The Gold Standard of Experimental Design | Targeted Maximum Likelihood Estimation, Double Robustness and Super Learner

1. Define the target parameter before choosing algorithms

Back to contents

State the population, treatment versions, outcome horizon and contrast, such as an average treatment effect or risk difference. Write the causal identification assumptions and the corresponding observed-data parameter. TMLE targets a parameter; it cannot repair an ambiguous intervention or silently decide whether the question is intention-to-treat, per-protocol, conditional or transported.

2. Identify the two nuisance functions

Back to contents

For a point-treatment average effect, the main nuisance functions are often the outcome regression—the expected outcome given treatment and covariates—and the treatment mechanism or propensity score. Their exact form changes with the parameter and data structure. They are called nuisance functions only relative to the target; they still carry much of the causal adjustment and must be diagnosed.

3. Use Super Learner as cross-validated ensemble prediction

Back to contents

Super Learner evaluates a prespecified library of candidate algorithms with cross-validation and finds a loss-based combination. Include simple, regularised and flexible learners appropriate to sample size and variable type. The library is a scientific choice, not a ritual list. Cross-validation reduces reliance on in-sample fit, but a library that omits plausible structure cannot learn it.

4. Understand the targeted fluctuation

Back to contents

TMLE starts with an outcome estimate and updates it using a submodel informed by the treatment mechanism and the efficient influence curve. For bounded outcomes, the updated predictions can remain within their natural range. The update aims to solve the estimating equation for the target parameter; it is not a general refit of every feature of the outcome distribution.

5. State double robustness precisely

Back to contents

In common settings, the estimator can remain consistent if either the outcome regression or treatment mechanism is consistently estimated, together with regularity and identification conditions. That asymptotic property is not a guarantee for a small sample, two severely misspecified models or practical positivity failure. Inference also depends on convergence rates, empirical-process conditions or cross-fitting strategies.

6. Worked example: a treatment effect with nonlinear confounding

Back to contents

An observational study estimates the effect of a tutoring programme on examination success. Enrolment depends nonlinearly on prior score and travel time; success also depends on their interaction. A simple outcome model misses the interaction, while a flexible Super Learner library captures it. TMLE updates predicted success using estimated enrolment probabilities and returns a population risk difference. When extreme probabilities are inspected, a small rural subgroup has almost no untreated observations. The estimate is reported for supported students and the rural extrapolation is shown separately rather than hidden under “double robust.”

7. Check positivity and the clever covariate

Back to contents

Plot treatment probabilities overall and for substantively important histories. Extreme inverse probabilities make the targeting step and influence curve unstable. Inspect the size of the fluctuation and influential observations. Truncation may stabilise variance but changes bias and should be analysed transparently. Redefining the population or intervention can be more honest than relying on unsupported counterfactuals.

8. Changed-context transfer check: a randomised trial

Back to contents

Move from observational tutoring data to a randomised treatment. The treatment mechanism may be known by design, which strengthens one nuisance component. Covariate-adjusted TMLE can improve precision while preserving the randomised estimand when implemented correctly. The changed context does not permit post-randomisation adjustment or rescue missing outcomes automatically; those require their own identification and nuisance models.

9. Know the limits and common misconceptions

Back to contents

TMLE is not one universal algorithm, and “using Super Learner” does not certify valid inference. Double robustness does not mean two chances to ignore diagnostics. Unmeasured confounding, interference, informative missingness, measurement error and ill-defined treatment remain. Highly adaptive learners may need cross-fitting or CV-TMLE for valid asymptotics, and bootstrap choices must respect the sampling design.

10. Audit the roadmap, learners and influence curve

Back to contents

Publish the causal model, estimand, identification formula, covariate timing, learner library, cross-validation folds, tuning, outcome bounds, propensity diagnostics, targeting update, influence-curve variance and any truncation. Compare prespecified alternative libraries, simple estimators and negative or falsification checks. Reproduce the estimate from code and record seeds and software versions.

11. Frequently asked questions

Back to contents

Why not use the best prediction model alone? Prediction loss is not the same as bias for a causal parameter. Does Super Learner always win? It performs relative to its library and data. What is double robustness? Consistency under one of two correctly estimated nuisance components in suitable settings. Can TMLE fix unmeasured confounding? No. Why target? To update nuisance estimates toward the parameter and its efficient influence curve.

12. The gold standard targets a defined question, not a fashionable method

Back to contents

A credible TMLE analysis begins with design and identification, uses diverse cross-validated learners, performs a transparent targeting step and lets positivity diagnostics limit the claim. The gold standard states double robustness narrowly and audits finite-sample behaviour. Machine learning helps estimate nuisance structure; it does not remove the obligation to show where the causal comparison comes from.