The Gold Standard of Experimental Design | Collaborative TMLE, Propensity-Score Selection and Positivity Diagnostics

1. Start with the estimand and identification assumptions

Back to contents

Define the treatment, outcome time, population and causal contrast before choosing C-TMLE. Exchangeability requires sufficient pre-treatment adjustment, consistency requires a meaningful treatment intervention and positivity requires treatment alternatives in the relevant covariate strata. Collaborative selection operates within these assumptions. It cannot recover an effect that the observed data do not identify.

2. Separate outcome regression from treatment mechanism

Back to contents

For a point treatment, the outcome regression estimates expected outcome given treatment and covariates, while the propensity score estimates treatment probability given covariates. Ordinary doubly robust estimators combine both. C-TMLE makes their estimation collaborative by choosing treatment-mechanism candidates according to how they support bias–variance performance for the target parameter, not treatment prediction alone.

3. Understand the collaborative sequence

Back to contents

A classical implementation begins with an initial outcome regression and constructs increasingly flexible propensity-score candidates, often by adding covariates or relaxing regularisation. Each candidate enters a targeting step that updates the outcome regression. Cross-validation or a targeted criterion selects along the path. Record the ordering, stopping rule and loss; the final propensity model is the result of an estimator algorithm, not an ordinary explanatory model.

4. Do not confuse confounder control with prediction accuracy

Back to contents

Variables strongly associated with treatment but weakly related to outcome can improve propensity prediction while pushing scores toward zero or one. Omitting a genuine confounder can bias the effect. Outcome-adaptive or collaborative selection seeks a useful trade-off, but background causal knowledge still determines the eligible adjustment set. Post-treatment variables, colliders and convenience features do not become safe because an algorithm selected them.

5. Diagnose practical positivity before targeting

Back to contents

Plot propensity distributions by treatment, summarise tails, inspect clever covariates and identify covariate patterns with little support. Report effective sample size and influential observations where relevant. Structural positivity failures require changing the target population or treatment contrast. Practical violations may motivate truncation, regularisation or collaborative dimension reduction, but every repair changes bias, variance or estimand interpretation.

6. Worked example: specialist referral and near-certain treatment

Back to contents

An observational study estimates the effect of a specialist referral on recovery. A rare severity pattern predicts referral for 97 of 100 comparable patients, while outcomes also depend strongly on severity. A highly flexible propensity learner produces extreme clever covariates and a volatile TMLE. C-TMLE selects a less aggressive propensity candidate that improves targeted validation loss, but repeated analyses still show one stratum with almost no untreated patients. The team restricts the target population and reports both estimates; collaboration stabilised fitting but could not create the missing comparison.

7. Choose truncation collaboratively but keep the target clear

Back to contents

Positivity-C-TMLE can select a propensity-score truncation level using a targeted criterion rather than a generic treatment likelihood. This may improve finite-sample bias–variance behaviour in studied settings. It is not a universal optimal rule. Prespecify the candidate grid, report the selected threshold and show fixed-threshold sensitivity. Severe truncation can make the analysis answer a more supported but less explicit question.

8. Changed-context transfer check: a randomised trial with dropout

Back to contents

Move from observational treatment selection to a randomised trial where treatment propensity is known but outcome observation depends on covariates and prior outcomes. The transferable principle is collaborative estimation of a nuisance mechanism for a target parameter. The changed mechanism is censoring rather than treatment assignment. Folds, time ordering and sequential positivity must match the longitudinal structure; a point-treatment C-TMLE cannot simply be relabelled.

9. Use cross-validation without leaking the collaborative path

Back to contents

Variable ordering, propensity candidates, outcome fitting, truncation selection and targeting must respect training and validation boundaries when cross-fitting is used. Preprocessing on the full data can leak information. Keep clusters or repeated observations together when they form the independent unit. Repeat prespecified folds to diagnose partition sensitivity rather than choosing the split that gives the preferred interval.

10. Know the limits and common misconceptions

Back to contents

C-TMLE is not a licence for automated confounder discovery. It does not make positivity optional, prove that the selected propensity variables are causal or guarantee nominal coverage. A less predictive treatment model can be better for the target estimator, but poor prediction alone is not the goal. Different C-TMLE implementations have different theory and computational paths; name the algorithm instead of treating the acronym as one procedure.

11. Audit the full collaborative estimator

Back to contents

Publish the causal graph or adjustment rationale, initial outcome learner, candidate propensity path, ordering rule, cross-validation scheme, targeting loss, selected candidate, truncation, clever covariates, influence values, standard-error formula and convergence failures. Compare a conventional TMLE, prespecified propensity models and relevant collaborative variants. Use simulations calibrated to sample size, overlap and misspecification when interval coverage matters.

12. The gold standard lets the nuisance models serve the question

Back to contents

Collaborative TMLE is valuable because nuisance estimation is organised around the causal parameter rather than isolated prediction scores. The gold standard begins with identification, shows the collaborative path, diagnoses unsupported comparisons, tests selection and truncation, and reports sensitivity across algorithms. Stability is credible only when the population, overlap and inference remain visible.