An Example of Fixed Design with Two Correlated Endpoints
In TrialSimulator: Clinical Trial Simulator

knitr::opts_chunk$set(
  collapse = TRUE,
  cache.path = 'cache/fixedDesign/',
  comment = '#>',
  dpi = 300,
  out.width = '100%'
)

library(dplyr)
library(kableExtra)
library(TrialSimulator)
set.seed(12345)

In this vignette, we illustrate how to use TrialSimulator to simulate a trial of two correlated endpoints. Under a fixed design, only one milestone is specified for final analysis.

Simulation Settings

Trial consists of two active arms of high and low dose, and a standard-of-care arm.
Patients are randomized into the trial with ratio 1:1:1.
30 patients are randomized per month for the first 10 months, and 50 patients per month after then until 1000 patients.
Dropout rate is 10% at month 18, which is modeled by an exponential distribution, with rate parameter -log(1 - 0.1)/18.
Modeled by a three-state ill-death model, two endpoints PFS and OS are simulated.
- Primary endpoint OS has a medians of 15, 18.5, 20 months in standard-of-care, low dose and high dose arms, respectively.
- Key secondary endpoint PFS has a median of 7, 9, 10 months in the three arms.
- Pearson's correlation between OS and PFS are 0.68, 0.65, and 0.60 in the three arms.
To ensures sufficient powers for both endpoints, final analysis is set when we have at least a total of 800 PFS events in the standard-of-care and high dose arm, meanwhile a total of 550 OS events are observed in three arms
- OS is tested using one-sided logrank test.
- PFS is tested using one-sided p-value from the Cox proportional hazard model as the PH assumption is assumed.
- The Bonferroni correction is adopted to split overall $\alpha = 5\%$, i.e., claiming significant effect for an endpoint when p-value is lower than $0.05/4$.

Transition Hazards of `PFS` and `OS`

We adopt the ill-death model to simulate the two endpoints. This ensures PFS $\leq$ OS with probability one, and makes no assumption on latent variables or copula parameters. TrialSimulator offers a function solveThreeStateModel() to convert endpoints' medians and correlation to the transition hazards, which are required by the built-in generator CorrelatedPfsAndOs3(). Refer to the vignette Simulate Correlated Progression-Free Survival and Overall Survival as Endpoints in Clinical Trials for more details.

pars_soc <- solveThreeStateModel(median_pfs = 7, median_os = 15, corr = .68, 
                                 h12 = seq(.07, .10, length.out = 50))
pars_soc

pars_low <- solveThreeStateModel(median_pfs = 9, median_os = 18.5, corr = .65, 
                                 h12 = seq(.04, .07, length.out = 50))
pars_low

pars_high <- solveThreeStateModel(median_pfs = 10, median_os = 20, corr = .60, 
                                  h12 = seq(.02, .06, length.out = 50))
pars_high

rbind(pars_soc, pars_low, pars_high) %>% 
  structure(class = 'data.frame') %>% 
  mutate(arm = c('soc', 'low', 'high')) %>% 
  select(arm, h01, h02, h12) %>% 
  kable(format = 'html', digits = 3, 
        caption = 'Transition hazards in three treatment arms')

Define Treatment Arms

We use generator CorrelatedPfsAndOs3() with endpoint() and arm() to define the three treatment arms.

#' define SoC
pfs_os_in_soc <- endpoint(name = c('pfs', 'os'), 
                          type = c('tte', 'tte'), 
                          generator = CorrelatedPfsAndOs3, 
                          h01 = 0.075, h02 = 0.024, h12 = 0.090)

soc <- arm(name = 'soc')
soc$add_endpoints(pfs_os_in_soc)

#' define low dose arm
pfs_os_in_low <- endpoint(name = c('pfs', 'os'), 
                          type = c('tte', 'tte'), 
                          generator = CorrelatedPfsAndOs3, 
                          h01 = 0.051, h02 = 0.026, h12 = 0.062)

low <- arm(name = 'low')
low$add_endpoints(pfs_os_in_low)

#' define high dose arm
pfs_os_in_high <- endpoint(name = c('pfs', 'os'), 
                           type = c('tte', 'tte'), 
                           generator = CorrelatedPfsAndOs3, 
                          h01 = 0.040, h02 = 0.030, h12 = 0.047)

high <- arm(name = 'high')
high$add_endpoints(pfs_os_in_high)

We can request for a summary report of, e.g., the high dose arm, by printing the arm object in R console. The medians of PFS and OS matches to the settings very well.

high

Define a Trial

With three arms, we can define a trial using the function trial(). Recruitment curve are specified through enroller with a built-in function StaggeredRecruiter of piecewise constant rate. We set duration to be an arbitrary large number (500) but controlling the end of trial through a pre-defined milestone later. Note that if seed = NULL, TrialSimulator will pick a seed for the purpose of reproducibility.

