Basic examples"

In adaptr: Adaptive Trial Simulator

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

This vignette provides examples of various trial specifications using different combinations of settings, including various randomisation strategies including fixed randomisation, response-adaptive randomisation (RAR), and combinations.

The general-purpose function for specifying a trial is setup_trial(), but because trials with binary, binomially distributed and continuous, normally distributed outcomes are so common, the package comes with two convenience functions for specifying such trial designs (using default priors only): setup_trial_binom() and setup_trial_norm().

To keep things simple, this vignette uses only the setup_trial_binom() function and focuses on settings that apply to trial designs regardless of outcome type. The code is heavily annotated, but comments focus on settings not touched on earlier in the vignette (e.g. we do not keep annotating the arm and true_ys arguments).

Keep in mind that the calibrate_trial() function can be used to calibrate a trial specification to obtain a specific value for a certain performance metric (e.g., the Bayesian type 1 error rate for trial specifications with no between-arm differences).

For a general overview of how to use the adaptr package, please see vignette("Overview", "adaptr").

An advanced example on how to specify a trial design with setup_trial(), including the use of custom functions for generating outcomes and yielding posterior draws, is provided in vignette("Advanced-example", "adaptr").

First, the package is loaded:

library(adaptr)

Trial designs without a common control arm

In this section, several examples for trials without a common control arm are provided. General settings applicable for all trial designs (including both trial specifications with and without a common control arm) are covered in this section.

Example 1: general settings

setup_trial_binom(
  # Four arms
  arms = c("A", "B", "C", "D"),
  # Set true outcomes (in this example event probabilities) for all arms
  true_ys = c(0.3, 0.35, 0.31, 0.27), # 30%, 34%, 31% and 27%, respectively

  # Set starting allocation probabilities
  # Defaults to equal allocation if not specified
  start_probs = c(0.3, 0.3, 0.2, 0.2),
  # Set fixed allocation probability for first arm
  # NA corresponds to no limits for specific arms
  # Default (NULL) corresponds to no limits for all arms
  fixed_probs = c(0.3, NA, NA, NA),
  # Set minimum and maximum probability limits for some arms
  # NA corresponds to no limits for specific arms
  # Default (NULL) corresponds to no limits for all arms
  # Must be NA for arms with fixed_probs (first arm in this example)
  # sum(fixed_probs) + sum(min_probs) must not exceed 1
  # sum(fixed_probs) + sum(max_probs) may be greater than 1, and must be at least
  # 1 if specified for all arms
  min_probs = c(NA, 0.2, NA, NA),
  max_probs = c(NA, 0.7, NA, NA),

  # Set looks - alternatively, specify both max_n AND look_after_every
  data_looks = seq(from = 300, to = 1000, by = 100),

  # No common control arm (as default, but explicitly specified in this example)
  control = NULL,

  # Set inferiority/superiority thresholds (different values than the defaults)
  # (see also the calibrate_trial() function)
  inferiority = 0.025,
  superiority = 0.975,

  # Define that the outcome is desirable (as opposed to the default setting)
  highest_is_best = TRUE,

  # No softening (the default setting, but made explicit here)
  soften_power = 1,

  # Use different simulation/summary settings than default
  cri_width = 0.89, # 89% credible intervals
  n_draws = 1000, # Only 1000 posterior draws in each arm
  robust = TRUE, # Summarise posteriors using medians/MAD-SDs (as default)

  # Trial description (used by print methods)
  description = "example trial specification 1"
)

Example 2: equivalence testing, decreasing softening

• No common control arm
• Equivalence testing
• Different softening powers (decreasing softening as the trial progresses)
• Default settings for many unspecified arguments

setup_trial_binom(
  # Specify arms and true outcome probabilities (undesirable outcome as default)
  arms = c("A", "B", "C", "D"),
  true_ys = c(0.2, 0.22, 0.24, 0.18),

  # Specify adaptive analysis looks using max_n and look_after_every
  # max_n does not need to be a multiple of look_after_every - a final look
  # will be conducted at max_n regardless
  max_n = 1250, # Maximum 1250 patients
  look_after_every = 100, # Look after every 100 patients

