Sample Size Re-Estimation in Adaptive Trials

Introduction

Sample size is one of the most important design parameters in a clinical trial. Before enrollment begins, investigators typically specify the number of participants needed to achieve a desired probability of detecting a clinically meaningful treatment effect.

That calculation, however, depends on assumptions.

For example, a two-sample comparison of means requires an assumption about the standard deviation. A survival trial requires assumptions about the event rate, accrual, follow-up, and sometimes the treatment effect. A binary endpoint requires assumptions about the response probabilities.

If those assumptions are substantially wrong, the original sample size may be too small or unnecessarily large.

Sample size re-estimation (SSR) provides a mechanism for modifying the planned sample size during an ongoing trial using information that becomes available after enrollment has begun.

Why Re-Estimate Sample Size?

The original sample size calculation is based on planning assumptions.

Suppose a trial comparing two treatments assumes a common standard deviation of:

but the blinded pooled data observed during the trial suggest that the actual standard deviation is closer to 14.

The original sample size may no longer provide the planned power.

Similarly, a survival trial might assume that 60% of participants will experience the required endpoint by the time of analysis, but the accumulating data may suggest that only 45% will experience the event.

The trial may then require more participants or more follow-up to achieve its original information target.

The Basic Idea

The concept can be summarized as:

What Is Being Re-Estimated?

Sample size re-estimation does not necessarily mean estimating the treatment effect.

In many designs, the adaptation is based on a nuisance parameter whose value affects the required sample size but is not itself the primary parameter of interest.

Examples include:

Blinded vs. Unblinded Re-Estimation

One of the most important distinctions in adaptive sample size planning is whether the person performing the re-estimation can see treatment-specific information.

Blinded Sample Size Re-Estimation

In a blinded SSR procedure, the nuisance parameter is estimated without revealing the treatment assignments or treatment-specific outcomes to the adaptation team.

For example, investigators might estimate the overall variance from pooled treatment groups:

without examining which observations belong to treatment A or treatment B.

Because the treatment effect is not used to determine the revised sample size, this approach is often attractive when the primary concern is uncertainty about a nuisance parameter.

Unblinded Sample Size Re-Estimation

In an unblinded SSR procedure, the adaptation uses treatment-specific interim information.

For example, the design might calculate the probability of eventually rejecting the null hypothesis given the interim treatment effect.

This is commonly described using conditional power.

Why Blinding Matters

Suppose a two-arm trial has a planned sample size based on:

where \(\Delta\) is the treatment difference investigators wish to detect.

The investigators may discover during the trial that the variance is larger than expected.

A blinded SSR can respond to the larger variance without asking whether the observed treatment difference is favorable or unfavorable.

This preserves a cleaner separation between:

• the operational decision to correct a sample-size assumption; and
• the inferential decision about treatment efficacy.

Sample Size for a Two-Sample Continuous Endpoint

Consider a parallel-group trial comparing two means.

Assume equal allocation and a common standard deviation \(\sigma\).

For a two-sided test with significance level \(\alpha\) and power \(1-\beta\), a commonly used planning approximation is:

where:

• \(\sigma\) is the assumed common standard deviation;
• \(\Delta\) is the treatment difference to detect;
• \(\alpha\) is the two-sided type I error;
• \(1-\beta\) is the target power.

The total sample size is approximately:

Why the Standard Deviation Matters So Much

Notice that the required sample size depends on:

Therefore, if the true standard deviation is 20% larger than assumed, the variance is approximately:

times as large.

The required sample size may therefore be approximately 44% larger.

A Worked Example: Blinded SSR for a Continuous Endpoint

Suppose a randomized Phase III trial compares two treatments using a continuous primary endpoint.

The investigators plan to detect a treatment difference of:

with an assumed standard deviation of:

The design uses:

During the trial, after approximately half of the planned participants have provided the required endpoint, a blinded estimate of the pooled standard deviation is obtained.

Suppose the observed estimate is:

The original design assumed 10, so the observed variability is substantially larger.

Step 1: Calculate the Original Sample Size

Using:

Using approximately:

gives:

Thus approximately 85 participants per group, or about:

are required before allowing for additional operational adjustments.

