Why a Near-Zero Variance Is a Boundary Case
Neyman allocation assigns more participants to the arm whose outcomes are more variable, because additional observations there reduce the variance of the estimated treatment contrast more. When one arm’s outcome variance is close to zero, this rule can assign that arm a very small sample-size share. The result may be mathematically optimal under a narrow variance-minimization objective, yet unsuitable for a clinical design that needs a meaningful comparison, reliable monitoring, or estimation in both arms.
The distinction is between the allocation target implied by a model and the allocation a practical trial can use. The target is calculated under assumptions about outcome variances and a fixed total sample size. Near a boundary, those assumptions and any minimum arm-size requirements matter especially strongly.
This is a planning result for a specified contrast and variance objective, not a general prescription for every trial. It assumes the variances are known for the calculation, outcomes are independent within and across arms, and each arm’s sample mean is the estimator of interest. It also treats the total sample size as fixed and does not include costs, ethical objectives, missing outcomes, or other design constraints.
Deriving the Target and Its Near-Zero Limit
Write \(n_H=N-n_L\). The variance to minimize over \(0<n_L<N\) is
For positive variances, differentiating gives
At an interior minimum, \(V'(n_L)=0\). Since the standard deviations are positive, taking positive square roots yields \(n_L/(N-n_L)=\sigma_L/\sigma_H\), or equivalently \(n_L/n_H=\sigma_L/\sigma_H\). The second derivative is positive throughout the feasible interior, so this stationary point is the unique minimum.
The target share for arm \(L\) is therefore \(\sigma_L/(\sigma_L+\sigma_H)\), not \(\sigma_L^2/(\sigma_L^2+\sigma_H^2)\). As \(\sigma_L\) approaches zero while \(\sigma_H\) stays positive, the target share of \(L\) approaches zero. The minimized variance, when both standard deviations are positive, is \((\sigma_L+\sigma_H)^2/N\). These expressions describe the formal unconstrained target; neither guarantees adequate representation in both arms.
Worked Example: A Small but Nonzero Variance
Suppose a hypothetical trial plans to compare mean outcomes in two arms with total sample size \(N=200\). For planning, assume \(\sigma_L=0.20\) outcome units and \(\sigma_H=2.00\) outcome units. Assume independent outcomes and use the fixed-sample variance objective above.
The Neyman ratio and shares are
Thus the continuous targets are \(n_L=200(0.20/2.20)=18.1818\) and \(n_H=200(2.00/2.20)=181.8182\). These sum to 200. The smaller target is not zero, but it represents only about 9.1% of the planned sample.
The target variance is
outcome-units squared. For comparison, equal allocation gives \(n_L=n_H=100\), with variance \(0.20^2/100+2.00^2/100=0.000400+0.04000=0.04040\) outcome-units squared. Under these assumed variances and this objective, the Neyman target has lower variance. The calculation does not establish that 18 participants in arm \(L\) are enough for every other purpose of the trial.
Minimum Arm Sizes and Constrained Targets
A design may require \(n_L\ge m_L\) and \(n_H\ge m_H\), where the positive integers \(m_L,m_H\) are specified for operational or inferential reasons. For the continuous allocation calculation, feasible values of \(n_L\) lie between \(m_L\) and \(N-m_H\). When both variances are positive, the unconstrained objective is strictly convex, so the constrained minimum is the unconstrained target if it lies in this interval; otherwise it is the nearest boundary. In other words, the constrained value is the unconstrained \(n_L\) limited to \([m_L,N-m_H]\). The resulting target can then be converted to integer sample sizes while preserving the total and both minimums.
This boundary rule has a direct interpretation. If the unconstrained target puts too few participants in arm \(L\), increasing \(n_L\) from that target raises the variance objective, but the minimum-size requirement does not permit the mathematical optimum. The best feasible choice is therefore \(n_L=m_L\), with the remaining participants in arm \(H\), provided the other bound is also satisfied. A minimum is not a claim that the variance objective itself prefers equal or balanced allocation; it is an added design constraint.
Worked Example: A Minimum Size Changes the Target
Consider a hypothetical two-arm design with \(N=100\), \(\sigma_L=0.10\) outcome units, and \(\sigma_H=1.00\) outcome units. Suppose the protocol requires at least 20 participants in each arm. Assume the variances are applicable to the planned outcomes and the independent-sampling variance formula is the planning criterion.
Without the minimum, the continuous Neyman target for arm \(L\) is
This violates \(n_L\ge20\). The feasible interval for \(n_L\) is \([20,80]\), so the constrained target is \(n_L=20\), \(n_H=80\). Its variance is
outcome-units squared. The unconstrained continuous target would have variance \((0.10+1.00)^2/100=0.01210\) outcome-units squared, smaller as expected, but it is not an admissible design under the stated minimum. Equal allocation would give variance \(0.01/50+1/50=0.02020\) outcome-units squared. Among allocations satisfying the 20-per-arm minimum, the constrained target is preferable for this variance objective; the minimum-size rule itself reflects requirements beyond that objective.
What Changes When a Variance Is Exactly Zero?
