The Conventional 3+3 Escalation Algorithm

Purpose and Scope of the Algorithm

The conventional 3+3 design is a rule-based method for escalating through a prespecified dose ladder. Participants are treated in small cohorts, and decisions depend on the number of dose-limiting toxicities (DLTs) observed at the current dose during the defined DLT observation window. The design is called “3+3” because a dose level begins with three evaluable participants and may be expanded by three more.

The algorithm is a sequence of operational rules, not a statistical model that estimates a dose–toxicity curve. Its instructions are meaningful only when the dose levels, DLT definition, observation window, evaluability criteria, and escalation limits have been specified in advance. The earlier material on defining DLTs and observation windows and constructing a dose-escalation ladder supplies those protocol elements; here, the focus is how the conventional algorithm uses the resulting DLT counts.

Definition: At a given dose level, an evaluable participant is one whose DLT outcome can be classified according to the protocol for the specified observation window. The DLT count is the number of evaluable participants in the current cohort with a DLT. The denominator in a rule such as \(1/3\) is the number of evaluable participants at that dose, not necessarily the number initially enrolled.

In its conventional form, the design starts at the lowest planned dose. A cohort of three evaluable participants is treated at that dose. If the count permits escalation, the next cohort is treated at the next planned level. The process is sequential: a dose decision should not be made from an incomplete DLT window unless the protocol contains an explicit rule for handling delayed or pending outcomes.

The Cohort Rules

The initial decision at each dose depends on the DLT count among the first three evaluable participants. Zero DLTs among three permits escalation. One DLT among three triggers expansion at the same dose to six evaluable participants. Two or three DLTs among the first three means escalation stops at that dose. After expansion, one or fewer DLTs among all six permits escalation; two or more means escalation stops.

Core rule: Start with three evaluable participants at a dose. With \(0/3\) DLTs, escalate to the next planned dose. With \(1/3\), treat three additional evaluable participants at the same dose. With \(2/3\) or \(3/3\), stop escalation. After expansion, escalate if the total is \(0/6\) or \(1/6\); stop escalation if the total is at least \(2/6\).

The expansion is not a new dose cohort: its three participants receive the same dose as the first three. The DLT count is combined across all six participants at that level. For example, if the first three include one DLT and the next three include none, the combined result is \(1/6\), which permits escalation under the conventional rule.

“Stop escalation” has a precise, limited meaning: do not move upward to the next planned dose under the algorithm. The conventional rule does not itself prescribe a new cohort at a lower dose. The immediately preceding dose is commonly treated as a candidate for further dose selection, but the status assigned to that dose and any requirement to add participants there must be set by the protocol. If the stopping dose is the lowest planned dose, there is no lower tested dose to nominate. These cases should not be improvised during conduct of the trial.

Worked Applications

Worked Example: Escalating After Zero DLTs

Suppose a hypothetical trial uses four planned doses, \(d_1\) through \(d_4\). At \(d_1\), all three participants complete the DLT window and none has a DLT. At \(d_2\), the first three include one DLT; among the three additional participants at \(d_2\), none has a DLT. What does the algorithm direct?

At \(d_1\): The evaluable count is \(0/3\). The core rule permits escalation, so the next cohort is treated at \(d_2\).

At \(d_2\): The initial count is \(1/3\), so this is not an immediate escalation decision. Three more evaluable participants are treated at \(d_2\). Combining the cohorts gives

\frac{1+0}{3+3}=\frac{1}{6}.

The total is \(1/6\), so escalation to \(d_3\) is permitted. This conclusion assumes that all six participants are evaluable under the protocol and that their DLT windows are complete. It does not claim that the true DLT probability at \(d_2\) is known to be low; it is the action specified by this rule for the observed count.

Worked Example: Stopping at the Starting Dose

In a hypothetical trial, three evaluable participants receive the lowest planned dose. Two experience a protocol-defined DLT during the observation window, and the third does not.

The observed count is \(2/3\). Because at least two of the initial three have DLTs, the algorithm stops escalation immediately at this dose; it does not enroll three more at the same dose. Since this is the lowest planned dose, there is no lower dose in the ladder to nominate as a candidate. The trial team follows the prespecified stopping and safety procedures rather than inventing a lower-dose cohort or continuing upward.

The decision follows from the count, not from a claim that the dose is certainly unsafe for every participant. The observed DLTs are safety outcomes in a small cohort, and the algorithm supplies a conservative next action in response to them.

Worked Example: Stopping After Expansion

At a hypothetical dose \(d_3\), one of the first three evaluable participants has a DLT. The protocol therefore expands the cohort. Of the three additional evaluable participants, one has a DLT and two do not. Determine the combined count and action.

The combined DLT count is

1+1=2 \text{ DLTs among } 3+3=6 \text{ evaluable participants}, \qquad \frac{2}{6}.

