Choosing Between AIPW and TMLE for an Analysis

Choosing an Estimator Is a Design Decision

AIPW and TMLE can target the same average treatment effect (ATE) using the same estimated outcome regression and propensity score. Their finite-sample estimates may differ, but neither is automatically more robust, more accurate, or more appropriate in every analysis. A defensible choice depends on the estimand, the outcome scale, the nuisance-learning strategy, the observed overlap, and the inferential plan.

The choice is made after specifying the causal question and checking whether the data support it. Neither estimator repairs unmeasured confounding, a lack of treatment overlap, or an outcome model that does not correspond to the stated estimand. The earlier material on AIPW and TMLE estimating equations provides useful diagnostics; here the emphasis is on using such considerations to decide which estimator to use.

Setup and assumptions: Let \(A\in\{0,1\}\) be treatment, \(Y\) the outcome, and \(W\) measured pre-treatment covariates. Define \(Q(a,W)=E[Y\mid A=a,W]\) and \(g(W)=P(A=1\mid W)\). For the ATE, the target is \(\psi=E[Q(1,W)-Q(0,W)]\). A causal interpretation requires consistency, conditional exchangeability \(Y(a)\perp A\mid W\), and positivity \(0<g(W)<1\) for covariate values in the target population. Independent sampling and suitable regularity conditions are also needed for the usual large-sample inference.

These assumptions are shared by AIPW and TMLE. In particular, “doubly robust” does not mean robust to unmeasured confounding: it refers to protection against misspecification of one of the two nuisance functions, under the identification assumptions and suitable estimation conditions. Severe practical positivity violations can still make either estimator unstable.

What the Estimators Have in Common

For observations \(i=1,\ldots,n\), let \(\hat Q(a,W_i)\) and \(\hat g(W_i)\) denote estimated nuisance functions. The AIPW estimator combines a model-based outcome contrast with weighted residual corrections:

\hat\psi_{\mathrm{AIPW}} = \frac{1}{n}\sum_{i=1}^n \left[ \hat Q(1,W_i)-\hat Q(0,W_i) +\frac{A_i}{\hat g(W_i)}\{Y_i-\hat Q(1,W_i)\} -\frac{1-A_i}{1-\hat g(W_i)}\{Y_i-\hat Q(0,W_i)\} \right].

TMLE begins with an initial outcome regression, updates it using information from the treatment mechanism, and then averages the updated predicted contrasts. In notation, its estimate has the plug-in form

\hat\psi_{\mathrm{TMLE}} = \frac{1}{n}\sum_{i=1}^n \{\hat Q^*(1,W_i)-\hat Q^*(0,W_i)\},

where \(\hat Q^*\) is the targeted outcome regression. For binary outcomes, an appropriate logistic fluctuation can keep updated predictions between zero and one. AIPW and TMLE often have similar large-sample behavior when they use comparable nuisance estimates and satisfy the conditions for asymptotic linearity. They need not agree in a finite sample.

Why Double Robustness Does Not Decide the Choice

A short expectation argument shows why both approaches are associated with double robustness. Consider the population version of the AIPW expression, with candidate outcome regression \(\bar Q\) and propensity score \(g\). If \(\bar Q=Q\), both residual terms have conditional expectation zero, so the expected expression is \(E[Q(1,W)-Q(0,W)]=\psi\), even if the propensity model is wrong, provided its weights are defined.

Alternatively, suppose the propensity score is correct but \(\bar Q\) is not. Conditional on \(W\), the treated residual correction has expectation \(Q(1,W)-\bar Q(1,W)\), and the control residual correction before its minus sign has expectation \(Q(0,W)-\bar Q(0,W)\). Thus the conditional expectation of the full expression is

\begin{aligned} &\bar Q(1,W)-\bar Q(0,W) +Q(1,W)-\bar Q(1,W) -\{Q(0,W)-\bar Q(0,W)\}\\ &\qquad=Q(1,W)-Q(0,W). \end{aligned}

A corresponding robustness property holds for TMLE under the conditions supporting its targeting and estimation procedure. This argument explains a shared protection; it does not guarantee good finite-sample performance, valid confidence intervals in every setting, or resistance to extreme weights. The choice between the estimators is therefore about implementation and analysis goals, not a claim that one has a universally stronger identification argument.

Worked Example: Binary Outcome With Natural Bounds

Suppose a hypothetical analysis estimates an ATE for a binary outcome, so the outcome lies in \([0,1]\). The analyst has fitted a logistic outcome regression and is considering whether a targeted logistic update is a good fit for the analysis.