Step 2: Recalculate Using the Interim Variance Estimate

Now suppose the blinded estimate is:

Substituting 14 for the original standard deviation:

This gives approximately:

or approximately:

The difference is substantial.

Sample Size Caps Are Important

In practice, investigators may not permit an unlimited increase.

For example, the protocol might specify:

Suppose the maximum permitted sample size is 300.

The calculated requirement of approximately 330 would then be capped:

This creates a transparent operational rule rather than allowing investigators to make an unrestricted decision after seeing interim data.

A Typical Blinded SSR Decision Rule

A protocol might specify:

The exact boundaries are design-specific.

The important point is that the rule should be established before the interim data are examined.

Blinded SSR Does Not Require Conditional Power

A blinded nuisance-parameter re-estimation and a conditional-power adaptation are different concepts.

Conditional Power

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

Let \(D_1\) represent the interim data and \(D_2\) represent future data. Then conditional power can be written conceptually as:

Conditional power is therefore inherently conditional on both the observed interim information and assumptions about what will happen after the interim analysis.

Conditional Power Under the Observed Effect

One common approach uses the observed interim treatment effect as the assumed effect for the remaining patients.

This is sometimes called conditional power under the observed effect.

If the observed treatment effect is smaller than the design effect, conditional power may fall below the original target.

The adaptation might then increase the sample size if conditional power falls within a prespecified intermediate range.

A Generic Conditional-Power Adaptation

A protocol might specify:

This is only an illustrative framework.

Actual clinical trial designs require a prespecified statistical procedure that defines the adaptation and demonstrates control of the relevant error rates.

Why Conditional Power Can Be Misleading

Suppose a trial was designed assuming a treatment effect of:

but the interim estimate is:

If the observed effect is assumed to continue unchanged, the conditional power may be relatively low.

But that does not mean the treatment has no effect.

The interim estimate is noisy, particularly when the information fraction is small.

Therefore, an adaptation rule based too directly on the observed treatment effect can produce unstable decisions.

Information Fraction

Adaptive sample size procedures are often described using the information fraction.

For a simple setting, the information fraction can be represented as:

where:

• \(I_1\) is the information available at the interim analysis;
• \(I_{\mathrm{final}}\) is the planned final information.

For some approximately normally distributed test statistics, information is closely related to effective sample size.

For example, in a simple equal-variance two-group comparison:

so the information fraction may be approximately the proportion of the total sample size already enrolled and evaluable.

Sample Size and Information

The relationship between information and sample size is central to adaptive design.

If the information required to achieve the desired power is underestimated, the trial may need additional participants.

For a simple normal approximation, the standardized treatment effect is:

and information increases as more observations become available.

This is why sample size re-estimation can be viewed as an attempt to correct the trial's planned information level.

Sample Size Re-Estimation in Survival Trials

Survival trials provide an especially important example.

For many proportional-hazards designs, the primary determinant of power is the number of observed events rather than simply the number of randomized patients.

A commonly used approximation for the required number of events in a two-group log-rank design is based on:

where \(D\) is the required number of events and \(p\) is the allocation proportion to one treatment group.

If the trial's event rate is lower than expected, the number of randomized participants required to obtain the target number of events may need to increase.

Event-Driven Sample Size Re-Estimation

Suppose the trial originally plans:

events.

The interim review shows that the event probability is lower than anticipated.

The adaptation might increase the number of randomized participants while retaining the target of 400 events.

The revised randomization target might therefore be:

while the primary analysis remains event-driven.

Sample Size Re-Estimation for Binary Endpoints

For a two-group binary endpoint, sample size depends on the assumed event probabilities.

Let:

and:

If the control response probability is uncertain, a blinded or otherwise appropriately designed adaptation may update the planning calculation.

However, using interim treatment-specific response differences to simply recalculate the sample size can have implications for type I error and must be handled through an appropriate adaptive inferential framework.

Blinded SSR for a Binary Endpoint

Suppose the original calculation assumes:

The investigators may be uncertain about the overall event rate.

A blinded approach can use the pooled event proportion:

without using treatment-specific response rates to determine the adaptation.

The sample-size calculation can then be updated if the nuisance parameter assumption has changed sufficiently.

Prespecified Increase Rules

A strong adaptive design does not simply say: "increase the sample size if the interim results look bad."

