Overlap Weights in Causal Inference

• Overlap weights are weighting functions that prioritize units in regions with balanced treatment probabilities, ensuring robust causal effect estimation.
• Defined using the propensity score, they produce bounded weights that minimize variance compared to traditional inverse probability weighting.
• Beyond causal inference, overlap weights extend to network and sequence analyses by quantifying similarity through set comparisons for stable, high-dimensional analysis.

Overlap weights are a class of weighting functions that reweight units in causal inference, network analysis, aggregation, and related fields according to explicit criteria of "overlap" between groups, probability distributions, or set intersections. In causal inference, the central application, overlap weights prioritize units that belong to regions of covariate space where both treatment and control groups are well represented (that is, where the propensity score is neither near 0 nor near 1). This ensures robust estimation of treatment effects in the presence of limited positivity. In network and combinatoric applications, overlap weights quantify the strength or redundancy of links or substrings via Jaccard-like or min/max-based set comparisons. Overlap weighting is now a foundational tool for statistical stability and efficiency in high-dimensional, non-experimental, and networked data contexts.

1. Formal Definition and Mathematical Properties

Let $Z_i \in \{0,1\}$ be the binary treatment, $X_i$ a vector of confounders, with the propensity score $e(X_i) = P(Z_i = 1|X_i)$. The (causal inference) overlap weight for unit $i$ is defined as

$$w_i^{\mathrm{OW}} = \begin{cases} 1 - e(X_i), & Z_i=1 \ e(X_i), & Z_i=0 \end{cases}$$

or, equivalently, $w_i^{\mathrm{OW}} = Z_i[1-e(X_i)] + (1-Z_i)e(X_i)$ (Matsouaka et al., 2022, Ben-Michael et al., 2022, Li et al., 2016, Zhou et al., 2020).

The target estimand associated with overlap weighting is the average treatment effect in the overlap, or equipoise, population (ATO/OWATE): $$\tau_{\mathrm{OW}} = \frac{E[e(X)(1 - e(X))[Y(1) - Y(0)]]}{E[e(X)(1-e(X))]}$$ where the expectation is over the covariate distribution $f_X(x)$ (Matsouaka et al., 2022, Zhou et al., 2020, Li et al., 2016).

Overlap weights are bounded: $w_i^{\mathrm{OW}}\in [0,1]$, with maximum at $e(X_i) = 0.5$. This ensures that units at the extremes of the propensity score distribution (where positivity is violated) receive nearly zero weight, limiting estimator variance and preventing instability seen in inverse probability weighting (IPW).

2. Rationale, Target Population, and Statistical Guarantees

Overlap weights target the subpopulation for which $e(X)\approx 0.5$, i.e., units whose treatment assignment is maximally uncertain. This "clinical equipoise" region is where both treated and control units are well represented, maximizing the empirical basis for estimating potential outcome contrasts, and minimizing reliance on model extrapolation (Matsouaka et al., 2022, Zigler et al., 2017, Ben-Michael et al., 2022). Unlike IPW, overlap weighting smoothly downweights propensity score tails without trimming, ensuring stability without loss of information due to arbitrary cutoffs.

Key statistical properties include:

• Minimal asymptotic variance: The overlap weighting function $h^{\mathrm{OW}}(x) = e(x)[1-e(x)]$ minimizes the (homoskedastic) asymptotic variance of the (weighted) difference-in-means estimator among all balancing weights (Li et al., 2016, Zhou et al., 2020).
• Exact covariate balance: When the propensity score is fitted by logistic regression, overlap weighting achieves exact mean balance for any covariate included in the linear predictor, even in finite samples (Li et al., 2016).
• No large weights: $\max_x w(x) \leq 0.5$, preventing domination by single units and ensuring effective sample size (ESS) remains large (Matsouaka et al., 2022).
• Robustness to positivity violations: Overlap-weighted estimators remain unbiased (in large samples) and have lower relative root mean squared error (RRMSE) compared to IPW or trimmed IPW under both overlap loss and model misspecification (Matsouaka et al., 2022, Zhou et al., 2020, Ben-Michael et al., 2022).

3. Comparison to Alternative Weighting Schemes

Overlap weights are a member of the general balancing weight family, which includes:

• IPW: $h(x) = 1$ (targets full population, unbounded weights).
• Matching weights (MW): $h(x) = \min\{e(x), 1-e(x)\}$ (targets samples with near-equal probability of either group).
• Entropy weights (EW): $h(x) = -[e(x)\ln e(x) + (1-e(x))\ln(1-e(x))]$ (smoothly downweights tails, like overlap). These methods differ primarily in the selection function $h(x)$ and thus in which subpopulation is emphasized.

Overlap, matching, and entropy weights all assign greatest mass to $e(x)=0.5$ and smoothly downweight tails, yielding estimators that numerically coincide in large $n$ (up to $O(n^{-1/2})$). Under positivity violations, IPW exhibits bias $>$20%, poor coverage ($<$0.70), while overlap/matching/entropy weights retain bias $<1\%$ and coverage near 0.95 (Matsouaka et al., 2022, Zhou et al., 2020, Ben-Michael et al., 2022). In practical terms, overlap weights dominate (or match) the efficiency and stability of these competitors.

