Seamless Phase II/III Adaptive Designs

Introduction

Traditional drug development often separates clinical development into distinct trials: a Phase II study is conducted to determine whether a treatment is sufficiently promising, followed by a new Phase III trial intended to provide confirmatory evidence.

That separation has an obvious advantage: the objectives and statistical properties of each trial are relatively straightforward.

It also has a substantial cost.

The Phase II trial may require hundreds of patients, months or years of follow-up, and extensive operational infrastructure. If the treatment is selected for further development, a new Phase III trial must then be initiated, with another protocol, another site activation process, another randomization system, and another population of patients.

A seamless Phase II/III adaptive design attempts to reduce this duplication by allowing the trial to transition from an exploratory development stage directly into a confirmatory stage without formally stopping the trial and starting an entirely separate Phase III study.

What Does "Seamless" Actually Mean?

The word seamless does not simply mean that a trial has an interim analysis.

A conventional adaptive Phase III trial may have an interim analysis at which sample size is changed or treatment arms are dropped.

A seamless Phase II/III design goes further: the interim analysis is explicitly part of a development strategy in which the trial transitions from an exploratory objective to a confirmatory objective.

For example, the first stage might evaluate four doses:

The interim analysis might conclude that doses B and C are sufficiently promising, discontinue A and D, and select C as the dose to carry forward.

Instead of terminating the study and starting a new Phase III trial, the same trial can continue enrolling additional patients under the selected treatment strategy.

Seamless vs. Separate Phase II and Phase III Trials

The Fundamental Statistical Problem

Suppose the treatment effect is represented by:

and the confirmatory hypothesis is:

The Phase II stage provides an estimate:

and the later Phase III stage provides additional information:

If the Phase II data are used only descriptively, the final analysis may be relatively simple.

But suppose the interim result is used to decide:

• which treatment is continued,
• which dose is selected,
• how many additional patients are enrolled, or
• which endpoint or population receives priority.

The final analysis is then conditional on an earlier data-dependent decision.

That is where ordinary textbook hypothesis testing can fail.

Why Type I Error Requires Special Attention

Suppose the nominal one-sided type I error is:

If the treatment selection decision were completely independent of the confirmatory data, a conventional Phase III analysis could often be used.

But in a seamless design, the treatment is frequently selected because it looked especially good in Stage 1.

Consequently, the selected treatment has been chosen partly because of the observed data.

This creates a form of selection-induced bias if the Phase II information is then treated as though the selected treatment had been fixed in advance.

A Simple Illustration of Selection Bias

Imagine two experimental doses, A and B, with identical true efficacy. Suppose Stage 1 contains enough patients that the estimated effects are:

The investigator naturally selects A.

But A was selected precisely because its observed result was larger.

If the Stage 1 estimate for A is then combined naively with the Stage 2 estimate as though A had been selected in advance, the final statistic benefits from a selection effect.

The treatment that happened to look best is being given the opportunity to carry its favorable random fluctuation into the confirmatory analysis.

This is one reason why seamless designs require specialized methods for combining evidence.

The Basic Seamless Architecture

Consider a trial with three experimental treatments and one control:

Stage 1 enrolls patients to explore the treatments.

At the interim analysis:

• \(T_1\) is dropped for insufficient activity.
• \(T_2\) is dropped for insufficient activity.
• \(T_3\) is selected.
• The control continues.
• Additional patients are enrolled for confirmatory evaluation.

The trial therefore looks like:

Two Broad Classes of Seamless Designs

It is useful to distinguish between two broad concepts.

Operationally Seamless Design

The trial uses one protocol and transitions from a Phase II-like stage into a Phase III-like stage, but the confirmatory analysis may rely primarily on information collected after treatment selection.

This can simplify the inferential problem.

Inferentially Seamless Design

Information collected during Stage 1 contributes directly to the final confirmatory test.

This can provide greater statistical efficiency, but the analysis must account for the dependence created by treatment selection and other adaptations.

Example 1: Dose Selection in a Seamless Oncology Trial

Suppose a randomized oncology study investigates three doses of an experimental drug:

The first stage enrolls 120 patients.

The primary Phase II objective is to identify a dose with an encouraging response profile while maintaining acceptable safety.

Suppose the interim results are:

A prespecified rule might select D2 because it provides the best overall benefit-risk profile.

D1 and D3 are discontinued.

The trial then continues with:

using an additional 300 patients.

The final trial therefore contains a Phase II component and a Phase III component without requiring the treatment-selection stage to be treated as a completely separate clinical trial.

What Was Learned During Phase II?

The first stage can provide several kinds of information.

• Preliminary treatment effect
• Dose-response information
• Safety information
• Variability estimates
• Event rates
• Control-group response rate
• Recruitment information
• Endpoint feasibility
• Subgroup behavior

Some of these quantities may be used to make adaptation decisions.

Others may be used for planning but excluded from the final confirmatory hypothesis test.

This distinction should be explicit.

Information Used for Adaptation vs. Information Used for Confirmation

Independent Increments

Many seamless designs are built around the idea of independent increments.

Let:

represent standardized information from Stage 1, and:

represent standardized information from the additional Stage 2 data.

Under appropriate assumptions, the two increments can be treated as independent:

This structure is extremely useful because the first-stage information can be combined with the second-stage information without treating the entire adaptation process as though it never occurred.

Combining Stage-Wise Evidence

One general approach is to convert each stage's evidence into a stage-specific p-value:

A combination function then combines these p-values.

