Why Nuisance Parameters Matter
A sample-size formula does not turn a treatment objective into a participant count without additional assumptions. It also needs quantities such as outcome variability, control-group event probability, the fraction of participants expected to experience an event, and the degree of similarity among participants in the same cluster. These are nuisance parameters: they affect the information available for the treatment comparison but are not the treatment effect the trial is designed to detect.
A useful planning distinction is between the target effect and the assumptions used to calculate the sample size. For a continuous endpoint, the target might be a mean difference \(\delta\), while the standard deviation \(\sigma\) is a nuisance parameter. For a binary endpoint, the target might be a difference in event probabilities, while the probabilities themselves determine the variability. A change in a nuisance parameter can change the required sample size even when the target effect and the desired error rates remain fixed.
The calculations below are benchmarks under stated assumptions, not guarantees. Their purpose is to reveal which assumptions drive the answer, so planners can check plausible values and examine how much the requirement changes when those values are wrong.
Continuous Outcomes: Variance and the Target Difference
Suppose a trial compares two independent groups with equal allocation using a continuous outcome. Assume a common population standard deviation \(\sigma\), a target difference \(\delta>0\) between the population means, a two-sided test at significance level \(\alpha\), and desired power \(1-\beta\). The usual large-sample normal approximation gives the required number per group as
where \(z_q\) is the \(q\)-quantile of the standard normal distribution. This formula assumes independent observations, equal group sizes, a correctly specified target difference and variance, and an analysis whose standard error is adequately approximated by the normal calculation. It does not automatically account for losses to follow-up, clustering, multiplicity, or other design features.
The formula shows the key dependence directly: doubling \(\sigma\) multiplies the required sample size by four; doubling \(\delta\) divides it by four. The error-rate choices also matter through the normal quantiles, but they do not remove uncertainty in the nuisance value \(\sigma\).
Worked Example: Sensitivity to the Standard Deviation
Suppose a hypothetical two-arm trial aims to detect a mean difference of 6 points with 90% power and a two-sided \(\alpha=0.05\). Assume independent participants, equal allocation, a common standard deviation, and no multiplicity adjustment. Use \(z_{0.975}=1.96\) and \(z_{0.90}=1.2816\), so their sum is approximately 3.242.
If the planning standard deviation is 12 points, the requirement per arm is
Rounding up gives 85 participants per arm before any allowance for missing outcomes. If a plausible alternative standard deviation is 14 points, while the target difference and other assumptions stay fixed, then
The higher variance raises the requirement to 115 per arm after rounding up. The increase is consistent with the variance ratio: \(14^2/12^2\approx1.361\), and \(84.08(1.361)\approx114.4\). This comparison changes a nuisance assumption, not the clinical target of a 6-point difference.
Binary Outcomes: Event Probabilities Shape Information
For a two-arm comparison of independent binary outcomes with equal allocation, let \(p_1\) and \(p_2\) be the event probabilities in the two groups. For a two-sided normal-approximation test of a difference in proportions, a common planning formula is
Here \(n\) is the number per group. The calculation assumes independent participants, a fixed outcome definition and assessment time, equal allocation, and an adequate normal approximation. Exact or other methods may be preferable for very small samples or probabilities near zero or one. The formula is sensitive to both the target difference \(p_1-p_2\) and the underlying probabilities: the same absolute difference can yield different precision at different baseline risks.
Worked Example: Planning for Two Event Probabilities
Consider a hypothetical trial with control event probability \(p_1=0.40\) and treatment event probability \(p_2=0.25\), giving a target absolute difference of 0.15. Assume independent participants, equal allocation, a two-sided \(\alpha=0.05\), 80% power, and use \(z_{0.975}=1.96\) and \(z_{0.80}=0.8416\). The average probability is \(\bar{p}=(0.40+0.25)/2=0.325\).
Substitution gives
The first variance term is \(2(0.325)(0.675)=0.43875\), so \(1.96\sqrt{0.43875}\approx1.2983\). The second is \(0.24+0.1875=0.4275\), so \(0.8416\sqrt{0.4275}\approx0.5503\). Therefore,
Round up to 152 participants per group before accounting for missing outcomes. These values are hypothetical planning inputs. If the control probability is uncertain, planners should calculate requirements under several plausible pairs of probabilities rather than treating one borrowed estimate as certain.
Time-to-Event Outcomes: Required Events and Censoring
In an event-driven time-to-event trial, the primary sample-size target is often the number of observed events, not simply the number randomized. For a two-group log-rank comparison under proportional hazards, with equal allocation, a common approximation for the required number of events is
where \(D\) is the total event count, \(HR\) is the target hazard ratio, and \(\pi\) is the fraction allocated to one group. Equal allocation gives \(\pi=0.5\). This approximation presumes proportional hazards, independent participants, the specified allocation ratio, and an appropriate large-sample log-rank framework. The target event count depends on the hazard ratio, not directly on the baseline event rate or censoring; those quantities help determine how many participants and how much follow-up will be needed to accrue the events.
Worked Example: From a Hazard Ratio to an Event Target
Suppose a hypothetical equal-allocation trial targets a hazard ratio of \(HR=0.75\), with two-sided \(\alpha=0.05\) and 80% power. Assume proportional hazards and use \(z_{0.975}=1.96\) and \(z_{0.80}=0.8416\). The event requirement is
The trial therefore needs at least 380 events under this approximation. This is not a target of 380 participants. If only a fraction of randomized participants are expected to experience the event during the planned observation period, the required number randomized will be larger. The expected fraction depends on event incidence, accrual, follow-up, and censoring; an incorrect assumption about any of these can make the calendar-time plan or enrollment target inadequate even when the event-count calculation is correct.