Table: Comparison of Weighting Schemes

Method	Weighting Function	Target Population	Variance Control
IPW	$1/e(x)$, $1/(1-e(x))$	Full	Unbounded
Overlap	$1-e(x)$, $e(x)$	Equipoise (ATO)	Bounded
Matching Weights	$\min\{e(x),1-e(x)\}/e(x)$	ATM	Bounded
Entropy Weights	$-\left[e(x)\ln e(x)+\ldots\right]/e(x)$	ATEN	Bounded

(Matsouaka et al., 2022, Li et al., 2016, Zhou et al., 2020)

4. Implementation Procedures and Diagnostics

Algorithmic steps for overlap weighting are as follows (Matsouaka et al., 2022, Li et al., 2016, Ben-Michael et al., 2022, Zhou et al., 2020):

1. Fit a (e.g., logistic) propensity score model $e(x;\hat\beta)$ for $P(Z=1|X)$.
2. Compute overlap weights:
   • For treated: $$\hat w_{1,o}(x) = [1 - \hat e(x)]/\sum_{i: Z_i=1} [1-\hat e(x_i)]$$
   • For controls: $$\hat w_{0,o}(x) = \hat e(x)/\sum_{i: Z_i=0} \hat e(x_i)$$
3. Estimate the ATO: $$\hat\tau_o=\sum_{i=1}^N [Z_i\hat w_{1,o}(x_i)Y_i - (1-Z_i)\hat w_{0,o}(x_i)Y_i]$$
4. Obtain sandwich variance via estimating equations or nonparametric bootstrap for accurate confidence intervals.
5. Report $\hat\tau_o$ with 95% CI; compare with ATE-IPW and ATT/ATC as context.

Diagnostics:

• Inspect propensity score histograms for separation.
• Compute effective sample size (ESS) under IPW and overlap weighting.
• If ESS$_{\mathrm{IPW}}\ll N$, but ESS$_{\mathrm{OW}}\approx N$, strongly prefer overlap weighting (Matsouaka et al., 2022).

5. Extensions and Applications: Survival, High-dimensional, and Time-varying Settings

Survival analysis: Overlap weighting generalizes to right-censored outcomes by combining with inverse probability of censoring weighting (IPCW). Estimators for survival curves and restricted mean survival times (RMST) with OW+IPCW display superior bias, efficiency, and empirical coverage compared to IPTW or trimmed IPTW, especially under limited overlap (Cheng et al., 2021, Cao et al., 2023).

Meta-learners and regularization: In high-dimensional/nuisance-rich environments, overlap-adaptive regularization (OAR) penalizes model complexity in proportion to local overlap, enforced via weighted penalties in the second-stage CATE estimator. This prevents overfitting in low-overlap regions and improves precision-recall for heterogeneous treatment effect estimation (Melnychuk et al., 29 Sep 2025).

Dynamic/longitudinal treatment effects: In time-varying (longitudinal) causal inference, overlap weighting is extended via joint probabilities of treatment sequences conditional on history, ensuring robust meta-learners for conditional ATE under diminishing support for certain trajectories (Hess et al., 22 Oct 2025).

6. Network and Combinatorial Interpretations

Beyond causal inference, overlap weights are used in graph theory and string analysis.

• Edge overlap in networks: The "overlap weight" of an edge is the ratio of common neighbors to the union of neighbors (Jaccard-type). Corrected definitions address the bias of classical overlap weights toward small cliques by normalizing with the maximal triangle count, yielding better identification of globally central links (Batagelj, 2019, Choumane, 2020).
• Overlap in weighted networks: Generalizations to edge-weighted graphs employ fuzzy set min/max to define overlap consistently for both weighted and unweighted topologies (Choumane, 2020).
• String/sequence analysis: The "overlap weight" between strings is the length of the maximal overlap (longest suffix of one matching prefix of another), essential for sequence assembly algorithms and applications such as bioinformatics (Cazaux et al., 2018).

7. Practical Considerations, Limitations, and Recommendations

Overlap weighting is recommended whenever positivity/overlap is limited, effective sample size under IPW collapses, or modelers wish to minimize variance and restrict inference to the most empirically identifiable population (Matsouaka et al., 2022, Zhou et al., 2020, Matsouaka et al., 2020). When overlap is moderate/poor, compare overlap-weighted ATO with alternative estimands (ATT, ATE); if they disagree, report ATO as the credible target effect (Ben-Michael et al., 2022). Under correct modeling, overlap-weighted estimates maintain nominal coverage and bias control, even with misspecified propensity models. Practitioners must, however, recognize that the ATO is an internal estimand—it does not represent the entire or the treated/control populations but rather the region with genuine equipoise, which may have distinct scientific or policy interpretations.

In summary, overlap weights operationalize the principle of clinical or empirical equipoise across treatment groups, yielding estimators with optimal efficiency, finite-sample stability, and transparent interpretability. Their adoption is especially critical in modern observational studies, high-dimensional settings, and networked data, where standard weighting methods can be unreliable or uninterpretable due to failings of positivity or the presence of unbalanced or redundant structures (Matsouaka et al., 2022, Li et al., 2016, Zhou et al., 2020, Batagelj, 2019, Melnychuk et al., 29 Sep 2025).