A general form is:

where \(w_1\) and \(w_2\) are prespecified or appropriately determined weights.

The final test rejects the null when the combination statistic crosses its critical value.

Inverse-Normal Combination Tests

One important approach uses the inverse normal transformation.

Suppose the stage-specific p-values are \(p_1\) and \(p_2\).

Define:

and:

A weighted combination statistic is:

with weights satisfying:

The final decision can then be based on the corresponding critical value.

For example, with equal information weights:

The combined statistic becomes:

This is a conceptual illustration. In an actual confirmatory trial, the exact combination procedure, weights, information fractions, and treatment-selection rules must be prespecified appropriately.

Why the Stage 1 Weight Matters

Suppose the information fraction at the interim analysis is:

where \(I_1\) is the information available at Stage 1 and \(I\) is the total information.

A natural weighting structure might therefore relate the contribution of each stage to its information.

If Stage 1 contributes 25% of the total information, its influence should not be treated as though it contributes 75%.

The exact weighting scheme depends on the statistical method and design.

Example 2: Continuous Endpoint

Suppose the primary endpoint is change from baseline in a continuous biomarker.

The Phase II stage estimates the treatment effect:

with standard error:

The corresponding z-statistic is approximately:

Suppose the resulting one-sided p-value is approximately:

The treatment is selected for continuation.

Stage 2 produces:

with:

Using an illustrative inverse-normal combination with equal weights:

The important point is not the numerical result itself but the structure: Stage 1 evidence and Stage 2 evidence are combined through a prespecified method rather than treating treatment selection as irrelevant.

What If the Treatment Is Selected Based on the Best Stage 1 Result?

Suppose three treatments are evaluated.

Their Stage 1 p-values are:

Suppose the protocol specifies that the treatment with the smallest p-value will be selected.

Then \(T_3\) is selected.

But the final test cannot simply pretend that \(T_3\) was the only treatment under consideration from the beginning.

The selection rule itself is part of the statistical design.

Prespecified Treatment Selection Rules

A treatment-selection rule might be based on:

• Largest estimated treatment effect
• Largest posterior probability of benefit
• Best benefit-risk profile
• Best dose-response profile
• Minimum efficacy threshold plus acceptable safety
• Composite utility function

For example:

If more than one treatment meets the criterion, a secondary prespecified selection rule may be applied.

Example 3: Multiple Treatments With a Futility Rule

Suppose four experimental treatments are being investigated.

At the interim analysis, the protocol specifies:

The design can therefore terminate early if no treatment is sufficiently promising, or continue with one selected treatment.

The important statistical issue remains the same: if the selected treatment will contribute Stage 1 information to the final confirmatory analysis, the method for doing so must account for the selection process.

Seamless Sample Size Re-Estimation

Another common adaptation is sample size modification.

Suppose the original design assumed a treatment effect:

After Stage 1, the observed effect may suggest that the true effect is smaller.

The trial might then increase the planned sample size.

Alternatively, if the effect is larger than anticipated, the trial might retain or reduce the planned sample size subject to prespecified constraints.

The general structure is:

where \(N_2\) is the additional sample size.

Example 4: Sample Size Adaptation

Suppose a Phase II/III trial initially plans:

The design assumes a standardized effect size of:

At the interim analysis, the observed standardized effect is only:

The protocol permits increasing the maximum sample size to:

The adaptation algorithm might calculate that approximately 820 patients are required under the updated assumptions.

The trial therefore increases the planned maximum enrollment from 600 to 820.

The exact inferential consequences depend on how the sample size decision is made and whether the interim treatment effect contributes to the final test.

Conditional Power

A common way to evaluate whether additional enrollment is likely to produce success is conditional power.

Conditional power asks: Given what has already been observed, what is the probability that the trial will ultimately reject the null under a specified assumption about the future data?

Symbolically:

The assumption about future treatment effect must be specified.

Possible choices include:

• Observed interim effect
• Originally assumed effect
• A clinically meaningful effect
• A design alternative

Illustrative Conditional Power Example

Suppose a trial has:

• planned total information corresponding to 800 patients,
• an interim analysis after 400 patients,
• a target effect size of 0.30.

Suppose the interim z-statistic is:

The trial is therefore not currently close to a conventional confirmatory boundary.

Under the originally assumed effect size, the conditional probability of eventual success may still be substantial because approximately half the information remains to be collected.

If the observed interim effect is instead used as the assumed future effect, conditional power may be much lower.

Predictive Power vs. Conditional Power

These concepts are often confused.

Conditional Power

Conditional power conditions on the interim data and evaluates future success under a specified effect assumption.

Predictive Probability of Success

Predictive probability integrates over uncertainty about the future treatment effect, often using a prior or posterior distribution.

Conceptually:

This is especially useful in Bayesian adaptive designs.

Why Seamless Designs Can Save Time

Consider a traditional development program.

A seamless design may allow the trial to continue recruiting while the interim decision is made, with sites and infrastructure already in place.

This can reduce development time, although the actual operational benefit depends heavily on the disease, endpoint, recruitment rate, database availability, and regulatory requirements.

Operational Seamlessness Is Not the Same as Statistical Seamlessness

Suppose Stage 1 finishes and the selected treatment is identified.

The sponsor could continue using the same trial infrastructure but analyze only patients randomized after treatment selection for the confirmatory hypothesis.

