In many oncology trials the control patients may switch to the experimental treatment, either at disease progression or after an interim analysis has shown a benefit on progression-free survival (PFS). Switching leaves the outcomes observed before the switch unchanged and alters only what happens afterward, typically prolonging the remaining overall survival (OS) of the switchers and diluting the intention-to-treat comparison of OS.
The TrialSimulator package (Zhang, 2026) expresses such designs with
regimen() (a switching rule fixed at enrollment) and
crossover() (a switch that opens at a milestone), and runs
the simulated trials one at a time. Because a switch never rewrites the
history before it, the same designs can also be simulated with the whole
set of trials at once. switch_fast() modifies the simulated
data after their generation, and cutoff_fast() provides the
per-trial calendar time at which crossover opens:
simdata_fast() generates PFS and OS for all trials from
an illness-death model;cutoff_fast() finds the calendar time of the PFS
analysis in each trial;analysis_fast() analyzes PFS at that time, and the
trials in which PFS is significant are selected;switch_fast() lets the control patients of those trials
switch, changing only their outcomes after the switch;The interim analysis is unaffected by step 4 by construction, which
is checked below. Switching rules that belong to the data-generating
process, such as a switch after or before progression with a given
hazard, are given to simdata_fast() directly (see the
section on switching rules in the data-generating model).
The trial randomizes 300 patients per group over 18 months. Progression, death without progression, and death after progression follow an illness-death model with exponential transition times, and patients drop out at 5 percent per year. PFS is analyzed once, at 380 PFS events. OS is analyzed at 350 deaths. Both tests are one-sided at level 0.025.
nsim <- 1000
seed <- 20261007
df <- simdata_fast(
nsim = nsim,
n = c(300, 300),
a.time = c(0, 18),
a.rate = 600 / 18,
h01.median = list(8, 12), # progression
h02.median = list(30, 36), # death without progression
h12.median = list(12, 12), # death after progression
d.hazard = -log(1 - 0.05) / 12,
seed = seed
)cutoff_fast() counts the PFS events (the columns
e1_tte and e1_event) to find the analysis time
of each trial, and analysis_fast() analyzes PFS at those
times, with the same columns given by tte.col and
event.col.
pfs_cut <- df |>
cutoff_fast(event.looks = 380, tte.col = "e1_tte", event.col = "e1_event")
pfs_res <- df |>
analysis_fast(control = 1, cutoff.looks = pfs_cut, side = 1,
tte.col = "e1_tte", event.col = "e1_event")
pfs_pos <- (pfs_res$reached & pfs_res$logrank.p <= 0.025) %in% TRUE
c(PFS_power = mean(pfs_pos),
mean_PFS_analysis_month = mean(pfs_cut[, 1], na.rm = TRUE))
#> PFS_power mean_PFS_analysis_month
#> 0.93900 21.55091Three scenarios are compared with the trial without switching.
when = "later", the rule of the worked example of the
milestone crossover in TrialSimulator). Their remaining survival time is
multiplied by 1.3.when = "intermediate"), for half of the control patients
who progress, with a post-progression median survival of 18 months
instead of 12.sw_gated <- df |>
switch_fast(group = 1, when = "later", cutoff = pfs_cut, sims = pfs_pos,
aft.factor = 1.3)
sw_all <- df |>
switch_fast(group = 1, when = "later", cutoff = pfs_cut, aft.factor = 1.3)
sw_prog <- df |>
switch_fast(group = 1, prob = 0.5, when = "intermediate", median = 18)The OS analysis at the time of the PFS analysis is the same with and without switching, because every switch happens after that time.
The OS analysis is triggered by 350 deaths in the modified data, since switching delays deaths and therefore the analysis.