For one hypothetical covariate profile, suppose the targeted linear predictors are \(-1.3863+0.05(10)\) under treatment and \(1.3863-0.05(1.1111)\) under control. The logistic inverse link is \(\operatorname{expit}(x)=1/(1+e^{-x})\). Substitution gives

\begin{aligned} \hat Q^*(1,W) &=\operatorname{expit}(-1.3863+0.05(10))\\ &=\operatorname{expit}(-0.8863)\approx0.2919,\\ \hat Q^*(0,W) &=\operatorname{expit}(1.3863-0.05(1.1111))\\ &=\operatorname{expit}(1.330745)\approx0.7910. \end{aligned}

The updated contrast for this profile is approximately \(0.2919-0.7910=-0.4991\), using the unrounded predictions for the subtraction. This is a profile-specific predicted contrast, not by itself the population ATE; the ATE estimate averages updated contrasts over the analysis sample.

A TMLE using this link has a practical advantage when valid predicted risks are important: the predictions remain within \([0,1]\), and their average contrast lies in \([-1,1]\). That is a reason to consider TMLE for a binary outcome, not proof that it will be more precise or less biased. AIPW may still be a reasonable choice, but its residual-corrected estimate is not constrained to the risk-difference bounds in every finite sample. The analyst must also check that the propensity estimates and targeting step are stable.

Four Criteria for a Practical Choice

Analysis feature	Reason to consider TMLE	Reason to consider AIPW	What neither method resolves
Outcome bounds	A compatible targeting link can keep updated predictions within the outcome range and produce a bounded plug-in estimate.	AIPW can be straightforward when the outcome scale is not naturally bounded or when a residual-correction estimate is the intended implementation.	Misspecified outcome support, poor measurement, or a causal estimand that was defined incorrectly.
Nuisance learning	Useful when the analysis is organized around updating initial predictions and targeting the parameter, including with data-adaptive fits.	Useful when a transparent estimating-equation implementation and direct residual diagnostics are priorities.	Inadequate nuisance estimates, leakage between training and evaluation data, or failure to account for data-adaptive fitting.
Finite-sample behavior	The plug-in form can respect natural bounds, but the update can still be unstable when weights are extreme.	Often simpler to compute and inspect; residual corrections can nevertheless be large and the estimate can exceed natural bounds.	Weak overlap, sparse treatment groups, or high-variance inverse-probability contributions.
Inference	Can be paired with influence-function-based standard errors and confidence intervals under suitable conditions.	Can also use influence-function-based inference under suitable conditions.	Nonregular behavior, inadequate sample size, or invalid standard errors caused by unaddressed nuisance-estimation issues.

A flexible nuisance-learning strategy does not by itself favor one estimator. Both approaches can use estimated propensity scores and outcome regressions produced by data-adaptive algorithms. The relevant questions are whether the fitting and evaluation scheme is appropriate, whether the nuisance estimates support the required rate and regularity conditions, and whether the targeting or correction step is implemented correctly. Cross-fitting can help control overfitting-related empirical-process problems, but it does not create overlap or guarantee valid inference.

Worked Example: Extreme Propensity Scores and Finite-Sample Stability

Consider a hypothetical sample of 200 people. Suppose 8 people, or \(8/200=0.0400\) (4.00%), have estimated propensity scores below 0.02. For any treated person with \(\hat g(W_i)<0.02\), the treated residual multiplier \(1/\hat g(W_i)\) exceeds \(1/0.02=50\). For a control person with \(\hat g(W_i)>0.98\), the control multiplier \(1/\{1-\hat g(W_i)\}\) likewise exceeds 50.

This pattern is a warning to inspect the actual treatment assignments, covariate profiles, and residual contributions. It is not enough to count the 8 scores: large weights matter most when they occur in observations whose residuals are influential. If several such observations contribute heavily, AIPW can be highly variable. TMLE is not automatically protected, because its targeting step also depends on the treatment mechanism and can become unstable under poor overlap.

The decision is therefore not “choose TMLE because weights are extreme.” First assess whether the target population and positivity assumption are credible. Then inspect the influence-function contributions, consider whether a restricted target population is scientifically appropriate, and report any prespecified sensitivity analysis for weight handling. Truncating or otherwise modifying propensity estimates changes the procedure and may change its bias; it should not be treated as a harmless numerical fix. If poor overlap is severe, neither estimator can recover precise information the data do not contain.

Match the Method to the Inferential Goal

For either estimator, a common large-sample approach estimates the standard error from the empirical variability of estimated influence-function values. In a simple independent-sample setting, if \(\hat D_i\) denotes the estimated influence-function value for observation \(i\), the standard error is approximately the sample standard deviation of the \(\hat D_i\)'s divided by \(\sqrt n\). This requires appropriate conditions, including adequate nuisance estimation and a valid approximation by the influence function. The calculation does not make a confidence interval reliable when those conditions fail.

When the goal is a risk difference or another bounded-outcome contrast, preserving the outcome range can make TMLE’s plug-in estimate easier to interpret. When a direct estimating-equation implementation, transparent residual contributions, or a particular existing analysis pipeline is central, AIPW may be convenient. If the primary goal is inference, the estimator label alone is not enough: specify the variance method, account for the actual sampling structure, and assess whether the standard-error approximation is credible.