That is operationally seamless.

Alternatively, the sponsor could use both pre- and post-selection patients in the final analysis through a combination-test framework.

That is inferentially more integrated.

These approaches have different statistical implications.

Example 5: "Fresh" Phase III Information

Suppose Stage 1 enrolls:

patients.

At the interim analysis, Treatment A is selected.

The sponsor then enrolls:

additional patients.

One conservative confirmatory strategy is to define the final primary hypothesis test using only the 600 post-selection patients.

The Stage 1 data were still crucial for treatment selection, but they do not necessarily contribute directly to the confirmatory test statistic.

This can simplify the inference but sacrifices some potential statistical efficiency.

Example 6: Combining Phase II and Phase III Evidence

Suppose the same trial instead uses both stages.

The first stage produces:

and the second stage produces:

A prespecified combination-test procedure can incorporate both pieces of evidence.

The advantage is that the information collected during Phase II does not become statistically irrelevant after treatment selection.

The price is greater statistical complexity.

Stage 1 and Stage 2 Need Not Have Identical Roles

A sophisticated seamless design can assign different objectives to each stage.

This is one reason the design should be considered a development strategy rather than merely a sample-size trick.

Endpoint Continuity Is Critical

A seamless Phase II/III trial is easiest to interpret when the endpoint used for Phase II naturally transitions into the confirmatory endpoint.

For example:

• Phase II: progression-free survival
• Phase III: progression-free survival

is conceptually much more seamless than:

• Phase II: biomarker response
• Phase III: overall survival

although the latter is possible.

If the endpoint changes, the protocol should explain exactly how the Phase II information relates to the confirmatory objective.

Example 7: Oncology Endpoint Transition

Suppose Stage 1 uses objective response rate to select a dose.

The confirmatory endpoint is progression-free survival.

Stage 1 may therefore be used for:

• dose selection,
• initial safety characterization,
• preliminary efficacy assessment, and
• sample size planning.

But the final confirmatory hypothesis concerns progression-free survival.

In this setting, the Stage 1 response rate should not automatically be treated as though it were direct confirmatory evidence for the progression-free survival hypothesis.

Control-Arm Continuity

Control-arm handling is another important consideration.

Suppose Stage 1 contains a control group:

and Stage 2 continues the same control:

The statistical treatment of the control information depends on the design.

Potential approaches include:

• Use all control information.
• Use only post-selection controls for confirmation.
• Combine independent increments.
• Use a prespecified combination test.

The appropriate choice depends on the adaptation and inferential framework.

Why Concurrent Controls Matter

Suppose Stage 1 is conducted in 2026 and Stage 2 extends into 2028.

Disease management, diagnostic criteria, supportive care, and background therapy may change.

If historical or nonconcurrent controls are used, temporal drift can complicate interpretation.

A seamless design using a continuing concurrent control can therefore have an important practical advantage.

Changing the Randomization Ratio

A seamless design may also alter randomization after treatment selection.

For example, Stage 1 might use:

randomization across three experimental arms and control.

After selecting one treatment, Stage 2 might use:

randomization between the selected treatment and control.

Such changes must be incorporated into the statistical design because the information contribution of each stage changes.

Example 8: Changing Randomization After Dose Selection

This is a natural operational design.

However, the final information fraction cannot simply be calculated from the total number of patients without considering the allocation and variance structure.

Information Fraction

In group sequential and adaptive designs, the relevant quantity is often information rather than sample size.

Define:

Here:

• \(I_1\) is the statistical information available at the interim analysis.
• \(I_{\max}\) is the planned maximum information.
• \(t_1\) is the information fraction.

For a simple normally distributed endpoint with equal allocation and known variance, information may be approximately proportional to sample size.

For survival endpoints, information is more closely related to the number of events than the number of randomized patients.

Example 9: Event-Driven Seamless Design

Suppose a Phase II/III trial has a maximum target of:

events.

The interim analysis occurs after:

events.

The information fraction is approximately:

Thus the interim analysis occurs at approximately 40% information.

If the trial then changes the target to 650 events, the remaining information and final analysis need to be recalculated under the adaptation framework.

Stopping for Futility

A seamless design may stop the entire trial if the interim evidence suggests that the treatment is unlikely to succeed.

One approach uses conditional power.

For example:

Alternatively, a predictive probability might be used:

The exact threshold is a design choice and must be prespecified.

Futility Does Not Always Require Type I Error Adjustment

A subtle but important point is that a nonbinding futility rule can often be handled differently from an efficacy rule.

For example, if the trial is stopped when there is insufficient evidence of benefit, the rule may not increase the probability of rejecting the null.

Nevertheless, the exact effect depends on the stopping rule and final analysis.

Therefore, the protocol and statistical analysis plan should explicitly describe whether futility boundaries are binding or nonbinding and how they enter the final operating-characteristic calculations.

Early Efficacy Stopping

A seamless Phase II/III design may also permit stopping for efficacy.

For example:

The critical value \(c_1\) must be selected so that the overall type I error remains controlled.

This is conceptually similar to group sequential methodology.

Seamless Designs and Group Sequential Designs

The two concepts overlap but are not identical.

A seamless Phase II/III design can therefore incorporate group sequential features, treatment selection, sample size adaptation, or multiple adaptive components.

A More Complete Worked Example

Consider a randomized cardiovascular trial evaluating two doses:

The trial uses a continuous primary endpoint.

