Targeted maximum likelihood estimation: Definition & Examples

Targeted maximum likelihood estimation (TMLE) is a framework that updates an initial statistical fit toward a specified target parameter, then evaluates that parameter using the updated fit. In quant research, the target might be the effect of a specified order-routing policy on a binary fill outcome. Accurate fill predictions alone do not guarantee an accurate policy-effect estimate: prediction loss and bias in an effect contrast are different objectives. This distinction motivates the original van der Laan–Rubin construction.

TMLE is a family of estimators. The construction below is a basic binary, single-treatment-time version for an average treatment effect.

Fix the question before targeting the fit

Observe baseline covariates $W$ , treatment $A\in\{0,1\}$ and binary outcome $Y$ . Define

Q(a,w)=\mathbb{E}[Y\mid A=a,W=w],\qquad e(w)=\mathbb{P}(A=1\mid W=w).

Here $e$ is the propensity score. The statistical target is

\psi=\mathbb{E}[Q(1,W)-Q(0,W)].

A causal interpretation as

\mathbb{E}[Y(1)-Y(0)]

additionally requires a well-defined intervention and consistency, conditional exchangeability given pretreatment

W

, and positivity in the target population:

0<e(W)<1

almost surely. The simple notation assumes no interference between units. In a routing study, order interactions or unrecorded urgency can invalidate these assumptions. Targeting cannot repair them. Gruber and van der Laan, §2, distinguish the statistical target from its causal interpretation.

The Conditional Expectation lesson develops the averaging operation used here. Its public page previews the lesson; the full lesson requires membership.

A small regression aimed at one contrast

Fit initial outcome predictions

\widehat Q

and obtain

\widehat e

from an assignment model or a known randomized design. Define the clever covariate

H(a,w)=\frac{a}{\widehat e(w)}-\frac{1-a}{1-\widehat e(w)}.

Fluctuate the predictions along

\operatorname{logit}Q_\epsilon(a,w) =\operatorname{logit}\widehat Q(a,w)+\epsilon H(a,w), \qquad \operatorname{logit}(p)=\log\frac{p}{1-p}.

Fit only

\epsilon

by Bernoulli maximum likelihood, holding the initial logit predictions fixed as an offset, with no additional intercept. Write

\widehat Q^*=Q_{\widehat\epsilon}

. This offset-update mechanism is explained in Rosenblum and van der Laan, §1. Initial probabilities must be strictly inside

(0,1)

for finite logits.

At an interior optimum, the likelihood score gives

\frac1n\sum_i H(A_i,W_i)\{Y_i-\widehat Q^*(A_i,W_i)\}=0.

Then average the two updated predictions over the same baseline population:

\widehat\psi_{\mathrm{TMLE}} =\frac1n\sum_i\{\widehat Q^*(1,W_i)-\widehat Q^*(0,W_i)\}.

This is a targeted plug-in, or substitution, estimate. For binary outcomes the fitted probabilities stay between zero and one, so their average difference stays within

[-1,1]

. Bounded continuous outcomes can use an appropriate rescaling and Bernoulli-type loss; arbitrary unbounded returns cannot be inserted unchanged into this binary construction. See Gruber and van der Laan, §§3.1–3.3.

Six observations: solve the fluctuation

Take six fictional independent observations from a randomized design with known $e=1/2$ , no baseline covariates and initial predictions $\widehat Q(1)=\widehat Q(0)=1/2$ .

Row	$A$	$Y$	$H$
1	1	1	2
2	1	1	2
3	1	0	2
4	0	1	−2
5	0	0	−2
6	0	0	−2

Let $p=\operatorname{expit}(2\epsilon)$ , where $\operatorname{expit}(z)=1/(1+e^{-z})$ . Then $Q_\epsilon(1)=p$ and $Q_\epsilon(0)=1-p$ . Multiplying the six Bernoulli likelihood contributions gives $p^4(1-p)^2$ , hence

\ell(\epsilon)=4\log p+2\log(1-p),\qquad \ell'(\epsilon)=8-12p.

The score vanishes at $p=2/3$ . Moreover, $\ell''(\epsilon)=-24p(1-p)<0$ at every finite $\epsilon$ , so this stationary point is the unique maximum:

\widehat\epsilon=\frac{\log 2}{2}\approx0.3466,\qquad \widehat Q^*(1)=\frac23,\qquad \widehat Q^*(0)=\frac13.

