Calculate dose-transition pathways for a CRM study

Calculate dose-transition pathways (DTPs, Yap et al, 2017) for a dose-finding trial using the continual reassessment method (CRM) design. DTPs are a glimpse into the future for an in-progress trial. They tell us what the model would advise for all feasible future outcomes. They can be used in the design stages to detect possible undesirable behaviour. They can be used during the trial to aid planning and understanding.

crm_dtps(
  skeleton,
  target,
  model,
  cohort_sizes,
  previous_outcomes = "",
  next_dose = NULL,
  user_dose_func = NULL,
  verbose = FALSE,
  i_am_patient = FALSE,
  ...
)

Arguments

skeleton	a vector of the prior guesses of toxicity at doses. This should be a monotonically-increasing vector of numbers between 0 and 1.
target	the target toxicity probability, a number between 0 and 1. This value would normally be one of the values in skeleton, but that is not a requirement.
model	Character string to denote desired model. One of empiric, logistic, logistic_gamma, or logistic2. The choice of model determines which extra parameters are required by .... See Details.
cohort_sizes	vector of future cohort sizes, i.e. positive integers. E.g. To calculate paths for the the next cohort of two followed by another cohort of three, use cohort_sizes = c(2, 3).
previous_outcomes	Outcomes observed hitherto in the syntax required by df_parse_outcomes.
next_dose	optional, integer (1-based) dose-level to be given to the next cohort. If omitted, the dose suggested by the model is used.
user_dose_func	optional delegate for deciding dose. A function that takes a crm_fit as the sole argument and returns the integer (1-based) dose-level to be given next, or NA to show that no dose should be chosen and the trial stopped. This function gives the user the opportunity to build in custom behaviour to tailor the dose selection decision in response to the insights garnered by the fit model, or recommend that a trial path be halted immediately. If omitted, the dose ordinarily chosen by the model is used. An example is given below.
verbose	logical, TRUE to get log messages.
i_am_patient	logical. The number of paths to analyse grows faster than linearly in the number of future cohorts to resolve. Fitting many models by MCMC can take a long time. This function will not proceed unless you signify your patience when the number of paths to reolve exceeds 100.
...	Extra parameters passed to stan_crm.

Value

A list of dose_finding_path_node objects.

Details

Different model choices require that different parameters are provided. See below.

Parameter requirements of empiric model

• beta_sd

Parameter requirements of logistic model

• a0

• beta_mean

• beta_sd

Parameter requirements of logistic_gamma model

• a0

• beta_shape

• beta_inverse_scale

Parameter requirements of logistic2 model

• alpha_mean

• alpha_sd

• beta_mean

• beta_sd

References

Yap C, Billingham LJ, Cheung YK, Craddock C, O’Quigley J. Dose transition pathways: The missing link between complex dose-finding designs and simple decision-making. Clinical Cancer Research. 2017;23(24):7440-7447. doi:10.1158/1078-0432.CCR-17-0582

See also

df_parse_outcomes, stan_crm, crm_path_analysis, dose_finding_path_node

Examples

if (FALSE) {
target <- 0.25
skeleton <- c(0.05, 0.15, 0.25, 0.4, 0.6)

# Run DTPs for the first two cohorts of two for new a trial:
paths <- crm_dtps(skeleton = skeleton, target = target, model = 'empiric',
                  cohort_sizes = c(2, 2), next_dose = 3, beta_sd = 1)
length(paths)  # 13

library(tibble)
df <- as_tibble(paths)
df


# Run DTPs for the next cohort of three in a trial that has already treated
# six patients, seeing some toxicity at dose-level 3:
paths2 <- crm_dtps(skeleton = skeleton, target = target, model = 'empiric',
                   cohort_sizes = c(3), previous_outcomes = '2NNN 3TTN',
                   beta_sd = 1)
length(paths2)  # 5
as_tibble(paths2)
# We see that de-escalation to dose-level 2 should occur now, and that any
# further toxicity will result in advice for further de-escalation to
# dose-level 1.


# An example with a custom dose selection function
paths3 <- crm_dtps(skeleton = skeleton, target = target, model = 'empiric',
                   cohort_sizes = c(3, 3), previous_outcomes = '2NN 3TN',
                   next_dose = 2, beta_sd = 1,
                   user_dose_func = function(x) {
                     careful_escalation(x, tox_threshold = target + 0.1,
                                        certainty_threshold = 0.7)
                   }, seed = 123, refresh = 0)
spread_paths(as_tibble(paths3) %>% select(-fit, -parent_fit, -dose_index))
# Stopping is recommended when the dose selection function returns NA.
}