The initial design is:

Stage 1

The first 240 patients are randomized among the three arms.

Suppose the interim estimated treatment effects versus control are:

The corresponding standardized statistics are approximately:

and:

Suppose the prespecified dose-selection rule chooses D2.

Stage 2

The trial drops D1 and continues with:

An additional 480 patients are enrolled.

Suppose the Stage 2 analysis produces:

A valid combination-test procedure can then combine the Stage 1 evidence associated with D2 and the independent Stage 2 evidence.

The final inference is not simply:

without considering the fact that D2 was selected because it looked better than D1.

The selection mechanism must be part of the confirmatory design.

What Happens to Patients Randomized to Dropped Treatments?

Patients randomized to D1 before the interim decision remain part of the clinical trial database.

Their data can contribute to:

• safety summaries,
• dose-response understanding,
• descriptive efficacy analyses,
• development decisions, and
• regulatory documentation.

But their data do not automatically contribute to the confirmatory hypothesis test for D2.

Whether and how they contribute is a statistical-design question, not simply a programming decision.

The Role of the Independent Statistical Team

Because treatment selection may require access to unblinded interim efficacy data, operational safeguards are often needed.

A typical structure separates:

• the unblinded interim analysis team,
• the sponsor personnel involved in trial operations,
• the independent statistical analysis group, and
• the clinical decision-making committee.

The exact governance structure depends on the trial and organizational requirements.

Blinded vs. Unblinded Adaptations

Sample Size Re-Estimation Based on Nuisance Parameters

Consider a continuous endpoint with variance:

Suppose the initial sample size was calculated assuming:

At interim, the blinded pooled variance estimate suggests:

The larger variance reduces statistical information.

A blinded sample size re-estimation can therefore increase the required sample size while avoiding direct use of treatment-effect information.

This is often statistically simpler than adapting based on an unblinded observed treatment effect.

Sample Size Formula Illustration

For a simple two-group comparison of means with equal allocation, a rough planning relationship is:

If the assumed standard deviation increases from 10 to 12 while the effect size remains fixed, the sample-size multiplier is approximately:

Thus a design that originally required approximately 200 patients might, under the simplified assumptions, require roughly:

patients.

This is an illustration of the statistical principle rather than a complete sample-size calculation.

Adaptive Enrichment

A seamless Phase II/III design may also select a target population.

Suppose the treatment appears particularly effective in biomarker-positive patients.

The trial might transition from:

to:

This is an example of population enrichment.

The adaptation can potentially improve efficiency if the treatment effect is truly concentrated in the selected subgroup.

But it introduces additional statistical considerations because the subgroup was selected using observed interim data.

Example 10: Biomarker Enrichment

Suppose Stage 1 contains 300 patients:

• 180 biomarker-negative
• 120 biomarker-positive

The interim estimated treatment effects are:

The prespecified rule might allow the trial to continue enrollment exclusively in the biomarker-positive population if the estimated effect and safety criteria meet predefined thresholds.

The final analysis then needs to reflect the population-selection procedure.

Multiplicity in Seamless Designs

Multiplicity can arise from several sources simultaneously.

• Multiple doses
• Multiple treatment arms
• Multiple interim analyses
• Multiple endpoints
• Multiple subgroups
• Early efficacy stopping
• Treatment selection

For example, a design with three doses and two interim analyses has a much more complex multiplicity structure than a single fixed two-arm comparison.

Familywise Error Rate

If several confirmatory hypotheses are tested, the relevant quantity may be the familywise type I error rate:

A seamless design may therefore require simultaneous consideration of:

• treatment multiplicity,
• interim analyses,
• dose selection, and
• endpoint multiplicity.

The multiplicity strategy should be established before the confirmatory trial begins.

Seamless Designs With Multiple Confirmatory Treatments

Suppose two treatments survive Stage 1:

Both are carried into Stage 2.

The final hypotheses might be:

If both hypotheses are tested confirmatorily, multiplicity must be addressed.

Possible approaches include:

• Bonferroni-type procedures
• Holm procedures
• Gatekeeping strategies
• Graphical multiple-testing procedures
• Closed testing

The appropriate method depends on the clinical claims being pursued.

Seamless Designs With Multiple Endpoints

Suppose the final confirmatory objectives include:

• overall survival,
• progression-free survival, and
• objective response.

A hierarchical testing strategy might specify:

The adaptive treatment-selection mechanism must still be incorporated into the overall inferential strategy.

Seamless Design With a Binary Endpoint

For a binary endpoint, suppose Stage 1 provides:

and Stage 2 provides:

If treatment is fixed throughout the trial, total responses can often be represented as:

But if treatment selection occurs between stages, the distribution of the selected treatment's Stage 1 result is no longer the same as if that treatment had been fixed in advance.

This is the key reason the naive binomial calculation can be inappropriate after adaptive treatment selection.

Example 11: Why Naive Pooling Can Be Misleading

Suppose two treatments have identical true response probability:

After Stage 1:

Treatment A is selected because it has the larger observed response rate.

If Stage 2 produces 15 responses among 50 additional patients for A, a naive analysis might pool:

responses among:

patients.

The pooled observed response rate is:

But the 8 Stage 1 responses are not statistically neutral with respect to selection: A was chosen precisely because it had the more favorable Stage 1 result.

Therefore, ordinary fixed-treatment inference may not preserve the intended type I error.

Combination Tests as a General Solution

Combination-test approaches separate the evidence into stage-specific pieces.