Because the total is at least \(2/6\), escalation stops at \(d_3\); the algorithm does not proceed to \(d_4\). If \(d_2\) is the immediately preceding tested dose, it may be considered under the protocol’s dose-selection rule. The 3+3 escalation instruction alone does not establish whether \(d_2\) has sufficient participants or is formally selected for a later purpose.

Why Expansion and Stopping Are Structured This Way

The logic can be justified directly from the algorithm’s decision structure. A result of \(0/3\) meets the rule for moving up without adding participants at the current dose. A result of \(1/3\) is treated as a signal to gather three more outcomes at that same dose. Once six outcomes are available, the combined count is compared with the stopping threshold: \(1/6\) permits escalation, whereas \(2/6\) or more stops it. Thus, expansion changes the denominator while preserving the same dose, and the second decision uses the total across both groups.

This logic is a design convention, not a derivation showing that a particular DLT probability is optimal. In particular, the labels “one of six” and “two of six” are count-based decision boundaries. They should not be mistaken for confidence limits or for proof that a dose’s underlying toxicity probability lies on one side of a specified target.

Conditions for Applying the Rules Reliably

Before applying a count rule, check that the trial has the conditions needed for that count to mean what the protocol intends.

DLTs are defined in advance. The event definition, severity criteria, attribution approach, and observation window must be specified. Otherwise, participants may be classified inconsistently.
Outcomes are ready for decision. The relevant participants should have completed the DLT window, or have a protocol-defined classification that permits a decision. An outcome still pending is not automatically “no DLT.”
The denominator is evaluable participants. If someone cannot be classified under the protocol’s evaluability rule, apply the prespecified replacement or handling procedure. Do not silently count that person as a non-DLT or include them in the denominator as if evaluable.
Cohorts are assigned to one dose level. Count DLTs at the current dose and combine the first and expansion cohorts only when both were treated at that same level.
Escalation stays within the planned ladder. Move only to the next permitted dose, and obey any prespecified maximum dose, enrollment cap, or safety review.

Clinical safety oversight may require a hold or other action when an unexpected serious event occurs. Such action does not need to wait for a 3+3 count threshold. The algorithm is a decision aid within a protocol, not a replacement for safety monitoring, ethics review, or clinical judgment.

Diagnostics, Alternatives, and Limitations

Several operational signs indicate that a decision should be paused and checked. A dose is being escalated while participants’ DLT windows are incomplete; the reported denominator includes participants whose outcomes are not evaluable; the count combines participants treated at different doses; or the team cannot explain how a participant with missing follow-up was handled. Each problem can change the apparent count and therefore change the prescribed action. Resolve it using the protocol before applying the rule.

A further concern is that observed DLT counts may not rise steadily with dose. A lower dose can show more DLTs than a higher dose simply because cohorts are small and outcomes vary. The conventional algorithm uses the count at the current dose to direct its next step; it does not fit a model that smooths noisy observations across doses. Do not silently rearrange counts or skip levels to make the pattern look orderly.

The main alternative is a model-based dose-finding approach, which uses a statistical model and accumulated outcomes to guide dose assignment or selection. Such methods can use information across dose levels, but require model choices and their own operating procedures. The 3+3 is simpler to administer and explain, but its small cohorts and fixed rules provide limited information about the dose–toxicity relationship. It can also stop at a dose without directly estimating a target toxicity probability. Neither an escalation decision nor the absence of a stopping signal is a guarantee of safety.

Common mistake: “One of the first three participants had a DLT, so the next cohort should move up because only one person was affected.” The correction is to expand at the same dose to three additional evaluable participants. Escalation is considered only after the combined count among six is known, unless a separate protocol safety action requires a hold or stop.

A second common error is to describe stopping at a dose as automatic “de-escalation.” In the conventional algorithm, reaching the stopping threshold ends upward escalation. The algorithm does not automatically send a new cohort to the preceding dose; any such enrollment or formal dose designation must follow a prespecified rule.

Check Your Understanding

For each question, apply the conventional rules to evaluable participants and distinguish an algorithmic action from a conclusion about the true toxicity probability.

At a dose, the first three evaluable participants have no DLTs. What is the next dose action?
At a dose, one of the first three evaluable participants has a DLT. What must happen before the algorithm can decide whether to escalate?
After expansion, the combined result is two DLTs among six evaluable participants. What is the action?
Why should a participant with an incomplete DLT observation window not automatically be counted as having no DLT?
If the stopping threshold is reached at the lowest planned dose, what does the algorithm say about selecting a lower tested dose?

Solutions: (1) Escalate to the next planned dose, subject to the protocol’s dose limits and safety procedures. (2) Treat three additional evaluable participants at the same dose, then combine their outcomes with the first three. (3) Stop escalation; the total is at least two of six. (4) The DLT outcome has not yet been classified over the required window, so treating it as no DLT can undercount events and change the decision. Apply the protocol’s rule for pending or incomplete outcomes. (5) There is no lower tested dose in the ladder. Follow the prespecified stopping and dose-selection provisions; do not invent a lower level or continue upward.