os_final <- function(d) {
d |>
analysis_fast(control = 1, stat = c("logrank", "coxph"), side = 1,
tte.col = "e2_tte", event.col = "e2_event",
cutoff.looks = \(x) cutoff_fast(x, event.looks = 350,
tte.col = "e2_tte",
event.col = "e2_event"))
}
scen <- list(
"no switching" = df,
"crossover after positive PFS" = sw_gated,
"crossover after PFS, all trials" = sw_all,
"50% switch at progression" = sw_prog
)
tab <- do.call(rbind, lapply(names(scen), function(nm) {
d <- scen[[nm]]
r <- os_final(d)
data.frame(
scenario = nm,
control_switched = round(mean(d$switched[d$group == 1] == 1), 3),
OS_power = round(mean(r$logrank.p <= 0.025, na.rm = TRUE), 3),
mean_HR = round(exp(mean(r$cox.coef, na.rm = TRUE)), 3),
mean_OS_month = round(mean(r$cutoff, na.rm = TRUE), 1)
)
}))
tab
#> scenario control_switched OS_power mean_HR
#> 1 no switching 0.000 0.399 0.834
#> 2 crossover after positive PFS 0.516 0.141 0.913
#> 3 crossover after PFS, all trials 0.550 0.134 0.918
#> 4 50% switch at progression 0.380 0.082 0.950
#> mean_OS_month
#> 1 34.4
#> 2 35.6
#> 3 35.7
#> 4 35.9Switching pulls the intention-to-treat hazard ratio for OS toward one and lowers the power of the OS test. The two crossover scenarios differ only in the trials where PFS was not significant, in which crossover is opened in the second scenario but not in the first, so opening crossover in every trial dilutes the OS comparison more.
The whole study, from data generation to the OS analysis of one scenario, runs in about a second.
system.time({
d0 <- simdata_fast(nsim = nsim, n = c(300, 300), a.time = c(0, 18),
a.rate = 600 / 18,
h01.median = list(8, 12), h02.median = list(30, 36),
h12.median = list(12, 12),
d.hazard = -log(1 - 0.05) / 12, seed = seed)
pc <- d0 |>
cutoff_fast(event.looks = 380, tte.col = "e1_tte", event.col = "e1_event")
pr <- d0 |>
analysis_fast(control = 1, cutoff.looks = pc, side = 1,
tte.col = "e1_tte", event.col = "e1_event")
pos <- (pr$reached & pr$logrank.p <= 0.025) %in% TRUE
r1 <- d0 |>
switch_fast(group = 1, when = "later", cutoff = pc, sims = pos,
aft.factor = 1.3) |>
os_final()
})
#> user system elapsed
#> 0.39 0.08 0.51The acceleration factor of switch_fast() is the causal
model of the rank-preserving structural failure time (RPSFT) method of
Robins and Tsiatis (1991). If the experimental treatment multiplies the
time spent on it by a factor f, both from randomization in
the experimental group and after the switch in the control group, the
RPSFT estimate of psi = -log(f) should be close to the true
value. The check below simulates 10 large trials in which the
experimental group has all transition hazards divided by f
(an accelerated failure time effect of f), lets 60 percent
of the control patients who progress switch with the same factor, and
estimates psi in each trial with the rpsftm package
(Allison, White, and Bond, 2017), ignoring censoring.
f <- 1.5
h01 <- log(2) / 8
h02 <- log(2) / 30
h12 <- log(2) / 12
n_rep <- 10
big <- simdata_fast(
nsim = n_rep, n = c(2000, 2000), a.time = c(0, 12), a.rate = 4000 / 12,
h01.hazard = list(h01, h01 / f), h02.hazard = list(h02, h02 / f),
h12.hazard = list(h12, h12 / f), seed = 7
) |>
switch_fast(group = 1, prob = 0.6, when = "intermediate", aft.factor = f,
seed = 8)
suppressWarnings(suppressPackageStartupMessages({
library(survival)
library(rpsftm)
}))
psi_hat <- vapply(seq_len(n_rep), function(s) {
one <- big[big$sim == s, ]
rp_dat <- data.frame(
time = one$e2_surv_time,
status = 1,
arm = one$group - 1,
cens = 1e6
)
# Proportion of the observed time spent on the experimental treatment.
rp_dat$rx <- ifelse(rp_dat$arm == 1, 1,
ifelse(one$switched == 1,
(one$e2_surv_time - one$switch_time) /
one$e2_surv_time, 0))
fit <- tryCatch(
rpsftm(Surv(time, status) ~ rand(arm, rx), data = rp_dat,
censor_time = cens),
error = function(e) NULL
)
if (is.null(fit)) NA_real_ else unname(fit$psi)
}, numeric(1))
round(c(true_psi = -log(f), mean_estimate = mean(psi_hat, na.rm = TRUE),
sd_estimate = sd(psi_hat, na.rm = TRUE),
se_of_mean = sd(psi_hat, na.rm = TRUE) / sqrt(sum(!is.na(psi_hat)))), 4)
#> true_psi mean_estimate sd_estimate se_of_mean
#> -0.4055 -0.3929 0.0326 0.0103The mean of the estimates is close to the true value relative to its standard error, so the RPSFT method recovers the acceleration factor used to generate the switched survival times.
