Why a Redesign Must Start With the Remaining Budget
Interim data can change which observations are most valuable. A sampling time that looked useful before enrollment may become less informative after the parameter uncertainty has shifted; another time or a different allocation of participants may now improve the design. But a redesign is not an unconstrained fresh start. Visits already completed cannot be recovered, assays already used cannot be reassigned, and the remaining study must fit within its actual resource limits.
The design decision is therefore conditional on both the interim data and the resources still available. The task is to choose future sampling times and participant allocations that optimize the chosen design criterion among plans that are feasible. The criterion might measure expected information, posterior uncertainty reduction, or robust performance over a set of plausible parameter values. Whatever the criterion, an attractive score does not make a plan feasible.
For each resource type \(k\), write \(B_k\) for the original limit and \(u_{Ik}\) for the amount consumed or reserved by the interim analysis. The remaining amount is \(R_k=B_k-u_{Ik}\). If a future allocation \(d\) uses \(c_k(d)\) units, feasibility requires \(c_k(d)\leq R_k\) for every resource type. Typical resources include participant visits, assay runs, staff time, and a total monetary or operational credit limit. A plan must satisfy these conditions simultaneously, not one at a time.
Conditional Optimization With Multiple Constraints
A useful general formulation is
Here \(\mathcal{D}\) is the set of allowable future designs, \(Q(d\mid D_I)\) is the chosen design criterion evaluated using what is known at interim, and \(c_k(d)\) is the use of resource \(k\). The design space may be discrete: for example, a participant may be assigned to one of a set of allowable sampling schedules, and the number assigned to each schedule must be a nonnegative integer. Integer constraints matter because a fractional participant allocation is not implementable.
The criterion must represent the study's actual inferential aim. For example, a local information criterion may evaluate expected information at an interim estimate, while a Bayesian criterion may average over the updated parameter distribution. A robust criterion can score a plan across several plausible parameter values and optimize a worst-case or other specified summary. These approaches differ in how they evaluate performance, but all must use the same remaining-budget constraints. Recomputing the criterion after interim data is not a reason to ignore assay or visit limits.
If the criterion is information from the full study, it should include both the interim contribution and the proposed future contribution. In a simplified additive information representation, this can be written as \(I_{\text{total}}=I(D_I)+I(d\mid D_I)\). The full criterion need not itself be additive: for instance, a log-determinant criterion applies a nonlinear function to the combined information. Consequently, selecting the future plan with the largest standalone score can be wrong if it fails to account for the information already obtained.
Why the Resource Ledger Belongs in the Optimization
The constraints follow directly from the accounting. Suppose a study has a total limit \(B_k\) for resource \(k\), of which \(u_{Ik}\) has already been used or reserved. The total use under a proposed future plan is \(u_{Ik}+c_k(d)\). Requiring this not to exceed the original limit gives
This simple subtraction is consequential. If several limits apply, the plan must pass each inequality. A plan that fits the remaining assay capacity but exceeds the visit limit is infeasible; unused credits in one category do not automatically compensate for excess use in another unless the protocol explicitly permits that conversion.
Record completed and irrevocably reserved visits, assays, and other resources, along with the data used to update the design criterion.
Subtract consumed or reserved amounts from the corresponding fixed total. Do not combine unlike resources unless an explicit conversion rule exists.
Use the updated criterion to compare allowable sampling schedules and integer participant counts, rejecting any plan that violates even one remaining limit.
Recalculate every resource total from the selected allocation and confirm that the final design includes the interim data and remains within all limits.
