Sample Size Re-Estimation Methods: A Practical Guide

may define the region in which the sample size is increased.

The exact boundaries depend on the trial design.

The conceptual structure is:

Why the Promising Zone Is Interesting

Suppose the interim data are extremely unfavorable.

Even a large sample size increase may not provide a reasonable probability of success.

Conversely, if the interim data are extremely favorable, the original sample size may already have sufficient conditional power.

The promising zone therefore focuses additional enrollment where it can be most useful.

Sample Size Re-Estimation and Futility

Sample size re-estimation and futility stopping can be combined.

For example:

However, these decisions should not be improvised at the interim analysis.

The stopping and adaptation boundaries should be established before the trial begins.

Sample Size Re-Estimation and Efficacy Stopping

The same principle applies to early efficacy stopping.

If the interim evidence is overwhelmingly favorable, an adaptive design might permit early rejection of the null.

This transforms the study into a broader adaptive or group-sequential design rather than a simple sample-size adjustment.

The statistical framework must then account for both:

• the possibility of stopping early for efficacy, and
• the possibility of changing the sample size.

Sample Size Re-Estimation Is Not a Substitute for Good Planning

A common misconception is that sample size re-estimation eliminates the need for careful planning assumptions.

It does not.

The initial design still determines:

• The original maximum sample size
• The interim timing
• The target effect
• The desired power
• The adaptation boundaries
• The statistical analysis

A poorly designed adaptive rule can create substantial operational and statistical problems.

When Sample Size Re-Estimation Is Most Useful

SSR is particularly useful when a key nuisance parameter is uncertain and can be estimated reliably during the trial.

Examples include:

• Continuous endpoints with uncertain variance
• Cluster-randomized trials with uncertain intraclass correlation
• Repeated-measures studies with uncertain covariance parameters
• Survival studies with uncertain event rates
• Binary endpoints with uncertain control-group event rates
• Trials with uncertain dropout or evaluability rates

Cluster-Randomized Trials

Cluster trials introduce another important nuisance parameter: the intraclass correlation coefficient, or ICC.

The design effect is often approximated by:

where:

• \(m\) = average cluster size
• \(\rho\) = intraclass correlation coefficient

If the planning ICC is too small, the original sample size may be insufficient.

An internal pilot can provide information about the actual ICC and permit a prespecified sample size adjustment.

Repeated-Measures Trials

Repeated-measures trials introduce uncertainty in within-patient correlations and covariance structures.

Suppose the variance of an estimated treatment effect depends on a correlation parameter:

where \(m\) is the number of repeated measurements.

If the actual correlation differs from the planning value, the effective information may differ from expectations.

A blinded internal pilot can sometimes be used to update the nuisance parameters and revise the sample size.

Sample Size Re-Estimation for a Single Proportion

Suppose the goal is to estimate a binary response probability with a desired margin of error.

A simple approximate sample-size formula is:

where \(E\) is the desired half-width of the confidence interval.

If the planning value of \(p\) is uncertain, an interim estimate can be used to update the sample size.

Again, whether such an adaptation affects the inferential properties depends on the exact procedure and the analysis being performed.

Bayesian Sample Size Re-Estimation

Sample size adaptation can also be incorporated into Bayesian trial designs.

Rather than focusing on frequentist type I error and power, a Bayesian design may use quantities such as:

• Posterior probability of treatment benefit
• Posterior probability that an effect exceeds a clinically meaningful threshold
• Predictive probability of trial success
• Posterior precision

For example:

may be used to characterize the evidence after interim data \(D\).

The same conceptual distinction remains important: a Bayesian adaptation should be specified prospectively, including the decision criterion and its operational consequences.

Predictive Probability vs. Conditional Power

Conditional power and predictive probability are related but conceptually different.

Conditional power usually conditions on an assumed future effect.

Predictive probability integrates over uncertainty about the future effect.

In Bayesian notation:

where the uncertainty in the future treatment effect is integrated according to the specified predictive distribution.