If \(\sigma_L=0\) and \(\sigma_H>0\), the interior Neyman formula no longer gives a positive target for \(L\): its formal limit is \(n_L=0,n_H=N\). The variance expression’s \(L\)-term is zero for any positive \(n_L\), while assigning more observations to \(H\) reduces \(\sigma_H^2/n_H\). Thus, if one considers only the algebraic variance objective and permits the boundary limit, it favors making \(n_L\) as small as possible.
That limit must not be confused with a usable two-arm experiment. With no participants in arm \(L\), the arm-\(L\) sample mean is unavailable, so the observed treatment contrast cannot be formed from the two arm means. Even if the variance is believed to be zero, the design still needs observations in each arm to estimate its mean from trial data. A positive minimum makes the feasible design explicit. If both variances are exactly zero, the variance expression alone cannot distinguish allocations, but it still does not answer how many observations are needed to establish the arm means or meet other design requirements.
Worked Example: Near-Zero Versus Zero
For a hypothetical trial, let \(N=80\), \(\sigma_L=0.02\) outcome units, and \(\sigma_H=1.00\) outcome units. Suppose each arm must have at least 10 participants. Assume the fixed-sample variance model applies.
The unconstrained share for arm \(L\) is \(0.02/(0.02+1.00)=0.01961\), so its continuous target is \(80(0.01961)=1.5686\), with \(78.4314\) for arm \(H\). This is below the required minimum. The constrained target is therefore \(n_L=10,n_H=70\), with variance
outcome-units squared. If the planning standard deviation for arm \(L\) were exactly zero instead, the same minimum-constrained allocation would have variance \(0/10+1/70=0.0142857\) outcome-units squared. The small difference in the variance objective does not remove the need to enroll participants in arm \(L\). The minimum remains part of the design because the contrast requires observations from both arms, not because the variance formula assigns a large precision benefit to that arm.
Diagnostics, Alternatives, and Practical Limits
A near-boundary target deserves review before it is used. In particular, check whether the small target is driven by a credible population variance or by a noisy pilot estimate, whether the continuous allocation can be rounded while retaining required arm sizes, and whether the intended analysis needs more information than the variance-minimization calculation represents. For binary outcomes, for example, a pilot with no observed events in an arm does not by itself establish that the population outcome variance is exactly zero. More generally, a variance estimate near zero may reflect a small pilot, a restricted range, or data-processing issues rather than a genuinely deterministic outcome.
The main alternatives answer different questions. Unconstrained Neyman allocation minimizes the variance of the specified mean contrast under the assumed variances and fixed total sample size. Equal allocation is simpler and gives both arms the same sample size, but it is not generally variance-minimizing when variances differ. A Neyman target subject to minimum arm sizes preserves the variance objective as far as possible within explicit feasibility constraints. The choice among these is not determined by the variance formula alone: the assumptions, minimums, estimand, and operational needs must be stated.
| Approach | What it does near a zero variance | Key caution |
|---|---|---|
| Unconstrained Neyman target | Assigns a very small share to the low-variance arm; the share tends to zero as its standard deviation tends to zero. | May yield too few participants for a usable two-arm comparison or other design needs. |
| Equal allocation | Keeps arm sizes equal regardless of the variance ratio. | May sacrifice precision under the specified variance objective when the variances differ substantially. |
| Neyman target with minimum sizes | Uses the unconstrained target when feasible and otherwise selects the nearest allowed boundary. | The minimum must be justified by design needs; it is an added constraint, not a consequence of the Neyman formula. |
A common mistake is to say, “The variance in arm \(L\) is almost zero, so that arm contributes almost nothing and can be omitted.” The variance contribution \(\sigma_L^2/n_L\) may indeed be small, but omitting the arm also removes its trial-based sample mean. The correction is to separate the precision objective from the requirement to observe both arms, and to impose and justify a positive minimum when the design requires one.
The target is also only as credible as its variance inputs and assumptions. If observations are dependent, outcome variances change over time, or the actual sampling process does not match the fixed-sample model, the displayed variance calculation may not describe the design’s uncertainty. These are reasons to treat a near-zero target as a diagnostic signal and assess the design under its actual assumptions, not to interpret the continuous ratio mechanically.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- For two arms with standard deviations \(0.25\) and \(1.25\), what are the continuous Neyman targets when \(N=120\)?
- With \(N=100\), standard deviations \(0.10\) and \(1.00\), and a minimum of 20 participants per arm, give the constrained target and its variance.
- If one arm has exactly zero variance and the other has standard deviation \(0.60\), why does the formal variance objective not justify assigning zero participants to the first arm in a two-arm trial?
- A small pilot observes identical outcomes in one arm and gives a sample variance of zero. What conclusion about the population variance is warranted?
Solutions. 1. The standard-deviation ratio is \(0.25:1.25=1:5\), so the shares are \(1/6\) and \(5/6\). The targets are \(120/6=20\) and \(100\) participants. 2. The unconstrained first-arm target is \(100(0.10/1.10)=9.0909\), below 20, so the constrained target is 20 and 80. Its variance is \(0.10^2/20+1.00^2/80=0.000500+0.01250=0.01300\) outcome-units squared. 3. With no observations in that arm, its sample mean and thus the observed two-arm contrast cannot be formed. A positive minimum is needed if the design requires trial observations in both arms. 4. A zero sample variance in a small pilot does not establish that the population variance is zero. It is a pilot result that should be treated cautiously when specifying the allocation target.