Worked Example: Choosing an Allocation Under a Credit Limit
Suppose interim data have been incorporated into a hypothetical design criterion. Three future sampling options are available: early (E), middle (M), and late (L). The table gives the cost and the criterion-score contribution assigned to each participant under this illustrative, additive scoring rule. The remaining total limit is 11 credits, at least one participant must be assigned to E for protocol coverage, and no more than three participants can be enrolled. Find the best feasible allocation.
| Option | Credits per participant | Score per participant |
|---|---|---|
| E | 5 | 1.5 units |
| M | 3 | 2.4 units |
| L | 2 | 1.8 units |
Let \(n_E,n_M,n_L\) be nonnegative integer participant counts. The objective and credit constraint are
First consider \(n_E=1\). This uses 5 credits, leaving 6 credits and room for at most two more participants. Two M participants use \(2(3)=6\) credits and give
Other possibilities with \(n_E=1\) do not score higher: one M and one L use \(5+3+2=10\) credits and score \(1.5+2.4+1.8=5.7\) units; two L participants score \(1.5+2(1.8)=5.1\) units. A single additional participant also gives a lower score.
If \(n_E=2\), the early participants use \(2(5)=10\) credits, leaving only 1 credit. Neither an M participant, costing 3 credits, nor an L participant, costing 2 credits, can be added. Thus the best feasible score for this case is \(2(1.5)=3.0\) units, not a score that includes an L participant. Three E participants are infeasible because they cost 15 credits. Therefore the best feasible allocation is one E participant and two M participants, with a score of 6.3 units.
The score values here are hypothetical and deliberately additive to make the resource arithmetic transparent. In a real design, information contributions can interact, so the criterion should be recalculated for each complete allocation rather than assuming that separate participant scores add exactly.
Interim Updating Can Change the Best Feasible Plan
Interim data can change the criterion without changing the resource ledger. For example, updated parameter uncertainty may change the relative value of early and late observations. A robust design may also be re-evaluated over the plausible parameter set that remains relevant after interim analysis. The essential separation is: the data determine how candidate plans are scored, while the resource record determines which plans are eligible for comparison.
Worked Example: A Robust Allocation After Interim Updating
Consider three future participants and three sampling options, again using hypothetical values. After interim updating, the design team evaluates each option under two plausible parameter scenarios. The entries below are score units per participant; the design criterion is the minimum total score across the two scenarios. Each participant uses one assay, the remaining assay limit is 3, and the remaining credit limit is 7. The credit costs are 3 for E and 2 each for M and L.
| Option | Scenario 1 score | Scenario 2 score | Credits |
|---|---|---|---|
| E | 6 | 1 | 3 |
| M | 3 | 3 | 2 |
| L | 1 | 6 | 2 |
For the allocation E+M+L, the credit use is \(3+2+2=7\), and assay use is \(1+1+1=3\), so it is feasible. Its scenario totals are \(6+3+1=10\) and \(1+3+6=10\); its worst-case score is therefore \(\min(10,10)=10\) units.
For comparison, M+M+M costs \(2+2+2=6\) credits and uses 3 assays. Its totals are \(3+3+3=9\) in both scenarios, giving a worst-case score of 9 units. E+L+L costs \(3+2+2=7\) credits and uses 3 assays; its totals are \(6+1+1=8\) and \(1+6+6=13\), so its worst-case score is 8 units. E+M+M also uses 7 credits and 3 assays, but its scenario totals are 12 and 7, giving a worst-case score of 7 units.
The remaining three-participant allocations that fit the credit limit are M+M+L, M+L+L, and L+L+L, with worst-case scores 7, 5, and 3 units, respectively. Allocations with two E participants and a third participant cost at least \(3+3+2=8\) credits and are infeasible. Thus E+M+L has the highest worst-case score among the feasible allocations. This conclusion depends on the stated scenario scores and robust criterion; a different updated criterion could select a different feasible allocation.
Multiple Limits and Complete Schedule Choices
A participant allocation may involve whole schedules rather than a single sampling time. Each schedule can consume a different number of visits and assays, and its score may reflect the value of the complete sampling pattern. In that setting, listing one cost per participant is not enough: the redesign must track all resource dimensions for every schedule.