R Implementation: Blinded Variance Re-Estimation

A simple illustration of a variance-driven sample-size calculation can be implemented in R.

alpha <- 0.05
power <- 0.90
delta <- 5

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

sample_size_continuous <- function(sigma) {

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

  ceiling(n)

}

sample_size_continuous(10)
sample_size_continuous(13)

The function recalculates the total sample size for a new standard deviation.

Adding Sample Size Bounds

min_n <- 80
max_n <- 300

bounded_sample_size <- function(sigma) {

  n <- sample_size_continuous(sigma)

  max(min_n, min(n, max_n))

}

bounded_sample_size(10)
bounded_sample_size(13)
bounded_sample_size(20)

This illustrates the common principle of imposing prespecified lower and upper bounds.

R Implementation: Event-Driven Re-Estimation

Suppose a trial requires a fixed number of events.

required_events <- 300

enrollment_for_events <- function(event_rate) {

  ceiling(required_events / event_rate)

}

enrollment_for_events(0.60)
enrollment_for_events(0.50)
enrollment_for_events(0.40)

The results illustrate how lower event rates require more enrolled patients to generate the same number of events.

R Implementation: A Simple Conditional-Power Calculation

For an illustrative one-sided normal framework, suppose the interim statistic, information, and assumed standardized effect are known.

conditional_power <- function(
  z1,
  I1,
  I2,
  theta,
  alpha = 0.025
) {

  zcrit <- qnorm(1 - alpha)

  numerator <-
    z1 * sqrt(I1) +
    theta * I2 -
    zcrit * sqrt(I1 + I2)

  numerator / sqrt(I2)

  pnorm(numerator / sqrt(I2))

}

conditional_power(
  z1 = 1.5,
  I1 = 50,
  I2 = 100,
  theta = 0.20
)

This is an illustrative implementation of one normal-theory formulation. Actual clinical-trial calculations should use the test statistic, information structure, variance estimator, and analysis model specified for the trial.

A Simple Promising-Zone Rule in R

A prespecified adaptation can be represented computationally as a decision function.

choose_sample_size <- function(
  conditional_power,
  n_original,
  n_increased,
  cp_lower = 0.20,
  cp_upper = 0.80
) {

  if (conditional_power < cp_lower) {

    return(n_original)

  }

  if (conditional_power < cp_upper) {

    return(n_increased)

  }

  n_original

}

choose_sample_size(
  conditional_power = 0.15,
  n_original = 300,
  n_increased = 420
)

choose_sample_size(
  conditional_power = 0.50,
  n_original = 300,
  n_increased = 420
)

choose_sample_size(
  conditional_power = 0.90,
  n_original = 300,
  n_increased = 420
)

Again, the numerical thresholds are illustrative rather than universal recommendations.

What Makes an SSR Procedure Valid?

A valid sample size re-estimation procedure should have a clearly defined statistical and operational structure.

At minimum, investigators should specify:

• The timing of the interim analysis
• The information available at the interim analysis
• Whether the treatment effect is blinded
• The nuisance parameters to be estimated
• The sample-size calculation used for re-estimation
• The minimum sample size
• The maximum sample size
• The rounding rule
• The adaptation boundaries
• The final analysis method
• The type I error-control strategy, when relevant
• The handling of missing data
• The operational decision-maker

Who Should See Unblinded SSR Results?

If the adaptation uses unblinded treatment-effect information, operational procedures should prevent inappropriate dissemination of interim results.

Depending on the trial structure, the information may be handled by an independent statistical group, Data Monitoring Committee, or another prespecified independent function.

The investigators responsible for routine trial conduct may remain blinded to the treatment-effect information.

Sample Size Re-Estimation and the Data Monitoring Committee

In trials with unblinded adaptive decisions, a Data Monitoring Committee may have a role in reviewing interim information.

The committee's responsibilities depend on the charter and protocol.

Possible responsibilities include:

• Reviewing unblinded safety data
• Reviewing interim efficacy information
• Confirming whether prespecified adaptation criteria have been met
• Recommending continuation, modification, or stopping

The exact governance structure should be defined prospectively.

