Defining Dose-Finding Estimands

Why Define the Estimand Before Comparing Doses?

A dose-finding study can produce many numbers: a mean response at each dose, a difference from a reference dose, a response rate, or a prediction for a patient with specified characteristics. Those numbers answer different questions. Before analyzing data or ranking dose options, specify which population quantity is the target. Otherwise, a calculation may be precise yet answer a question that was never intended.

An estimand is the precise quantity a study seeks to learn. In dose finding, it identifies who the conclusion concerns, which treatment regimens are compared, what outcome is measured, and when that outcome is assessed. A separate decision rule may then use the estimand, along with other information and preferences, to rank doses. The estimand describes the treatment effect or outcome of interest; it does not by itself say which tradeoffs are acceptable.

The Components of a Dose-Finding Estimand

Target population. Name the people to whom the result is intended to apply, using clinically meaningful eligibility and baseline characteristics. “Patients in the study” is not always sufficient: a study sample may differ from the intended population, and a trial that enrolls only people with a particular disease severity cannot automatically answer a question about everyone with that disease. The target population also determines how patient characteristics are represented in a population average.

Treatment conditions and contrast. A dose is part of a regimen, not just a number. Specify the dose, route, schedule, duration, and any relevant background treatment. State the comparator and the contrast, such as a difference in mean outcome or a ratio of means. The scale matters: a difference is expressed in outcome units, whereas a ratio is unitless. For a dose-specific mean, the reference dose may be placebo, standard of care, or another active dose.

Outcome and time horizon. Define what is measured and when. “Efficacy” is too broad to be an outcome definition. A specific outcome might be change from baseline in a symptom score at week 8, or the number of exacerbations accumulated over 12 weeks. For a repeated measure, specify which visit or summary over time is intended. A measurement at week 8 and an average over weeks 4 through 8 are different outcomes, even if both use the same scale.

Intercurrent events. Treatment discontinuation, rescue medication, switching therapy, or death after treatment begins can affect how an outcome is interpreted. State how such events are handled. For example, a treatment-policy strategy can define the outcome as measured regardless of discontinuation, while a hypothetical strategy asks what the outcome would have been under a specified scenario in which the event did not occur. These target different quantities; neither is automatically preferable in every setting. The event and strategy need to be clear enough that the outcome is well-defined.

For clarity, let \(Y^d(\tau)\) denote the outcome that a member of the target population would have at horizon \(\tau\) under regimen \(d\), with the stated strategy for intercurrent events. The dose-specific population mean is \(\mu_d(\tau)=E[Y^d(\tau)]\). One possible contrast against reference regimen \(d_0\) is \(\Delta(d,d_0;\tau)=E[Y^d(\tau)]-E[Y^{d_0}(\tau)]\). These definitions are not a commitment to a particular analysis model: they define the target quantity first.

A Short Justification: Population Composition Is Part of the Target

Suppose \(X\) denotes a baseline characteristic, such as renal-function group. The population mean under dose \(d\) averages the conditional mean outcome at each value of \(X\) over the distribution of \(X\) in the target population. For a difference in means, the same averaging applies to the conditional differences:

The second equality follows from linearity of expectation and averaging over the target distribution \(F_{\text{target}}\). Thus, even if conditional dose contrasts are the same in two studies, their unstandardized population averages can differ when the studies represent different populations. This is why “the population” is not background wording: it is part of the mathematical definition.

Worked Examples: Population and Event Strategy

Estimands, Estimates, Models, and Utilities

These terms refer to different parts of a decision process. The estimand is the target population quantity, defined before looking at the analysis result. An estimate is a value calculated from observed data to approximate that target, and ordinarily has sampling uncertainty. A model is a representation used to describe data or predict outcomes; its predictions need not automatically equal the estimand unless the model, data, and target population align. A utility expresses how decision makers value outcomes or options for a decision.

A utility may combine outcomes or account for burdens and preferences, but it does not retroactively define what a study measured. Two decision makers can agree on the estimand and estimated dose effects while preferring different doses because they assign different value to benefit, risk, burden, or uncertainty. Conversely, a stated utility cannot repair an ambiguous population, outcome, or time horizon. The dose-ranking step uses a utility or other decision criterion; specifying the estimand is a prior task.

Assumptions, Diagnostics, and Limits

An estimand is defined independently of whether the available data identify it. To estimate a causal contrast from a randomized comparison, the assigned regimens must correspond to the specified treatment conditions, outcomes must be measured and interpreted consistently, and missing outcomes or intercurrent events must be handled in a way compatible with the chosen strategy. In a nonrandomized comparison, causal interpretation also requires adequate control of confounding. For standardization across baseline groups, the data must support the relevant groups and the conditional contrasts must be applicable to the target population.

For potential-outcome interpretations, consistency means that a person's observed outcome under the regimen actually received corresponds to the potential outcome defined for that same regimen. This requires treatment conditions to be specified well enough that materially different versions are not silently combined. No interference means one person's regimen does not alter another person's outcome. In nonrandomized settings, exchangeability requires no uncontrolled confounding of treatment and potential outcomes, conditional on the variables used, and positivity requires a nonzero probability of receiving each compared regimen within relevant covariate patterns. These are identification conditions for causal estimation, not additional words that make an ill-defined estimand precise.

Useful checks begin with the specification: can a clinician identify who is in the target population, what exactly each regimen entails, which outcome is recorded, and the assessment time? Then check whether data collection matches that definition. Look for missing outcome measurements, differential follow-up, treatment versions that were pooled despite different schedules, and insufficient representation of target-population subgroups. A large sample cannot fix a mismatch between the study's measured outcome and the outcome named in the estimand.

A common pitfall is to report “the best dose” without saying best for whom, on which outcome, and over what period. Another is to treat an analysis model's predicted mean as the estimand itself. The correction is to write the target quantity in words and, where useful, as a population expectation and contrast before discussing how it will be estimated. Specify the utility or decision rule separately.

No single estimand answers every clinical question. A mean can conceal variation across patients; one time horizon does not describe the entire trajectory; and a treatment-policy strategy may answer a different practical question than a hypothetical strategy. These are reasons to define the target carefully, not to assume that one summary is universally sufficient. The estimand gives a disciplined basis for interpreting estimates and for subsequent dose decisions.