Worked Example: Reallocating Complete Schedules Under Three Limits
Suppose exactly four future participants must be assigned to one of three hypothetical schedules. After interim data, the current score contribution for each schedule is as follows. The remaining limits are 8 visits, 6 assays, and 14 credits.
| Schedule | Visits per participant | Assays per participant | Credits per participant | Score units |
|---|---|---|---|---|
| A | 2 | 2 | 4 | 5 |
| B | 1 | 1 | 2 | 3 |
| C | 1 | 2 | 3 | 4 |
Let \(a,b,c\) denote the numbers assigned to A, B, and C. The participant total gives \(a+b+c=4\). The assay use is \(2a+b+2c\), which, after substituting \(b=4-a-c\), becomes \(4+a+c\). Therefore the assay limit requires \(a+c\leq2\). The credit use becomes \(4a+2b+3c=8+2a+c\), so the credit limit requires \(2a+c\leq6\). The visit use is \(2a+b+c=4+a\), at most 8 for any allocation of four participants.
The score is \(5a+3b+4c=12+2a+c\). Since \(a+c\leq2\) and \(a,c\) are nonnegative, the score is at most \(12+2(2)=16\). This bound is attained with \(a=2,c=0,b=2\). The allocation of two A participants and two B participants uses \(2(2)+2(1)=6\) visits, \(2(2)+2(1)=6\) assays, and \(2(4)+2(2)=12\) credits. Its score is \(2(5)+2(3)=16\) units, so it is feasible and optimal under the stated score and limits.
This calculation illustrates why a plan should be checked against the complete resource vector. The score-maximizing allocation uses all remaining assays but not all remaining visits or credits; unused capacity in those latter categories does not justify exceeding the assay limit.
Diagnostics, Alternatives, and Limitations
A redesign calculation is only useful if both its inputs and its feasibility checks are reliable. Before implementation, check the resource ledger against operational records, including reservations that are not yet completed. Recompute resource use from the final integer allocation rather than relying only on an optimizer's summary. If schedules have shared or batch costs, represent those costs explicitly; a simple per-participant cost may otherwise misstate what the plan consumes.
Also inspect the objective. Confirm that it is conditional on the interim data and uses the intended updated parameter uncertainty or robustness criterion. If a scalar score is only an approximation to a matrix-valued information objective, label it as such and do not treat a linear ranking as exact. Recalculate the selected and nearby feasible plans with the full criterion when possible. A sudden score change caused by a small change in a parameter estimate, a nearly singular information matrix, or a boundary allocation can signal sensitivity that deserves attention.
Three common approaches differ in what they optimize. Keeping the original future schedule avoids adaptation but can ignore information learned at interim. Optimizing without resource constraints may reveal an attractive benchmark, but it is not an implementable plan. A greedy rule that repeatedly assigns the next participant to the option with the largest immediate score can be convenient, yet it can miss the best joint allocation when schedules interact or when early assignments consume scarce resources needed later. Conditional constrained optimization compares complete feasible plans using the selected criterion.
No method removes the need to specify the criterion, the allowable schedules, and the accounting rules carefully. An incorrect remaining budget produces an incorrect feasible set. A poorly chosen criterion can optimize a quantity that does not reflect the study objective. And if model assumptions used to score designs are inadequate, satisfying every resource limit does not make the resulting design scientifically sound.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A study has a total assay limit of 40, with 27 assays used and 5 reserved at interim. How many assays remain available for redesign?
- A proposed plan uses 6 visits, 4 assays, and 12 credits. Remaining limits are 7 visits, 5 assays, and 11 credits. Is the plan feasible? Explain.
- In the first worked example, why is a score that includes two E participants and one L participant not a feasible comparison under the stated credit limit?
- Give one reason a greedy participant-by-participant assignment can differ from the best joint allocation.
- When interim data change the design criterion, must the resource limits also change? Explain the distinction.
Solutions. 1. The remaining assay capacity is \(40-27-5=8\) assays. 2. No. Although visits and assays are within their limits, 12 credits exceed the remaining 11-credit limit. All limits must be met simultaneously. 3. That allocation costs \(2(5)+1(2)=12\) credits, which exceeds the remaining limit of 11; it must be excluded. 4. A greedy assignment can consume a scarce resource early or overlook interactions among schedules, preventing a higher-scoring complete allocation. 5. Not necessarily. Interim data may change the scores of candidate plans, while the resource limits follow from the original limits minus resources used or reserved. They change only if the study's operational budget itself is formally revised.