The plug-in effect moves from zero to

2/3-1/3=1/3

, or about 33.33 percentage points. The six values of

H(Y-\widehat Q^*)

, in row order, are

\frac23,\quad\frac23,\quad-\frac43,\quad-\frac43,\quad\frac23,\quad\frac23.

Their sum is exactly zero. Here the updated predictions also equal the two observed group proportions. That coincidence reflects this simple balanced fixture; it does not make general covariate-adjusted TMLE a raw difference of group means.

Fictional six-row example. Starting at epsilon zero, both predictions equal 1/2 and the mean clever score is 1/3. At epsilon ln(2)/2, treated prediction is 2/3, control prediction is 1/3, and the mean clever score is zero.

Original targeting paths for the six fictional observations. The fitted fluctuation sets the score to zero; it does not establish out-of-sample prediction quality or an uncertainty interval.

This example verifies the update and its score equation. Six rows do not establish asymptotic normality, confidence-interval coverage or the true population effect; no interval is inferred here.

How it relates to AIPW and prediction

Using the same

H

, an augmented inverse-probability-weighted (AIPW) estimate based on the initial fit is

\widehat\psi_{\mathrm{AIPW}} =\frac1n\sum_i\left[\widehat Q(1,W_i)-\widehat Q(0,W_i) +H(A_i,W_i)\{Y_i-\widehat Q(A_i,W_i)\}\right].

AIPW adds a residual correction directly to an effect estimate. TMLE updates the outcome fit and then plugs it in. If AIPW instead uses the targeted fit and the targeting score is zero, its correction vanishes, giving the same estimate as TMLE. The original TMLE paper explains this estimating-equation connection. Equality with AIPW using the initial fit is not guaranteed. In this fixture it happens: the initial average correction is

1/3

The update's in-sample likelihood improvement does not certify better held-out predictions. Predictive validation still has its own job.

A zero score is not an inference guarantee

For this target the estimated efficient influence function is

\widehat D_i=H(A_i,W_i)\{Y_i-\widehat Q^*(A_i,W_i)\} +\widehat Q^*(1,W_i)-\widehat Q^*(0,W_i)-\widehat\psi_{\mathrm{TMLE}}.

Targeting makes its empirical mean zero; it does not remove its sampling variation. Under suitable positivity, convergence and regularity conditions, double robustness permits consistency when either the outcome regression or propensity model is consistently estimated. It does not guarantee valid inference whenever one fitted model looks plausible. Rosenblum and van der Laan describe the correctly specified parametric case.

With flexible nuisance learners, standard efficient inference typically needs a negligible remainder: an appropriate product of outcome and propensity estimation errors must be

o_p(n^{-1/2})

, alongside influence-function convergence, finite nonzero variance and control of overfitting. Cross-fitting can address own-observation overfitting; it does not supply the required convergence rates. Propensities near zero or one amplify the clever covariate, and capping them cannot create missing support. Repeated orders and shared market shocks also require dependence-aware estimation and variance theory, rather than treating every row as independent. Kennedy, §§4.1–4.3, develops the expansion, sample splitting and product-error conditions.

Sources

van der Laan, M. J., and Rubin, D. (2006), Targeted Maximum Likelihood Learning, abstract and §1: targeted updates and the connection to efficient estimating equations.
Gruber, S., and van der Laan, M. J. (2012), Targeted Minimum Loss Based Estimator that Outperforms a given Estimator, §§2–3.4: target, influence function, logistic loss and bounded outcomes. Its additional efficiency-enhancing updates are beyond this basic construction.
Rosenblum, M., and van der Laan, M. J. (2010), Estimating Causal Effects Using Targeted Maximum Likelihood Estimation, §1: fixed-offset logistic targeting and substitution estimation.
Kennedy, E. H. (2023), Semiparametric Doubly Robust Targeted Double Machine Learning: A Review, §§4.1–4.3: influence-function corrections, cross-fitting and second-order remainder conditions; version 2, 26 January 2023.

Fix the question before targeting the fit

A small regression aimed at one contrast

Six observations: solve the fluctuation

How it relates to AIPW and prediction

A zero score is not an inference guarantee

Sources

Related terms

Causal overlap and positivity

Cross-fitting