Instead, the protocol should define explicit rules.

For example:

Example of a Capped Adaptation Rule

Suppose:

is the original total sample size.

The protocol might specify:

where:

• \(N_{\mathrm{calc}}\) is the recalculated sample size;
• \(N_0\) is the original sample size;
• \(N_{\max}\) is the maximum permitted sample size.

Suppose:

and:

Then:

Why Sample Size Decreases Are More Complicated

It may seem reasonable to decrease the sample size if the interim nuisance parameter estimate is more favorable than expected.

However, reducing the sample size can have operational and statistical consequences.

For example, an adaptation rule that allows both increases and decreases may produce:

• Different final information levels;
• Different stopping probabilities;
• Different conditional error probabilities;
• More complicated operating characteristics.

Therefore, many practical SSR procedures are designed primarily to allow increases when the original sample size appears inadequate.

Type I Error Control

The most important statistical issue in unblinded adaptive sample size re-estimation is preservation of the type I error rate.

Suppose:

versus:

If the sample size is changed using information related to the treatment effect, the final test statistic may no longer have the same distribution as it would under a fixed sample size.

Therefore, the original fixed-design p-value calculation cannot automatically be assumed to retain the desired type I error.

Conditional Error Principle

One important framework for adaptive clinical trials is the conditional error principle.

At the interim analysis, the originally planned design has a conditional probability of ultimately rejecting the null under the null hypothesis.

The adaptation can be constructed so that the revised procedure does not exceed this conditional error.

Conceptually, if the original design has conditional error:

then an adaptive continuation rule can be constructed so that:

This allows substantial flexibility while maintaining control of the overall type I error under the specified assumptions.

Combination-Test Approaches

Another important strategy is to combine evidence from different stages using a prespecified combination function.

Suppose the interim and final stage p-values are:

A combination test constructs a statistic:

whose distribution under the null is known or controlled under the design.

Combination tests can provide flexibility in adaptive designs because the sample size decision can be based on interim information while the final inference is constructed to preserve the desired type I error.

Why the Final Analysis Cannot Simply Ignore the Adaptation

Consider a fixed-design trial with:

If an unblinded interim analysis increases the sample size to:

the final data are no longer generated according to the original fixed sampling plan.

If the adaptation was based on interim treatment information, the final inferential method must account for that adaptation.

This is why adaptive designs should be planned as integrated statistical procedures rather than treating sample size changes as an informal operational decision.

Futility and Sample Size Re-Estimation

Sample size re-estimation can be combined with futility rules.

For example:

This can create an adaptive decision structure with three regions:

Illustrative Conditional-Power Example

Suppose a trial has an original planned sample size of:

At the interim analysis, the observed information fraction is:

Suppose the interim evidence produces a conditional power of 38%.

The protocol might specify:

The new sample size could then be calculated by a prespecified procedure and capped at, for example:

The key point is that the decision rule was established before observing the interim result.

Predictive Power vs. Conditional Power

These terms are sometimes confused.

Conditional Power

Conditional power usually conditions on the observed interim data and assumes a particular value or scenario for the future treatment effect.

Predictive Power

Predictive power integrates over uncertainty about the future treatment effect or other unknown quantities using a specified predictive distribution.

Conceptually:

where the expectation averages conditional power over the posterior or predictive distribution specified by the design.

Sample Size Re-Estimation Workflow

Simulation Is Often Essential

For simple blinded nuisance-parameter re-estimation, analytical calculations may sometimes be sufficient.

For more complicated adaptive designs, simulation is often the most practical way to evaluate operating characteristics.

Simulation can estimate:

• Type I error
• Power
• Probability of sample size increase
• Probability of sample size decrease
• Probability of early stopping
• Expected sample size
• Distribution of the final sample size
• Bias of the treatment-effect estimator
• Coverage of confidence intervals

A Generic Simulation Algorithm

Suppose a trial has a planned sample size \(N_0\), an interim analysis, and a maximum sample size \(N_{\max}\).

