Normalized Weights and the Hájek Estimator

Why Normalize Inverse Probability Weights?

Inverse probability weighting uses each observation’s treatment status and propensity score to construct a weighted estimate of a potential-outcome mean. In a finite sample, however, the weights assigned to a treatment group need not sum to the sample size. The resulting weighted outcome total can therefore reflect both the outcome values and a random discrepancy in the total weight. The Hájek estimator divides by the observed sum of weights, turning that weighted total into a weighted mean.

This normalization is a finite-sample choice, not a change to the intended causal estimand. Under suitable assumptions, both the normalized and unnormalized estimators target the same population mean. They can nevertheless differ in a particular sample, and neither is automatically reliable when propensity scores produce extreme weights or important covariate imbalance.

The weights are nonnegative when the fitted propensity scores lie strictly between zero and one. Within each treatment group, the normalized weights sum to one. Consequently, the Hájek estimate is a weighted average of the observed outcomes in that group. The ATE estimate is the difference between the two normalized weighted means, \(\widehat{\tau}_{H}=\widehat{\mu}_{1,H}-\widehat{\mu}_{0,H}\).

Assumptions and the Normalization Step

Let \(Y_i(1)\) and \(Y_i(0)\) be the potential outcomes, so consistency gives \(Y_i=A_iY_i(1)+(1-A_i)Y_i(0)\). Let \(X_i\) denote measured pre-treatment covariates and \(e(X_i)=P(A_i=1\mid X_i)\). The usual identification conditions are consistency, conditional exchangeability, \(Y(a)\perp A\mid X\) for \(a=0,1\), and positivity, \(0<e(X)<1\) in the target population. In practice, the propensity-score model must also be adequate enough to estimate \(e(X)\) consistently, or otherwise provide suitable weights for the target estimand.

For known propensity scores, exchangeability implies that, conditional on \(X\), the treated outcome \(Y\) represents \(Y(1)\), and the untreated outcome represents \(Y(0)\). In particular,

These identities motivate the unnormalized inverse probability weighted means: each weighted outcome total is divided by \(n\). The Hájek estimator instead divides each arm’s weighted outcome total by that arm’s own total weight. With a correctly specified propensity score, the expected weight in each arm is one per observation in the population, so the two denominators have the same large-sample target after scaling. In any particular finite sample, they need not be equal.

The ratio form also gives a direct justification for consistency. For arm \(a\), write \(U_{ai}\) for the inverse-weighted outcome contribution and \(W_{ai}\) for its corresponding weight contribution. If the observations are independent and identically distributed, the relevant expectations are finite, and \(E[W_{ai}]>0\), sample averages converge to their expectations. Thus,

Under the identification conditions, this ratio is the target potential-outcome mean. With estimated propensity scores, the same conclusion requires suitable consistency of their estimation and regularity conditions for the weighted sums. Normalizing weights does not repair a misspecified propensity model or a failure of positivity.

Calculating Normalized Weights

For the ATE, the normalization is performed separately in the two treatment groups. The treated and untreated weighted means need not use the same total weight. This is not an error: each denominator normalizes the weights used to estimate its own potential-outcome mean.

Finite-Sample Behavior: A Comparison

For the ATE, the unnormalized inverse probability weighted estimator is

The Hájek version replaces the common denominator \(n\) with the observed sum of weights in each arm. If a group’s total weight is larger than expected, its unnormalized outcome total can be correspondingly large even when its weighted outcome average is moderate. Normalization removes that scale variation by dividing by the group’s total weight.

This finite-sample difference does not imply that Hájek estimates are always closer to the truth. A ratio of random quantities is generally not exactly unbiased, even if its numerator and denominator each have favorable expectation properties. With independent observations, finite second moments, and \(E[W]>0\), a first-order expansion of a generic ratio \(\bar U/\bar W\) around \(E[U]/E[W]\) gives the approximate bias

This approximation shows why the ratio can have finite-sample bias: variability in the denominator and its covariance with the numerator matter. At the same time, by eliminating random variation in the overall scale of the weights, normalization can reduce variance or improve stability in some settings. The balance depends on the data-generating process; no universal finite-sample ranking follows from normalization alone.

Diagnostics, Limitations, and Alternatives

Normalization addresses the scale of weights, not their extremity. A very small estimated propensity score among treated people, or a very large one among untreated people, can create a large inverse probability weight. The resulting normalized weight may dominate its arm’s estimate. Useful checks include inspecting propensity-score ranges by treatment group, examining the largest weights and their contribution to the total, assessing weighted covariate balance, and checking that the sum of weights in each arm is not near zero. Effective sample size, often summarized as \((\sum_i w_i)^2/\sum_i w_i^2\) within an arm, can flag concentration: it decreases when a few observations carry most of the weight. It is a diagnostic summary, not a replacement for checking balance or overlap.

The weighted means are sensitive to the propensity-score model and to lack of overlap. If positivity fails, normalization cannot create information about outcomes in covariate regions where one treatment is essentially absent. Trimming or modifying weights may improve stability but can change the target population or estimand; such changes must be stated rather than treated as a purely computational fix. The earlier material on Estimating Weighted Potential-Outcome Means covers the broader calculation of weighted means; the focus here is the finite-sample normalization step.

The unnormalized estimator has a direct weighted-total form and, with known probabilities under the usual sampling setup, has an unbiasedness property for the relevant population total or mean. The Hájek estimator is a ratio and is generally not exactly unbiased in finite samples, but its scale is anchored by the observed weight total. Which is preferable depends on the estimand, sampling design, and finite-sample behavior. Outcome regression and matching are other approaches to adjustment; they rely on their own modeling or design choices and are not made unnecessary by normalized weighting.