Inverse Probability of Treatment Weighting

• Inverse Probability of Treatment Weighting (IPTW) is a causal inference method that reweights subjects by the inverse of their treatment probabilities to construct a balanced pseudo-population.
• It computes normalized weights to match covariate distributions across treated and control groups, ensuring self-normalization and improved finite-sample efficiency.
• Enhanced with covariate balancing techniques and doubly robust estimators, IPTW unifies approaches for estimating both average treatment effects and treatment effects on the treated.

Inverse Probability of Treatment Weighting (IPTW) is a central methodology in the identification and estimation of causal effects from observational data. By constructing a pseudo-population in which measured covariates are independent of treatment assignment, IPTW enables consistent estimation of population-level or local causal estimands, such as the average treatment effect (ATE) and the average treatment effect on the treated (ATT), under the core assumptions of unconfoundedness and positivity. The method’s ongoing evolution incorporates robust weighting schemes, covariate balancing techniques, sensitivity analysis, and generalizations to complex longitudinal and recurrent-event data structures.

1. Principles and Definitions of IPTW

Let $T_i\in\{0,1\}$ denote the binary treatment indicator, $Y_i$ the observed outcome, and $X_i$ a vector of covariates for subject $i$. The propensity score is $e(X_i)=P(T_i=1|X_i)$. The classic inverse-probability-of-treatment weights for estimating the ATE reweight treated subjects by $1/e(X_i)$ and controls by $1/(1-e(X_i))$, generating for each unit: $w_i = \frac{T_i}{e(X_i)} + \frac{1-T_i}{1-e(X_i)}.$ Normalized (Hajek) weights further divide these by their group totals, enforcing finite-sample stability and exact covariate balance moments in the sample (Słoczyński et al., 2023).

The IPTW estimator for the ATE is then: $$\hat\tau_{IPW} = \frac{1}{n}\sum_{i=1}^{n}w^*_i(1)Y_i - \frac{1}{n}\sum_{i=1}^{n}w^*_i(0)Y_i,$$ where $w^*_i(a)$ are normalized weights for treatment $Y_i$0.

For ATT, the estimator uses reweighted controls to match the covariate distribution of the treated: $Y_i$1 with weights chosen so that the reweighted covariate mean for controls matches that of treated units (Słoczyński et al., 2023).

2. Construction and Properties of Covariate-Balancing Weights

Covariate balancing propensity score (CBPS) and inverse probability tilting (IPT) methods select propensity score models by directly solving moment conditions to balance covariate distributions rather than relying on maximum likelihood. For the ATE, IPT imposes: $Y_i$2 where $Y_i$3. These moment conditions ensure self-normalization: $Y_i$4 (Słoczyński et al., 2023).

For the ATT, CBPS enforces: $Y_i$5 providing automatic normalization of the resulting weights (Słoczyński et al., 2023).

These approaches guarantee that covariate balance holds exactly in the sample, improving finite-sample efficiency and reducing sensitivity to PS model misspecification (Słoczyński et al., 2023).

3. Doubly Robust Estimation and Equivalences

Doubly robust estimators—such as the augmented inverse probability weighting (AIPW) and IPW regression adjustment (IPWRA)—combine propensity score weighting with regression adjustment for outcomes. When the propensity score is estimated by IPT (for ATE) or CBPS (for ATT) and the outcome regression is linear, all three estimators (IPW, AIPW, IPWRA) are exactly equivalent: $Y_i$6

$Y_i$7

and similarly for $Y_i$8 (Słoczyński et al., 2023). This algebraic identity arises because covariate balancing weights enforce the sufficient balance that makes all three estimating equations equivalent.

This equivalence implies the unification of the major approaches and indicates that, in large samples and with suitable covariate balancing weights, the issue of which specific "doubly robust" estimator to use becomes moot (Słoczyński et al., 2023).

4. Practical Consequences: Self-Normalization, Estimator Unification, and Robustness

Self-normalization arises because the balancing constraints ensure the sum of weights matches (sub)sample sizes, so normalized and unnormalized IPW (or AIPW) coincide. Estimator unification eliminates the ambiguity in choosing among semiparametric estimators or in selecting normalizing constants.

Covariate balancing estimators inherit the double robustness familiar from AIPW: consistent estimation is achieved if either the propensity model or the linear outcome model is correct, not necessarily both (Słoczyński et al., 2023). In practice, covariate balancing methods (such as IPT and CBPS) enable

• Automatically normalized weights (algebraic normalization by construction)
• Elimination of ambiguity among different semiparametric estimators for ATE/ATT
• Preservation of consistency and efficiency under linear working models, even when the outcome regression may be misspecified (Słoczyński et al., 2023).

5. IPTW Workflow and Core Assumptions

The standard IPTW estimator workflow is:

1. Estimate the propensity score $Y_i$9 (preferably via IPT/CBPS for covariate balance).
2. Compute (normalized or unnormalized) inverse probability weights for treated/controls.
3. Form weighted means for potential outcomes in each arm.
4. Take differences to estimate ATE or ATT.

Key identification and estimation assumptions for IPTW are:

• Unconfoundedness/exchangeability: $X_i$0
• Positivity: $X_i$1 for all $X_i$2 in the empirical sample (prevents extreme or undefined weights)
• Correct specification or balancing of the propensity score, especially when aiming for equivalence among IPW, AIPW, and IPWRA (Słoczyński et al., 2023).

6. Relationship to Alternative Weighting and Robustness Methods

The algebraic equivalence results position IPT/CBPS-based IPTW at the center of a broad class of doubly robust estimators. The use of balancing weights stands in contrast to maximum-likelihood-based estimation of the propensity score, which does not guarantee sample covariate balance and, in finite samples, yields less efficient or less robust estimators.

In simulation and empirical study, balancing-based weights provide improved finite-sample efficiency, reduced variance, and diminished sensitivity to moderate PS model misspecification relative to ML-based weighting (Słoczyński et al., 2023).

Weighting Approach	Key Property	Consequence
ML-based PS	May not balance $X_i$3	Larger variance, possible bias
IPT/CBPS PS	Balances $X_i$4 in sample	Equivalence of IPW/AIPW/IPWRA estimators

Practical guidance advocates the use of covariate balancing propensity score methods—namely inverse probability tilting for ATE and CBPS for ATT—whenever attainable, as they unify the major semi-parametric estimators of average treatment effects, deliver automatically normalized weights, and guarantee exact sample covariate balance (Słoczyński et al., 2023).