A generic simulation proceeds as follows:

R Example: Blinded Variance Re-Estimation

The following example illustrates the basic computational idea for a two-sample continuous endpoint.

alpha <- 0.05
power_target <- 0.90

delta <- 5

sigma_initial <- 10
sigma_interim <- 14

z_alpha <- qnorm(1 - alpha / 2)
z_beta  <- qnorm(power_target)

n_initial_per_group <- ceiling(
  2 * sigma_initial^2 *
  (z_alpha + z_beta)^2 /
  delta^2
)

n_reestimated_per_group <- ceiling(
  2 * sigma_interim^2 *
  (z_alpha + z_beta)^2 /
  delta^2
)

n_initial <- 2 * n_initial_per_group
n_reestimated <- 2 * n_reestimated_per_group

n_initial
n_reestimated

This calculation illustrates the fundamental mechanism:

As the estimated standard deviation increases, the required sample size increases.

Applying a Maximum Sample Size

n_max <- 300

n_final <- min(
  n_max,
  max(n_initial, n_reestimated)
)

n_final

The resulting value is the revised planned sample size after applying the prespecified cap.

A More General Re-Estimation Function

reestimate_n <- function(
  delta,
  sigma,
  alpha = 0.05,
  power = 0.90,
  n_min,
  n_max
) {

  z_alpha <- qnorm(1 - alpha / 2)
  z_beta  <- qnorm(power)

  n_per_group <- ceiling(
    2 * sigma^2 *
    (z_alpha + z_beta)^2 /
    delta^2
  )

  n_total <- 2 * n_per_group

  min(
    n_max,
    max(n_min, n_total)
  )
}

For example:

reestimate_n(
  delta = 5,
  sigma = 14,
  n_min = 170,
  n_max = 300
)

The function applies the basic re-estimation calculation and then imposes the prespecified lower and upper bounds.

Why the Adaptation Must Be Prespecified

Suppose investigators calculate ten different possible sample sizes after seeing the interim results and choose whichever one gives the most attractive final result.

That is fundamentally different from a prespecified adaptive design.

A valid adaptive procedure should specify:

• When the adaptation occurs
• Who performs it
• What data are available
• Whether treatment assignments are visible
• What parameter is estimated
• How the estimate is used
• Minimum sample size
• Maximum sample size
• Exact decision thresholds
• Final inferential method

Operational Firewalls

Unblinded SSR creates an important operational challenge.

If personnel involved in day-to-day trial conduct learn the interim treatment effect, their behavior could potentially change.

For example, knowledge that treatment appears highly effective could influence:

• Enrollment behavior
• Investigator enthusiasm
• Patient recruitment
• Protocol adherence
• Assessment behavior

For this reason, unblinded adaptations are commonly handled by a limited group such as an independent statistical team or an independent data monitoring structure, with appropriate information barriers.

Independent Statistical Review

A robust adaptive-trial process often separates:

The exact organizational structure depends on the design and trial governance plan.

Sample Size Re-Estimation Is Not Sample Size "Peeking"

Another common misconception is that any adaptive sample-size change is statistically invalid.

That is not correct.

Adaptive designs can be statistically valid when:

• The adaptation rule is prespecified;
• The information used for adaptation is appropriately controlled;
• The final inferential procedure accounts for the adaptation when required;
• The operating characteristics are demonstrated.

The problem is not adaptation itself.

The problem is making unplanned data-dependent decisions without establishing their statistical consequences.

Sample Size Re-Estimation and Bias

Changing the sample size does not automatically create bias in the treatment effect estimate.

However, some adaptive procedures can create bias or affect confidence-interval coverage if the adaptation depends strongly on the observed treatment effect and the final analysis does not account for the resulting selection mechanism.

This is particularly important when:

• The treatment effect directly determines the sample size;
• The sample size can change substantially;
• The trial can stop early;
• Multiple adaptations occur;
• The final analysis ignores the adaptation.

Sample Size Re-Estimation vs. Group Sequential Design

These designs are related but not identical.

A trial can combine group sequential monitoring with sample size re-estimation, but the combined operating characteristics must be evaluated as a single design.

Multiple Adaptive Features

Modern adaptive trials may include several adaptations simultaneously.

For example:

• Early efficacy stopping
• Early futility stopping
• Sample size increase
• Treatment-arm dropping
• Allocation changes
• Endpoint selection

When multiple adaptations interact, the design can become substantially more complex.

The probability of each possible path should therefore be evaluated through appropriate analytical methods or simulation.

