Rare Events And Boundary Corrections In Binary Re-Estimation

Why Rare Events Challenge Binary Re-Estimation

When an outcome is uncommon, a modest number of participants can produce no observed events in one or both groups. A normal approximation to a binary sample-size formula may then be unstable: the estimated event rate can lie at the boundary, and a small change in that estimate can cause a large change in the calculated sample size. The issue is not simply that a rare event is inconvenient. The approximation treats a discrete count as if it were sufficiently close to a continuous normal variable.

In blinded re-estimation, the analyst typically uses a pooled event rate to update a nuisance quantity while retaining the prespecified treatment effect for planning. This tutorial focuses on diagnosing boundary sensitivity and applying three safeguards: a continuity correction, an expected-count check, and exact planning based on a specified test. None automatically repairs a misspecified treatment effect or guarantees that the full adaptive design has the desired operating characteristics.

Definition: A rare-event boundary problem occurs when the binary outcome rate is sufficiently close to zero or one that the estimated rate, expected event count, or normal approximation used in a sample-size update is sensitive to small changes in the observed data. A boundary estimate such as zero observed events is not proof that the underlying event probability is zero.

For the calculations below, assume independent participants, a binary outcome observed once per participant, complete outcome data, and the stated allocation. The planning rates \(p_0\) and \(p_1\) are the event probabilities in the control and treatment groups under the planning alternative. The formula shown uses equal group sizes, a two-sided significance level \(\alpha\), and target power \(1-\beta\). It is an approximate risk-difference calculation, not an exact test.

Normal Approximation and Its Boundary Diagnostics

With \(n\) participants per group, the standard normal-approximation sample-size calculation for testing \(H_0:p_1=p_0\) against a two-sided alternative is

n_{\mathrm{normal}} = \frac{ \left[ z_{1-\alpha/2}\sqrt{2\bar p(1-\bar p)} + z_{1-\beta}\sqrt{p_0(1-p_0)+p_1(1-p_1)} \right]^2 }{ (p_1-p_0)^2 }, \qquad \bar p=\frac{p_0+p_1}{2}.

Here \(z_q\) is the \(q\)-quantile of the standard normal distribution. The first variance term corresponds to the null calculation and the second to the planning alternative. This formula presumes that the normal approximations behind the test and power calculation are adequate. It becomes especially sensitive when the rates are very small, because the information depends on a small number of expected events and the squared risk difference appears in the denominator.

A useful diagnostic is to inspect the expected event and non-event counts in each group at the proposed \(n\): \(np_0\), \(n(1-p_0)\), \(np_1\), and \(n(1-p_1)\). If a relevant expected count is small, the normal calculation warrants scrutiny. Rules such as “at least five” or “at least ten” expected events are rough screening conventions, not mathematical guarantees that the approximation is accurate. A second diagnostic is sensitivity: recalculate across plausible pooled rates or nuisance-rate estimates. A large change in \(n\) from a small change in the rate signals that a plug-in update is fragile.

Worked Example: Normal and Continuity-Corrected Planning

Suppose a hypothetical equal-allocation trial plans to compare event rates \(p_0=0.005\) and \(p_1=0.010\), with two-sided \(\alpha=0.05\) and 80% power. Assume independent binary outcomes and complete follow-up. Then \(\bar p=0.0075\), and use \(z_{0.975}=1.96\) and \(z_{0.80}=0.8416\).

The two numerator terms are

1.96\sqrt{2(0.0075)(0.9925)}\approx 0.2391481, \qquad 0.8416\sqrt{(0.005)(0.995)+(0.010)(0.990)} \approx 0.1026442.

Thus their sum is \(A\approx0.3417923\), and the uncorrected calculation is

n_{\mathrm{normal}} \approx \frac{(0.3417923)^2}{(0.010-0.005)^2} = \frac{0.116822}{0.000025} \approx 4672.9.

Rounding up gives about 4,673 participants per group. At this size the planning alternative implies about \(4673(0.005)=23.4\) and \(4673(0.010)=46.7\) expected events per group. These expected counts are not at the boundary, although a blinded interim estimate from a much smaller internal pilot may be.

For example, with only 20 participants per group and a pooled event probability near 0.005, the probability of observing zero events among all 40 participants is approximately \((1-0.005)^{40}=0.8183\). A zero-event pilot is therefore unsurprising under that rate; directly substituting a zero estimate into a formula that requires positive rates would be inappropriate.

