Sample Size Changes During a Trial

Why Planned Enrollment May Change

A trial’s planned sample size is based on assumptions made before all its outcomes are known. Those assumptions may not hold during conduct: outcomes may be more variable than expected, recruitment may be slower, or the information already being collected may be more than the decision requires. A sample size adaptation changes the planned number of participants, usually in response to an interim review or an updated estimate of a design input.

Increasing enrollment can add information and help achieve a prespecified level of power when an assumption such as outcome variability was too optimistic. Decreasing enrollment can avoid collecting data that appear unnecessary or infeasible to obtain. Neither direction is automatically beneficial. More participants generally require more time and resources and expose more people to trial procedures; fewer participants generally mean less information and may leave an important treatment difference harder to detect.

Definition: A sample size change is a planned revision to the number of participants to be enrolled or analyzed in a trial. The change may increase or decrease the original target. It is an adaptation when it is made according to a trial-design rule, rather than simply reflecting an accidental shortfall or unplanned protocol deviation.

A planned target is not a promise that a trial will produce a particular result. It is a design choice intended to make a specified analysis informative under stated assumptions. In particular, increasing sample size does not guarantee statistical significance, and reducing it does not prove that the treatment difference is unimportant.

How Sample Size Relates to Information

For a simple illustration, suppose a trial compares the mean of a continuous outcome between two treatment groups. Assume equal allocation, independent participants, a common outcome standard deviation $\sigma$, and a difference in group means of interest $\delta$. For a two-sided test of $H_0$ that the mean difference is zero, an approximate fixed-design sample size per group is

n \approx \frac{2\sigma^2\left(z_{1-\alpha/2}+z_{1-\beta}\right)^2}{\delta^2}.

Here, $\alpha$ is the significance level, $1-\beta$ is the target power, and $z_q$ is the $q$-quantile of the standard normal distribution. The formula is an approximation for a two-sample normal comparison; it assumes the stated common variance and equal allocation, and it does not account for losses to follow-up, clustering, or other design features. In practice, the sample size may need adjustment for those features.

The relationship has a useful interpretation: if variability is larger, the groups need more participants to distinguish the same difference from noise. If the difference worth detecting is smaller, more participants are needed. When the other quantities stay fixed, the required $n$ is proportional to $\sigma^2$ and inversely proportional to $\delta^2$.

The same idea appears in the standard error of the estimated difference in means. With $n$ participants per group and common standard deviation $\sigma$, its standard error is $\sigma\sqrt{2/n}$. More participants reduce this standard error, but with diminishing returns: doubling $n$ reduces the standard error by a factor of $1/\sqrt{2}$, not by half.

Key interpretation: Increasing $n$ can improve precision and power under the design assumptions. It does not change the treatment effect itself, remove bias, or make an inappropriate outcome measure suitable.

Worked Sample Size Calculations

Worked Example: Increasing the Target After Greater Variability Is Estimated

Suppose a hypothetical trial is designed to compare two group means. The planning assumptions are a common standard deviation of 10 outcome units, a difference of 5 units worth detecting, a two-sided significance level of 0.05, and 80% power. Assume equal allocation and use $z_{0.975}=1.96$ and $z_{0.80}=0.842$. These assumptions meet the illustrative formula’s normal-outcome, common-variance, and equal-allocation conditions.

The initial approximate sample size per group is

n \approx \frac{2(1.96+0.842)^2(10)^2}{5^2} = \frac{2(2.802)^2(100)}{25} = \frac{1570.2408}{25} = 62.809632.

Rounding up gives a target of 63 participants per group, or 126 participants in total. Now suppose an appropriate interim estimate suggests that the standard deviation is 12 units, while the target difference, significance level, and power remain the same. Substituting 12 for $\sigma$ gives

n \approx \frac{2(2.802)^2(12)^2}{5^2} = \frac{2261.146752}{25} = 90.44587008.

Rounding up gives 91 participants per group, or 182 in total: 28 additional participants per group, 56 overall. Under this simple formula, higher variability calls for a larger target to retain the same planned power for a 5-unit difference. Whether the increase is worthwhile depends on feasibility and the trial’s decision needs, not just on the arithmetic.

Worked Example: Quantifying the Precision Gain and Its Resource Cost

Continue with the hypothetical design, and temporarily hold the standard deviation at 10 units to isolate the effect of sample size. With 63 participants per group, the standard error of the estimated difference is

SE = 10\sqrt{\frac{2}{63}} \approx 1.782 \text{ outcome units}.

With 91 participants per group, it is

SE = 10\sqrt{\frac{2}{91}} \approx 1.483 \text{ outcome units}.

Thus, increasing enrollment from 63 to 91 per group increases the sample size per group by $91/63 \approx 1.444$, or about 44.4%. The standard error falls by about 16.8%, since $1-1.483/1.782 \approx 0.168$. This is a real precision gain, but not a 44.4% reduction in uncertainty.

Suppose, hypothetically, that a site can enroll eight additional participants per month across both groups and that each additional participant costs $4,000 to recruit and follow. Enrolling 56 more participants would require about $56/8=7$ additional months and cost $56 \times \$4{,}000=\$224{,}000$. The adaptation may be justified if the information could materially improve a clinical decision; it may not be justified if that decision would be unchanged or if the extra time delays needed evidence. These are planning assumptions for illustration, not general recruitment or cost estimates.

