AIPW Estimation of the Average Treatment Effect

Why Augment an Average Treatment Effect Estimate?

An average treatment effect compares the mean outcome that would occur under treatment with the mean outcome that would occur without treatment in the same target population. In observational data, treatment groups may differ in their covariates, so a raw difference in observed means need not estimate this causal contrast. The augmented inverse probability weighted estimator, abbreviated AIPW, combines predictions from an outcome model with weighted residual corrections based on a propensity score.

The point of this tutorial is to make the parts of that estimator explicit and show how to calculate them. The estimator uses outcome-regression predictions for both treatment conditions, the estimated probability of the treatment actually received, and the difference between an observed outcome and its corresponding prediction. It does not make causal interpretation automatic: the study design and assumptions still matter.

Assumptions and the Three Components

The estimator is calculated from observed data, but interpreting it as an estimate of a causal ATE requires assumptions. First, consistency means that the observed outcome equals the potential outcome under the treatment actually received: \(Y=A Y(1)+(1-A)Y(0)\). Second, no interference means one individual's treatment does not change another individual's potential outcomes. Third, conditional exchangeability requires \(\{Y(1),Y(0)\}\) to be independent of \(A\) given \(X\); in words, the measured pretreatment covariates are sufficient to remove confounding conditional on those covariates. Finally, positivity requires \(0<e(X)<1\) for covariate values in the target population. Both treatment options must be possible across the covariate patterns whose effects are being averaged.

For the sample calculation, the fitted propensity scores must also be usable: treated observations require \(\hat e(X_i)>0\), and control observations require \(\hat e(X_i)<1\). A specified outcome model produces \(\hat m_1(X_i)\) and \(\hat m_0(X_i)\) for every individual. In practice, the quality of those fits and of the propensity-score fit matters. The formula alone does not ensure either model is adequate, nor does it guarantee stable estimates in a small sample.

The expression inside the square brackets has three interpretable parts. The first, \(\hat m_1(X_i)-\hat m_0(X_i)\), is the outcome-regression contrast: the fitted difference between the two treatment-specific outcomes at that individual's covariates. Averaging this part alone gives a model-based estimate over the sample's covariate distribution.

The second and third parts together are the residual correction. For a treated individual, the correction is the observed-minus-predicted treated outcome, \(Y_i-\hat m_1(X_i)\), multiplied by \(1/\hat e(X_i)\). For a control individual, it is the observed-minus-predicted control outcome, \(Y_i-\hat m_0(X_i)\), multiplied by \(-1/\{1-\hat e(X_i)\}\). Thus the propensity-score component supplies the inverse-probability multipliers, while the residuals determine the size and direction of the adjustments. Only the residual from the treatment actually received appears for each individual.

This decomposition follows directly by grouping the terms in the estimator. If \(B_i=\hat m_1(X_i)-\hat m_0(X_i)\), \(R_{1i}=Y_i-\hat m_1(X_i)\), and \(R_{0i}=Y_i-\hat m_0(X_i)\), then each individual contributes \(B_i+A_iR_{1i}/\hat e(X_i)-(1-A_i)R_{0i}/\{1-\hat e(X_i)\}\). Averaging gives the outcome-regression contrast average plus the treated and control residual-correction averages. If the observed-arm residuals happen to be zero for every individual, both corrections are zero and the estimate equals the average fitted contrast. This is an algebraic statement about the calculation, not a claim that fitted residuals will generally vanish.

How the Propensity Score Changes a Residual Correction

The inverse-probability multiplier is not a multiplier on every outcome or on the fitted contrast. It acts specifically on the observed-arm residual. When a treated person has a small fitted treatment probability, the factor \(1/\hat e(X_i)\) is large. When a control person has a fitted treatment probability close to one, \(1/\{1-\hat e(X_i)\}\) is large. In either case, even a modest residual can make a large adjustment to the estimate.

Diagnostics, Alternatives, and Limits

A practical calculation should be accompanied by checks of the inputs and the resulting contributions. Inspect the distribution of fitted propensity scores separately by treatment group, paying particular attention to values near zero for treated individuals and near one for controls. Also inspect the weighted residual corrections and individual AIPW contributions. An estimate driven by a small number of unusually large contributions deserves investigation, even if the overall average appears plausible.

Check that the outcome predictions are defined for both treatment conditions across the covariate values in the sample, and that the fitted propensity scores correspond to the probability of receiving treatment. Confirm that the coding of \(A=1\) and \(A=0\), the sign of the control correction, and the target population are consistent throughout. A coding error can change the sign or meaning of the estimate without producing an obvious numerical warning.

AIPW differs from outcome regression alone, which averages \(\hat m_1(X_i)-\hat m_0(X_i)\) without residual corrections. It also differs from inverse-probability weighting alone, which uses propensity-score weights to construct a contrast without this outcome-regression augmentation. AIPW includes both fitted contrasts and weighted residuals; neither set of terms should be mistaken for the complete estimator. Matching is another adjustment strategy, with its own construction of comparisons and target. These methods do not automatically share the same finite-sample behavior or target, so comparisons require attention to their estimands and assumptions.

Several limitations remain. Poor covariate measurement can undermine conditional exchangeability; fitted models cannot adjust for unmeasured confounding simply by being combined. Weak overlap can produce highly variable corrections, and the formula does not resolve a lack of support in the data. AIPW is also an estimate, not an uncertainty interval: assessment of sampling variability requires an appropriate inferential procedure that accounts for the data and estimation process. Finally, an estimate outside the outcome's possible range may occur with some outcomes and finite samples; it signals that the estimator is not constrained to that range, not that the arithmetic must be wrong.