Continuity Correction: What It Does and Does Not Do

A continuity correction adjusts a normal approximation for the fact that event counts change in whole-number steps. For this tutorial, use the following conventional correction to the equal-allocation risk-difference calculation: require the planning separation in expected event counts to exceed the normal-approximation requirement by one-half event,

n\delta \ge A\sqrt{n}+\frac{1}{2}, \qquad \delta=|p_1-p_0|,

where \(A\) is the sum of the two \(z\)-weighted standard-deviation terms in the formula above. This convention yields a corrected approximate sample size. Other software or references may implement continuity corrections differently, so a protocol should name the formula being used rather than treating all corrected calculations as interchangeable.

Set \(x=\sqrt n\). The inequality becomes \(\delta x^2-Ax-\tfrac12\ge0\). Since \(\delta>0\), the positive root is the smallest nonnegative value satisfying equality; hence round its square upward:

x = \frac{A+\sqrt{A^2+2\delta}}{2\delta}, \qquad n_{\mathrm{cc}}=\left\lceil x^2\right\rceil.

The correction addresses discreteness in the normal approximation. It does not make a normal model reliable when expected counts are extremely small, and it does not replace an exact method when exact operating characteristics are required.

Worked Example: Applying the Continuity Correction

Continue the hypothetical planning calculation above, using the unrounded component values to obtain \(A\approx0.3417923\), and \(\delta=0.005\). The positive root is

x = \frac{0.3417923+\sqrt{(0.3417923)^2+2(0.005)}}{2(0.005)} \approx 69.7913.

Therefore

n_{\mathrm{cc}}=\left\lceil (69.7913)^2\right\rceil =\left\lceil 4870.8\right\rceil =4871 \quad\text{participants per group}.

The corrected result is slightly larger than the uncorrected result of 4,673 per group. The difference is due to the half-event adjustment, not a change to the assumed rates, power, or significance level. As before, a continuity-corrected result should still be checked against expected event counts and the intended analysis.

Minimum-Count Safeguards and Blinded Rate Reconstruction

A minimum-count safeguard is a prespecified diagnostic or operational rule that prevents an update from relying on implausibly sparse information. For example, a protocol might require the planned sample size to imply at least ten expected events in the lower-rate group under the planning rates. If the lower planning rate is \(p_{\min}\), that particular rule requires \(n p_{\min}\ge10\), or \(n\ge10/p_{\min}\). Ten is an illustrative choice, not a universal threshold, and satisfying it does not prove that a normal approximation is accurate.

In a blinded design, the observed pooled rate alone does not identify the two group rates. A treatment effect must be supplied for planning; one option is a prespecified odds ratio. If \(\tilde p\) is the pooled rate and allocation is equal, then the planned rates must satisfy \((p_0+p_1)/2=\tilde p\). For an assumed treatment-to-control odds ratio \(\theta\), the corresponding treatment rate is

p_1=\frac{\theta p_0}{1-p_0+\theta p_0}.

This conversion can be numerically delicate near zero. Small rounding differences in \(\tilde p\) or in the solved rates can affect a subsequent sample-size calculation, particularly when the planned risk difference is small. The rates must also be feasible probabilities between zero and one.

Worked Example: Reconstructing Rates From a Rare Pooled Rate

Suppose a hypothetical equally allocated internal pilot has one event among 202 participants, so the pooled estimate is \(\tilde p=1/202\approx0.004950495\). For planning, assume an odds ratio \(\theta=2\), with treatment event odds twice control event odds. Under equal allocation, \(p_0+p_1=2\tilde p=1/101\). The odds-ratio relation gives \(p_1=2p_0/(1+p_0)\), so solve

p_0+\frac{2p_0}{1+p_0}=\frac{1}{101}.

Multiplying by \(101(1+p_0)\) gives \(101p_0^2+302p_0-1=0\). The positive root is

p_0=\frac{-302+\sqrt{302^2+404}}{202} \approx 0.00330760, \qquad p_1=\frac{2p_0}{1+p_0}\approx0.00659339.

To six decimal places, the planned rates are \(p_0=0.003308\) and \(p_1=0.006593\). Their average is approximately \(0.0049505\), consistent with the pooled rate, and their odds ratio is 2 by construction. These are planning rates implied jointly by the pooled estimate and the assumed effect; they are not separately observed group rates. A minimum expected-count check or an exact planning calculation is still needed if the resulting expected counts are sparse.

Exact Planning and Discrete Outcomes

When normal approximations are doubtful, exact planning evaluates the rejection probability using the distribution of the integer-valued outcomes under the chosen test. The method must match the planned analysis. For a conditional Fisher exact test, for example, one identifies the rejection set at the chosen \(\alpha\) and sums the probabilities of outcomes in that set under the planning alternative. Because the set of possible outcomes is discrete, the achieved type I error may be below \(\alpha\), and the attainable power can change in steps as \(n\) changes.