For example:

and:

The final test uses a prespecified combination function.

The attraction is that the adaptation can be quite flexible provided the necessary independence or conditional validity properties are satisfied.

Fisher's Combination Method

One classic combination method is Fisher's method:

Under appropriate assumptions, this statistic has a chi-square distribution with four degrees of freedom when the two p-values are independent and valid under the null.

Thus:

The method is attractive because it allows the stage-specific evidence to be combined without requiring the final analysis to behave like a simple fixed-design test.

Inverse-Normal vs. Fisher Combination

Neither method should be selected solely because it is mathematically convenient. The choice should reflect the exact adaptation, endpoint, desired inference, and regulatory strategy.

Conditional Error

Another important framework for adaptive designs is the conditional error principle.

Suppose an original trial has a rejection region \(R\).

After interim information \(X_1=x_1\), the conditional probability under the null of eventually rejecting the hypothesis can be represented as:

An adaptation can then be constructed so that the new design does not exceed the original conditional error.

Conceptually, this allows considerable flexibility while preserving the original type I error guarantee.

Why Conditional Error Is Powerful

Conditional-error methodology can accommodate adaptations such as:

• sample size modification,
• treatment selection,
• population selection,
• changes in allocation, and
• other design modifications.

The exact implementation depends on the original test and the adaptation.

Example 12: Conceptual Conditional-Error Adaptation

Suppose an original confirmatory design has a one-sided type I error:

At interim, the observed data are \(x_1\).

Suppose the original design's conditional error at that point is:

The sponsor decides to modify the second-stage sample size and allocation.

A valid adaptive continuation must be constructed so that its conditional probability of rejecting under the null does not exceed the permitted conditional error.

The exact critical value depends on the adaptation and test statistic.

Sample Size Adaptation and Conditional Error

Suppose the trial originally plans:

At interim, the sponsor increases the sample size to:

If the final test simply pools all observations and applies the original critical value without considering the adaptation, type I error may not be preserved in every adaptive scenario.

Conditional-error or combination-test methods can provide a principled way to construct the adapted final test.

Bayesian Seamless Designs

Bayesian methods can also be used in seamless Phase II/III designs.

Suppose the treatment effect has prior distribution:

After Stage 1, the posterior becomes:

The interim decision may use:

For example:

might make a treatment eligible for continuation.

A Bayesian design can also calculate the predictive probability of eventual success.

However, Bayesian adaptation does not eliminate the need to consider the regulatory and inferential framework applicable to a confirmatory claim.

Example 13: Bayesian Dose Selection

Suppose three doses have posterior probabilities of exceeding the clinically important effect:

If the prespecified rule requires a probability of at least 0.80, both D2 and D3 qualify.

A secondary rule could select the dose with the highest posterior expected utility.

D2 would therefore be carried forward if the utility rule also favors D2.

Again, the selection rule is part of the design.

Frequentist vs. Bayesian Seamless Designs

Seamless Designs and Confidence Intervals

A confirmatory trial should generally report effect estimates with confidence intervals rather than relying only on a p-value.

For a treatment effect estimate:

a conventional interval may be represented as:

However, after adaptive treatment selection, the interpretation of an ordinary confidence interval may require additional care.

Selection can cause the selected treatment's effect estimate to be optimistically biased.

Therefore, confirmatory estimation procedures should be aligned with the adaptive inferential framework.

Selection Bias in Effect Estimates

Suppose three treatments all have true effect:

Random variation might produce Stage 1 estimates:

A rule that selects the largest estimate will choose A.

Even though all treatments have true effect zero, the selected treatment has a positive observed effect.

This is a simple illustration of the winner's curse or selection-induced optimism.

Selection Does Not Automatically Invalidate the Trial

Adaptive treatment selection is not inherently a statistical error.

The problem arises when the analysis fails to account for the selection.

A well-designed seamless trial can preserve valid confirmatory inference if the adaptation and final testing procedure are appropriately constructed.

Interim Data Access

The more flexible the adaptation, the more important control of interim information becomes.

For example, a treatment-selection committee may need to know:

• treatment-specific efficacy,
• safety,
• biomarker information, and
• operational feasibility.

The committee may not need access to all detailed patient-level data.

A carefully designed interim process can therefore provide the information needed for the adaptation without unnecessarily exposing the broader study team to unblinded information.

Data Cutoff and Timing

The interim decision should specify exactly which patients contribute to the interim analysis.

For example:

What Is a "Stage"?

A stage does not necessarily mean a conventional regulatory phase.

In a statistical design, Stage 1 and Stage 2 are periods of information accumulation separated by an adaptation.

The first stage may be described as Phase II-like and the second as Phase III-like, but the statistical stages are defined by the information and decision structure.

Seamless Does Not Mean That the Trial Has No Phase Boundary

The trial can have a clear transition in objectives even if there is no operational gap.

For example:

When a Seamless Design May Be Attractive

Potential advantages include:

• Reduced duplication between Phase II and Phase III.
• Potentially faster transition to confirmatory development.
• Ability to use information accumulated during the early stage.
• Potentially lower overall development cost.
• Potentially fewer patients exposed to ineffective doses.
• More efficient use of trial infrastructure.
• Ability to adapt treatment selection based on emerging evidence.

Potential Disadvantages

The design also introduces challenges.

