Defining Joint Efficacy And Toxicity Outcomes

Why Outcome Definitions Come Before Analysis

A dose-finding trial may ask whether a patient benefits, experiences toxicity, or does both. Those descriptions are not yet outcomes that can be recorded consistently. “Benefit” might mean tumor response by a fixed visit, while “toxicity” might mean any treatment-related adverse event, a dose-limiting event, or the most severe event in a specified period. Different choices produce different patient records, even when the observed clinical course is identical.

The first task is therefore to specify what counts as efficacy and toxicity for each patient, when it is assessed, and how repeated or simultaneous events are handled. This tutorial focuses on those definitions and the resulting records. It does not prescribe a dose-selection rule or a model for the resulting outcomes.

Definition: A patient-level outcome is a prespecified summary of one patient’s observed clinical course for a defined endpoint and assessment period. A joint patient-level outcome records the efficacy and toxicity outcomes for the same patient, using rules that make both components interpretable. The definition must specify the outcome criterion, the time origin, the assessment window, and the handling of multiple events and incomplete observation.

Let \(Y_i\) denote the efficacy outcome and \(Z_i\) the toxicity outcome for patient \(i\). These labels do not imply a particular measurement scale: each may be binary, ordinal, or time-to-event. The paired record \((Y_i,Z_i)\) keeps the two outcomes linked to the same patient. It does not, by itself, assume that efficacy and toxicity are independent or specify how either varies with dose.

Binary Outcomes and Their Windows

A binary outcome has two defined states. For example, efficacy may be coded as response versus no response by a prespecified day, and toxicity as at least one qualifying adverse event versus none by a prespecified day. The criterion must say what qualifies: a response threshold, an adverse-event grade, attribution requirements if relevant, and any confirmation rule.

Definition: For a specified assessment window, a binary efficacy outcome \(Y_i\) can be coded \(1\) if patient \(i\) meets the prespecified efficacy criterion within that window and \(0\) otherwise. A binary toxicity outcome \(Z_i\) can be coded \(1\) if patient \(i\) meets the prespecified toxicity criterion within its window and \(0\) otherwise. A zero means the criterion was not met under adequate observation; it must not be used to disguise missing or incomplete assessment.

An assessment window needs an anchor and an endpoint. For example, “by day 28 after first dose” defines both the time origin and the deadline. If efficacy is assessed at day 56 but toxicity is assessed through day 28, those are different windows and should be stated separately. The definition should also state whether an event exactly on the deadline counts; a common convention is to include events occurring on or before the deadline.

The binary coding rule reduces a clinical course to one bit of information. It is useful when the clinical question is naturally threshold-based, but it can hide differences in severity, timing, recurrence, and duration. For instance, a mild event and a life-threatening event both become \(Z_i=1\) under an “any event” rule. That loss of detail is a consequence of the definition, not a defect that later computation can repair.

Worked Example: Binary Outcomes in Separate Windows

Hypothetical trial. Suppose efficacy is defined as a confirmed response on or before day 56 after first dose. Toxicity is defined as at least one treatment-emergent grade 3 or higher adverse event on or before day 28. Consider four patients with complete observation through both deadlines:

Patient	Confirmed response by day 56?	Grade 3+ event by day 28?
A	Yes	No
B	No	Yes
C	Yes	Yes
D	No	No

Apply the stated criteria directly: A has \((Y_i,Z_i)=(1,0)\), B has \((0,1)\), C has \((1,1)\), and D has \((0,0)\). Thus the four records are \((1,0),(0,1),(1,1),(0,0)\). The windows differ intentionally: a response first confirmed on day 40 can qualify, while a grade 3 event first occurring on day 40 does not qualify under this toxicity definition.

The coding is justified only because all four patients were observed through the relevant windows, and because the response confirmation and event-grade criteria have been specified. If patient D had left follow-up on day 10, “no event recorded” would not establish \(Z_i=0\) through day 28.

Ordinal Outcomes Preserve Ordered Categories

An ordinal outcome has categories with a meaningful order but does not require equal spacing between adjacent categories. Efficacy might be classified as progressive disease, stable disease, partial response, or complete response. Toxicity might be recorded using ordered severity grades. The order must be clinically interpretable, and the coding rule must identify which observation determines the patient’s category.

Definition: An ordinal patient-level outcome assigns each patient to one category in a prespecified ordered set. The ordering is meaningful, but the difference between successive categories is not assumed to be numerically equal. If repeated assessments or events occur, the summary rule—such as status at a fixed visit, best response, or maximum toxicity grade—must be specified in advance.

A common toxicity summary is the maximum grade observed during a defined period. If a patient has events of grades 1, 2, and 3 during the window, the maximum-grade summary is grade 3. This rule is straightforward, but it discards the number, duration, and timing of events. A fixed-visit efficacy category instead describes status at that visit and may differ from a “best response during follow-up” category. Neither definition is universally correct; each answers a different clinical question.

