Binary DLT Likelihoods in Escalation Trials

Why the Likelihood Starts With the Outcome Definition

A dose-escalation trial may record many clinical details, but a binary DLT likelihood uses a particular summary: for each patient, whether a dose-limiting toxicity occurred within a defined observation window. The likelihood then describes how plausible the observed pattern of DLT and non-DLT outcomes is under specified dose-specific toxicity probabilities. Its construction depends first on defining the outcome and deciding which patients have enough follow-up to classify it.

Let \(Y_{ij}\) denote the binary outcome for patient \(i\) treated at dose level \(j\). Set \(Y_{ij}=1\) if that patient experiences a DLT during the prespecified assessment window, and \(Y_{ij}=0\) if the patient completes that window without a DLT. Write \(p_j\) for the probability of a DLT within that window at dose level \(j\). In this tutorial, each \(p_j\) is a probability associated with a dose, not a model linking different doses.

This definition makes the assessment window part of the endpoint. For example, “DLT by day 28” and “DLT by day 42” are not interchangeable binary outcomes: a patient might have no DLT by day 28 but experience one later. The DLT criteria, observation window, and rules for assigning outcomes should therefore be specified consistently before assembling the likelihood.

From Patient Outcomes to a Bernoulli Likelihood

Suppose the observed outcomes at one dose level are \(y_1,\ldots,y_n\), each equal to zero or one. If the patients’ outcomes are conditionally independent given the common dose-specific probability \(p\), the joint probability of this particular labeled sequence is the product of the individual Bernoulli probabilities. As a function of \(p\), that joint probability is the likelihood.

The product form follows because \(p^{y_i}(1-p)^{1-y_i}\) equals \(p\) for an observed DLT and \(1-p\) for an observed non-DLT. Multiplying across patients gives one factor of \(p\) for each of the \(y\) DLTs and one factor of \(1-p\) for each of the \(n-y\) non-DLTs. A likelihood is a function of the unknown parameter for the data that were observed; it is not a probability distribution over possible values of that parameter.

Aggregating Outcomes by Dose

At dose level \(j\), let \(n_j\) be the number of patients with classifiable binary outcomes and let \(y_j\) be the number of those patients with DLTs. Under conditional independence and a common probability \(p_j\) for patients at that dose, the count has a binomial distribution. Its probability mass includes a combinatorial coefficient because it counts all possible patient-level arrangements with \(y_j\) DLTs.

For several dose levels, if patient outcomes are conditionally independent given their dose-specific probabilities, the likelihood for the observed counts is

Here \(\boldsymbol{p}\) denotes the collection of dose-specific probabilities. For parameter estimation or comparing values of \(\boldsymbol{p}\), the coefficients \(\binom{n_j}{y_j}\) can be omitted when they are constant across the parameter values being considered. They must be retained when the goal is the probability of the observed count itself. The distinction is between the probability of a count and the probability of one specific labeled sequence, not between two different Bernoulli assumptions.

Why the Count Likelihood Has This Form

The binomial likelihood can be justified directly from patient-level outcomes. For any one specific arrangement with \(y\) DLTs and \(n-y\) non-DLTs, independence gives probability \(p^y(1-p)^{n-y}\). There are \(\binom{n}{y}\) distinct arrangements of \(y\) DLTs among \(n\) labeled patients. These arrangements are mutually exclusive and each has the same probability, so summing their probabilities gives the binomial probability of the count.

A useful related result is that the likelihood for one dose is maximized at the observed DLT proportion when the count is strictly between zero and \(n\). To see this, take the logarithm of the likelihood without its parameter-independent count coefficient:

For \(0<y<n\), differentiating and setting the derivative to zero gives

The second derivative, \(-y/p^2-(n-y)/(1-p)^2\), is negative for \(0<p<1\), so this stationary point is a maximum. If \(y=0\), the likelihood is maximized at \(p=0\); if \(y=n\), it is maximized at \(p=1\). These boundary estimates describe the maximum of this dose-specific likelihood. They do not by themselves establish that the underlying toxicity probability is known precisely, particularly when \(n\) is small.

Assumptions, Diagnostics, and Alternatives

The basic likelihood is appropriate only when its assumptions match the data and endpoint. In particular, the analysis should make clear which DLT definition and time window apply, which patients have classifiable outcomes, and why a common probability is reasonable within each dose level. Conditional independence means that, given the dose-specific probability, one patient’s binary outcome does not provide additional dependence information about another patient’s outcome. If outcomes are clustered or otherwise dependent, the simple product likelihood may understate uncertainty.

Escalation trials often assign later patients’ doses using earlier outcomes. This adaptive allocation does not automatically invalidate a likelihood for the observed outcomes. When the allocation rule is based on recorded trial history and its assignment probabilities do not involve the unknown toxicity probabilities beyond that history, the outcome likelihood is constructed conditional on the assigned doses and observed history. The dose assignment rule still matters for the design and interpretation of the trial; it is not an extra Bernoulli DLT factor.

Useful checks include verifying that every included patient has a defined dose and a consistent outcome window, that \(0\leq y_j\leq n_j\), and that the total number of classifiable outcomes agrees with the patient records. Compare the patient-level outcome list with the dose-level counts; a count that does not equal the sum of the indicators signals a data or coding error. Also inspect whether a zero DLT count means all patients completed the window without DLT, or instead includes patients whose follow-up is still pending. Those are not equivalent data states.

The simple binary likelihood does not use the timing of DLTs within the window. It also does not, by itself, model differences between patients, dependence, or partially observed follow-up. A time-to-event approach can use event times and handle censoring under its own assumptions; it is an alternative when follow-up timing is central, not a reason to label an incomplete patient as a non-DLT. A dose-response model can relate probabilities across ordered doses, whereas separate dose-specific Bernoulli likelihoods leave each \(p_j\) unrestricted. The latter is transparent but may be imprecise when there are few patients per dose. The logistic dose-toxicity model is treated separately in the next topic.

The main practical advantage of an explicit Bernoulli likelihood is that each factor has a clear interpretation and the aggregated count form is easy to verify. Its limitations are equally direct: the binary summary discards event timing, a small count can leave substantial uncertainty about \(p_j\), and an unrestricted set of dose-specific probabilities does not borrow information across doses. The appropriate likelihood is therefore determined by the endpoint, follow-up, and assumptions—not by which formula is easiest to compute.