Sample Size Re-Estimation For Noninferiority Trials

Why Re-Estimate Sample Size in a Noninferiority Trial?

A noninferiority trial asks whether a new treatment’s effect is not unacceptably worse than that of an active comparator. Its design depends on a clinically justified noninferiority margin and an assumed treatment effect, but also on nuisance parameters such as the outcome variance or event rate. If interim data show that a nuisance parameter differs from its planning value, the original sample size may no longer deliver the intended power.

This tutorial focuses on updating sample size from nuisance-parameter estimates while keeping the noninferiority margin, hypotheses, and target treatment effect fixed. The update is not a chance to redefine what counts as clinically acceptable or to select a more favorable treatment effect after seeing interim outcomes. Those choices must remain part of the prespecified design.

Definition: Let \(\theta\) denote the treatment effect, oriented so that larger values favor the experimental treatment. Let \(\Delta>0\) be the prespecified noninferiority margin. The noninferiority hypotheses are \(H_0:\theta\leq-\Delta\) and \(H_a:\theta>-\Delta\). A noninferiority conclusion requires rejecting \(H_0\); failing to reject it does not establish noninferiority.

For example, if \(\theta\) is the experimental-minus-control difference in mean symptom improvement, measured in points, then larger differences favor the experimental treatment. A margin of \(\Delta=5\) points means that differences at or below \(-5\) points are considered unacceptably worse under the specified criterion. The margin is a clinical design choice, not a parameter to estimate from the interim data.

Continuous Outcomes: Deriving the Update

Suppose the endpoint is continuous and the design compares two independent groups with equal allocation. Assume normally distributed outcomes with a common variance \(\sigma^2\), independent observations, and a large enough sample for the planned one-sided normal test to be appropriate. Write \(\bar{X}_T-\bar{X}_C\) for the estimated treatment effect. For \(n\) participants per group, its standard error is \(\sigma\sqrt{2/n}\).

Let \(\theta_A\) be the treatment effect assumed for power calculations. The distance between that assumed effect and the null boundary is \(d=\theta_A+\Delta\), which must be positive for the planned alternative to lie above the noninferiority boundary. For one-sided significance level \(\alpha\) and target power \(1-\beta\), the normal-approximation sample size is:

n=\frac{2\sigma^2\left(z_{1-\alpha}+z_{1-\beta}\right)^2}{(\theta_A+\Delta)^2}

Here \(z_q\) is the \(q\)-quantile of the standard normal distribution. The result is the required number per group before rounding. If equal allocation is used, round up to a whole participant per group.

Assumptions for this calculation: the endpoint variance is common across groups; the treatment groups are independent; the normal approximation is adequate; allocation is equal; \(\Delta\), \(\alpha\), target power, and \(\theta_A\) are prespecified; and the stated variance is appropriate for the planned analysis. If these conditions do not fit the trial, this formula may not provide the intended power.

The formula follows from the rejection rule for a one-sided normal test. Under \(H_0\), the test rejects for a standardized estimate above \(z_{1-\alpha}\). Under the planning alternative, the standardized estimate has mean \(d/SE\) and variance one. To obtain power \(1-\beta\), its mean must exceed the critical value by \(z_{1-\beta}\); rearranging gives the sample-size expression.

P(\text{reject }H_0\mid\theta=\theta_A) \approx 1-\Phi\left(z_{1-\alpha}-\frac{d}{\sigma\sqrt{2/n}}\right) $$ $$ 1-\Phi\left(z_{1-\alpha}-\frac{d}{\sigma\sqrt{2/n}}\right)\geq 1-\beta \quad\Longleftrightarrow\quad \frac{d}{\sigma\sqrt{2/n}}\geq z_{1-\alpha}+z_{1-\beta} $$ $$ n\geq\frac{2\sigma^2(z_{1-\alpha}+z_{1-\beta})^2}{d^2}

The derivation also explains the role of nuisance re-estimation: with the margin and target effect held fixed, a revised estimate of \(\sigma^2\) changes the required \(n\) in direct proportion to that variance estimate. It does not change the distance \(d\).

Worked Example: Updating for a Larger Continuous-Outcome Variance

Consider a hypothetical trial with equal allocation, a continuous endpoint measured in points, \(\Delta=5\) points, assumed effect \(\theta_A=0\) points, one-sided \(\alpha=0.025\), and target power \(90\%\). The initial planning standard deviation is \(10\) points. Use \(z_{1-\alpha}=1.960\) and \(z_{1-\beta}=1.282\); use those displayed values consistently in the arithmetic.