  # Assess equivalence between all arms: stop if >90 % probability that the
  # absolute difference between the best and worst arms is < 5 %-points
  # Note: equivalence_only_first must be NULL (default) in designs without a
  # common control arm (such as this trial)
  equivalence_prob = 0.9,
  equivalence_diff = 0.05,

  # Different softening powers at each look (13 possible looks in total)
  # Starts at 0 (softens all allocation probabilities to be equal) and ends at
  # 1 (no softening) for the final possible look in the trial
  soften_power = seq(from = 0, to = 1, length.out = 13)
)

Trial designs with a common control arm

In this section, several examples for trials with a common control arm are provided and focus mostly on options specific to trial designs with a common control arm.

Example 3: common control and sqrt-based fixed allocation

• A common control arm
• square-root-transformation-based fixed allocation probabilities (see description in setup_trial())
• Assessment of both equivalence and futility compared to the initial control only (not assessed for superior arms that become subsequent controls)

setup_trial_binom(
  arms = c("A", "B", "C", "D"),
  # Specify control arm
  control = "A",

  true_ys = c(0.2, 0.22, 0.24, 0.18), 

  data_looks = seq(from = 100, to = 1000, by = 100),

  # Fixed, square-root-transformation-based allocation throughout
  control_prob_fixed = "sqrt-based fixed",

  # Assess equivalence: drop non-control arms if > 90% probability that they
  # are equivalent to the common control, defined as an absolute difference of
  # < 3 %-points
  equivalence_prob = 0.9,
  equivalence_diff = 0.03,
  # Only assess against the initial control (i.e., not assessed if an arm is
  # declared superior to the initial control and becomes the new control)
  equivalence_only_first = TRUE,

  # Assess futility: drop non-control arms if > 80% probability that they are
  # < 10 %-points better (in this case lower because outcome is undesirable in
  # this example) compared to the common control
  futility_prob = 0.8,
  futility_diff = 0.1,
  # Only assessed for the initial control, as described above
  futility_only_first = TRUE
)

Example 4: sqrt-based initial allocation and restricted RAR

• Square-root-transformation-based initial allocation probabilities
• Square-root-transformation-based allocation to the control arm (including subsequent controls, if a non-control arm is declared superior to the initial control)
• Restricted response-adaptive randomisation to the non-control arms

setup_trial_binom(
  arms = c("A", "B", "C", "D"),
  control = "A",

  true_ys = c(0.2, 0.22, 0.24, 0.18), 

  data_looks = seq(from = 100, to = 1000, by = 100),

  # Square-root-transformation-based control arm allocation including for
  # subsequent controls and initial equal allocation to the non-control arms,
  # followed by response-adaptive randomisation
  control_prob_fixed = "sqrt-based",

  # Restricted response-adaptive randomisation
  # Minimum probabilities of 20% for non-control arms, must be NA for the
  # control arm with fixed allocation probability
  # Limits are ignored for arms that become subsequent controls
  # Limits are rescaled (i.e., increased proportionally) when arms are dropped
  min_probs = c(NA, 0.2, 0.2, 0.2),
  rescale_probs = "limits",

  # Constant softening of 0.5 (= square-root transformation)
  soften_power = 0.5
)

Example 5: sqrt-based allocation only to initial control arm

This example is similar to that above (with different restriction settings), but only use square-root-transformation-based allocation probabilities to the initial control arm. Hence, this will not apply if another arm is declared superior and becomes the new control.

setup_trial_binom(
  arms = c("A", "B", "C", "D"),
  control = "A",

  true_ys = c(0.2, 0.22, 0.24, 0.18), 

  data_looks = seq(from = 100, to = 1000, by = 100),

  # Square-root-transformation-based control arm allocation for the initial
  # control only and initial equal allocation to the non-control arms, followed
  # by response-adaptive randomisation
  control_prob_fixed = "sqrt-based start",

  # Restrict response-adaptive randomisation
  # Minimum probabilities of 20% for all non-control arms
  # - must be NA for the initial control arm with fixed allocation probability
  min_probs = c(NA, 0.2, 0.2, 0.2),
  # Maximum probabilities of 65% for all non-control arms
  # - must be NA for the initial control arm with fixed allocation probability
  max_probs = c(NA, 0.65, 0.65, 0.65),

  soften_power = 0.75
)