Worked Example: Similar Estimates, Similar Inference

Suppose a hypothetical study estimates a risk difference in the probability of an adverse outcome. The prespecified analysis uses the same covariates and compatible nuisance-learning procedures for both methods. Assume the causal assumptions are plausible, overlap diagnostics show no influential extreme contributions, and influence-function-based standard errors are considered appropriate. The two analyses report:

AIPW estimate: \(0.120\), with \(SE=0.045\).
TMLE estimate: \(0.115\), with \(SE=0.043\).

For illustration, a normal-approximation 95% confidence interval is the estimate plus or minus \(1.96\) standard errors. For AIPW,

0.120\pm1.96(0.045) =0.120\pm0.0882 =[0.0318,\ 0.2082].

For TMLE,

0.115\pm1.96(0.043) =0.115\pm0.0843 =[0.0307,\ 0.1993].

Both intervals are on the risk-difference scale; under the stated coding, the estimates correspond to estimated increases of 0.120 and 0.115 in outcome probability, or 12.0 and 11.5 percentage points. These illustrative intervals are conditional on the stated standard-error method and its assumptions. Their overlap is not a formal test that the estimators are equivalent. Because the estimates and uncertainty are similar, the analyst should base the final choice on the prespecified outcome-scale and implementation goals, and report the other estimate as a useful robustness comparison if appropriate.

Diagnostics, Limitations, and a Decision Rule

A practical selection process begins with the scientific target and ends with diagnostics, not with a preference for a familiar estimator. The following sequence makes the reasoning explicit.

Confirm the target and identification conditions.
State the estimand, covariates, and causal assumptions. If exchangeability or positivity is implausible, neither estimator alone addresses that problem.

Consider the outcome scale.
For binary or otherwise bounded outcomes, decide whether respecting bounds in the reported plug-in estimate is important and whether a compatible targeting link is appropriate.

Plan nuisance estimation and inference.
Specify how the outcome and propensity functions will be estimated, whether cross-fitting is needed, and how uncertainty will be estimated under the sampling design.

Inspect stability and report limitations.
Examine propensity-score overlap, influential residual or influence-function contributions, targeting convergence, and sensitivity to defensible analytic choices. Explain what these checks cannot establish.

Diagnostics that should prompt investigation include propensity scores near zero or one, unusually large weighted residual contributions, a targeting procedure that fails to converge, predictions outside a known outcome range, and major changes in the estimate under reasonable nuisance-model choices. These findings do not mechanically dictate AIPW or TMLE. They may indicate that the design, estimand, or data do not support a stable answer.

The estimators also have limitations beyond overlap. Their robustness is not protection against unmeasured confounding, incorrect treatment timing, measurement error, or selection bias. AIPW can produce finite-sample estimates outside plausible outcome bounds; TMLE’s bounded plug-in property depends on using an appropriate link and correctly carrying out the targeting procedure. Both rely on adequate nuisance estimation and appropriate uncertainty calculations. Comparing both methods is informative, but agreement does not prove that the causal assumptions hold.

Common Mistake

Common mistake: “TMLE is always better because it is bounded and doubly robust.” The boundedness claim applies only when the outcome scale and targeting link support it, and AIPW also has a double-robustness property under suitable conditions. Neither method repairs poor overlap or unmeasured confounding. Correction: choose based on the outcome scale, nuisance-learning plan, finite-sample diagnostics, and inferential goal; state the shared assumptions and assess whether they are plausible.

Key takeaway: AIPW and TMLE can target the same ATE and share core identification requirements. TMLE is especially attractive when bounded plug-in predictions matter; AIPW may be attractive for a direct estimating-equation implementation. Neither choice removes the need to diagnose overlap, validate nuisance estimation, and justify inference.

Check Your Understanding

Answer each question, then compare with the brief solutions.

For a binary outcome, what practical property may favor TMLE, and why does that property not establish that TMLE is more accurate?
AIPW and TMLE estimates agree closely. Does that demonstrate conditional exchangeability? Explain.
In a sample of 250 people, 10 have estimated propensity scores below 0.025. What percentage is this, and what is the corresponding lower bound on \(1/\hat g(W)\) for those observations?
Why can extreme propensity scores destabilize both AIPW and TMLE?
Name two conditions that are shared by the methods and are not guaranteed by using a doubly robust estimator.

Solutions. 1. A compatible TMLE can keep updated predictions within the outcome bounds, which can keep a plug-in contrast within its natural range; boundedness alone does not guarantee low bias or variance. 2. No. Agreement is a comparison of estimates, not evidence that unmeasured confounding is absent or that exchangeability holds. 3. \(10/250=0.0400\), or 4.00%; since \(\hat g(W)<0.025\), \(1/\hat g(W)>1/0.025=40\). 4. The residual correction or targeting step can be strongly influenced by observations with small treatment probabilities, creating high variability. 5. Examples include conditional exchangeability and positivity; neither is supplied by double robustness.