Sequential Design as Conditional Optimization

Why Design Becomes Conditional

A design chosen before any observations must anticipate uncertainty about both the parameters and the data that will be collected. At an interim decision, some observations have already accrued and some sampling opportunities remain. The design question is then conditional: given what is known now, which feasible future observations best serve the stated objective?

This is not the same as simply repeating the original optimization with a smaller sample size. The accrued data, the remaining resources, and the current distribution of uncertain parameters all help define the decision. A sequential design is therefore most naturally described as a rule that maps the information available at each interim point to a future design choice. The rule must be specified so that it uses only information available when the choice is made.

State, Feasibility, and Conditional Utility

Write \(D_t\) for the design and data accrued by interim time \(t\), \(q_t(\theta)\) for the current distribution over uncertain model parameters \(\theta\), and \(r_t\) for the remaining sampling opportunities or resources. The interim state is \(S_t=(D_t,q_t,r_t)\). The future design \(d\) might specify sampling times, the number of observations at each time, or another permitted sampling decision. Its feasible set, \(\mathcal{F}(S_t)\), contains exactly the choices allowed by the remaining resources and operational constraints.

A design criterion must make the candidates comparable. Let \(U(S_t,d,\theta,Y)\) denote the utility assigned to choosing \(d\), where \(Y\) represents any future observations relevant to that utility. The conditional expected utility of \(d\) is the average of this utility over the current uncertainty in \(\theta\) and the future data distribution implied by the model. A one-step conditional optimization chooses

The maximizer is a feasible choice with the largest conditional expected utility; if several choices tie, each is optimal under this criterion. A loss criterion can be handled equivalently by minimizing conditional expected loss. The utility, its units, and the way future outcomes enter it must be set before comparing candidates. For example, one might score an expected reduction in parameter uncertainty, but the score must be defined consistently across possible sampling choices.

The current distribution \(q_t(\theta)\) is an input to this decision, not a synonym for a single best-fitting parameter value. In a Bayesian analysis it could be a posterior distribution; in another framework it could represent a different justified distribution over uncertain parameters. The details of how that distribution is obtained are a separate question. Here, the essential point is that the redesign is conditional on the distribution currently available, rather than on an assumed point value alone.

The conditional expectation also accounts for future data. If the model gives a sampling distribution \(f(y\mid\theta,d)\), then the predictive distribution of a future observation is

Thus, the expected utility of a candidate averages over both parameter uncertainty and the variation in observations expected under that uncertainty. When utility depends only on \(\theta\) and the selected design, the expectation over \(Y\) is unnecessary. When utility depends on what the future data will be, omitting that expectation can change the ranking of designs.

Why the Optimization Is Conditional

The basic justification follows directly from the definition of expected utility. For each feasible choice \(d\), calculate its expected utility under the current state. Since the decision is made at state \(S_t\), only the distribution conditional on \(S_t\) is relevant. Choosing a design with a lower conditional expected utility cannot improve the stated objective at this decision point.

With multiple remaining opportunities, a policy can be better than a single fixed sequence because a later choice can depend on data that arrive in the meantime. To write this precisely, let \(V_t(S_t)\) be the largest expected terminal utility achievable from state \(S_t\) onward. If the next choice is \(d\), and a future observation \(Y\) leads to state \(S_{t+1}\), the value of that choice is its expected immediate utility plus the expected value of the resulting state. In a general additive-utility formulation,

At the final opportunity, \(V_{t+1}\) is the terminal utility. This recursion is a justification for optimizing over future consequences, not merely the immediate score. It assumes that the state contains the information needed to evaluate future choices, that future outcomes follow the specified model, and that the utility and resource constraints are consistently defined. If the objective is not additive over stages, the terminal utility can instead be evaluated directly over complete future policies.

Worked Examples

Assumptions, Diagnostics, and Alternatives

A conditional optimization is only as credible as its inputs and constraints. Before interpreting an optimizer, check that:

• The state is adequate. Accrued observations, current parameter uncertainty, and remaining opportunities are represented without omitting information that could change future feasibility or utility.
• The predictive model is appropriate for the decision. Candidate scores that depend on future data require a defensible sampling model and parameter distribution. The same model and uncertainty representation should be used when comparing candidates.
• The feasible set is operationally accurate. Sample counts, allowable times, assay capacity, participant burden, and other applicable constraints must be included. An unconstrained optimizer may propose a plan that cannot be implemented.
• The criterion is aligned with the objective. A score in information units, expected loss, or utility points answers only the question encoded by that score. It need not optimize a different scientific or clinical objective.
• The policy obeys the decision timeline. A choice at an interim may depend on accrued data, but not on future observations that have not yet occurred.

Useful diagnostics include checking that all candidate designs satisfy constraints, recalculating scores to catch arithmetic or coding errors, and examining whether the selected design is separated meaningfully from close competitors. If small changes in scenario probabilities or utility values reverse the ranking, report that the decision is sensitive rather than presenting the selected design as unambiguously superior. Also inspect whether the choice changes because of a genuine difference in expected utility or because candidates were scored with inconsistent units or definitions.

A fixed design is a useful alternative when decisions cannot be changed after enrollment or when a prespecified schedule is operationally essential. A robust fixed design can account for parameter uncertainty before data accrue, but it does not condition future choices on the realized interim state. A one-step, or greedy, redesign maximizes the immediate conditional score at each decision. It is simpler, but it can differ from the policy that maximizes terminal utility when an immediate gain affects later options or when the value of a choice depends on future outcomes. The appropriate comparison is therefore determined by the decision objective and the permitted adaptation, not by a universal preference for one design form.

Common Mistake and Limitations

Common mistake: “The original design selected time \(B\), so time \(B\) remains optimal after interim data arrive.” This reasoning treats an earlier decision as if it were unconditional and permanently binding. The correction is to define the current state, update the feasible set to reflect remaining opportunities, and compare future choices under the current parameter distribution and the stated objective. If the protocol requires the original design to remain fixed, that is a constraint—not evidence that it is still the conditional optimizer.

Conditional optimization does not guarantee a good decision if the model, utility, or current uncertainty distribution is poorly specified. It can also be computationally demanding when many possible interim states and future designs must be considered. Finally, maximizing an expected criterion need not protect against every unfavorable scenario or guarantee a useful realized dataset. The result is an optimum relative to its assumptions, feasible choices, and objective; those qualifications belong in its interpretation.