Deriving AIPW from Outcome Regression and Weighting

Why Add a Residual Correction?

Outcome regression and inverse-probability weighting organize the same causal comparison in different ways. Outcome regression predicts each person's outcome under each treatment and averages those predictions over the covariate distribution. Weighting instead uses observed outcomes, giving more weight to observations whose treatment assignment was less likely given their covariates. The augmented inverse-probability-weighted (AIPW) estimator combines these ideas: it starts with treatment-specific regression predictions and adds weighted residuals, the differences between observed outcomes and the predictions for the treatment actually received.

The goal here is to derive that estimating expression for the average treatment effect (ATE). The earlier material on AIPW estimation of the average treatment effect introduced its use; the derivation makes clear which term comes from regression, which comes from weighting, and why the correction has its particular form.

Assumptions and Identification

The derivation concerns identification of the population ATE before considering how to estimate it from a sample. It requires consistency, conditional exchangeability, and positivity. Consistency means that the observed outcome equals the potential outcome under the treatment actually received: \(Y=Y(A)\). Conditional exchangeability means \((Y(1),Y(0))\perp A\mid X\); after conditioning on the measured covariates, treatment assignment is independent of both potential outcomes. Positivity means \(0<e(X)<1\) for covariate values in the target population. These are causal identification assumptions, not properties that can be established merely by fitting a regression or checking a balance summary.

Under consistency and conditional exchangeability, the conditional mean outcome among people with \(A=a\) identifies the conditional mean potential outcome under \(a\). Consequently, for each treatment level,

Thus, the regression route to the ATE is \(E[m_1(X)-m_0(X)]\). The expectation averages over the covariate distribution of the target population, rather than comparing only the treated and control groups' unadjusted outcome averages.

From Weighting to the Augmented Expression

Weighting gives another representation of the same treatment-specific mean. Write \(I(A=a)\) for the indicator that equals 1 when \(A=a\) and 0 otherwise. If the probability of receiving treatment \(a\), conditional on \(X\), is \(p_a(X)\), then conditioning on \(X\) gives

The equality uses \(p_a(X)=P(A=a\mid X)\), which requires the propensity model used in this identity to represent the actual treatment probability. Taking expectations over \(X\) shows that the inverse-probability-weighted outcome also has expectation \(E[m_a(X)]\).

To obtain the augmented expression, start with the regression prediction \(m_a(X)\) and add a weighted residual for observations who received treatment \(a\). When \(m_a(X)\) is the true conditional mean, that residual correction has conditional expectation zero:

The second equality follows because \(m_a(X)=E[Y\mid A=a,X]\). Therefore, adding the correction preserves the target mean under the stated conditions. Equivalently, combining the regression prediction and the weighted residual gives a quantity whose expectation is \(E[m_a(X)]\):

This derivation explains the two parts of the estimator. The first part supplies a prediction for everyone in the target population. The second part uses observed outcomes to correct those predictions, with each residual scaled by the probability of the treatment actually received. For the ATE, subtract the control expression from the treatment expression. The resulting sample estimator is

Here \(\hat{m}_a(X_i)\) and \(\hat{e}(X_i)\) are fitted values for observation \(i\). The minus sign before the control residual term is essential: the estimator subtracts the estimated control mean from the estimated treatment mean. This tutorial derives the expression under correctly represented outcome and treatment models; a fuller analysis of what happens when a nuisance model is misspecified is a separate question.

Calculating the Sample Estimate

In practice, the population functions are unknown. An analyst fits an outcome model to obtain \(\hat{m}_1\) and \(\hat{m}_0\), and a treatment model to obtain \(\hat{e}\), then substitutes those fitted values into the estimating expression. The following hypothetical data illustrate the arithmetic. Assume the listed fitted values came from models and that the estimated propensity scores are positive and below 1 for every row.

Diagnostics, Alternatives, and Limitations

The algebra is simple, but its practical behavior depends on fitted values and the data. Before interpreting an estimate, inspect whether the estimated propensity scores approach 0 or 1. A small \(\hat{e}(X_i)\) can make a treated residual receive a large multiplier; a small \(1-\hat{e}(X_i)\) can do the same for a control residual. Such observations can make the estimate highly sensitive to a few outcomes. Also inspect the outcome-model predictions and residuals by treatment and covariate patterns, and check covariate balance using appropriate diagnostics. A good-looking fitted outcome model or propensity score alone does not verify exchangeability, positivity, or correct model specification.

AIPW is not a cure for unmeasured confounding, lack of overlap, or poor data. Positivity violations cannot be repaired by the formula: when some covariate profiles have essentially no chance of receiving one treatment, the data provide little information about that treatment's outcome for those profiles. Truncating extreme weights may reduce instability but changes the estimator and can introduce bias. Likewise, a finite-sample estimate can be unstable even when the population assumptions are plausible. Uncertainty quantification requires an appropriate variance procedure; the point-estimate derivation alone does not supply a standard error or confidence interval.

A common mistake is to read the word “augmented” as meaning that the estimate is automatically protected from every modeling or data problem. The derivation establishes why the correction has its form under the stated identification conditions; it does not guarantee that the fitted models are adequate, that all confounders were measured, or that extreme weights are harmless. Assess overlap, inspect influential weighted residuals, and keep the target population and causal assumptions explicit.