Worked Example: Maximum Grade and Fixed-Visit Response

Hypothetical trial. Efficacy is the disease-status category at day 42, ordered as progressive disease < stable disease < partial response < complete response. Toxicity is the maximum adverse-event grade through day 28, with grades 0, 1, 2, 3, and 4 ordered from none to most severe. A patient has toxicity grades 1, 3, and 2 on days 4, 12, and 20, respectively, and has stable disease at day 42.

For efficacy, the day-42 category is stable disease; an earlier partial response, if any, would not replace the fixed-visit status under this rule. For toxicity, the observed grades are \(1,3,2\), so the maximum is \(\max(1,3,2)=3\). The patient’s joint ordinal record is therefore (stable disease, grade 3).

The definition assumes that the day-42 disease assessment is available and that adverse events through day 28 have been captured. It does not imply that the distance from grade 1 to grade 2 is the same as the distance from grade 3 to grade 4, nor does the maximum-grade summary describe the duration or recurrence of toxicity.

Time-to-Event Outcomes and Incomplete Follow-Up

A time-to-event outcome records how long it takes from a defined origin to a specified event. The event must be defined precisely—for example, time from first dose to first qualifying toxicity, or time from randomization to first documented response. The event date and time origin should be specified consistently, including how same-day events are handled.

Definition: Let \(U_i\) be the time from the prespecified origin to patient \(i\)’s event of interest, and let \(C_i\) be the time from that origin to the end of ascertainable follow-up. With an administrative window of length \(W\), the recorded time is \(\tilde{U}_i=\min(U_i,C_i,W)\). The event indicator is \(\delta_i=1\) if the event is observed by both the end of follow-up and the end of the window, and \(\delta_i=0\) otherwise. A record with \(\delta_i=0\) is censored at \(\tilde{U}_i\); it does not assert that the event could never occur.

This representation assumes that event times and follow-up limits are measured from the same origin, and that event status up to the recorded time is known. In a fixed-window binary analysis, a patient known to be event-free through the entire window can be coded as no event. A patient whose follow-up ends early cannot be coded as event-free through the window merely because no event was observed before follow-up ended.

Death before an efficacy event illustrates why endpoint rules matter. Depending on the clinical question, death might be defined as an efficacy failure in a composite endpoint, retained as a distinct event type, or handled under another prespecified rule. Treating it automatically as ordinary censoring would claim that the patient’s later efficacy-event time is simply unknown, which may not represent the intended clinical outcome. The rule should be decided before outcomes are summarized.

Worked Example: Recording Event Times and Censoring

Hypothetical trial. The time origin is first dose. The toxicity endpoint is time to first grade 3 or higher event, with a 28-day window. The efficacy endpoint is time to first confirmed response, with a 56-day window. Follow-up ends at day 18 for patient A, day 28 for patient B, and day 56 for patient C. A has no qualifying toxicity or response before follow-up ends; B has a qualifying toxicity on day 11 and no response by day 28; C has a confirmed response on day 35 and no qualifying toxicity through day 28.

For toxicity, A is recorded at \((\tilde{U}_i,\delta_i)=(18,0)\): follow-up ended before the toxicity window ended, so A is censored at day 18. B is recorded at \((11,1)\), since the qualifying event occurred within the window. C has no qualifying toxicity through day 28, so the toxicity record is \((28,0)\), administratively censored at the window limit.

For efficacy, A is censored at day 18, giving \((18,0)\). B is observed without response through day 28 but has not completed the 56-day window, so the record is \((28,0)\), not a confirmed “no response by day 56.” C’s first confirmed response is observed on day 35, giving \((35,1)\). The same patient can therefore have different observation times for efficacy and toxicity because the endpoints have different windows.

Combining Multiple Events Into One Patient Record

A patient may have several adverse events, repeated disease assessments, or both efficacy and toxicity events. To produce a single endpoint, state which event or assessment controls the summary. Common choices include first qualifying event, maximum severity, occurrence of any qualifying event, event count, or category at a fixed visit. These summaries are not interchangeable: a maximum-grade rule emphasizes worst severity, while a count emphasizes recurrence.

When both outcomes are binary, the joint record has four possible categories: efficacy without toxicity, toxicity without efficacy, both, or neither. Define the categories explicitly:

\begin{aligned} C_{11} &= \{Y_i=1,\ Z_i=1\},\\ C_{10} &= \{Y_i=1,\ Z_i=0\},\\ C_{01} &= \{Y_i=0,\ Z_i=1\},\\ C_{00} &= \{Y_i=0,\ Z_i=0\}. \end{aligned}

These categories are mutually exclusive because a binary variable cannot equal both 0 and 1 at once. They are exhaustive because each of \(Y_i\) and \(Z_i\) must be either 0 or 1 when both outcomes are fully ascertained. Thus each patient contributes to exactly one category. In a sample of \(n\) fully classified patients, the four counts must sum to \(n\). This is a bookkeeping identity, not an assumption about how efficacy and toxicity are related.