Worked Example: Decreasing the Target After Lower Variability Is Estimated

Suppose instead that a prespecified, suitable blinded procedure gives a standard-deviation estimate of 8 outcome units at an interim review. Assume the same two-group comparison, equal allocation, target difference of 5 units, two-sided $\alpha=0.05$, and 80% power as above. If the revised estimate is considered reliable for planning, the calculation is

n \approx \frac{2(2.802)^2(8)^2}{5^2} = \frac{1004.954112}{25} = 40.19816448.

Rounding up gives 41 participants per group, 82 in total. Compared with the original target of 63 per group, this would reduce planned enrollment by 22 per group, or 44 participants overall. Under the formula’s assumptions, lower variability means fewer participants are needed for the same target difference and power.

This calculation alone does not authorize a reduction. The interim estimate must be suitable for the design and the revision must follow a plan that preserves valid inference. For example, a variance estimate can be unstable when based on limited data. A blinded estimate may also need careful handling because the combined data can reflect both within-group variability and differences between groups. The example assumes that the stated estimate is appropriate; a real trial must justify that choice.

Why Adaptations Need Rules

Changing the target based only on the observed treatment difference can create a problem for statistical inference. If investigators look at an interim effect estimate and increase enrollment when the result seems promising, the final analysis is affected by that selection. A conventional analysis that treats the final sample size as if it had been fixed in advance may then fail to maintain the intended type I error rate, the probability of rejecting $H_0$ when $H_0$ is true.

For this reason, a sample size adaptation should be described in advance: what information can be reviewed, when the review occurs, what quantity triggers a change, and what the maximum or minimum target is. The analysis method must also be appropriate for the adaptation. An adjustment based on a nuisance parameter such as outcome variability may be less dependent on the treatment effect than one based directly on the interim effect, but it still requires a justified rule and valid inference. The details of the adaptation rule matter; “we will change the sample size if needed” is not a sufficiently precise plan.

Identify the uncertainty.
Specify which planning input may be wrong, such as the outcome variability or recruitment feasibility.

Set a decision rule.
Define the interim information and conditions that would lead to an increase, decrease, or no change, including any enrollment bounds.

Assess consequences.
Compare added or avoided information with delay, cost, participant burden, and the consequences of a less precise result.

Use valid final inference.
Ensure the analysis accounts for the adaptation so that its uncertainty and error rates are interpreted appropriately.

Balancing Information Against Burden

An increase is most compelling when additional information could resolve an important uncertainty and the trial can obtain it without disproportionate delay or burden. A reduction may be attractive when the initial target appears unnecessarily large, the remaining information is adequate for the intended decision, or continued recruitment is infeasible. But operational difficulty alone does not make a smaller trial scientifically adequate.

Consideration	Potential gain from increasing enrollment	Potential cost or risk
Precision	Smaller standard error and typically narrower confidence intervals.	Diminishing returns: each additional participant contributes less precision than earlier participants.
Power	Greater ability to detect the target difference under the model assumptions.	Power is not guaranteed if assumptions are wrong or the true difference is smaller.
Time and resources	More data may support a more informative decision.	More recruitment, follow-up, analysis, and cost; evidence may arrive later.
Participants	More people contribute information to the trial.	More people undergo trial procedures and may face inconvenience or risk.

A smaller target reverses these trade-offs: it can save time and resources and avoid enrolling additional participants, but typically widens uncertainty and can reduce power. The relevant comparison is not simply “more data versus less data.” It is whether the expected value of the additional information is worth the burden and delay, while maintaining a sound design and valid inference.

Common Mistakes and Diagnostics

Assuming a larger trial is always better. More observations can improve precision, but they do not remove systematic bias or guarantee a clinically useful conclusion. Consider what decision the extra information could change.
Rounding a calculated target down. In the examples, 62.81 requires 63 participants per group, not 62; round upward to meet or exceed the calculated target.
Treating an interim estimate as exact. Estimates from partial data have uncertainty. Check how they were obtained and whether the adaptation rule accounts for that uncertainty.
Changing enrollment because a result looks promising, then using an ordinary fixed-sample analysis without adjustment. This can distort error control. Use a prespecified design and inference method appropriate to the change.
Equating a smaller target with evidence of no effect. A reduction is a design decision, not a finding that the treatment difference is zero or clinically irrelevant.

Useful warning signs include a large change in the variance estimate from a small interim sample, a target that exceeds the recruitment or budget limit, and a proposed rule that depends on the observed treatment effect but does not specify how the final analysis will account for selection. Each calls for review before enrollment is changed.

Key takeaway: A sample size change trades information against time, cost, and participant burden. Its rationale, decision rule, and final inference must fit together; a recalculation by itself does not establish that the adaptation is appropriate.

Check Your Understanding

Answer each question, then compare with the solutions below.

Under the illustrative two-group formula, what happens to the required sample size if the anticipated standard deviation doubles while all other inputs stay fixed?
A planned target is 50 participants per group. A justified recalculation gives 50.2. How many participants per group are needed to meet that target?
Why might an increase in sample size be valuable, and what is one cost of that increase?
Why is it risky to increase enrollment only because an interim treatment difference looks favorable?
If a valid procedure suggests lower outcome variability, does a reduced target prove that the treatment has no effect? Explain.

Solutions. (1) Since $n$ is proportional to $\sigma^2$, doubling $\sigma$ multiplies the required sample size by four. (2) Round up to 51 per group. (3) It can improve precision and power under the design assumptions; it may require more time, money, and participant involvement. (4) Selecting the sample size based on interim results can affect the operating characteristics of a conventional fixed-sample analysis, including its type I error rate; the rule and final inference need to account for the adaptation. (5) No. It is a planning decision based on variability and other design inputs, not evidence that the treatment effect is zero.