Treatment Switching

Treatment Switching in Randomised Trials and Health Technology Assessment

In a randomised controlled trial of a new medicine, treatment switching most often means that patients in the control arm start the experimental drug during follow-up, usually once their disease has progressed. If those patients gain from the drug, the control arm lives longer than it would have on standard care, so an intention-to-treat analysis of overall survival understates what the new drug adds and an economic model built on it overstates the ICER. Appraisal bodies therefore ask whether the trial's switching would happen in routine care and, if not, how survival can be adjusted without adding new bias. The term belongs to trials and is distinct from switching in routine care, where a prescriber or payer changes a patient's medicine.

Why switching happens and why it distorts survival

NICE Decision Support Unit Technical Support Document (TSD) 16 finds switching common in cancer trials and possible elsewhere. It is permitted for ethical reasons, when interim analyses suggest benefit and no other non-palliative treatment exists, and because trials that forbid it can be hard to recruit to. Regulators may accept progression-free survival as a primary endpoint, which weakens the incentive to keep groups intact after progression.

Health technology assessment compares a world in which the new treatment is available with one in which it is not, and a control arm that has received the new drug no longer represents the second world. The NICE manual (PMG36, section 4.2.23) says a lifetime time horizon is needed when technologies lead to differences in survival, so the model depends on overall survival, the endpoint most exposed to switching. When switching follows progression, progression-free survival is unaffected but post-progression survival in the control arm lengthens, so the observed gain shrinks and the ICER rises.

TSD 16 reviewed 45 NICE appraisals in advanced or metastatic cancer completed by December 2009 and found switching in the main trials of 25 (55.6%). In the appraisal of sunitinib for gastrointestinal stromal tumours (TA179), the intention-to-treat ICER of about GBP 77,000 per QALY fell to about GBP 27,000 once a rank preserving structural failure time model was used to adjust for switching.

Which switches an appraisal should adjust for

TSD 24 (2024), which supplements rather than replaces TSD 16, widens the definition to any move onto a treatment other than the one randomised, in either arm. Adjustment is usually appropriate for switches that would not occur in standard practice in the jurisdiction making the decision, and usually not for those that would. Control patients moving onto the experimental drug are usually adjusted for, unless that drug is already used at a later line. Experimental patients moving onto the control treatment usually are not, because the control is normally standard care. For moves onto other treatments the test is whether the treatment is available at that line in England and Wales, and TSD 24 notes that treatments funded through managed access, such as the Cancer Drugs Fund, may warrant adjustment because they are not routinely commissioned.

A standard treatment started at a non-standard time, for example after an interim analysis, may also need adjustment. TSD 24 adds that deciding adjustment is relevant does not mean any particular adjusted analysis will be accepted.

Estimands and the ICH E9(R1) addendum

The ICH E9(R1) addendum on estimands, adopted in November 2019, treats switching between treatments of interest as an intercurrent event, an event after treatment starts that affects how outcomes are interpreted. Its treatment policy strategy uses outcomes regardless of the switch, matching the intention-to-treat comparison, while its hypothetical strategy envisages the event not occurring, which is what switching adjustment estimates. For survival in NICE appraisals, TSD 24 recommends presenting both, rejects the while-on-treatment strategy because it amounts to censoring switchers, and asks sponsors to plan trial data collection with HTA questions in mind.

Simple approaches and their bias

The intention-to-treat analysis respects randomisation and TSD 16 says it should always be presented, but in the presence of switching it dilutes the survival benefit. The tempting shortcut, a per-protocol analysis that censors switchers at the time of switch or excludes them, breaks randomisation because clinicians decide who switches on patient characteristics that are linked to prognosis. PMG36 (section 4.6.26) states that such simple adjustment methods should be avoided because they are very susceptible to selection bias.

Other simple approaches sidestep the confounded data, for instance by modelling progression-free survival alone or assuming equal overall survival as a worst case that gives a maximum ICER. In the 25 affected appraisals reviewed in TSD 16, censoring or exclusion was used in 11 (44%), switching was not addressed in 7 (28%) and only TA179 used a recognised complex method.

