Targeted Maximum Likelihood Estimation as Targeted Updating

Why Update an Outcome Regression?

A fitted outcome regression can predict outcomes well on average without estimating a particular causal contrast accurately. Targeted maximum likelihood estimation (TMLE) addresses this distinction by starting with an initial outcome regression and then updating it in a way chosen for a specified parameter. The update is not an attempt to improve every prediction equally: it is intended to make the fitted outcome regression more informative for the causal quantity being estimated.

For a binary treatment, let \(O=(W,A,Y)\) denote one observation, where \(W\) is a vector of pretreatment covariates, \(A\in\{0,1\}\) is treatment, and \(Y\) is the outcome. Write \(Q_0(a,w)=E(Y\mid A=a,W=w)\) for the initial outcome regression, and \(g_0(w)=P(A=1\mid W=w)\) for the treatment mechanism. In practice, their fitted versions are estimated from observed data. Here the focus is how an initial outcome fit is targeted, not how to build those nuisance fits.

For example, the average treatment effect (ATE) in a target population with covariate distribution \(P_W\) is

An estimate of this parameter can be obtained by predicting each person's outcome under both treatment levels and averaging the predicted differences. TMLE changes the outcome regression before making these predictions. Thus the final estimate remains a plug-in estimate: it is calculated by substituting the updated regression into the parameter formula.

What Makes an Update Targeted?

The update has two essential ingredients. First, a fluctuation model specifies a restricted way to modify the initial regression. Second, the update parameter is estimated from the observed data using a loss or estimating equation related to the target. The fluctuation is deliberately narrower than refitting an unrestricted prediction model: it changes the initial fit along a chosen direction that is relevant to the target parameter.

In a simple additive illustration, let \(H(a,w)\) denote a prescribed update direction and define

The scalar \(\epsilon\) controls the size of the change. One possible fitting rule chooses \(\epsilon\) to make the empirical residual score in this direction equal to zero:

This equation is an illustrative fluctuation, not a universal specification of every TMLE. Actual choices of fluctuation model and fitting loss depend on the outcome type, parameter, and statistical setup. The direction must be chosen to connect the update to the target; an arbitrary adjustment is not made a TMLE merely by calling it a targeting step.

The basic arithmetic behind the additive example follows directly by substitution. Since \(Q_{\epsilon}=Q_0+\epsilon H\), the residual score equals

Setting this expression to zero gives \(\epsilon=\sum_i H_i(Y_i-Q_{0i})/\sum_i H_i^2\), provided \(\sum_i H_i^2>0\). The update therefore removes the empirical residual component in the selected direction. In a parameter-specific TMLE, the direction and fitting procedure are chosen so that this kind of calibration is relevant to the target's estimating information.

Assumptions and Interpretation

Targeting changes estimation; it does not by itself make a causal parameter identifiable. For the ATE to have a causal interpretation from observed data, common assumptions include consistency, conditional exchangeability, and positivity. Consistency says that the observed outcome equals the outcome under the treatment actually received. Conditional exchangeability says that, given measured pretreatment covariates \(W\), treatment assignment is independent of the potential outcomes. Positivity requires that each treatment level have positive probability at covariate values in the target population.

for covariate values \(w\) relevant to that population. These assumptions, together with suitable estimation conditions, support the causal interpretation of the target; the update does not establish that they hold. In particular, a fluctuation cannot create observed outcome information for treatment-covariate combinations that are absent from the data.

It is also important to distinguish population positivity from finite-sample support. If a sample stratum contains no treated participants, that sample cannot directly estimate the treated outcome there without additional modeling or data. This alone does not prove that the population treatment probability is zero in the stratum: a finite sample can omit treated participants even when population positivity holds. If treatment is structurally never assigned in that stratum, however, the ATE there is not identified from the observed data without additional assumptions or a change in target population.

Under appropriate conditions, TMLE is designed to retain important properties of augmented estimators, including double robustness: broadly, consistency may hold if either the outcome regression or the treatment mechanism is consistently estimated, subject to the required identification and regularity conditions. This is not a guarantee against every modeling or data problem. The earlier material on double robustness emphasizes that misspecification, poor overlap, and finite-sample instability can still matter.

Why the Target Parameter Matters

The same initial outcome regression can require different updates for different targets. For the ATE, the plug-in formula averages treatment-specific prediction differences over the target covariate distribution. For the average treatment effect on the treated (ATT), the target averages a contrast over the covariate distribution of treated individuals instead. Since the averaging population changes, the information that the update should emphasize also changes.

This is why an update chosen for one estimand should not automatically be reused for another. The target parameter determines which parts of the outcome regression matter and how the update should be constructed. More generally, targeting can apply to parameters other than a mean contrast, but each target requires a compatible fluctuation and plug-in calculation.

Diagnostics, Alternatives, and Limitations

Targeted updating is useful because it combines an initial outcome model with a parameter-focused adjustment while retaining a plug-in form. But the fitted update and its implications should be checked rather than assumed to work automatically.

• Check support and overlap. Examine treatment probabilities and the representation of treatment levels across important covariate regions. Extreme probabilities can make target-relevant updates unstable; an empty sample cell calls for caution, not an automatic conclusion of structural nonpositivity.
• Check the fluctuation fit. Confirm that the intended empirical score or loss condition is satisfied and that the update converged. A numerical failure or near-zero variation in the update direction may mean the proposed fluctuation is not estimable as specified.
• Check predictions. Ensure updated predictions are plausible for the outcome scale. A simple additive update can produce values outside a known outcome range; a suitable model and link function may be needed for bounded outcomes.
• Check the target definition. Confirm that the empirical averaging distribution matches the intended population. Averaging over all participants when the target is the treated population changes the estimand.

Standardization uses an outcome regression to predict outcomes under alternative treatment levels and averages the resulting contrasts, without necessarily adding a targeted update. AIPW combines outcome regression and treatment weighting through an augmented estimating expression. TMLE also uses outcome and treatment information, but modifies the outcome regression and then evaluates the target as a plug-in parameter. In suitable implementations, the methods can have closely related large-sample behavior while differing in finite-sample estimates, constraints, and computational details.

The update is not a general cure for poor nuisance fits, lack of overlap, unmeasured confounding, or an ill-defined target. A targeted fit may shift the estimate substantially, and that shift should be interpreted alongside the data support and modeling choices. Finally, a fitted score equation does not alone provide valid uncertainty quantification: inference requires appropriate variance estimation and regularity conditions.