Common Mistakes

1. Changing sample size informally. An adaptive decision should be defined before the interim results are observed.
2. Confusing blinded and unblinded SSR. A variance estimate from pooled data is fundamentally different from a treatment-effect-based conditional-power calculation.
3. Using the observed treatment effect without accounting for adaptation. This can affect type I error, estimator bias, or confidence-interval coverage.
4. Ignoring sample-size caps. A protocol should generally specify operational limits.
5. Assuming the original p-value remains valid automatically. This is not guaranteed when the sample size depends on unblinded interim treatment information.
6. Failing to simulate complicated designs. Simple formulas may not adequately describe complex adaptive procedures.
7. Ignoring the information fraction. The meaning of an interim estimate depends strongly on how much information has accumulated.
8. Using conditional power without defining the future-effect assumption. Observed-effect, planned-effect, and predictive approaches answer different questions.
9. Allowing unrestricted sample-size increases. Large increases can have substantial operational and statistical consequences.
10. Ignoring trial integrity. Unblinded interim information should be tightly controlled.

What Should Be Specified in the Protocol?

A sample size re-estimation procedure should be described prospectively.

At minimum, specify:

• Original sample size
• Primary endpoint
• Primary estimand
• Type I error
• Target power
• Original nuisance-parameter assumptions
• Interim timing
• Information fraction
• Blinded or unblinded status
• Who will perform the re-estimation
• Data available for the calculation
• Re-estimation formula
• Minimum sample size
• Maximum sample size
• Rounding rules
• Rules for sample-size increases
• Rules for sample-size decreases, if permitted
• Futility rules, if applicable
• Final inferential procedure
• Methods for preserving type I error
• Simulation methodology

Statistical Analysis Plan Considerations

The Statistical Analysis Plan should make the adaptation reproducible.

It should identify:

• The precise data cut used for the interim calculation;
• The analysis population used to estimate the nuisance parameter;
• The exact mathematical formula;
• The rounding convention;
• The sample-size ceiling;
• The sample-size floor;
• The final analysis method;
• The treatment of participants enrolled before and after adaptation.

The SAP should also describe how deviations from the adaptation procedure will be documented.

Regulatory and Governance Considerations

Adaptive sample-size procedures are generally most defensible when the statistical rationale is clear before trial initiation.

The design documentation should explain:

• Why the adaptation is scientifically necessary;
• Why the selected interim timing is appropriate;
• Which information is used;
• How trial integrity is protected;
• How type I error is controlled;
• How the final analysis accounts for the adaptation;
• How operating characteristics were evaluated.

The more complex the adaptation, the more important it becomes to document the complete statistical mechanism rather than simply describing the trial as "adaptive."

Design Operating Characteristics

A sample size re-estimation design should be evaluated over a range of possible true parameter values.

Important quantities include:

Expected Sample Size Under the Null

Suppose the original sample size is \(N_0\), the revised sample size is \(N_1\), and the probability of triggering an increase under the null is \(q_0\).

A simple expected-sample-size expression is:

If:

then:

The expected sample size is therefore 420 even though the maximum is 500.

Why the Probability of Increase Matters

A design with a very large maximum sample size may still have a reasonable expected sample size if the increase occurs rarely.

Conversely, a small maximum increase may materially affect expected enrollment if it occurs frequently.

Therefore, both quantities should be reported.

Adaptation Regions

A useful way to visualize an SSR procedure is through adaptation regions.

The precise regions should be defined using statistical criteria rather than post hoc judgment.

Sample Size Re-Estimation Does Not Guarantee Higher Power

Increasing sample size does not automatically guarantee that the final trial will have the target power under every possible interim scenario.

For example, if the interim treatment effect is extremely unfavorable, increasing \(N\) may have little practical value.

This is one reason why some designs combine sample-size adaptation with futility rules.

The Difference Between Precision and Power

Sample size can be increased for different reasons.

Power Motivation

The goal is to increase the probability of rejecting the null when a specified effect exists.

Precision Motivation

The goal is to reduce the width of the confidence interval around the treatment effect.

These are related but not identical objectives.

An adaptive trial should clearly identify which objective motivates the sample size re-estimation.