Methods that adjust survival for switching

Three families of methods recur in TSD 16 and TSD 24, each resting on a different untestable assumption, and they can disagree. In the NICE appraisal of pazopanib for metastatic renal cell carcinoma, TSD 16 reports an ICER of about GBP 49,000 per QALY using inverse probability of censoring weights and about GBP 33,000 using the rank preserving structural failure time model.

Rank preserving structural failure time model

The RPSFTM was developed by Robins and Tsiatis for non-compliance in randomised trials. It links each patient's observed survival to the survival they would have had without the experimental drug.

$$U_i = T_i^{\text{off}} + e^{\psi} \, T_i^{\text{on}}$$

where $U_i$ is the counterfactual untreated survival time of patient $i$, $T_i^{\text{off}}$ and $T_i^{\text{on}}$ are the observed times spent off and on the experimental treatment, and $e^{\psi}$ is the treatment effect expressed as a time ratio, the factor that converts time on treatment into equivalent untreated time. G-estimation searches for the value of $\psi$ at which counterfactual survival is independent of randomised group. Because randomisation acts as an instrument, the method is a form of instrumental variable analysis and keeps the p-value of the intention-to-treat analysis.

Its weak point is the common treatment effect assumption, since a patient who switches after progression may benefit less per unit of time than one treated from randomisation. Counterfactual censoring times can become informative, so data in the arm at risk of switching are recensored at the earlier of the administrative censoring time and $e^{\psi}$ times it, which discards long-term information needed for extrapolation. Branson and Whitehead's iterative parameter estimation uses the same model with a parametric fit.

Inverse probability of censoring weights

IPCW censors switchers at the switch and gives extra weight to non-switchers whose baseline and time-varying characteristics resemble those of switchers, removing the selection bias of plain censoring if every confounder is measured. Trials rarely record some predictors of switching, such as patient preference, and TSD 16 reports that the method became highly prone to error when about 90% of control patients switched in a trial of 500.

Two-stage estimation

Two-stage methods treat the trial as randomised up to a secondary baseline, usually progression, and observational after it. The simple version fits an accelerated failure time model to post-progression survival in the control arm, comparing switchers with non-switchers and adjusting for covariates measured at progression, then uses the resulting time ratio to shrink switchers' survival times. It needs no common treatment effect, but assumes no time-dependent confounding between progression and switch. TSD 24 describes a newer version using g-estimation (Latimer and colleagues, 2020) that accounts for time-dependent confounding between progression and switching, provided time-varying covariates are measured after progression.

Worked example of an RPSFTM adjustment and its effect on the ICER

The numbers below are illustrative. A control-arm patient progresses at 6 months, switches at once to the experimental drug and dies at 16 months, so $T^{\text{off}} = 6$ and $T^{\text{on}} = 10$. Suppose g-estimation gives $e^{\psi} = 0.6$, meaning the drug stretches survival time by a factor of 1/0.6.

1. Counterfactual survival for the switcher.

$$U = 6 + 0.6 \times 10 = 12$$

where $U$ is the survival in months the patient would have had without the experimental drug. Switching added 4 months.

2. Recensoring. With administrative censoring at 36 months, control-arm data are recensored at 0.6 × 36 = 21.6 months, so any counterfactual death after 21.6 months is treated as censored.

3. Effect on the ICER. Suppose a model gives mean survival of 26 months with the new drug and 20 months in the control arm as observed, but 16 months once control survival is adjusted. Assume a utility of 0.7 throughout, no discounting and an incremental cost of GBP 30,000 in both analyses, with the cost of the switched-to drug excluded and other cost changes from shorter adjusted survival ignored for simplicity. The intention-to-treat gain is 6 months, or 0.5 × 0.7 = 0.35 QALYs, and the adjusted gain is 10 months, or (10/12) × 0.7 = 0.5833 QALYs.

$$\text{ICER}_{\text{ITT}} = \frac{30000}{0.35} = 85714$$

where the numerator is the incremental cost in GBP and the denominator the intention-to-treat QALY gain.

$$\text{ICER}_{\text{adjusted}} = \frac{30000}{0.58333} = 51429$$

where the denominator is the adjusted QALY gain. The adjustment cuts the ICER from about GBP 85,700 to about GBP 51,400 per QALY, so the credibility of $e^{\psi}$ and of the common treatment effect assumption moves the ICER by about GBP 34,000 per QALY.