• Greater statistical complexity.
• More complicated simulation requirements.
• Potential regulatory complexity.
• More complicated trial operations.
• Need for careful interim-data governance.
• Potential selection bias.
• More complex final estimation.
• Potential difficulty maintaining endpoint consistency.
• Greater programming and validation requirements.

Simulation Is Essential

Analytical calculations alone are often insufficient for a complex seamless design.

Simulation can evaluate the complete adaptive decision process under many possible scenarios.

For example, investigators can simulate:

• Null treatment effects
• Clinically meaningful treatment effects
• Intermediate effects
• Different dose-response patterns
• Different control rates
• Different dropout rates
• Different recruitment patterns
• Different event rates
• Different treatment-selection outcomes

Simulation-Based Type I Error

Suppose the trial is simulated 100,000 times under the global null hypothesis.

If 2,420 simulations produce a false positive, the estimated type I error is:

This provides an empirical estimate of the complete adaptive design's type I error.

Simulation is particularly useful when the analytical distribution is complicated by treatment selection, changing randomization, multiple adaptations, or survival endpoints.

Simulation-Based Power

Suppose the clinically meaningful treatment effect is simulated.

If 82,100 of 100,000 simulated trials reject the null:

The estimated power is therefore approximately 82.1%.

Simulation Across Multiple Scenarios

R Simulation Framework

A simplified simulation structure might look like:

set.seed(12345)

nsim <- 100000

success <- logical(nsim)

for (i in seq_len(nsim)) {

  # Generate Stage 1 data
  stage1 <- rnorm(
    n = 240,
    mean = 0,
    sd = 10
  )

  # Prespecified interim decision
  # treatment selection occurs here

  # Generate Stage 2 data
  stage2 <- rnorm(
    n = 480,
    mean = 0,
    sd = 10
  )

  # Apply the prespecified
  # final combination rule

  # success[i] <- TRUE/FALSE
}

mean(success)

A real simulation should explicitly model randomization, treatment effects, endpoint distributions, treatment selection, missing data, interim timing, sample-size adaptation, and the exact final analysis.

Simulating a Dose-Selection Rule

A simplified dose-selection simulation might be:

select_dose <- function(effect_d1,
                         effect_d2,
                         effect_d3) {

  effects <- c(
    D1 = effect_d1,
    D2 = effect_d2,
    D3 = effect_d3
  )

  names(which.max(effects))
}

The actual implementation should normally include uncertainty, safety criteria, statistical thresholds, and the complete prespecified selection algorithm.

Assessing Selection Probabilities

One useful operating characteristic is:

For example, if D2 is truly the best treatment, investigators may want to know the probability that the design selects D2.

These probabilities should be estimated through simulation when the selection algorithm is sufficiently complex.

Selection Probability Is Different From Power

This distinction is extremely important.

Selection probability asks:

Power asks:

A design can have high probability of selecting the best treatment but relatively modest probability of successfully demonstrating the selected treatment's confirmatory effect.

Conversely, a design may have high confirmatory power after selecting a treatment but frequently select the wrong treatment when treatment effects are similar.

Example 14: Two Nearly Identical Doses

Suppose the true effects are:

These treatments are clinically very similar.

Even a well-designed interim analysis may select either treatment because of sampling variability.

That does not necessarily indicate a poor design.

If the treatments have essentially equivalent benefit-risk profiles, the clinical consequence of selecting the "wrong" one may be minimal.

Benefit-Risk Treatment Selection

Suppose dose D3 has the largest efficacy estimate but substantially more toxicity.

A reasonable selection rule might use a utility function:

where:

• \(E_j\) represents efficacy,
• \(S_j\) represents a safety measure, and
• \(w_E,w_S\) represent prespecified weights.

The dose selected for Phase III is then the one maximizing:

In practice, the utility function may be considerably more sophisticated.

Safety Should Not Be Reduced to a Single Number

A seamless design may involve multiple safety considerations:

• Serious adverse events
• Discontinuations
• Laboratory abnormalities
• Organ toxicity
• Dose interruptions
• Treatment-related mortality

The treatment-selection algorithm should therefore be developed jointly by clinical, statistical, pharmacological, and safety experts.

Interim Decision Committees

A seamless trial may establish an independent committee responsible for the interim decision.

The committee may receive a controlled package containing:

• blinded safety summaries,
• unblinded efficacy summaries where necessary,
• dose-response analyses,
• biomarker analyses, and
• operational information.

The committee then applies the prespecified decision rules.

Adaptation Rules Should Be Algorithmic Whenever Possible

A rule such as:

"Select the dose that appears most promising."

is generally too vague for a complex confirmatory design.

A stronger rule is:

and, if multiple doses qualify:

The exact framework may be frequentist or Bayesian, but the key principle is that the rule should be sufficiently precise to reproduce the decision.

Unplanned Adaptations

A major risk occurs when investigators make an unplanned decision after reviewing interim efficacy.

For example:

"Initially we planned to select the highest dose, but because D2 looked much better than expected, we decided to continue D2 instead."

If the final statistical method assumes a different adaptation, the operating characteristics may no longer hold.

Protocol Elements for a Seamless Phase II/III Trial

The protocol should clearly define:

• Stage 1 objectives
• Stage 2 objectives
• Interim analysis timing
• Primary endpoint
• Key secondary endpoints
• Treatment-selection criteria
• Futility criteria
• Efficacy stopping criteria
• Sample-size adaptation rules
• Maximum sample size
• Randomization changes
• Population-selection rules
• Final hypothesis-testing strategy
• Multiplicity adjustment
• Interim data-access procedures
• Database-lock procedures
• Handling of missing data

Statistical Analysis Plan Considerations

The SAP should go beyond simply repeating the protocol.