The planning distance is \(d=0+5=5\) points, and the squared quantile sum is \((1.960+1.282)^2=3.242^2=10.510564\). Thus:

n=\frac{2(10^2)(10.510564)}{5^2} =\frac{200(10.510564)}{25} =84.0845

The initial design therefore requires 85 participants per group, or 170 total. Suppose a prespecified interim nuisance-parameter assessment estimates the standard deviation as \(12\) points. Holding all the design quantities other than the variance estimate fixed gives:

n_{\mathrm{updated}} =\frac{2(12^2)(10.510564)}{5^2} =\frac{288(10.510564)}{25} =121.0817

Rounding up gives 122 participants per group, or 244 total. If 40 participants per group have already been enrolled and remain in the final analysis, the arithmetic implies 82 additional participants per group. Whether this increase is permitted, and whether the final test remains calibrated after the adaptation, must be addressed in the prespecified design.

Binary Outcomes: Updating Event-Rate Nuisance Parameters

For a binary endpoint, let \(\theta=p_T-p_C\), the experimental-minus-control difference in success probabilities, so that a larger value favors the experimental treatment. Under equal allocation and a normal approximation for the difference in sample proportions, the variance is approximately \(\{p_T(1-p_T)+p_C(1-p_C)\}/n\), where \(n\) is the number per group. The corresponding planning formula is:

n= \frac{(z_{1-\alpha}+z_{1-\beta})^2 \left[p_T(1-p_T)+p_C(1-p_C)\right]} {(\theta_A+\Delta)^2}

This is a planning approximation for a specified analysis, not a claim that every final noninferiority test uses this exact variance estimator. It assumes independent participants, a sufficiently accurate normal approximation, and nuisance rates representative of the planned population. In practice, investigators should use a sample-size method aligned with the prespecified final test, including any continuity correction or alternative variance convention if required by that method.

Worked Example: Updating a Binary-Outcome Design

In a hypothetical trial, success is the favorable outcome and \(\theta=p_T-p_C\). Set \(\Delta=0.10\), \(\theta_A=0\), one-sided \(\alpha=0.025\), and target power \(90\%\). Assume both success probabilities are initially \(0.70\). Use the displayed quantiles from the preceding example, so the squared quantile sum is \(10.510564\).

The assumed planning distance is \(0+0.10=0.10\), and the variance contribution per participant per group-size unit is \(0.70(0.30)+0.70(0.30)=0.42\). Therefore:

n=\frac{10.510564(0.42)}{(0.10)^2} =\frac{4.4144369}{0.01} =441.4437

The initial design rounds up to 442 participants per group. Suppose the prespecified nuisance-rate update uses an appropriate blinded pooled estimate and gives a common success-rate estimate of \(0.60\). Keeping the treatment-effect assumption and margin fixed, the updated variance contribution is \(2(0.60)(0.40)=0.48\), so:

n_{\mathrm{updated}}= \frac{10.510564(0.48)}{0.01} =\frac{5.0450707}{0.01} =504.5071

The updated requirement is 505 participants per group. The estimate \(0.60\) changes the variance term, not the assumed treatment difference or the margin. The particular pooled estimator must be specified carefully: its suitability depends on the design and final analysis, and data-dependent adaptation does not automatically guarantee a valid final test.

What Must Stay Fixed, and What Can Be Updated?

A nuisance-parameter update changes quantities that affect precision or the variance of the estimator, such as a continuous-outcome variance or binary event rate. The noninferiority margin \(\Delta\), hypotheses, one-sided significance level, target power, and planning treatment effect \(\theta_A\) remain fixed in the procedure described here. Fixing them prevents the design objective from shifting in response to interim results.

In particular, an observed interim treatment difference is not automatically a new value for \(\theta_A\). Replacing the planned effect with an interim estimate can make the sample size depend on a noisy treatment-effect estimate and can introduce bias or alter operating characteristics. The earlier material on observed treatment effects and sample size bias addresses that distinct issue. The focus here is updating nuisance parameters under a fixed effect assumption.