Some switching rules are part of the data-generating process rather
than decisions taken at an analysis. simdata_fast()
generates them in the illness-death model together with the other
transitions:
switch.prop: a switch at progression with a given
probability;h1s.hazard: a switch after progression, for the
patients who did not switch at it, with a (piecewise) constant hazard
that competes with death;h0s.hazard: a switch before progression, after which
progression and death follow h01.switch.hazard and
h02.switch.hazard; with
e1.censor.switch = TRUE, PFS is censored at the
switch;h12.switch.hazard or switch.aft.factor:
the effect of a switch on survival, as a new hazard or as an
acceleration factor of the remaining time.The post-progression hazards are functions of the time since
progression by default (switch.clock = "reset") or of the
time since randomization (switch.clock = "forward"), and
h12.switch.clock sets whether the post-switch hazard is
measured from the switch or on the same clock. The rules follow the
multistate model of the DesignComplexSurvival
package; the script validation/validate_switching.R of the
installed package compares the data of simdata_fast() with
its analytic values and with independent simulations in Python.
In the trial above, suppose that 30 percent of the control patients
switch at progression, that the others switch after progression at 4
percent per month, and that 1 percent per month switch before
progression, in which case PFS is censored at the switch. After a switch
at or after progression the median survival is 18 months. After a switch
before progression, death before progression keeps the control hazard
(h02.median of 30 months), and death after a later
progression has the median of 18 months, on the clock of
switch.clock (from progression by default).
df_rules <- simdata_fast(
nsim = nsim,
n = c(300, 300),
a.time = c(0, 18),
a.rate = 600 / 18,
h01.median = list(8, 12),
h02.median = list(30, 36),
h12.median = list(12, 12),
switch.prop = list(0.3, 0), # at progression
h1s.hazard = list(0.04, 0), # after progression, per month
h0s.hazard = list(0.01, 0), # before progression, per month
h12.switch.median = 18, # survival after a switch
e1.censor.switch = TRUE, # PFS censored at an early switch
d.hazard = -log(1 - 0.05) / 12,
seed = seed
)
# Kind of switch in the control group, from the switch time and the time of
# leaving the progression-free state
ctl <- df_rules[df_rules$group == 1, ]
kind <- ifelse(ctl$switched == 0, "no switch",
ifelse(ctl$switch_time < ctl$e1_surv_time,
"before progression",
ifelse(ctl$switch_time == ctl$e1_surv_time,
"at progression", "after progression")))
round(prop.table(table(kind)), 3)
#> kind
#> after progression at progression before progression no switch
#> 0.192 0.209 0.080 0.519The OS analysis of these data can be compared with the scenarios in the table above.
switch_fast() covers switching rules whose timing is a
function of the patient’s own outcomes (progression), of an opening time
per trial, or of both, with an effect on the remaining time to the
terminal event given by an acceleration factor or by a new hazard. It
can also be applied to data with the switching rules of
simdata_fast(), for example for a crossover after an
interim analysis; the patients who already switched are left unchanged.
Designs in which an interim decision changes the enrollment or the
randomization itself, such as dropping an arm and re-randomizing new
patients, change the data-generating process and are outside this
approach.
Allison, A., White, I. R., & Bond, S. (2017). rpsftm: an R package for rank preserving structural failure time models. The R Journal, 9(2), 342-353.
Robins, J. M., & Tsiatis, A. A. (1991). Correcting for non-compliance in randomized trials using rank preserving structural failure time models. Communications in Statistics - Theory and Methods, 20(8), 2609-2631.
Zhang, H. (2026). TrialSimulator: Clinical Trial Simulator. R package version 1.39.0. https://CRAN.R-project.org/package=TrialSimulator