accrual_rate <- data.frame(end_time = c(10, Inf),
                           piecewise_rate = c(30, 50))
trial <- trial(
  name = 'Trial-3415', n_patients = 1000,
  seed = 1727811904, duration = 500,
  enroller = StaggeredRecruiter, accrual_rate = accrual_rate,
  dropout = rexp, rate = -log(1 - 0.1)/18, ## 10% by month 18
  silent = TRUE
)

trial$add_arms(sample_ratio = c(1, 1, 1), soc, low, high) ## 1:1:1
trial

Define Milestone and Action for Final Analysis

To ensure sufficient powers of testing PFS and OS, final analysis is performed when we have at least 700 events for OS and 800 events for PFS. In the action function, we compute one-sided p-value of PFS using the proportional hazard Cox model, and one-sided p-value of OS using the logrank test. This is consistent with the assumption of ill-death model implemented in data generator CorrelatedPfsAndOs3(). Note that five columns are available in locked data: arm, pfs, 'os', 'pfs_event', and 'os_event', which are used to construct model formula. Estimates of hazard ratio are also computed. Built-in functions fitCoxph and fitLogrank return data frames. Refer to their help documants for more details.

action <- function(trial){

  locked_data <- trial$get_locked_data('final')

  pfs <- fitCoxph((Surv(pfs, pfs_event) ~ arm), placebo = 'soc', 
                  data = locked_data, alternative = 'less', 
                  scale = 'hazard ratio')

  os <- fitLogrank((Surv(os, os_event) ~ arm), placebo = 'soc', 
                   data = locked_data, alternative = 'less')

  ## Bonferroni test is applied to four hypotheses: 
  ## PFS_low, PFS_high, OS_low, and OS_high
  pfs$decision <- ifelse(pfs$p < .05/4, 'reject', 'accept')
  os$decision <- ifelse(os$p < .05/4, 'reject', 'accept')

  trial$save(
    value = pfs %>% filter(arm == 'low') %>% select(estimate, decision, info), 
    name = 'pfs_low')

  trial$save(
    value = pfs %>% filter(arm == 'high') %>% select(estimate, decision, info), 
    name = 'pfs_high')

  trial$save(
    value = os %>% filter(arm == 'low') %>% select(decision, info), 
    name = 'os_low')

  trial$save(
    value = os %>% filter(arm == 'high') %>% select(decision, info), 
    name = 'os_high')

}

Now we can define and register the milestone to a listener, which monitors the trial for us through a controller

final <- milestone(name = 'final', action = action, 
                   when = eventNumber(endpoint = 'pfs', n = 450, 
                                      arms = c('soc', 'high')) & 
                     eventNumber(endpoint = 'os', n = 550)
                   )

listener <- listener()
listener$add_milestones(final)

controller <- controller(trial, listener)

We can run a massive number of replicates in simulation to study operating characteristics of a trial design by specifying n in Controller$run(). We can set plot_event = FALSE to turn off plotting to save running time. The simulation results can be accessed by calling the member function get_output() of the controller.

controller$run(n = 1000, plot_event = FALSE, silent = TRUE)
output <- controller$get_output()

output <- TrialSimulator:::getFixedDesignOutput()

output %>% 
  head(5) %>% 
  kable(escape = FALSE) %>% 
  kable_styling(bootstrap_options = "striped", 
                full_width = FALSE,
                position = "left") %>%
  scroll_box(width = "100%")

For example, we can compute the powers and summarize the estimates of hazard ratio for PFS.

output %>% 
  summarise(
    across(matches('_<decision>$'), ~ mean(. == 'reject') * 100, .names = 'Power_{.col}'), 
    across(matches('_<estimate>$'), ~ mean(.x), .names = 'HR_{.col}')
  ) %>%
  rename_with(~ sub('_<decision>$', '', .), starts_with('Power_')) %>%
  rename_with(~ sub('_<estimate>$', '', .), starts_with('HR_')) %>% 
  kable(col.name = NULL, digits = 3, align = 'r') %>% 
  add_header_above(c('Low', 'High', 'Low', 'High', 'Low', 'High'), align = 'r') %>% 
  add_header_above(c('PFS' = 2, 'OS' = 2, 'PFS' = 2)) %>% 
  add_header_above(c('Power (%)' = 4, 'Hazard Ratio' = 2)) %>% 
  kable_styling(full_width = TRUE)

Note that OS does not satisfy the proportional hazard assumption, and a composite condition on event numbers is used to trigger the final analysis. Thus, the powers in the table above would not match to the output from power calculation packages, which is as expected.