Example 6: restricted RAR, matched control-arm allocation

• Restricted response-adaptive randomisation
• Control-arm allocation probability matched to that of the highest non-control arm (with re-scaling as necessary)
• Applies to both the initial and subsequent control arms

setup_trial_binom(
  arms = c("A", "B", "C", "D"),
  control = "A",

  true_ys = c(0.2, 0.22, 0.24, 0.18), 

  data_looks = seq(from = 100, to = 1000, by = 100),

  # Specify starting probabilities
  # When "match" is specified below in control_prob_fixed, the initial control
  # arm's initial allocation probability must match the highest initial
  # non-control arm allocation probability
  start_probs = c(0.3, 0.3, 0.2, 0.2),

  control_prob_fixed = "match",

  # Restrict response-adaptive randomisation 
  # - these are applied AFTER "matching" when calculating new allocation
  #   probabilities 
  # - min_probs must be NA for the initial control arm when using matching
  min_probs = c(NA, 0.2, 0.2, 0.2),

  soften_power = 0.7
)

Example 7: follow-up and data collection lag

This example uses the randomised_at_looks argument to specify follow-up and/or data collection lag. In real use cases, this should usually be considered, as this may affect the relative performance of different trial designs and the extent to which the 'final' results after all patients have reached follow-up and are analysed may differ from the results from the adaptive analyses with some randomised patients not included due to outcome data not being available yet for these patients.

setup_trial_binom(
  arms = c("A", "B", "C", "D"),
  control = "A",
  true_ys = c(0.2, 0.22, 0.24, 0.18), 

  # Analyses conducted every time 100 patients have follow-up data available
  data_looks = seq(from = 100, to = 1000, by = 100),
  # Specify the number of patients randomised at each look - in this case, 200
  # more patients are randomised than the number of patients that
  # have follow-up data available at each look
  randomised_at_looks = seq(from = 300, to = 1200, by = 100)
)

Example 8: different probability thresholds over time

In this example, we specify different probability thresholds for superiority and inferiority stopping rules at different adaptive analyses. Varying probability thresholds may similarly be specified for stopping rules for equivalence and futility. Importantly, all probability thresholds must be specified such that each subsequent threshold is never stricter than the previous threshold. Varying thresholds may also be used to make some stopping rules first function at later analyses (e.g., as long as the stopping threshold for superiority is 1 and the stopping threshold for inferiority is 0, trials will not be stopped and arms will not be dropped due to these rules).

setup_trial_binom(
  arms = c("A", "B", "C", "D"),
  control = "A",
  true_ys = c(0.2, 0.22, 0.24, 0.18), 

  # Analyses conducted every time 100 patients have follow-up data available
  data_looks = seq(from = 100, to = 1000, by = 100),

  # Specify varying inferiority/superiority thresholds
  # When specifying varying thresholds, the number of thresholds must match
  # the number of analyses, and thresholds may never be stricter than the
  # threshold used in the previous analysis
  # Superiority threshold decreasing from 0.99 to 0.95 during the first five
  # analyses, and remains stationary at 0.95 after that
  superiority = c(seq(from = 0.99, to = 0.95, by = -0.01), rep(0.95, 5)),
  # Similarly for inferiority thresholds, but in the opposite direction
  inferiority = c(seq(from = 0.01, to = 0.05, by = 0.01), rep(0.05, 5)),
)

Example 9: minimum allocation probabilities rescaled when arms are dropped

In this example, a trial design with four arms and restricted RAR (minimum allocation limits) is specified, with additional specification that the minimum allocation limits should be rescaled proportionally when arms are dropped (rescaling can similarly be applied to fixed allocation probabilities):

setup_trial_binom(
  arms = c("A", "B", "C", "D"),
  control = "A",
  true_ys = c(0.2, 0.2, 0.2, 0.2),

  min_probs = rep(0.15, 4), # Specify initial minimum allocation probabilities
  # Rescale allocation probability limits as arms are dropped
  rescale_probs = "limits", 

  data_looks = seq(from = 100, to = 1000, by = 100)
)

adaptr documentation built on May 29, 2024, 7:48 a.m.