Constructing the Augmented Inverse-Probability-Weighted Estimator

Why Augment Inverse-Probability Weighting?

Inverse-probability weighting estimates an average treatment effect by reweighting observed outcomes according to treatment probabilities. Outcome regression takes a different route: it predicts each unit’s outcome under treatment and under control, then averages the predicted differences. The augmented inverse-probability-weighted (AIPW) estimator combines these ideas. It starts with predicted treatment effects and adds weighted residual corrections that account for the difference between observed outcomes and those predictions.

The augmentation is not simply an extra adjustment to an already-computed effect. Its form follows from an identity for the average treatment effect under standard causal identification assumptions. This tutorial derives that identity, turns it into a sample estimator, and works through the calculation. The focus is constructing and interpreting the estimator, not conducting inference about it.

Setup and Identification Assumptions

For unit \(i\), let \(A_i\) equal 1 if the unit receives treatment and 0 otherwise, let \(X_i\) denote measured pre-treatment covariates, and let \(Y_i\) be the observed outcome. Write \(Y_i(1)\) and \(Y_i(0)\) for the potential outcomes under treatment and control. The target is the average treatment effect (ATE), \(E[Y(1)-Y(0)]\), for the population represented by the data.

The causal interpretation of the estimator relies on assumptions about how treatment is assigned and outcomes are observed. Consistency means the observed outcome is the potential outcome corresponding to the treatment actually received: \(Y=A Y(1)+(1-A)Y(0)\). The usual no-interference condition means one unit’s treatment does not change another unit’s potential outcomes. Conditional exchangeability means \((Y(1),Y(0))\) is independent of \(A\) given \(X\): after conditioning on the measured covariates, treatment assignment carries no additional information about the potential outcomes. Positivity requires \(0<e(X)<1\) throughout the target population, so units with each covariate pattern have a chance of receiving either treatment.

Together, these assumptions identify the ATE from observed data. They do not guarantee that a particular finite-sample estimate is accurate: in practice, propensity scores and outcome regressions must be estimated, and limited overlap can make weighted residual corrections unstable.

Deriving the Augmentation

Begin with the regression-based contrast \(m_1(X)-m_0(X)\). It predicts the difference between treatment and control outcomes for a unit with covariates \(X\). The observed treated outcome \(Y\), when \(A=1\), can provide a correction to the prediction \(m_1(X)\); the observed control outcome, when \(A=0\), can correct \(m_0(X)\). To use those residuals to represent the corresponding population differences, weight them by the inverse probability of observing that treatment status.

With the true propensity score \(e(X)\), conditional exchangeability and consistency imply the two identities below. For the treated residual, conditioning on \(X\) gives \(E[A\{Y-m_1(X)\}/e(X)\mid X]=E[Y(1)-m_1(X)\mid X]\). For the control residual, it gives \(E[(1-A)\{Y-m_0(X)\}/\{1-e(X)\}\mid X]=E[Y(0)-m_0(X)\mid X]\). These equalities follow because the factor \(A/e(X)\) selects treated units and corrects for their conditional treatment probability; the control factor works analogously.

Add the first residual correction to the predicted treatment outcome and subtract the second correction from the predicted control outcome. The outcome-regression predictions then cancel with the corresponding parts of the residual expectations, leaving the conditional mean treatment effect. Averaging over \(X\) gives the ATE.

This is the construction’s central justification: a predicted contrast is paired with weighted residual corrections, and their conditional expectation recovers the conditional causal contrast. Averaging that expression over the covariate distribution targets the population ATE.

Each unit contributes one estimated effect score: a predicted difference plus a treated residual correction or a control residual correction, but not both. The estimator is the ordinary sample average of those scores. In particular, this expression does not divide the treated and control weights by their separate sums; that would be a different normalization.

Calculating the Estimator

Seeing the Treated and Control Corrections Separately

The formula’s signs are easiest to check by writing the two cases separately. A treated unit has score \(\hat m_1-\hat m_0+(Y-\hat m_1)/\hat e\). A control unit has score \(\hat m_1-\hat m_0-(Y-\hat m_0)/(1-\hat e)\). The minus sign on the control residual is essential because the target contrast subtracts the control outcome.

How AIPW Relates to Inverse-Probability Weighting

The AIPW expression can also be rearranged to show the inverse-probability-weighted outcomes explicitly. For a treated unit, the outcome part becomes \(\hat m_1+(Y-\hat m_1)/\hat e\); for a control unit it becomes \(\hat m_0+(Y-\hat m_0)/(1-\hat e)\), with the control quantity subtracted. The predictions thus serve as baselines, while the weighted residuals adjust those baselines using observed outcomes.

Practical Checks and a Common Pitfall

Before interpreting an AIPW estimate, verify that the inputs and calculations match the target and data structure. The outcome regressions must predict the outcome under each treatment status using pre-treatment covariates; the propensity score must model treatment probability given those covariates. Predictions and observed outcomes must use the same outcome scale. Propensity estimates close to zero or one produce large inverse weights, so overlap and weight behavior matter even when every estimated probability is technically within the interval \((0,1)\).

A common mistake is to add the treated and control residual corrections with the same sign, or to use \(1/\hat e\) for both groups. The treated correction uses \(A_i/\hat e(X_i)\), while the control correction uses \((1-A_i)/\{1-\hat e(X_i)\}\) and is subtracted. Another error is to treat the average predicted contrast alone as the AIPW estimate; that omits the augmentation. Conversely, an estimate of the propensity score does not make the estimator automatically causal: the identification assumptions, covariate choice, and overlap still require substantive justification.