Choosing and reporting an adjustment in a NICE appraisal

TSD 16 sets out a stepwise framework. The analyst relates the switching mechanism to the decision problem, checks the switching proportion among eligible patients, tests each method's key assumption against external data and clinical opinion, inspects diagnostics such as recensoring and IPCW weights, and only then extrapolates following TSD 14, with sensitivity analysis across methods when the choice is unclear. The intention-to-treat result can be the least biased option, for example with a hazard ratio of about 0.75 to 1.00 and switchers gaining around 15% less.

TSD 24 reviewed 40 NICE cancer appraisals from January 2020 to April 2022. Seven applied adjustment methods to the main trial, none used all four complex methods from TSD 16, and key diagnostics were absent from the main text of every company submission.

How other HTA bodies treat switching

The EU Guidance on Validity of Clinical Studies, adopted by the HTA Coordination Group in September 2024, warns that the term crossover is sometimes used incorrectly for treatment switching. It notes that switching of control patients after progression is common in oncology and can influence estimated effects on survival, and that statistical adjustment can reduce this bias but has limitations.

IQWiG's General Methods (version 8.0, 2025), discussing adverse events and patient-reported outcomes, require analyses based on the treatment policy estimand and the intention-to-treat principle, and ask for outcome data to be collected completely after treatment is discontinued or switched.

Boundaries with crossover trials and routine-care switching

A crossover trial moves participants to the alternative intervention after a period fixed by the protocol, so that in most cases each patient receives each intervention. Treatment switching happens inside a parallel-group trial and is selective, and TSD 16 avoids the word crossover because it evokes crossover trials, a different design. Switching in routine care, including non-medical switching for formulary reasons, is part of the real care pathway and is usually left unadjusted when it occurs in a trial.

Common misreadings of switching-adjusted evidence

Switching adjustment changes the number that drives a reimbursement decision, so errors in reading it matter. Several recur.

• Adjusted survival is not observed survival. Every method rests on an assumption that the trial data cannot fully test, so an adjusted hazard ratio carries structural uncertainty beyond its confidence interval.
• Censoring or excluding switchers is not a safe compromise. Because switching depends on prognosis, a per-protocol analysis introduces selection bias whose size the trial data cannot reveal.
• Survival adjustment alone is incomplete. Costs and quality of life measured after a switch are also affected, and the cost of the switched-to drug should leave the adjusted comparison.

Sources

• Latimer NR, Abrams KR. NICE DSU Technical Support Document 16: Adjusting survival time estimates in the presence of treatment switching. NICE Decision Support Unit; July 2014. https://doi.org/10.15131/shef.data.33847621
• Bell Gorrod H, Latimer NR, Abrams KR. NICE DSU Technical Support Document 24: Adjusting survival time estimates in the presence of treatment switching, an update to TSD 16. NICE Decision Support Unit; April 2024. https://doi.org/10.15131/shef.data.33850897
• National Institute for Health and Care Excellence. NICE technology appraisal and highly specialised technologies guidance: the manual (PMG36). Published 31 January 2022, last updated 13 October 2026. https://www.nice.org.uk/process/pmg36/chapter/economic-evaluation
• International Council for Harmonisation. ICH E9(R1) Addendum on estimands and sensitivity analysis in clinical trials to the guideline on statistical principles for clinical trials. Adopted 20 November 2019. https://database.ich.org/sites/default/files/E9-R1_Step4_Guideline_2019_1203.pdf
• Member State Coordination Group on Health Technology Assessment. Guidance on Validity of Clinical Studies, version 1.0. Adopted 19 September 2024 under Regulation (EU) 2021/2282. https://health.ec.europa.eu/publications/guidance-validity-clinical-studies-joint-clinical-assessments_en
• Institute for Quality and Efficiency in Health Care (IQWiG). General Methods, version 8.0 of 19 December 2025. https://doi.org/10.60584/General-Methods_V8.0