Clustering, Attrition, and Other Inflations
When participants are randomized or observed in clusters, outcomes within a cluster may be correlated. For equal cluster size \(m\) and intracluster correlation coefficient \(\rho\), a common approximate design effect is
The design effect approximates the factor by which the independent-participant sample size must be multiplied to maintain comparable precision. It assumes a simple cluster structure and equal cluster sizes; unequal sizes and more complex designs can require different adjustments. Even a modest \(\rho\) can matter when clusters are large because the factor includes \(m-1\).
Worked Example: Clustering and Incomplete Outcomes
Suppose a hypothetical individually randomized calculation requires 100 analyzable participants per group. A planned cluster design has 20 participants per cluster and an assumed intracluster correlation of 0.05. Assume equal cluster sizes and, separately, that 10% of randomized participants will lack the primary outcome.
First calculate the design effect:
The cluster-adjusted analyzable target is \(100(1.95)=195\) participants per group. To allow for 10% without an observed outcome, inflate by dividing by the expected observed fraction, 0.90:
If clusters contain 20 participants, the number of clusters per group must be rounded up from \(216.7/20=10.84\) to 11 clusters. That gives 220 randomized participants per group. The calculation treats the 10% loss as individual-level loss and assumes the cluster sizes remain 20; cluster-level dropout or variable cluster size would need separate consideration.
Other common adjustments include unequal allocation, multiple primary comparisons, and nonadherence. These should not be folded into a generic safety factor without a model. For example, an unequal allocation ratio changes the variance of a two-group contrast, while multiplicity can change the critical value and nonadherence can reduce the treatment contrast that the analysis is expected to observe.
Checking Assumptions and Comparing Planning Approaches
A point estimate for a nuisance parameter is not the same as a robust planning choice. Historical studies may differ in eligibility criteria, outcome definitions, follow-up, or measurement methods. Small pilot studies can estimate variability or event rates imprecisely. A useful planning report therefore gives the source and rationale for each nuisance value and shows how the requirement changes across plausible scenarios.
| Planning approach | What it does well | Main limitation |
|---|---|---|
| Single-value calculation | Provides a transparent benchmark for a specified set of assumptions. | Can conceal sensitivity to an uncertain nuisance parameter. |
| Scenario or sensitivity analysis | Shows how sample size changes over plausible nuisance values. | Does not by itself identify which scenario is most likely. |
| Conservative planning value | Can protect against a plausible unfavorable value, such as higher variance. | May produce an unnecessarily large trial if the value is overly cautious. |
| Simulation under a design model | Can represent complex allocation, follow-up, and analysis features. | Results depend on the simulation model and the assumptions supplied to it. |
For each calculation, check whether the nuisance parameter corresponds to the planned population, endpoint, and analysis. For variance, confirm the measurement scale and whether the assumed value is a standard deviation rather than a variance. For binary outcomes, confirm the outcome time point and whether the two probabilities imply the intended treatment contrast. For survival outcomes, separate the event target from the enrollment and follow-up assumptions. For cluster designs, examine both cluster size and the assumed intracluster correlation.
A diagnostic warning is a sample-size result that changes sharply under small, credible changes to a nuisance input. That is not an arithmetic failure; it is evidence that the design is sensitive. A second warning is an implausible conversion between events and enrollment, or between analyzable participants and randomized participants. Recheck the assumptions and units before rounding. When necessary, use scenarios or simulation to assess the design rather than presenting one precise-looking count as certain.
Common Pitfall and Practical Interpretation
Wrong reasoning: “The treatment effect is our main scientific question, so the sample size is determined by that effect; uncertainty in the event rate or variance is a secondary detail.” Correction: The target effect is only one input. The variance determines the noise around a continuous contrast, probabilities determine binary-outcome variability, censoring affects how quickly event targets are reached, and clustering reduces the effective information per participant. A design can miss its intended power if these inputs are materially wrong, even when the target effect was chosen appropriately.
Likewise, do not interpret an event target as a randomized-participant target or apply a dropout inflation before applying a cluster design effect without checking the exact design assumptions. State what each number counts, which population it applies to, and whether it refers to enrollment or analyzable data. Sample-size calculations are conditional on their assumptions; they do not ensure that a trial will attain a particular power in the realized data.
Check Your Understanding
Answer each question, then compare with the solutions.
- If the standard deviation in an equal-allocation continuous-outcome calculation rises by 10%, by what factor does the approximate required sample size change, holding other inputs fixed?
- Why can two binary-outcome scenarios with the same absolute treatment difference require different sample sizes?
- In a log-rank design, what does the event-count formula determine, and why is it not itself the number of participants to randomize?
- For cluster size \(m=15\) and intracluster correlation \(\rho=0.02\), calculate the approximate design effect.
- Name two checks that should be made before using a historical event-rate estimate for planning.
Solutions. (1) Sample size is proportional to \(\sigma^2\), so the factor is \(1.10^2=1.21\), a 21% increase. (2) The probabilities affect the variance terms in the formula, so equal differences need not have equal precision at different baseline risks. (3) It determines the approximate number of events needed under the stated hazard ratio, allocation, error rate, and power; only some randomized participants may experience an event during follow-up, so enrollment must also reflect event incidence, censoring, accrual, and follow-up. (4) \(DE=1+(15-1)(0.02)=1.28\). (5) Check that the population and endpoint definition match the planned trial, and that the estimate refers to a comparable time horizon and follow-up scheme; also assess the estimate's uncertainty and the effects of censoring.