The Gold Standard of Experimental Design | Cloning–Censoring–Weighting, Dynamic Adherence Strategies and Per-Protocol Effects

1. Write the target trial before choosing the method

Back to contents

Specify eligibility, time zero, strategies, assignment, follow-up, outcome, causal contrast and analysis. The same timestamp should determine eligibility, strategy compatibility and start of outcome follow-up. A vague goal such as “effect of adherence” is not enough. State a per-protocol estimand for concrete strategies in a defined population and time horizon. Cloning is a tool for emulation, not a replacement for the protocol.

2. Define strategies as executable rules

Back to contents

A dynamic strategy maps observed history to an action: for example, continue treatment while toxicity remains below a threshold, pause after severe toxicity and restart after recovery. Every decision point, grace period, dose tolerance and allowed deviation must be explicit. If two strategies overlap during a grace period, one person can initially be compatible with both. Ambiguous rules make censoring irreproducible and the estimand hard to interpret.

3. Clone only across genuinely compatible strategies

Back to contents

At baseline, create one record per person per strategy that their observed history has not yet ruled out. The clones share the same pre-assignment data and outcome until their paths diverge, so analyses must account for induced duplication. Cloning does not create new independent people or balance confounders. It represents the fact that early behaviour may be consistent with more than one longitudinal strategy.

4. Censor at the first incompatible action

Back to contents

Follow each clone through decision times and artificially censor it when observed treatment becomes inconsistent with its assigned strategy. Natural loss to follow-up and competing events must remain distinguishable from artificial censoring. The deviation rule must use information available at that time. Looking ahead to decide whether someone adhered can introduce future information and distort risk sets.

5. Weight to repair selection created by censoring

Back to contents

Artificial censoring preferentially removes people whose evolving covariates make deviation more likely. Estimate the probability of remaining uncensored conditional on relevant past treatment, outcome predictors and time-varying covariates, then construct inverse-probability-of-censoring weights. Stabilisation and truncation can improve precision but change the practical analysis. Report numerator, denominator, time scale and truncation rule.

6. Worked example: two initiation windows

Back to contents

Eligible patients may start therapy within 7 days or within 30 days after diagnosis. Everyone is cloned into both strategies at day zero. A patient starting on day 5 remains compatible with both; one starting on day 20 is censored from the 7-day clone but remains in the 30-day clone; one untreated by day 30 is censored from both. Weighted risks at one year compare the defined initiation policies. The result is not simply “early versus late starters”, because the 30-day strategy includes the observed timing distribution allowed within its window.

7. Diagnose positivity and unstable weights

Back to contents

Positivity requires a non-zero chance of remaining compatible with each strategy for relevant covariate histories. If almost every severely ill patient initiates immediately, a delayed strategy has little support in that subgroup. Inspect numbers at risk, censoring by strategy and time, weight distributions, effective sample size and covariate balance after weighting. Extreme weights signal reliance on a few trajectories, not extra information.

8. Changed-context transfer check: a biomarker-triggered rule

Back to contents

Compare “intensify when biomarker exceeds 8” with “intensify when it exceeds 10”. Before either threshold is crossed, a person may be compatible with both rules. Later measurements determine the required action and censor incompatible clones. The transferable steps are protocol, compatibility, censoring and weighting. The changed context demands reliable measurement times and clear handling of missed tests, because an unobserved threshold crossing cannot be treated as a known decision.

9. Know the identification assumptions and limits

Back to contents

Interpretation requires consistency, exchangeability for censoring and treatment decisions, positivity and correct measurement of strategy-relevant history. Unmeasured severity can bias the result even when weights look stable. Informative visit schedules, measurement error and treatment versions can break the emulation. Weight models are not automatically correct because they converge. Negative controls, quantitative bias analysis and alternative specifications can probe but not erase these limitations.

10. Audit the emulation from eligibility to estimate

Back to contents

Publish the protocol, time-zero construction, person-strategy cloning code, interval structure, compatibility rules, reasons and times for artificial and natural censoring, weight models, diagnostics, outcome model and variance method. Re-run with plausible grace periods, censoring definitions, model forms and truncation levels. Show unweighted and weighted risk sets so readers can see where each strategy loses support.

11. Frequently asked questions

Back to contents

Why clone people? Early observed treatment can satisfy several strategies. Does cloning solve confounding? No. Why censor clones? To keep each copy aligned with its assigned strategy. Why weight after censoring? Artificial censoring creates selection. Are dynamic strategies always better? No; they are appropriate when actions depend on evolving history. Is this an intention-to-treat effect? No; it targets adherence to specified strategies under identification assumptions.

12. The gold standard keeps the estimand ahead of the machinery

Back to contents

A credible cloning–censoring–weighting study begins with executable strategies and aligned time zero, not a convenient dataset transformation. It makes compatibility rules reproducible, distinguishes artificial from natural censoring, diagnoses support and carries weight uncertainty into the estimate. The gold standard treats the final per-protocol effect as a claim about defined interventions—not about people labelled adherent after the fact.