It should describe the exact statistical implementation.

For example:

• How the interim dataset is constructed
• How treatment effects are estimated
• How p-values are generated
• How treatment selection is determined
• How Stage 1 and Stage 2 evidence are combined
• How information fractions are calculated
• How sample size is adapted
• How missing data are handled
• How confidence intervals are constructed
• How sensitivity analyses will be performed

Database Locks in Seamless Designs

A seamless trial can require multiple database freezes.

The precise terminology and procedures should be defined in the study's data management and statistical plans.

Missing Data

Missing data can be particularly complicated because the Stage 1 decision may depend on the observed interim data.

Potential approaches include:

• Predefined missing-as-failure rules for binary endpoints
• Multiple imputation
• Mixed models for repeated measures
• Inverse probability weighting
• Sensitivity analyses

The method should be specified before the final analysis and should be consistent with the estimand.

Estimands in Seamless Phase II/III Trials

The estimand framework is particularly useful because the clinical question can change between the exploratory and confirmatory stages.

The protocol should clarify:

• Population
• Treatment condition
• Endpoint
• Intercurrent-event strategy
• Summary measure

For example, the confirmatory estimand might be:

The fact that the treatment was selected adaptively does not remove the need to define the clinical question precisely.

Intercurrent Events

Suppose a patient discontinues treatment and begins rescue medication.

The final estimand must specify how this event affects the treatment effect.

Possible strategies include:

• Treatment-policy strategy
• Hypothetical strategy
• While-on-treatment strategy
• Composite strategy
• Principal-stratum strategy

These choices should be aligned between the Phase II learning objectives and the Phase III confirmatory objective when possible.

Seamless Designs With Survival Endpoints

For a time-to-event endpoint, the statistical information is often related to the number of observed events.

Let:

be the number of events at interim and:

the final event target.

The approximate information fraction is:

A seamless design may therefore transition from exploratory treatment selection into an event-driven confirmatory survival analysis.

Example 15: Seamless Survival Trial

Suppose a trial evaluates two doses:

against control.

The interim analysis occurs after 180 events.

The maximum trial requires 450 events.

Thus:

Suppose the interim hazard-ratio estimates are:

Suppose D2 is selected.

The trial then continues toward 450 events using D2 and control.

If the Stage 1 information contributes to the final hypothesis test, the analysis must account for the adaptive selection.

Event-Driven Sample Size Adaptation

Suppose the original design assumes a hazard ratio:

At interim, the observed effect appears closer to:

The protocol may permit increasing the final event target.

This can preserve power if the true effect is smaller than originally assumed, provided the adaptation and final analysis are constructed appropriately.

Seamless Designs and Accrual

Operational efficiency depends heavily on recruitment.

Suppose a trial recruits:

If a separate Phase III trial requires another 12 months for activation and enrollment, a seamless design could potentially save substantial calendar time.

However, if the primary endpoint requires long follow-up, eliminating the administrative gap may not eliminate the biological waiting time.

Trial Duration vs. Sample Size

Adaptive designs should therefore be evaluated on at least two dimensions:

• Number of patients
• Calendar time

A design with a smaller sample size can still be slower if the endpoint requires long follow-up.

Conversely, a slightly larger design may produce an earlier development decision if it allows more efficient recruitment and fewer trial transitions.

Cost Considerations

A seamless design may reduce:

• protocol start-up costs,
• site initiation costs,
• database-build duplication,
• vendor transitions,
• patient re-enrollment, and
• time between development stages.

But it may increase:

• statistical programming complexity,
• simulation requirements,
• interim-analysis infrastructure,
• independent statistical oversight, and
• operational planning requirements.

Therefore, economic efficiency should be evaluated across the entire development program.

Common Mistakes

1. Calling any two-stage trial "seamless." A true seamless Phase II/III strategy has an explicit transition from exploratory development objectives to confirmatory objectives.
2. Selecting the best treatment and then using a naive final test. Selection based on interim efficacy must be incorporated into the inferential strategy.
3. Assuming prespecification automatically solves type I error. Prespecification is necessary but does not guarantee statistical validity if the final analysis does not properly account for the adaptation.
4. Ignoring treatment-selection bias. The selected treatment tends to have a favorable random fluctuation in the interim data.
5. Confusing sample size with information. This is especially problematic for survival endpoints.
6. Ignoring the control arm. Control data collected before and after treatment selection may require careful treatment.
7. Changing the endpoint without revisiting the estimand. A Phase II endpoint and a Phase III endpoint may answer different clinical questions.
8. Using conditional power without defining the assumed future effect. Conditional power is conditional on assumptions about future treatment effects.
9. Failing to simulate the complete design. Complex adaptive designs often have operating characteristics that are difficult to derive analytically.
10. Underestimating operational complexity. A seamless design requires sophisticated data, randomization, interim-analysis, and governance infrastructure.
11. Allowing informal treatment-selection decisions. Rules such as "select the dose that looks best" are inadequate for a confirmatory adaptive design.
12. Ignoring multiplicity. Multiple doses, treatments, endpoints, and interim analyses can interact.

A Practical Seamless Phase II/III Workflow

Worked Example Summary

Consider the following hypothetical seamless Phase II/III trial.

