Defining the Dose-Selection Estimand

Why the Estimand Must Come Before the Dose Comparison

A dose cannot be called adequate without specifying for whom, over what treatment period, and according to which outcome at what time. A simulation may produce many concentration profiles and responses, but those outputs do not by themselves determine what “adequate” means. The estimand is the precise quantity or criterion that the dose-selection exercise is intended to address.

This tutorial focuses on defining that quantity. It does not determine an exposure-response target or a decision threshold. Those are separate choices: the target must be justified for the clinical question, and a decision rule must state how much evidence or attainment is required before a dose is selected.

An estimand is not the same as a simulation result. The estimand describes the quantity of interest under the stated assumptions; a simulation estimate is a numerical approximation to that quantity, based on a fitted model, parameter values or draws, and a finite number of simulated individuals. Keeping these distinct helps identify whether a disagreement concerns the clinical question, the model, or Monte Carlo precision.

Specify the Population, Regimen, Endpoint, and Decision Time

Let \(d\) denote a candidate dose, and let \(X\) denote an individual’s measured covariates. Let \(b\) represent the individual random effects in the NLME model, and let \(\varepsilon\) represent any relevant residual or outcome variability. A model-generated outcome under dose \(d\), treatment duration \(\tau\), and assessment time \(t^*\) can be written generally as \(Y(d,\tau,t^*;X,b,\varepsilon)\). The notation is a reminder that the outcome depends on the dose and schedule as well as on the individual and the time definition.

The target population is described by a distribution \(G\) for the covariates and other between-individual characteristics to be represented. The estimand averages or otherwise summarizes outcomes over that specified population. Random effects should represent between-individual variability as prescribed by the model; they are not interchangeable with a sample of empirical Bayes estimates. If the intended population has a particular covariate distribution, the simulated population needs to represent that distribution rather than merely reproduce the study sample by default.

Treatment duration and decision time are related but not identical. Duration \(\tau\) states how long the regimen is given before the outcome is considered. Decision time \(t^*\) states when, or over what window, the endpoint is evaluated. For example, “response at the end of 14 days of treatment” specifies both a duration and an assessment time. A trough concentration immediately before a dose on day 14 is a different endpoint and time definition from an average exposure over the preceding dosing interval, even if both come from the same simulated concentration profile.

The endpoint must also be operational. For a continuous response, specify its definition, scale, and units. For an attainment-type goal, specify the event that counts as meeting the criterion, including its direction and timing. The criterion itself should be clinically justified; merely choosing a convenient model output does not make it a suitable endpoint.

Population-Average and Minimum-Subgroup Goals

For an attainment-type endpoint, define \(A(d)\) as an indicator that equals 1 when a simulated individual meets the specified criterion under dose \(d\), and 0 otherwise. A population-average attainment estimand is the expected value of this indicator over the target population and the outcome variability included in the model:

This is the proportion of the target population expected to meet the stated criterion under the specified regimen, duration, endpoint, and decision time. It is a population summary; it does not say that every subgroup has the same attainment.

If the decision is intended to protect several prespecified subgroups, let \(g=1,\ldots,K\) index mutually exclusive and exhaustive groups, and let \(w_g\) be the target-population proportion in group \(g\). Let \(\psi_g(d)\) be the attainment proportion within group \(g\). Then the overall average is the weighted average of the subgroup proportions, whereas a minimum-subgroup goal focuses on the lowest subgroup value:

These answer different questions. The average describes the expected attainment for a randomly selected member of the target population. The minimum describes the least-attaining prespecified subgroup. A minimum-subgroup rule is not the same as requiring every individual to attain the criterion: the subgroup proportion itself can be less than 1.

Why the Two Summaries Differ

The relation between the summaries follows directly from the weighted-average definition. Let \(m=\min_g\psi_g(d)\). Since each subgroup proportion is at least \(m\), and the weights are nonnegative and sum to one,

Thus the population average is never below the minimum subgroup attainment. It equals the minimum when all groups with positive population weight have the same attainment; otherwise it is larger. This simple result explains why a favorable population average cannot establish that every subgroup has a favorable result.

The subgroup definition is part of the estimand, not a cosmetic way of presenting results. Groups should be chosen for the decision question and specified before comparing candidate doses. Changing the group boundaries or adding many small groups can change the minimum and make it unstable. Conversely, using broad groups may conceal meaningful variation within them.

Worked Examples of Estimand Choices

Assumptions, Diagnostics, and Failure Modes

An estimand is meaningful only relative to its assumptions. The target population must be relevant to the decision, the regimen must be defined well enough to simulate, and the endpoint must be measurable from the model outputs. The fitted NLME model and the distributions used for covariates and random effects must be suitable for projecting the population and treatment conditions named in the estimand. These assumptions do not become true merely because a large number of simulations is run.

Before interpreting simulated summaries, check that subgroup weights are nonnegative and sum to one, that the subgroup definitions are exhaustive and nonoverlapping if a weighted decomposition is used, and that each subgroup has enough simulated individuals for a stable summary. Confirm that units, assessment windows, treatment duration, and endpoint direction are consistent in the simulation code and in the written estimand. If simulated individuals have different inclusion rules from the intended population, the results may describe a different population.

A useful diagnostic is to compare subgroup summaries alongside the overall summary. If the average is strong but a subgroup value is much lower, that is not a calculation error by itself; it may be the substantive feature the minimum-subgroup goal was designed to reveal. If results change substantially under plausible target-population weights, report that sensitivity rather than treating one weighting scheme as universal.

Parameter uncertainty is also distinct from population variability. A population estimand is defined under a model and its parameters, while uncertainty about those parameters concerns how precisely the estimand has been learned. A simulation that incorporates parameter uncertainty may summarize uncertainty in the estimated estimand, but it does not remove the need to specify the estimand first. Likewise, random-effect variability describes differences among individuals under the model; it should not be replaced by uncertainty about the estimated parameter values.

These approaches are alternatives only in the sense that they summarize the goal differently. Separate subgroup reporting can accompany either an average or a minimum-subgroup criterion. None is automatically best: the summary should match the population-level decision being made. A further limitation is that even a carefully defined estimand remains conditional on the model, endpoint definition, and target population specification.

Common Mistake

Wrong reasoning: “The overall simulated attainment is high, so the dose works for all relevant patients.” This conclusion confuses a weighted average with a guarantee for each subgroup or individual. A high average can coexist with a materially lower subgroup proportion, especially when that subgroup is a small fraction of the target population.

Correction: State whether the goal is population-average attainment or a subgroup rule, then report the summaries that correspond to that goal. If subgroup protection matters, define the groups and minimum-subgroup criterion explicitly. Do not infer individual certainty from either population summary.