SSR Does Not Mean "Peek and Decide"

A trial should not simply inspect the interim treatment effect and ask:

"Do we need more patients?"

That approach can produce uncontrolled selection effects.

Instead, the design should specify a mathematical decision rule such as:

where \(D_1\) represents the prespecified interim information.

The function \(g(\cdot)\) is established before the interim data are observed.

Common Sample Size Re-Estimation Rules

Advantages of Sample Size Re-Estimation

• Protects against important planning-assumption errors.
• Can improve the probability that the intended information level is achieved.
• Can use information from the actual trial population.
• Can be particularly valuable when nuisance parameters are uncertain.
• Can provide flexibility while preserving a prospectively defined design.
• May prevent unnecessarily large enrollment when the original assumptions were conservative.

Potential Disadvantages

• Additional statistical complexity.
• Additional operational complexity.
• Potential inflation of maximum sample size.
• Possible extension of trial duration.
• Need for independent handling of unblinded interim information.
• Potential regulatory scrutiny of adaptive procedures.
• Risk of invalid inference if the adaptation is not incorporated into the final analysis.

Common Mistake: Treating Blinded and Unblinded SSR as Equivalent

They are not equivalent.

A blinded variance re-estimation based on pooled data answers:

"Was our assumption about the nuisance parameter reasonable?"

An unblinded conditional-power procedure answers something closer to:

"Given what we have observed about the treatment effect, how likely is the trial to succeed under specified future assumptions?"

These are fundamentally different questions.

Common Mistake: Increasing N Because the Effect Looks Smaller

Suppose the trial was designed around:

but the interim estimate is:

It may be tempting to simply plug 6 into a conventional sample-size formula and increase the trial.

That is not sufficient.

Because the observed effect was used to select the new sample size, the relationship between the interim statistic and final statistic must be accounted for.

An appropriate adaptive inference procedure is required.

Common Mistake: Assuming the Original P-Value Remains Valid

Suppose:

• Stage 1 produces a treatment-effect estimate.
• The estimate determines the new sample size.
• The trial then combines all patients and performs an ordinary final t-test.

The resulting p-value is not automatically guaranteed to have the same distribution as it would under a fixed design.

The sample size is now a function of interim data.

That dependence must be incorporated into the inferential procedure.

Common Mistake: Forgetting the Maximum Sample Size

A sample-size adaptation rule without a maximum sample-size cap can produce unexpected operational consequences.

For example, if a variance estimate becomes unusually large, a formula might produce:

when the original study was intended to enroll only 300 patients.

A protocol should define an upper bound when appropriate.

Common Mistake: Re-Estimating Too Early

An interim sample must contain enough information to estimate the relevant nuisance parameter reasonably well.

If the interim sample is extremely small, the estimate may be unstable.

For example, a standard deviation estimated from only a handful of patients may be substantially more variable than the original external estimate.

Therefore, the timing of SSR is itself a design parameter.

Common Mistake: Re-Estimating Too Late

The opposite problem can also occur.

If 95% of the trial has already been enrolled before the adaptation, there may be little opportunity to change the total sample size.

An SSR analysis should therefore occur early enough for the adaptation to have a meaningful operational effect.

Choosing the Timing of the Interim Analysis

The interim analysis should balance:

• Precision of the nuisance-parameter estimate
• Amount of information available about the treatment effect
• Time required to implement the adaptation
• Potential operational disruption
• Remaining enrollment

A very early analysis may be unstable.

A very late analysis may be operationally useless.

Sample Size Re-Estimation Workflow

Simulation Is Often Essential

For simple blinded variance re-estimation, analytic calculations may be sufficient.

For complex unblinded adaptive procedures, simulation is often the most practical way to evaluate operating characteristics.

A simulation can evaluate:

• Type I error
• Power
• Average sample size
• Probability of sample size increase
• Probability of stopping for futility
• Probability of stopping for efficacy
• Distribution of the final sample size
• Bias of treatment-effect estimates
• Confidence-interval coverage

Simulation Logic

A generic simulation proceeds as follows:

Estimating Type I Error by Simulation

Suppose 100,000 trials are simulated under the null hypothesis.