For ordinal outcomes, a joint record is an ordered pair of categories, such as (partial response, grade 2). For time-to-event outcomes, the record includes the event time and event indicator for each endpoint. These forms preserve different features of the clinical course; collapsing them to binary categories should be done only with an explicit threshold and window.

Worked Example: Checking Joint Binary Categories

Hypothetical trial. Among 12 patients with complete observation under specified binary efficacy and toxicity definitions, 5 meet the efficacy criterion without toxicity, 2 meet the toxicity criterion without efficacy, 3 meet both, and 2 meet neither.

The four counts are \(n_{10}=5\), \(n_{01}=2\), \(n_{11}=3\), and \(n_{00}=2\). Check the total:

n_{10}+n_{01}+n_{11}+n_{00}=5+2+3+2=12.

The number meeting efficacy is \(5+3=8\); the number meeting toxicity is \(2+3=5\). The “both” group is counted in each marginal total because those totals answer separate yes-or-no questions. It would be incorrect to add 8 and 5 and interpret 13 as the number of distinct patients with at least one outcome: the 3 patients meeting both criteria have been counted twice. The number meeting at least one criterion is \(5+2+3=10\).

Diagnostics, Alternatives, and Limitations

Before analyzing or reporting outcomes, check that the definitions can be applied consistently to the clinical records. A practical review should ask whether:

each endpoint has an explicit criterion, time origin, and window, including a rule for events on the deadline;
each patient’s follow-up is sufficient for the assigned binary status, or incomplete follow-up is marked as missing or censored rather than as “no event”;
repeated events and assessments have a prespecified summary rule, applied the same way to every patient;
death, treatment discontinuation, and other events that prevent later assessment have an explicit handling rule; and
the joint categories are possible, nonoverlapping, and exhaustive for patients with complete outcome information.

Investigators often choose among binary, ordinal, and time-to-event definitions. A binary endpoint is simple to communicate but discards severity and timing. An ordinal endpoint preserves ordering and severity categories but does not by itself express the time at which a category arose. A time-to-event endpoint preserves timing and can represent incomplete observation through censoring, but requires a clear event definition and adequate follow-up information. These are alternative summaries of clinical experience, not merely different calculations on an identical outcome.

No single patient-level summary captures every clinically relevant feature. “Any toxicity” can treat mild and severe events alike; maximum grade ignores recurrence and duration; first-event time ignores later events; and a fixed-visit response can miss an earlier response that was not present at that visit. A definition should therefore match the clinical question and the available assessment schedule. It should also be stated clearly enough that two reviewers would assign the same patient record from the same source data.

Common Mistake

Wrong reasoning: “The patient had no recorded toxicity, so code toxicity as 0,” even though the patient stopped attending on day 9 and the toxicity window runs through day 28. Correction: absence of a recorded event is not evidence of absence throughout a period that was not observed. Record incomplete follow-up, use a time-to-event representation with censoring when appropriate, or apply a prespecified missing-data rule. Do not silently convert incomplete observation into a negative outcome.

A related mistake is to use “maximum grade” and “any grade 3+ event” as if they were the same definition. They coincide for the binary question of whether any grade 3+ event occurred, but the maximum-grade outcome retains additional ordered information. State which form is intended and retain enough source information to verify the coding.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

A toxicity window is 28 days after first dose. A patient has complete follow-up and no qualifying event through day 28. What is the binary toxicity code, and what would change if follow-up ended on day 12?
A patient’s toxicity grades during a window are 1, 2, and 4. Under a maximum-grade rule, what is the ordinal toxicity outcome? Name one feature of the clinical course it does not retain.
A response is first confirmed on day 40. Efficacy is defined as confirmed response by day 56, while toxicity is defined as a qualifying event by day 28. Does the response meet the efficacy criterion? Does its timing affect the toxicity code?
In a complete binary joint classification of 20 patients, the counts are 6 efficacy only, 4 toxicity only, and 5 both. How many meet neither criterion?
Why is a patient who is event-free at day 15 but lost to follow-up before a day-56 efficacy window not automatically a binary “no response by day 56”?

Solutions. (1) The code is 0 because the patient was observed through the full 28-day window without a qualifying event. If observation ended on day 12, the patient’s status through day 28 would be unknown; do not code 0 solely from the available record. (2) The outcome is grade 4, since \(\max(1,2,4)=4\). The rule does not retain the number, timing, or duration of events. (3) Yes, the confirmed response by day 56 meets the efficacy criterion. Its day-40 timing does not itself determine toxicity, which uses a separate window and criterion. (4) The remaining count is \(20-(6+4+5)=5\) patients meeting neither criterion. (5) The patient was not observed for the entire window, so later response status is unknown. Lack of an observed response by day 15 does not establish lack of response by day 56.