The toxicity probability interval (TPI)is a dose-escalation design by Ji et al.
get_tpi(
num_doses,
target,
k1,
k2,
exclusion_certainty,
alpha = 0.005,
beta = 0.005,
...
)Number of doses under investigation.
We seek a dose with this probability of toxicity.
The K1 parameter in TPI determines the upper bound of the equivalence interval. See Details.
The K2 parameter in TPI determines the lower bound of the equivalence interval. See Details.
Numeric, threshold posterior certainty required to exclude a dose for being excessively toxic. The authors discuss values in the range 0.7 - 0.95. Set to a value > 1 to suppress the dose exclusion mechanism. The authors use the Greek letter xi for this parameter.
First shape parameter of the beta prior distribution on the probability of toxicity.
Second shape parameter of the beta prior distribution on the probability of toxicity.
Extra args are passed onwards.
an object of type selector_factory that can fit the
TPI model to outcomes.
The design seeks a dose with probability of toxicity \(p_{i}\) close to a target probability \(p_{T}\) by iteratively calculating the interval $$p_{T} - K_{2} \sigma_{i} < p_{i} < p_{T} + K_{1} \sigma_{i}$$ In this model, \(K_{1}\) and \(K_{2}\) are specified constants and \(\sigma_{i}\) is the standard deviation of \(p_{i}\) arising from a Bayesian beta-binomial conjugate model $$p_{i} | data \sim Beta(\alpha + x_{i}, \beta + n_{i} - x_{i}),$$ where \(x_{i}\) is the number of toxicities observed and \(n_{i}\) is the number of patients treated at dose \(i\), and \(\alpha\) and \(\beta\) are hyperparameters for the beta prior on \(p_{i}\). A dose is excluded as inadmissible if $$P(p_{i} > p_{T} | data) > \xi$$ The trial commences at a starting dose, possibly dose 1. If dose \(i\) has just been evaluated in patient(s), dose selection decisions proceed by calculating the posterior probability that the true toxicity rate at dose \(i\) belongs to the three partition regions \(p_{i} < p_{T} - K_{2} \sigma_{i}\), \(p_{T} - K_{2} \sigma_{i} < p_{i} < p_{T} + K_{1} \sigma_{i}\), and \(p_{i} > p_{T} + K_{2} \sigma_{i}\), corresponding to decisions escalate, stay, and de-escalate dose, respectively. Further to this are rules that prevent escalation to an inadmissible dose. In their paper, the authors demonstrate acceptable operating performance using \(\alpha = \beta = 0.005\), \(K_{1} = 1\), \(K_{2} = 1.5\) and \(\xi = 0.95\). See the publications for full details.
Ji, Y., Li, Y., & Bekele, B. N. (2007). Dose-finding in phase I clinical trials based on toxicity probability intervals. Clinical Trials, 4(3), 235–244. https://doi.org/10.1177/1740774507079442
Ji, Y., & Yang, S. (2017). On the Interval-Based Dose-Finding Designs, 1–26. Retrieved from https://arxiv.org/abs/1706.03277
target <- 0.25
model1 <- get_tpi(num_doses = 5, target = target, k1 = 1, k2 = 1.5,
exclusion_certainty = 0.95)
outcomes <- '1NNN 2NTN'
model1 %>% fit(outcomes) %>% recommended_dose()
#> [1] 2