-Michael et al., 2021, Wang et al., 2017, Assaad et al., 2020, Kitazawa, 2022).