Exact planning is not a universal correction factor to multiply into a normal sample size. It can require more computation, can be conservative because of discreteness, and depends on the chosen exact test and handling of nuisance parameters. The planning procedure should specify those choices. Exact computation is especially useful when candidate sample sizes imply very few events or when the normal approximation gives a result that changes sharply under small perturbations of the event rate.

Worked Example: Recognizing Discreteness in Exact Planning

Consider a hypothetical two-arm trial with two participants per group and a one-sided Fisher exact test at \(\alpha=0.05\). Conditional on the total number of events, the most favorable possible allocation for the treatment group occurs when there are two total events and both occur in treatment. Its one-sided conditional probability is

\frac{\binom{2}{2}\binom{2}{0}}{\binom{4}{2}} =\frac{1}{6} \approx0.1667.

This is larger than 0.05. With zero or four total events, there is no evidence of a treatment advantage; with one or three total events, the smallest possible one-sided tail probability is 0.5. Thus no outcome falls in a 0.05 rejection region for this design, and the exact test has power zero for any specified alternative rates. This tiny design is not a clinical recommendation: it shows why an exact planning calculation must enumerate attainable outcomes rather than assume that a nominal \(\alpha\) level guarantees useful power.

As another diagnostic, suppose a hypothetical pilot observes 0 events among 40 control participants and 3 among 40 treatment participants. Conditional on three total events, the one-sided Fisher probability of all three falling in treatment is

\frac{\binom{40}{3}}{\binom{80}{3}} = \frac{9880}{82160} \approx0.1202.

This calculation illustrates the discreteness of rare-event evidence; it is not a sample-size recommendation. For planning, enumerate the exact rejection region at each candidate \(n\), then sum its probability under the assumed alternative. Do not infer exact power from this single observed allocation.

Choosing a Safeguard and Avoiding Common Errors

Approach	Useful role	Main limitation
Uncorrected normal formula	Efficient approximation when expected counts support it	Can be unstable near a boundary and ignores count discreteness
Continuity correction	Adjusts a normal approximation for integer event counts	Does not ensure adequate normal approximation in extremely sparse settings
Minimum expected-count rule	Flags a calculation that relies on too few planned events	Threshold is a design convention, not proof of validity
Exact planning	Computes power for the specified discrete test	Can be conservative, computationally demanding, and test-dependent

A common mistake is to treat zero observed events in a small blinded pilot as a zero population event rate, then conclude that the required sample size is infinite or that no events will occur. The correction is to recognize the uncertainty in the rate estimate, use a prespecified method for handling boundary estimates, and examine plausible rates or exact planning. Another mistake is to assume that adding a continuity correction makes any normal-based update safe. It only modifies the approximation; sparse expected counts still call for a stronger diagnostic or an exact calculation.

Finally, a sensible safeguard does not by itself establish that the re-estimation preserves the trial’s type I error. The full rule—including how the nuisance rate is estimated, how the sample size is changed, and which final test is used—must be assessed under the design’s assumptions. The pooled estimate also does not reveal the treatment effect: rate reconstruction requires the prespecified planning effect, as in the odds-ratio example. If rates, allocation, follow-up, or dependence assumptions differ from those used in the calculation, the proposed sample size may not deliver its nominal power.

Key takeaway: At rare-event boundaries, inspect expected counts and sensitivity before trusting a normal sample-size update. A continuity correction addresses discreteness, a minimum-count rule flags sparse planning information, and exact planning evaluates the chosen discrete test; none substitutes for stating and checking the design assumptions.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

Why is a zero event rate from a small pilot not a sound estimate that the population event probability is exactly zero?
Under the illustrative rule of at least ten expected events in the lower-rate group, what minimum \(n\) per group is required if that planning rate is 0.002?
What does the continuity correction change in the approximate calculation, and what problem does it not solve?
Why can an exact test have power zero for a particular sample size even when the planning alternative differs from the null?

Solutions. (1) A zero count is a possible random outcome when the true rate is positive, especially with a small sample; it does not identify the population rate. (2) The condition is \(0.002n\ge10\), so \(n\ge5000\) participants per group. (3) It adds a half-event allowance to the normal-approximation requirement to reflect integer counts; it does not ensure a good normal approximation when expected counts remain very small. (4) The exact test can have no attainable outcome in its rejection region at the selected \(\alpha\); then its rejection probability is zero under every alternative for that sample size.