Common TMLE Implementation Errors

Why Implementation Details Matter

TMLE combines an initial estimate of the outcome regression with an update designed to reduce bias for a particular estimand. This targeting step is precise: its covariate, link scale, and predictions must correspond to the estimand and model being fitted. A program can return plausible-looking probabilities and still fail to carry out the intended update—for example, if it targets only one treatment arm or reports the initial regression as though it had been targeted.

This tutorial focuses on practical checks for the binary-treatment average treatment effect (ATE). Let \(A\in\{0,1\}\) be treatment, \(W\) measured covariates, and \(Y\in\{0,1\}\) the outcome. Write \(g(W)=P(A=1\mid W)\) for the propensity score and \(Q(a,W)=E(Y\mid A=a,W)\) for the outcome regression. The target is \(\psi=E\{Q(1,W)-Q(0,W)\}\), interpreted as the population average causal effect under consistency, conditional exchangeability given \(W\), and positivity: \(0<g(W)<1\) for covariate values in the target population. Estimating it also requires suitable nuisance estimation and sampling conditions; targeting does not repair unmeasured confounding or a violation of positivity.

Audit the Clever Covariate Before Fitting

The two terms in \(H(A,W)\) represent the two treatment arms. For a treated observation, \(H=1/g(W)\); for a control observation, \(H=-1/\{1-g(W)\}\). The negative sign is part of the ATE contrast. A frequent coding error is to use only \(A/g(W)\). That quantity is zero for every control, so it cannot contribute the residual information needed to target the control mean.

A useful first diagnostic is to compute the covariate for a treated and a control observation by hand, then compare those values with the values passed to the fluctuation fit. Also check that propensity predictions are aligned to the same rows as \(A,Y,W\), are probabilities rather than logits, and are strictly between zero and one. Clipping extreme estimated propensities may sometimes be a prespecified numerical or statistical choice, but it changes the fitted covariates and should be recorded; it does not establish that positivity holds.

Check the Scale of the Update

For binary outcomes, the usual fluctuation is logistic with the initial prediction as an offset. It moves the prediction on the log-odds scale, then transforms it back to a probability. Adding \(\epsilon H\) directly to a probability is a different model: it can produce values outside \([0,1]\) and is not the logistic fluctuation. Conversely, using a logit update for an unbounded continuous outcome without a suitable bounded-outcome model is a scale mismatch. The fluctuation model must match the outcome scale and the procedure being implemented.

Keep Initial, Targeted, and Counterfactual Predictions Distinct

The initial regression predictions \(Q(1,W_i)\) and \(Q(0,W_i)\) are inputs to the update. The targeted predictions \(Q^*(1,W_i)\) and \(Q^*(0,W_i)\) are the updated quantities used in the plug-in estimate. For the signed, one-parameter logistic fluctuation, compute both counterfactual predictions using the same fitted \(\epsilon\): \[ \operatorname{logit}Q^*(1,W_i)=\operatorname{logit}Q(1,W_i)+\epsilon/g(W_i), \qquad \operatorname{logit}Q^*(0,W_i)=\operatorname{logit}Q(0,W_i)-\epsilon/\{1-g(W_i)\}. \] Do not update only the prediction corresponding to the observed treatment and then treat it as though both counterfactual predictions were available. Each individual contributes both targeted counterfactual predictions to the plug-in ATE.

Different valid fluctuation parameterizations exist. For example, an arm-specific update can use separate coefficients and arm-specific clever covariates. Its predictions and coefficient structure differ from a one-parameter signed update. Code should implement one stated parameterization consistently: the covariate, the fitted coefficient or coefficients, the counterfactual update, and the final estimate must all match. Combining the covariate from one parameterization with the prediction formula from another is not a harmless change in notation.

What the Score Check Does—and Does Not—Establish

For a Bernoulli outcome and the logistic offset model, let \(Q_\epsilon(A,W)=\operatorname{expit}\{\operatorname{logit}Q(A,W)+\epsilon H(A,W)\}\). The derivative of the log likelihood with respect to \(\epsilon\) is \[ \frac{\partial \ell(\epsilon)}{\partial\epsilon} =\sum_{i=1}^{n}H(A_i,W_i)\{Y_i-Q_\epsilon(A_i,W_i)\}. \] Consequently, if the likelihood has a finite interior maximizer, its derivative there is zero and the empirical targeting equation holds: \[ \frac{1}{n}\sum_{i=1}^{n}H(A_i,W_i)\{Y_i-Q^*(A_i,W_i)\}=0. \] This is a useful implementation check: evaluate the score using the same rows, covariates, and targeted observed-treatment predictions used in the fit. It should be close to zero up to numerical tolerance when the stated conditions hold.

A zero score is not guaranteed. A finite maximizing coefficient may fail to exist, including under separation. For instance, consider one treated observation with \(Y=1\), one control observation with \(Y=0\), initial predictions both \(0.50\), and propensity scores both \(0.50\). Their signed covariates are \(2\) and \(-2\). At every finite \(\epsilon\), the treated fitted probability is below 1 and the control fitted probability is above 0, so both likelihood contributions increase as \(\epsilon\) increases. The likelihood approaches its supremum as \(\epsilon\) tends to positive infinity; no finite coefficient attains it and the finite score is not zero. A program that reports convergence merely because an optimizer stopped is not evidence of a finite interior solution.

When a fit fails or the coefficient grows without bound, record the failure and inspect the score, fitted probabilities, and optimizer status. A penalty or other numerical safeguard may be used only if it is explicitly specified and appropriate to the analysis; it changes the fitting procedure, so do not report the unmodified score equation as exactly solved. Likewise, extreme clever covariates can signal practical positivity problems. Truncation or stabilization should be documented rather than silently inserted as a coding fix.

Cross-Validation and a Compact Audit

In cross-validated TMLE, predictions used for targeting and estimation must respect the chosen fold structure. If nuisance fits generate out-of-fold predictions, retain the row-to-fold mapping and use the designated out-of-fold predictions consistently; do not accidentally replace them with in-sample predictions in one step of the calculation. The targeting fit, counterfactual predictions, and aggregation must follow the particular cross-validated TMLE construction being used. Mixing fold-specific predictions with an unrelated full-sample update can yield a number without implementing that estimator.

Before interpreting an estimate, check the following in order:

• Confirm the treatment coding, outcome coding, row alignment, and definition of the target population.
• Recompute \(H(A_i,W_i)\) for selected treated and control observations; check both signs and both denominators.
• Verify the offset is on the intended scale, and confirm all binary-outcome predictions remain probabilities after transformation.
• Check that the fluctuation fit returned a finite estimate and inspect convergence and score diagnostics rather than relying only on a software success flag.
• Generate both targeted counterfactual predictions for each target-population row, then calculate the plug-in contrast from those predictions.
• Keep cross-validation folds and prediction provenance intact; compare the reported estimate with a separately coded calculation from stored predictions.

TMLE's targeting step is distinct from simply reporting the initial outcome-regression estimate and from AIPW's direct use of an influence-function correction. Those approaches can be compared as separate estimators, but neither is a reason to skip the TMLE update when reporting a TMLE estimate. Targeting also does not guarantee good finite-sample behavior: poor overlap, unstable nuisance fits, or a failed fluctuation can remain consequential. Variance estimation and confidence intervals require their own procedures and are not established just by a successful targeting score check.