If 4,870 reject the null:

The Monte Carlo standard error is approximately:

For 100,000 simulations, this is small enough to estimate the type I error reasonably precisely.

R Simulation Skeleton

set.seed(12345)

nsim <- 100000

reject <- logical(nsim)
final_n <- numeric(nsim)

for (i in seq_len(nsim)) {

  # Generate Stage 1 data
  # Perform interim analysis
  # Determine revised sample size
  # Generate remaining data
  # Perform final analysis
  # Store rejection decision
  # Store final sample size

}

mean(reject)
mean(final_n)

The exact simulation code depends on the endpoint and adaptive procedure.

Operating Characteristics to Examine

A sample-size re-estimation procedure should be evaluated across a range of true parameter values rather than only at the design point.

Bias After Adaptive Sample Size Selection

An important issue is that adaptive sampling can affect the distribution of estimators.

Suppose a trial increases its sample size when the observed interim effect is small.

The trials receiving more patients are therefore systematically different from those that do not.

The final treatment-effect estimate may have properties different from those of a conventional fixed-sample estimator.

This does not mean adaptive designs are invalid.

It means that estimation, confidence intervals, and hypothesis testing should be evaluated under the specific adaptive design.

Confidence Intervals After SSR

A particularly important question is:

Can the ordinary confidence interval be used after sample size re-estimation?

The answer depends on the type of re-estimation.

For some blinded nuisance-parameter adaptations, the conventional final analysis can retain its usual properties under appropriate conditions.

For unblinded adaptations based on treatment-effect information, special methods may be required.

Sample Size Re-Estimation and Estimands

SSR changes the amount of information collected.

It should not silently change the primary estimand.

The protocol should continue to define the same primary scientific question, such as:

• Difference in mean change from baseline
• Odds ratio
• Hazard ratio
• Risk difference
• Response probability

The adaptive procedure determines how much information is collected to address that question.

Regulatory and Protocol Considerations

For a confirmatory clinical trial, sample size re-estimation should be prospectively described in the protocol and statistical analysis plan.

The documentation should explain:

• Why re-estimation is needed
• What information is used
• When it is evaluated
• Who performs the calculation
• Who receives the result
• What sample-size values are possible
• What happens at each boundary
• How the final analysis preserves the intended inferential properties

What Should Be Prespecified?

Blinded SSR: A Useful Mental Model

Think of blinded SSR as answering:

The treatment comparison remains hidden.

The adaptation corrects an assumption about the amount of statistical information available per patient.

Unblinded SSR: A Useful Mental Model

Think of unblinded SSR as answering:

This can be more powerful, but it requires a more sophisticated inferential framework.

A Practical Decision Tree

Sample Size Re-Estimation vs. Group-Sequential Design

These concepts are related but not identical.

The two approaches can also be combined.

Adaptive Designs Can Combine Multiple Features

A sophisticated clinical trial might allow:

• Early efficacy stopping
• Early futility stopping
• Sample size increase
• Arm dropping
• Sample-size reallocation
• Event-driven follow-up changes

The more adaptations are included, the more important it becomes to define the joint decision framework before the trial begins.

Illustrative Complete Design

Consider a two-arm trial with:

Suppose the interim pooled standard deviation is estimated as 13.

The revised sample size is calculated from the same original design equation.

The adaptation therefore changes the amount of enrollment without changing the scientific target.

What Happens If the Interim Variance Is Smaller?

Suppose instead:

The recalculated sample size would be smaller than the original requirement.

Whether the study is actually allowed to reduce enrollment should be prespecified.

Some protocols permit both increases and decreases.

Others only permit increases.

Increasing vs. Decreasing Sample Size

Why an Increase-Only Rule May Be Operationally Attractive

Suppose the original design was carefully powered and the interim estimate suggests a smaller sample size.

Reducing the trial could save enrollment, but it can also create practical complexity.

An increase-only rule can instead preserve the original sample size as a minimum while allowing the trial to compensate for unfavorable nuisance parameter estimates.

