TMLE's Targeting Equation and Its Link to AIPW

From an Outcome Update to an Estimating Equation

For a binary treatment and binary outcome, TMLE updates an initial estimate of the outcome regression in a direction chosen to target the average treatment effect (ATE). The update is not just a way to obtain new outcome predictions: its score equation is designed to make the empirical residual part of the efficient influence function equal zero. This gives a precise connection to AIPW while preserving the plug-in form of the TMLE estimate.

Let \(O_i=(X_i,A_i,Y_i)\), \(i=1,\ldots,n\), denote independent observations, with \(A_i\in\{0,1\}\) and \(Y_i\in\{0,1\}\). Write \(g(X)=P(A=1\mid X)\) for the propensity score and \(Q(a,X)=E(Y\mid A=a,X)\) for the outcome regression. The empirical average of a quantity \(f(O_i)\) is \(P_n f=n^{-1}\sum_{i=1}^n f(O_i)\). The target parameter is \(\psi_0=E\{Q(1,X)-Q(0,X)\}\).

Here \(\operatorname{logit}(u)=\log\{u/(1-u)\}\), and \(\operatorname{expit}(v)=1/(1+e^{-v})\). The update moves the prediction upward for treated observations and downward for controls when \(\epsilon\) is positive. To form the plug-in ATE, apply the fitted fluctuation to both treatment levels for every observed \(X_i\): \(\hat Q^\epsilon(a,X_i)=\operatorname{expit}[\operatorname{logit}\hat Q(a,X_i)+\epsilon\{a/\hat g(X_i)-(1-a)/(1-\hat g(X_i))\}]\). Then \(\hat\psi_{\mathrm{TMLE}}=P_n\{\hat Q^{\hat\epsilon}(1,X)-\hat Q^{\hat\epsilon}(0,X)\}\).

Why the Score Makes the Empirical Influence Function Vanish

For the ATE, the efficient influence function evaluated at nuisance functions \(Q\) and \(g\), and at a candidate parameter \(\psi\), is \[ D^*(Q,g,\psi)(O)=\frac{A}{g(X)}\{Y-Q(1,X)\} -\frac{1-A}{1-g(X)}\{Y-Q(0,X)\} +Q(1,X)-Q(0,X)-\psi. \] The first two terms are the weighted outcome residuals; the third is the plug-in contrast. This formula is the bridge between the targeting score and the AIPW estimator.

In the empirical data, the targeting score at the updated regression is exactly the empirical average of the two weighted residual terms. To see this, substitute the definition of \(H_{\hat g}\): for \(A=1\) the score contribution is \(\{Y-\hat Q^{\hat\epsilon}(1,X)\}/\hat g(X)\); for \(A=0\) it is \(-\{Y-\hat Q^{\hat\epsilon}(0,X)\}/\{1-\hat g(X)\}\). Thus, when the score equation is solved and \(\hat\psi_{\mathrm{TMLE}}\) is the empirical plug-in contrast,

The second zero follows from the definition of the empirical plug-in estimate, not from an additional model fit. This is an algebraic finite-sample identity when the targeting score is solved exactly. If numerical optimization stops with a nonzero score, the empirical influence-function average will generally be nonzero by that same amount.

The Exact Link to AIPW

For any fitted outcome regression \(Q\) and propensity score \(g\), the empirical AIPW estimate is the empirical average of the plug-in contrast plus the weighted residual correction. If the outcome regression is the targeted one, the correction is zero by the targeting equation. Consequently, TMLE equals the AIPW expression evaluated at the updated outcome regression and the same propensity score:

The equality follows because the last two terms have empirical average zero. It does not say that TMLE must equal AIPW computed with the initial \(\hat Q\). Updating \(\hat Q\) changes both the plug-in contrast and the residual correction; the targeting score constrains the correction at the updated fit, not at the initial fit. Nor does the identity require the propensity model to be correct as an algebraic matter. Causal interpretation and statistical validity require the relevant identification and estimation conditions.

Conditions and Numerical Diagnostics

The score-to-influence-function identity is direct algebra, but estimating a causal ATE requires more than solving that equation. The usual identification conditions include consistency of the observed outcome with the relevant potential outcome, conditional exchangeability of treatment given \(X\), and positivity: both treatment levels must have positive probability at covariate values represented in the target population. The observations are assumed independent for the simple empirical formulas here. For the binary-outcome logistic fluctuation, the initial predictions should be strictly between 0 and 1, and the estimated propensities should avoid 0 and 1 so that logits and weights are defined and stable.

A finite maximum-likelihood fluctuation estimate is not guaranteed. Complete or near separation can make the logistic likelihood increase toward a boundary rather than attain a finite maximizer. Extreme propensity scores can also make the clever covariate very large, so a small prediction residual may contribute strongly to the score. In computation, check the fitted \(\hat\epsilon\), the range of \(\hat g\), and the post-update score \(P_n[H_{\hat g}\{Y-\hat Q^{\hat\epsilon}(A,X)\}]\). A small score is evidence that the numerical targeting equation was solved; it is not proof that positivity, exchangeability, or the nuisance fits are adequate.

If optimization reports nonconvergence, or the score remains appreciably away from zero, do not describe the empirical influence-function equation as having been made to vanish. Check the likelihood optimization, the prediction bounds, and the propensity values. Altering or truncating extreme weights can improve stability but changes the procedure and may introduce bias; it should not be treated as an automatic cure.

How to Interpret the Connection

AIPW starts with a plug-in contrast and adds a weighted residual correction. TMLE instead updates the outcome regression so that the correction, evaluated at that updated fit, is zero in the observed sample; the resulting plug-in estimate is therefore also an AIPW estimate at the updated fit. AIPW is commonly expressed directly as a one-step correction to an initial regression, whereas TMLE uses a targeted update and retains an estimate in the parameter’s plug-in form. Neither method’s finite-sample identity alone establishes causal identification or guarantees good inference.

The empirical equation is also distinct from a population statement. A zero empirical influence-function average is imposed by the targeting procedure, subject to optimization accuracy. It does not mean the population influence function is identically zero, that the estimator has no sampling variability, or that a confidence interval has been established. Inference requires appropriate variance and nuisance-estimation considerations.