Skip to content
← Writing

Clone-censor-weight, explained: per-protocol effects without immortal-time bias

Jun 27, 2026 · 3 min read causal inferencetarget trial emulationmethods

In the post on target trial emulation I drew a line between two questions: the effect of starting a strategy (the intention-to-treat analogue) and the effect of sustaining it (the per-protocol effect). The per-protocol question is usually the one clinicians actually mean — “what happens if patients keep taking it?” — and it’s the one that quietly breaks naive analyses. Clone-censor-weight (CCW) is how you answer it honestly.

When you need it: sustained strategies

Reach for CCW when the “exposure” isn’t a single decision at baseline but a rule applied over time: treat to target for ≥2 years vs stop at one year; stay on the SGLT2 inhibitor vs switch; continue augmentation vs discontinue. The thing you’re comparing is a strategy sustained through follow-up.

Why the naive per-protocol analysis fails

The tempting move is to classify people by what they actually did over follow-up — “completers” vs “stoppers” — and compare them. But that conditions on the future. Patients who tolerate and respond tend to stay on; those who deteriorate or have side effects stop. So “stayers” look healthier for reasons that have nothing to do with the drug. You’ve re-introduced exactly the selection that randomisation was meant to prevent — plus immortal time, the stretch a patient must survive to be counted as having “sustained” the strategy.

The three moves

1. Clone

At time zero, copy each eligible patient into every strategy they’re compatible with at baseline. Someone who could plausibly start either arm becomes two clones — one assigned “sustain”, one “stop” — identical in every baseline covariate. This hard-codes the trial’s defining feature: assignment can’t depend on anything that happens later.

2. Censor

Follow each clone under its assigned rule. The moment a clone’s real data contradict its rule — the “sustain” clone stops the drug, or the “stop” clone keeps taking it — censor that clone at that time. Now every clone’s observed follow-up is consistent with its strategy by construction, so no immortal time can accumulate.

3. Weight (inverse-probability-of-censoring)

Censoring isn’t random: clones that deviate differ from those that don’t (side effects, worsening disease). Dropping them naively would re-introduce selection. So at each time you model the probability of remaining uncensored given measured, time-varying covariates, and weight each surviving clone by the inverse of that probability. The weights rebuild the population the censoring removed. The crucial assumption: no unmeasured confounders of deviation — you’ve measured the things that drive staying versus stopping.

A worked example

The sustained-strategy question shows up constantly in pharmacoepidemiology. In work on augmentation strategies for treatment-resistant depression, “what if patients kept the augmentation?” is a per-protocol question by definition. Cloning at the augmentation start, censoring at deviation, and applying inverse-probability-of-censoring weights lets the intention-to-treat analogue (effect of starting) and the per-protocol estimate (effect of sustaining) sit side by side — two genuinely different clinical questions, each answered on its own terms rather than conflated.

Pitfalls and practical tips

  • Grace period. Allow a short window to “start” before you censor, or you’ll censor almost everyone on day one. Pick it on clinical grounds and report a sensitivity analysis.
  • Positivity. If some covariate pattern always deviates, its weights explode — check for it before trusting the estimate.
  • Stabilise and truncate weights. Use stabilised weights; truncate extremes (e.g. 1st/99th percentile) and report what you did.
  • Sanity-check. Stabilised weights should average about 1; inspect the largest few. A big gap between the ITT-analogue and per-protocol results is informative, not embarrassing — but it deserves scrutiny.

The payoff

Clone-censor-weight lets you ask the sustained-strategy question — usually the clinically relevant one — without the immortal-time and selection traps that sink naive per-protocol analyses. It’s the natural partner to target trial emulation: TTE writes the protocol of the trial you can’t run; CCW emulates its per-protocol contrast.

This is the method behind several of my studies and my recurring ACPE educational session. Want the slides or a worked template? Get in touch.