The choice depends on the objectives of the trial and the statistical framework.

Sample Size Re-Estimation and Multiplicity

SSR is not inherently a multiplicity problem.

The issue arises because an adaptive procedure can alter the distribution of the final test statistic.

If multiple hypotheses, endpoints, treatment arms, or interim decisions are also involved, then multiplicity and adaptive inference may interact.

The complete design should therefore consider:

• Number of hypotheses
• Number of interim looks
• Adaptation rules
• Multiplicity adjustment
• Final test statistic

Sample Size Re-Estimation and Missing Data

Missing data can complicate re-estimation because the information available at the interim analysis may not represent the final evaluable population.

For example, if the internal pilot includes many participants with incomplete outcomes, the nuisance-parameter estimate may be unstable.

The protocol should therefore specify:

• Which participants contribute to the interim estimate
• How missing measurements are handled
• Whether incomplete participants are replaced
• Whether the final sample size is based on enrolled or evaluable participants

Simulation of a Variance Re-Estimation Procedure

A simple simulation can demonstrate why an internal pilot may be useful.

set.seed(2026)

nsim <- 10000

true_sigma <- 13
pilot_n <- 40

sigma_hat <- numeric(nsim)

for (i in seq_len(nsim)) {

  x <- rnorm(
    pilot_n,
    mean = 0,
    sd = true_sigma
  )

  sigma_hat[i] <- sd(x)

}

summary(sigma_hat)

The simulation demonstrates that the interim estimate itself has sampling variability.

Therefore, the re-estimated sample size is also a random quantity.

The Re-Estimated Sample Size Is a Random Variable

This is an important conceptual point.

In a fixed design:

In an adaptive design:

where \(D_1\) represents interim information.

Therefore:

for all possible sample-size outcomes.

Distribution of the Final Sample Size

A useful operating characteristic is:

for each possible final sample size.

For example:

The expected sample size would then be:

These numbers are illustrative.

Expected Sample Size Under Different True Effects

For an unblinded SSR procedure, expected sample size can depend strongly on the true treatment effect.

For example:

This illustrates why the operating characteristics should be evaluated over a range of plausible effects.

SSR Does Not Guarantee the Target Power

Suppose a trial was designed for 90% power at a treatment effect of:

If the true effect is actually:

a sample size increase may improve power, but it cannot make a very small treatment effect automatically become clinically important.

Power is always conditional on the true effect.

SSR and the Minimum Clinically Important Difference

The clinically meaningful effect should remain distinct from the statistical sample-size adaptation.

Suppose:

If the observed effect is 3, investigators should not automatically redefine 3 as the new clinically important effect simply because it is what the trial happened to observe.

Clinical interpretation and sample-size adaptation are separate decisions.

Protocol Language Should Be Precise

A weak protocol statement might say:

"The sample size may be increased if necessary based on interim data."

That is insufficiently specific for a rigorous adaptive design.

A stronger specification defines:

• Exactly which data are examined
• Exactly when they are examined
• Exactly how the relevant quantity is calculated
• Exactly what sample sizes are allowed
• Exactly who makes the decision
• Exactly how the final analysis is performed

Example Protocol Structure

The exact language should be tailored to the actual statistical design.

What a Statistical Analysis Plan Should Contain

The SAP should specify:

• Primary analysis model
• Original sample-size calculation
• Interim timing
• Re-estimation parameter
• Blinding status
• Re-estimation equation
• Bounds
• Rounding
• Allocation rules
• Stopping rules
• Final inferential method
• Handling of deviations and missing data
• Simulation or analytic validation

Final Worked Example: Putting Everything Together

Consider a randomized clinical trial with a continuous primary endpoint.

The investigators assume:

The original calculation produces a total sample size of approximately 86.

An internal pilot of 40 patients is included.

The blinded pooled standard deviation is then estimated.

Three possible outcomes illustrate the adaptation:

The actual protocol would apply the prespecified minimum, maximum, and rounding rules to these calculations.

