Skip to content

Protocol Generator · CCW

Clone-Censor-Weight protocol

Draft a target-trial-emulation protocol for treatment strategies that are only distinguishable after time zero — grace periods, sustained use, treat-to-target. Cloning + artificial censoring + inverse-probability-of-censoring weights (IPCW) remove the immortal-time and selection bias that a naïve analysis introduces. Everything runs in your browser.

← All designs
1 · Identity

📚 Databases / sites (fills “Data source(s)”)

Taiwan

South Korea

Japan

Hong Kong

Australia

Finland

United Kingdom

Netherlands

United States

2 · Causal question
Population
Outcome
3 · Time zero & strategies
4 · Censoring & weights
In one sentence: everyone is copied into both arms, and each copy is dropped the moment it stops matching its arm.

Because a patient's data are compatible with both strategies until they act, the same person-time counts in both arms during the grace period — that is what removes immortal-time bias. But the dropping (“artificial censoring”) is not random: the clones that survive in an arm are the ones who happened to behave that way, so they must be up-weighted to stand in for the ones removed. That up-weighting is the IPCW. Two consequences follow, and both matter for the analysis: model-based standard errors are too small (the number of events is artificially inflated), so confidence intervals come from a non-parametric bootstrap that re-runs cloning, censoring and weighting inside every replicate; and the two arms are identical at time zero by construction and separate only as patients act, so the hazard ratio starts at 1 and moves with follow-up time — proportional hazards is violated by design (Maringe 2020, doi:10.1093/ije/dyaa057).

Censoring events weighted by IPCW

Weights are estimated separately for each arm, because the two arms need not deviate on the same schedule. Which arm's model is time-fixed and which is time-varying depends on how you have written the strategies, not on which one you typed first: a strategy that can only be broken by missing a deadline has a single deviation time, while one that can be broken by an action at any moment has many — and both of yours may be of the second kind, since discontinuation, switch and add-on are all continuing requirements.

Exclusion criteria (name | look-back days, one per line)
Covariate types (name | look-back days, one per line)

Each exclusion / covariate line is drawn as its own labelled window in the diagram — drag its edge to adjust that look-back.

5 · Follow-up & analysis
Subgroup analyses (tick to perform)
Sensitivity analyses (tick to perform)
Auto-saved to this browser.

Live protocol preview

 
 
 
 
Consistency checks · target-trial 防呆
 
 
 
 
Study design diagram
 
Time zero, the grace period, the ascertainment windows that precede time zero, and follow-up. The figure is one timeline: it marks where cloning happens, but it does not draw the two arms separately. Drag the ● handles to adjust the covariate look-back, grace, and follow-up — the fields above update automatically.