The trial therefore integrates:

• dose selection,
• interim efficacy assessment,
• sample-size adaptation,
• futility monitoring,
• continued concurrent control, and
• confirmatory hypothesis testing.

What Should Be Simulated?

For a realistic seamless design, the simulation program should usually report at least:

A More Complete Simulation Skeleton in R

set.seed(20260821)

nsim <- 10000

results <- data.frame(
  selected = character(nsim),
  stopped = logical(nsim),
  success = logical(nsim),
  total_n = numeric(nsim)
)

for (i in seq_len(nsim)) {

  ## -----------------------------
  ## Stage 1
  ## -----------------------------

  # Generate hypothetical treatment effects
  # and patient-level outcomes here.

  effect_d1 <- rnorm(1, mean = 0.20, sd = 0.15)
  effect_d2 <- rnorm(1, mean = 0.30, sd = 0.15)
  effect_d3 <- rnorm(1, mean = 0.25, sd = 0.15)

  effects <- c(
    D1 = effect_d1,
    D2 = effect_d2,
    D3 = effect_d3
  )

  ## -----------------------------
  ## Prespecified selection rule
  ## -----------------------------

  selected <- names(which.max(effects))

  ## -----------------------------
  ## Futility rule
  ## -----------------------------

  if (max(effects) < 0.10) {

    results$selected[i] <- selected
    results$stopped[i] <- TRUE
    results$success[i] <- FALSE
    results$total_n[i] <- 240

    next
  }

  ## -----------------------------
  ## Stage 2
  ## -----------------------------

  stage2_effect <-
    rnorm(1, mean = 0.30, sd = 0.10)

  ## -----------------------------
  ## Final analysis
  ## -----------------------------

  # Apply the prespecified
  # combination-test or
  # conditional-error procedure.

  final_success <-
    (stage2_effect > 0.20)

  results$selected[i] <- selected
  results$stopped[i] <- FALSE
  results$success[i] <- final_success
  results$total_n[i] <- 720
}

mean(results$success)
mean(results$stopped)
mean(results$total_n)

table(results$selected)

This code is intentionally a framework rather than a validated confirmatory trial-analysis program.

A production simulation should model the exact endpoint distribution, randomization, treatment effects, interim information, selection rule, final test, missing data, and stopping rules.

Why Simulation Should Include the Global Null

For a multiple-treatment seamless design, the global null might be:

The simulation should repeatedly run the entire adaptive procedure under this scenario.

The estimated rejection probability represents the empirical type I error of the complete design.

This is more informative than calculating the type I error of only the final fixed-treatment test.

Simulation Under Partial Null Scenarios

It may also be important to simulate scenarios such as:

This scenario asks whether the design can successfully identify the one effective treatment.

Another scenario might be:

This evaluates selection when treatment effects are relatively close.

Scenario-Based Design Evaluation

Interpreting the Final Result

Suppose the seamless trial selects D2 and ultimately obtains:

with a two-sided 95% confidence interval:

The final clinical interpretation should focus on:

• magnitude of treatment effect,
• precision,
• clinical relevance,
• safety,
• estimand definition, and
• the prespecified confirmatory analysis.

The fact that the treatment was selected adaptively should be explicitly considered in the statistical interpretation.

What a Successful Seamless Design Accomplishes

A well-designed seamless Phase II/III trial can allow investigators to answer two fundamentally different questions within one development framework:

The statistical challenge is to connect those questions without allowing the adaptive selection process to compromise the validity of the confirmatory claim.

The Most Important Concept

The most important idea in seamless Phase II/III design is not simply combining two clinical phases into one trial.

It is the distinction between:

and:

The first objective encourages adaptation.

The second requires statistical control.

A seamless design must therefore connect the two through a formal statistical framework.

Quick Reference

References

Bauer, P. & Köhne, K. (1994). Evaluation of experiments with adaptive interim analyses. Biometrics, 50, 1029–1041.

Bauer, P. & Kieser, M. (1999). Combining different phases in the development of medical treatments within a single trial. Statistics in Medicine, 18, 1833–1848.

Bretz, F., Koenig, F., Brannath, W., Glimm, E. & Posch, M. (2009). Adaptive designs for confirmatory clinical trials. Statistics in Medicine, 28, 1181–1217.

Cui, L., Hung, H.M.J. & Wang, S.J. (1999). Modification of sample size in group sequential clinical trials. Biometrics, 55, 853–857.

Fisher, R.A. (1932). Statistical Methods for Research Workers. Oliver and Boyd.

Jennison, C. & Turnbull, B.W. (2000). Group Sequential Methods with Applications to Clinical Trials. Chapman & Hall/CRC.

Mehta, C.R. & Pocock, S.J. (2011). Adaptive increase in sample size when interim results are promising: A practical guide with examples. Statistics in Medicine, 30, 3267–3284.

Posch, M., Bauer, P. & Brannath, W. (2008). Assessing the impact of selection in adaptive designs. Statistics in Medicine.

Proschan, M.A. & Hunsberger, S.A. (1995). Designed extension of studies based on conditional power. Biometrics, 51, 1315–1324.

Tsiatis, A.A., Mehta, C., Hayter, A., et al. (2008). Inference following adaptive group sequential designs. Biometrika.

Wassmer, G. & Brannath, W. (2016). Group Sequential and Confirmatory Adaptive Designs in Clinical Trials. Springer.