Within-Patient Dose Escalation in Oncology Trials

Why a Dose Increase Within One Patient Needs Careful Interpretation

In an oncology escalation trial, a patient may begin at a low dose and move to a higher dose after tolerating one or more treatment intervals. This can provide information about how toxicity changes as that patient’s dose changes. But the observation at the higher dose does not stand alone: it follows earlier treatment, occurs later in time, and is available only because the patient remained eligible to escalate. The higher-dose observations therefore combine current dose, prior exposure, elapsed time, and selection of patients who tolerated earlier treatment.

The key question is not simply whether toxicity was more common at a higher recorded dose. It is: which risk is being described, for which patients, over what interval, and under what treatment history? Answering it requires distinguishing a conditional risk at the next dose from the probability of toxicity over an entire escalating course.

Representing Dose and Treatment History

Let \(i\) index patients and \(t\) index treatment intervals. Let \(d_{it}\) be the dose in interval \(t\), and let \(C_{it}\) summarize cumulative exposure before that interval. The exact definition of \(C_{it}\) must be stated: it might be cumulative administered dose, a time-weighted exposure measure, or a pharmacokinetic exposure summary. Those quantities are not interchangeable. Let \(Y_{it}=1\) indicate a toxicity event in the interval and \(Y_{it}=0\) indicate no event. A risk for \(Y_{it}\) is defined only among patients who are event-free and at risk at the start of the interval.

One illustrative model for an interval-specific risk is

Here \(H_{it}\) denotes the recorded treatment and event history available before interval \(t\); \(d_{\mathrm{ref}}\) is a reference dose; \(X_i\) contains measured patient characteristics; and \(b_i\) represents patient-specific susceptibility in a model with a patient random intercept. The logistic function maps the linear predictor \(\eta\) to risk \(p=1/(1+\exp(-\eta))\). This equation is an example of how to separate terms, not a universal toxicity model.

The model’s assumptions must match the intended interpretation. It assumes that the defined interval outcome and dose history are measured consistently, that the chosen functional forms adequately describe the conditional risks, and that patients with relevant histories provide information for the comparisons of interest. A causal interpretation of changing dose requires more: after conditioning on the recorded history, there must be no unmeasured factor that both influences dose escalation and predicts toxicity; the compared dose strategies must be sufficiently represented; and the hypothetical dose strategy must be well-defined. Without those conditions, a fitted dose coefficient is an adjusted association, not necessarily the effect of intervening to change dose.

Dose and history can be strongly related by design. A patient at a high dose has generally received previous doses and spent more time on treatment than a patient observed at a low dose. If dose, cumulative exposure, and time have little independent variation, their separate coefficients may be imprecise or difficult to interpret. Within-patient observations do not automatically remove this problem: a patient’s susceptibility may be stable, but time-varying health, disease status, supportive care, and accumulated exposure can still differ across intervals.

From Interval Risk to Risk Over an Escalating Course

A sequence of interval risks can be combined to obtain a cumulative probability, but only when the conditioning is made explicit. Suppose a patient follows a specified dose strategy and is followed completely through \(K\) intervals. Define \(q_j\) as the probability of toxicity in interval \(j\), conditional on no earlier toxicity and on the treatment and risk history specified by that strategy up to interval \(j\). By the chain rule, the probability of no toxicity through interval \(K\) is the product of the conditional probabilities of avoiding toxicity in each interval.

This is an accounting identity for conditional risks along a defined event-free pathway, not an assumption that interval outcomes are independent. The relevant \(q_j\) may change with dose, time, cumulative exposure, and patient history. If follow-up ends before the interval outcome is known, or observation ends for reasons related to toxicity risk, the observed risks among those still observed cannot automatically be multiplied to recover the original population’s cumulative risk. Censoring and selection need appropriate handling.

Escalation Selects the Patients Seen at Higher Doses

In many designs, reaching a higher dose depends on tolerating earlier doses. Consequently, the patients observed at a high dose are selected from the patients who began treatment. If patients differ in underlying susceptibility, those who remain eligible may have lower average susceptibility than the starting population. A lower-than-expected toxicity rate among escalated patients therefore need not show that the higher dose is intrinsically safe. Conversely, a high rate could reflect a susceptible group, prior exposure, or time on treatment as well as the current dose.

Interpretation, Diagnostics, and Alternatives

A within-patient dose increase is informative, but it does not by itself make the comparison causal or remove confounding. Before interpreting an estimated dose-response pattern, specify the target: an interval risk among patients who are currently event-free and eligible, a risk under a particular escalation strategy, or a risk for the population that started treatment. These are different quantities. A conditional comparison among escalators generally does not answer what would have happened if every patient in the original population had received the higher dose.

Useful diagnostics focus on the data-generating process as well as the fitted model. Check whether dose increases follow the stated eligibility rules, whether toxicity timing and cumulative exposure are correctly recorded, and whether patients have meaningful variation in dose beyond a common cycle trend. Compare the characteristics and prior histories of patients who do and do not reach each dose. Examine whether predicted interval risks align with observed risks in relevant history groups, while recognizing that small risk sets make such checks unstable. Large changes in estimates after adding time or exposure terms, wide uncertainty, or little overlap in histories across doses are warnings that separate dose effects are weakly identified.

Alternative designs answer different questions. Conventional between-patient escalation compares groups assigned to different doses, but group differences and small cohorts can limit precision. Within-patient escalation provides repeated observations and can control implicitly for some stable patient characteristics, but dose is entangled with treatment history and eligibility selection. A randomized dose comparison can strengthen causal interpretation when ethically and clinically feasible, but may not be appropriate when toxicity is delayed, persistent, or irreversible; carryover and treatment changes also complicate crossover comparisons. No design eliminates the need to define the target population, outcome window, and treatment history.

The limitations are especially important when toxicity is cumulative or delayed. A single cumulative-dose summary may fail to represent timing, interruptions, recovery, or peak exposure. A model can also extrapolate to dose-history combinations rarely observed in the trial. Inferences should therefore be limited to histories with adequate support, and conclusions about a particular dose should state the population and treatment strategy to which they apply. Detailed sensitivity analysis for alternative toxicity-model assumptions is a separate problem; the first requirement here is to avoid confusing a selected, history-dependent risk with a dose-only risk.