Dose Selection and the Treatment Condition

Why Dose Selection Changes the Treatment Condition

A trial may randomize participants to several doses, examine interim data, and use a prespecified rule to choose one dose for continued study. The dose administered after that decision is not necessarily a fixed treatment condition: it can depend on the random interim results. The scientific question must therefore distinguish two targets: the effect of a particular dose, and the effect of following a strategy that selects a dose.

This distinction matters even if the same dose is ultimately selected in a particular trial. A fixed-dose effect refers to a specified dose in the target population. A strategy effect refers to a rule, including how interim information determines which dose is administered. They answer different questions and need not have the same value.

A dose number alone may not define a treatment condition completely. For example, “20 mg” is ambiguous if formulation, frequency, titration, or treatment duration differs between trial stages. The treatment versions and stage-to-stage consistency matter: if the intended condition is “20 mg once daily for 12 weeks,” a later stage using a different schedule is not automatically the same condition.

The outcome and its handling must also be held clear. Let \(Y(d)\) denote the outcome that a member of the target population would have under the specified fixed-dose condition \(d\), and let \(Y(0)\) denote the outcome under the specified comparator condition. The outcome should be defined at a stated time and with an intercurrent-event strategy appropriate to the question. Earlier material on estimands and intercurrent events provides the vocabulary for specifying those features; they cannot be left implicit when comparing doses.

Fixed-Dose Effects and Strategy Effects

The fixed-dose estimand holds \(d\) constant. The strategy estimand does not: the dose is determined by the rule \(g\). As a result, a strategy effect depends on the selection rule, the interim data available to it, the timing of selection, and the treatment conditions before and after selection. Changing any of these may change the strategy even when the list of candidate doses stays the same.

A useful way to define the strategy effect is to consider a future target patient whose outcome is measured after the strategy has selected a dose using interim data. If those interim data \(S\) are independent of the future patient’s potential outcomes, then the strategy’s mean outcome is \(E[Y(g(S))]\). By conditioning on the selected dose, this becomes a mixture of fixed-dose means:

Here \(\mathcal{D}\) is the set of candidate doses. This identity follows by partitioning outcomes according to which dose the rule selects. Independence is important: it allows the future patient’s mean potential outcome under dose \(d\) to remain \(E[Y(d)]\) within each selection event. If the same participants both contribute to dose selection and supply the outcomes used to estimate the effect, this simple mixture need not describe the estimand or the estimator. The joint relationship between selection and outcomes must then be considered.

What Is Included in a Dose-Selection Strategy?

A strategy should be specific enough that two teams applying it to the same interim information would make the same treatment decision. Its definition ordinarily includes:

• the candidate doses and comparator;
• the interim data and analysis time used for selection;
• the rule mapping those data to a selected dose, including how ties or inconclusive results are handled;
• which participants receive each dose before selection and which treatment condition applies afterwards; and
• the target population, outcome, assessment time, and intercurrent-event strategy used to define the effect.

Not every strategy has to use a complicated algorithm. A rule such as “select the dose with the largest estimated mean improvement among 10, 20, and 40 mg” is a strategy, provided the analysis population and tie rule are specified. A rule such as “select the best dose” is not yet operational: “best” needs a defined criterion.

Conditional Effects Are Not Automatically Strategy Effects

A trial may select 20 mg in one realization. Reporting the effect of 20 mg after that decision can be useful, but the interpretation depends on the target. It may be a fixed-dose effect in a future cohort, an effect conditional on the event that 20 mg was selected, or an estimate using participants whose data also drove selection. These are not interchangeable.

In particular, selection tends to favor doses with unusually favorable interim estimates. Random error can contribute to that favorable ranking, so the selected estimate can overstate the selected dose’s true mean effect. This is a selection effect, not evidence that the treatment caused a larger effect. If a later cohort is independent of the selection data and is randomized to the selected fixed dose and comparator, its contrast addresses that selected dose in that cohort; it still does not by itself describe what would happen on average across repetitions of the dose-selection strategy.

Assumptions, Diagnostics, and Failure Modes

The potential-outcome definitions require well-defined treatment conditions: a participant receiving dose \(d\) must receive the version specified by that condition. For a randomized comparison to identify a fixed-dose contrast, assignment must be randomized as intended, outcome measurement and follow-up must correspond to the specified estimand, and the analysis must address missing outcomes and intercurrent events according to the chosen strategy. The treatment conditions must also be available to the relevant participants under the design.

For a strategy effect, the rule must be fully defined and the target must specify how randomness in interim data is handled. If the target is an average across repeated applications, the selection probabilities are part of its meaning. If the target is conditional on selection of a particular dose, say so explicitly. When selection and outcome data overlap, avoid assuming independence or using the probability-weighted formula without justification.

Practical checks include comparing protocol versions and case-report forms across stages, verifying that dose, schedule, formulation, and outcome timing did not drift, and confirming that the implemented decision rule matches the prespecified one. Review how many candidate doses were considered and whether the reported estimate came from the same data used to select the dose. A large difference between interim and later estimates can be a warning sign of selection noise or changing treatment conditions; it is not, by itself, proof of either explanation.

The main alternatives answer different questions:

A strategy effect may be more relevant when evaluating the whole adaptive policy, while a fixed-dose effect may be more relevant when deciding whether a specific regimen benefits the target population. Neither is universally preferable. The limitation is that a strategy effect can be harder to interpret and estimate if the rule is underspecified, if the design rarely assigns some doses, or if treatment conditions vary across stages. A fixed-dose effect, in turn, cannot summarize the consequences of a policy that may choose different doses.

Common Mistake: Naming the Realized Dose as the Strategy

Incorrect reasoning: “The interim analysis selected 20 mg, so the trial’s treatment condition was fixed 20 mg, and the reported estimate is the effect of the selection strategy.” This treats one realized decision as if it were the rule that generated it. It also ignores any different dose the same rule might have selected under another interim sample.

Correction: Report the fixed-dose contrast as an effect of the specified 20 mg condition when that is the target. Report a strategy effect only when the treatment policy—including candidate doses, interim information, decision rule, and target averaging or conditioning—is defined. Keep the two interpretations separate, even when they coincide numerically in one trial.