Worked Example: Why the Target Effect Is Not Replaced by the Interim Difference

Continue the hypothetical continuous-outcome design with \(\Delta=5\) points, \(\alpha=0.025\), \(90\%\) power, and an updated standard deviation of \(11.2\) points. The prespecified planning effect remains \(\theta_A=0\), so \(d=5\) points. Using \(1.960\) and \(1.282\), the squared quantile sum is \(10.510564\):

n= \frac{2(11.2^2)(10.510564)}{5^2} =\frac{250.88(10.510564)}{25} =105.4756

The correct rounded requirement under this rule is 106 per group. Suppose instead that an interim estimate of the treatment effect is \(1.5\) points. Substituting it for the prespecified planning effect would make the distance \(1.5+5=6.5\) points and give:

n_{\mathrm{wrong}}= \frac{2(11.2^2)(10.510564)}{6.5^2} =\frac{2636.89029632}{42.25} \approx62.4116

That calculation yields 63 per group after rounding, but it answers a different planning question: it assumes a more favorable treatment effect than the one fixed in the design. It is not the nuisance-only re-estimation rule. The interim effect may be relevant to a separate prespecified decision, but it should not be silently substituted here.

Assumptions, Diagnostics, and Protection of the Final Test

A sound update requires more than inserting an interim estimate into a formula. Before using the rule, check that the estimator targets the nuisance parameter used in the power calculation, that the interim data are sufficiently informative, and that the outcome definition and analysis remain unchanged. For a continuous endpoint, inspect whether the variance estimate is distorted by outliers, unequal variances, or departures from the planned population. For a binary endpoint, check that the estimated rates and event counts support the normal approximation; sparse data may make it unreliable.

Also check the operational rule: when the estimate is calculated, who can see it, how the updated total is rounded, and what floors and caps apply. State whether the interim participants count toward the final total. If the calculated requirement is smaller than the number already enrolled, the final sample size cannot be reduced below that number.

Most importantly, a fixed nominal \(\alpha\) in the formula does not by itself prove that the final adaptive test controls Type I error. The final test must be prespecified and valid for the adaptation actually used. A nuisance estimate based on internal data can be statistically dependent on the final test statistic; consequently, simply recalculating \(n\) and then applying an ordinary fixed-sample test may not have its nominal operating characteristics. The design should use an established valid approach or calibrate the complete adaptation and final analysis under the null. Earlier material on maintaining a fixed final test after re-estimation addresses this broader requirement.

A useful comparison is with keeping the original sample size unchanged. That option avoids a data-dependent sample-size change but may leave power lower or higher than intended if the nuisance parameter differs from its planning value. Nuisance-only re-estimation can align the sample size more closely with the intended power, at the cost of added operational complexity and a need to protect the final test. A different option is to revise the target effect or the margin; that is not a nuisance update and changes the scientific design question.

Common mistake: “The interim results look favorable, so I can use the observed treatment effect in the power formula and reduce the sample size.” This changes the planning alternative using treatment-effect data, rather than updating a nuisance parameter under a fixed design. Keep the margin and target effect fixed; update only the prespecified nuisance estimate, apply the stated rounding and cap rules, and verify that the final test remains valid.

These calculations have limitations. The normal formulas are approximations; actual power depends on the endpoint distribution, estimator, missingness, allocation, and final analysis. A single interim estimate may also be imprecise. Re-estimation does not repair a poorly chosen margin, an inappropriate endpoint, or a mismatch between the assumed and actual treatment effect. It is a controlled adjustment to sample size, not a substitute for sound trial design.

Check Your Understanding

Answer each question using the fixed-margin, fixed-effect nuisance-update framework. Short solutions follow.

For a treatment effect oriented so that larger values favor treatment, write the noninferiority hypotheses using margin \(\Delta>0\).
If the standard deviation in a two-group continuous-outcome design rises from \(10\) to \(12\), with all other design quantities fixed, by what factor does the unrounded sample size change?
In a binary-outcome design with \(\Delta=0.08\) and \(\theta_A=0\), what is the planning distance from the null boundary?
Why is substituting an interim treatment-effect estimate for the planned effect not a nuisance-parameter update?
Does using a nominal one-sided \(\alpha\) in the sample-size formula alone establish Type I error control after re-estimation?

Solutions. (1) \(H_0:\theta\leq-\Delta\) versus \(H_a:\theta>-\Delta\). (2) Sample size is proportional to \(\sigma^2\), so the factor is \(12^2/10^2=1.44\). (3) \(d=\theta_A+\Delta=0.08\). (4) It changes the planning alternative and makes the calculation depend on an interim estimate of the treatment effect; the nuisance-only rule keeps \(\theta_A\) fixed. (5) No. The final test must be valid for the actual adaptation, including any dependence between interim data and the final statistic.