Sample Size Re-Estimation for Precision

Suppose the goal is to estimate a treatment difference with a desired confidence-interval half-width \(d\).

For a simple mean:

If the interim estimate of \(\sigma\) is larger than expected, the sample size required to achieve the desired precision also increases.

This type of adaptation is conceptually different from increasing sample size because the observed treatment effect is disappointing.

Sample Size Re-Estimation in Cluster Trials

Cluster randomized trials provide another important application.

The required sample size depends on the intracluster correlation coefficient (ICC), often denoted:

A common design-effect approximation is:

where \(m\) is the average cluster size.

If the true ICC is larger than assumed, the effective sample size is reduced.

An adaptive design may therefore consider whether additional clusters are needed.

Sample Size Re-Estimation in Repeated-Measures Trials

Longitudinal trials may depend on:

• Within-subject correlation;
• Residual variance;
• Correlation across visits;
• Dropout;
• Covariance structure.

Interim information can therefore be used to refine the planning assumptions.

However, the sample-size calculation should be consistent with the planned mixed-model or repeated-measures analysis.

Dropout and Sample Size Re-Estimation

Another common reason for adaptation is uncertainty about dropout.

Suppose the required number of evaluable participants is:

and the expected retention rate is:

Then the enrollment target can be approximated by:

If interim retention is lower than expected, the number randomized may need to increase.

This is usually an operational adjustment rather than an efficacy-based adaptation, but it should still be prespecified when possible.

Blinded Dropout Re-Estimation

Suppose a trial needs 500 evaluable participants.

The original planning assumption is 90% retention:

Then:

If interim data suggest retention is only 80%:

A prespecified cap might then determine the maximum permitted enrollment.

Adaptive Sample Size and Multiplicity

If a trial includes multiple hypotheses, arms, or endpoints, sample size adaptation may interact with multiplicity control.

For example, consider a trial with:

• Two doses;
• One control;
• Two primary comparisons;
• An interim sample-size adaptation.

The adaptation cannot be evaluated independently of the multiplicity procedure.

The full design should preserve the desired family-wise error rate or other prespecified error criterion.

Sample Size Re-Estimation and Estimands

The adaptation should also be consistent with the trial's primary estimand.

For example, if the primary estimand is based on a treatment-policy strategy for intercurrent events, the nuisance parameter used in sample-size re-estimation should be relevant to the analysis actually planned.

Changing the analysis estimand after the interim adaptation can create a different statistical problem.

Common Misinterpretations

Several statements are frequently misleading.

A Practical Adaptive Trial Workflow

What Should Be Simulated?

A thorough simulation study should examine the design under many scenarios.

For example:

Final Worked Example Summary

The continuous-endpoint example can be summarized as follows.

The Most Important Concept

The most important conceptual point is that sample size re-estimation is an adaptation rule, not simply a new sample-size calculation performed halfway through a trial.

A valid procedure connects four components:

• The original design assumptions;
• The interim information;
• The prespecified adaptation rule;
• The final inferential procedure.

When the problem is uncertainty about a nuisance parameter such as variance, event rate, ICC, or dropout, a blinded re-estimation can often provide a relatively clean solution.

When the adaptation uses treatment-specific interim evidence, the statistical problem becomes more complicated.

Conditional power, predictive power, conditional error, combination tests, and other adaptive-inference methods may then be appropriate depending on the design.

References

Bauer, P. & Köhne, K. (1994). Evaluation of experiments with adaptive interim analyses. Biometrics, 50, 1029–1041.
Cui, L., Hung, H.M.J. & Wang, S.J. (1999). Modification of sample size in group sequential clinical trials. Biometrics, 55, 853–857.
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.
Proschan, M.A. & Hunsberger, S.A. (1995). Designed extension of studies based on conditional power. Biometrics, 51, 1315–1324.
Jennison, C. & Turnbull, B.W. (2000). Group Sequential Methods with Applications to Clinical Trials. Chapman & Hall/CRC.
Chow, S.C. & Chang, M. (2011). Adaptive Design Methods in Clinical Trials. Chapman & Hall/CRC.
ICH E9 (1998). Statistical Principles for Clinical Trials.
ICH E9(R1) (2019). Addendum on Estimands and Sensitivity Analysis in Clinical Trials.