The Core Mathematical Structure

Most sample-size re-estimation procedures can be represented abstractly as:

where:

• \(D_1\) = interim data
• \(\theta\) = prespecified design quantities
• \(g(\cdot)\) = the re-estimation rule

For blinded SSR:

where \(\widehat{\eta}_{\text{blind}}\) is a nuisance-parameter estimate.

For unblinded SSR:

where the adaptation may use interim treatment-effect information.

The Most Important Statistical Distinction

The central distinction is not simply between "small" and "large" sample-size changes.

It is between:

The first can often be handled using a blinded approach.

The second requires careful adaptive-inference methodology.

Common Mistakes

1. Calling every sample-size adjustment "blinded SSR." A treatment-effect-based adjustment is fundamentally different.
2. Using the interim treatment effect without an adaptive inference framework. The final p-value may no longer have the intended null distribution.
3. Failing to prespecify the adaptation. The mathematical rule should be established before interim data are observed.
4. Ignoring the distinction between total and remaining sample size. Always define exactly what the calculated number represents.
5. Ignoring event counts in survival studies. Enrollment and information are not necessarily equivalent.
6. Allowing unlimited sample-size increases. A maximum sample size should generally be considered and prespecified where appropriate.
7. Re-estimating too early. An unstable interim nuisance-parameter estimate can produce unstable adaptations.
8. Re-estimating too late. There may not be enough enrollment remaining to implement the change.
9. Assuming a smaller observed effect automatically justifies a larger N. The adaptation must follow the prespecified statistical rule.
10. Ignoring confidence intervals and estimation. Adaptive sampling can affect more than the hypothesis-test p-value.
11. Failing to simulate a complex design. Analytic intuition is often insufficient for complicated adaptive procedures.
12. Changing the clinical target after observing interim data. The clinical estimand and meaningful effect should remain distinct from the adaptive rule.

Practical Sample Size Re-Estimation Workflow

What Should Be Reported When the Trial Is Completed?

The final clinical study report should make it possible to understand how the adaptive procedure operated.

Depending on the design, useful information includes:

• Original planned sample size
• Interim analysis timing
• Whether SSR was triggered
• Re-estimated sample size
• Final enrolled sample size
• Number of evaluable patients
• Number of events, if applicable
• Primary analysis method
• Handling of the adaptation in the final inference

Summary Table

The Most Important Concept

The most important idea in sample size re-estimation is that the sample size is a function of information learned during the trial.

In a conventional design:

In an adaptive design:

The statistical challenge is therefore not merely calculating a new number of patients.

The challenge is ensuring that the entire procedure remains scientifically interpretable and statistically valid after the sample size becomes dependent on interim information.

For blinded nuisance-parameter re-estimation, the primary goal is often to correct uncertainty in quantities such as variance or event rate.

For unblinded re-estimation, the primary goal may be to respond to evidence about the treatment effect, conditional power, or predictive probability.

These approaches require different statistical considerations.

References

Gould, A.L. (1992). Interim analyses for monitoring clinical trials that do not materially affect the type I error rate. Statistics in Medicine.
Wittes, J. & Brittain, E. (1990). The role of internal pilot studies in increasing the efficiency of clinical trials. Statistics in Medicine.
Birkett, M.A. & Day, S.J. (1994). Internal pilot studies for estimating sample size. Statistics in Medicine.
Proschan, M.A. & Hunsberger, S.A. (1995). Designed extension of studies based on conditional power. Biometrics.
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.
Jennison, C. & Turnbull, B.W. (2000). Group Sequential Methods with Applications to Clinical Trials. Chapman & Hall/CRC.
Chow, S.-C. & Chang, M. (2008). Adaptive Design Methods in Clinical Trials. Chapman & Hall/CRC.
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.
Gao, P., Ware, J.H. & Mehta, C. (2008). Sample size re-estimation for adaptive sequential design in clinical trials. Journal of Biopharmaceutical Statistics.
Bretz, F., Koenig, F., Brannath, W., Glimm, E. & Posch, M. (2009). Adaptive designs for confirmatory clinical trials. Statistics in Medicine, 28, 1189–1222.