Allocation Ratios and Contrast Dependence

Why Allocation Ratios Matter

When several treatments are compared with the same control group, every treatment-control contrast includes the control sample mean. That shared quantity links the contrasts: a control mean that happens to be high shifts each treatment-minus-control contrast downward. The strength of this dependence is affected not only by sharing a control, but also by how many observations are allocated to each treatment relative to the control.

This tutorial studies that effect under a common-variance model. The key result is a simple formula: the correlation depends on the treatment-to-control allocation ratios, not on the absolute group sizes separately. The result describes dependence between contrasts; it does not by itself tell us how precise the contrasts are or how much power a trial has.

Model, Ratios, and Contrast Correlation

Consider two treatment groups, indexed by 1 and 2, and one control group, indexed by 0. Let \(n_i\) be the sample size in group \(i\), and let \(\bar{X}_i\) be its sample mean. Assume observations are independent within and between groups, and that each group has a finite mean \(\mu_i\) and the same variance \(\sigma^2\). The group means may differ; equality of means is not required. For the correlation result, require \(\sigma^2>0\), and require all three sample sizes to be positive.

Definition: The treatment-to-control allocation ratio for treatment \(i\) is \(r_i=n_i/n_0\). The treatment-control contrast is \(D_i=\bar{X}_i-\bar{X}_0\). The correlation between the two contrasts is their covariance divided by the product of their standard deviations.

Under the stated assumptions, the group means are independent across groups and have variances \(\sigma^2/n_i\). Consequently, the covariance between \(D_1\) and \(D_2\) comes from the control mean shared by both contrasts. Their variances also include the variability of their respective treatment means.

\operatorname{Cov}(D_1,D_2)=\frac{\sigma^2}{n_0}, \qquad \operatorname{Var}(D_i)=\sigma^2\left(\frac{1}{n_i}+\frac{1}{n_0}\right).

Since \(\sigma^2>0\) and the sample sizes are positive, both contrast variances are positive, so their correlation is defined. Substituting the covariance and variances into the definition gives the allocation-ratio formula.

\begin{aligned} \operatorname{Corr}(D_1,D_2) &=\frac{\sigma^2/n_0} {\sqrt{\sigma^2(1/n_1+1/n_0)\,\sigma^2(1/n_2+1/n_0)}}\\ &=\frac{\sqrt{n_1n_2}}{\sqrt{(n_0+n_1)(n_0+n_2)}}\\ &=\sqrt{\frac{r_1r_2}{(1+r_1)(1+r_2)}}. \end{aligned}

The final equality follows by dividing each \(n_i\) in the middle expression by \(n_0\). Because the same positive scale factor appears in numerator and denominator, multiplying all three group sizes by the same amount leaves the correlation unchanged.

Worked Example: Equal Allocation to All Three Groups

Suppose a hypothetical trial allocates \(n_0=50\) participants to control, \(n_1=50\) to treatment 1, and \(n_2=50\) to treatment 2. Assume independent observations and a common variance \(\sigma^2>0\). The treatment-to-control ratios are \(r_1=50/50=1\) and \(r_2=50/50=1\).

Using the ratio formula,

\operatorname{Corr}(D_1,D_2) =\sqrt{\frac{1\cdot1}{(1+1)(1+1)}} =\sqrt{\frac{1}{4}} =0.5000.

Thus, under this model, the two treatment-minus-control sample mean contrasts have correlation \(0.5000\). The same result follows from the group-size formula: \(\sqrt{50\cdot50}/\sqrt{(50+50)(50+50)}=50/100=0.5000\). This correlation does not mean that either contrast has a standard error of 0.5; it is a dimensionless measure of their joint variation.

How Unequal Ratios Change Dependence

With equal treatment-to-control ratios, \(r_1=r_2=r\), the formula simplifies to a useful special case. In particular, equal group sizes give \(r=1\), while treatment groups twice the size of control give \(r=2\).

\operatorname{Corr}(D_1,D_2) =\sqrt{\frac{r^2}{(1+r)^2}} =\frac{r}{1+r}, \qquad r>0.

For unequal ratios, each contrast contributes its own factor. A treatment group with a large sample relative to control has a more precisely estimated treatment mean. Its contrast therefore contains a larger relative contribution from the common control mean, which can make its correlation with the other shared-control contrast larger.

Worked Example: Unequal Treatment-to-Control Ratios

Suppose a hypothetical study uses \(n_0=80\) control participants, \(n_1=40\) participants in treatment 1, and \(n_2=120\) in treatment 2. Assume independent observations, a common variance \(\sigma^2>0\), and positive group sizes. Then \(r_1=40/80=0.5\) and \(r_2=120/80=1.5\).

The correlation is

\begin{aligned} \operatorname{Corr}(D_1,D_2) &=\sqrt{\frac{0.5(1.5)}{(1+0.5)(1+1.5)}}\\ &=\sqrt{\frac{0.75}{3.75}} =\sqrt{0.2} =0.4472. \end{aligned}

Checking directly from the group sizes gives \(\sqrt{40\cdot120}/\sqrt{(80+40)(80+120)}=\sqrt{4800}/\sqrt{24000}=\sqrt{0.2}\), the same result. The contrasts are positively correlated, with correlation about \(0.4472\). They need not have the same precision: treatment 1 has fewer participants than treatment 2, so their contrast variances differ.

Why the Correlation Increases with a Ratio

The formula also makes the direction of the allocation effect explicit. Holding \(r_2\) fixed, increasing \(r_1\) increases the correlation. To see this without relying only on examples, take the logarithm of the positive correlation formula and differentiate with respect to \(r_1\):

\frac{\partial}{\partial r_1}\log\{\operatorname{Corr}(D_1,D_2)\} =\frac{1}{2r_1}-\frac{1}{2(1+r_1)} =\frac{1}{2r_1(1+r_1)}>0.

The correlation itself is positive, so a positive derivative of its logarithm means the correlation increases as \(r_1\) increases. By symmetry, the same argument applies to \(r_2\). This monotonicity is about changing a treatment-to-control ratio while holding the other ratio fixed; it should not be confused with changing sample sizes under any arbitrary fixed-total constraint.

The limiting values clarify the interpretation. As one ratio approaches zero while the other stays positive, the formula approaches zero. As both ratios become large, the correlation approaches one. In the latter case, the treatment means are relatively precise compared with the common control mean, so the shared control contributes a large fraction of the variation in both contrasts. These are limiting descriptions; in an actual design the group sizes are positive and finite.

Worked Example: Comparing Equal and Treatment-Heavy Allocation

Compare two hypothetical allocation plans under the same independent, common-variance model with \(\sigma^2>0\). In Plan A, each treatment group and the control group have 60 participants. In Plan B, control has 60 participants and each treatment group has 120.

For Plan A, \(r_1=r_2=60/60=1\), so

\operatorname{Corr}_A(D_1,D_2)=\frac{1}{1+1}=0.5000.

For Plan B, \(r_1=r_2=120/60=2\), so

\operatorname{Corr}_B(D_1,D_2)=\frac{2}{1+2}=0.6667.

The treatment-heavy plan has a higher correlation between its two contrasts. This does not mean that Plan B is less informative overall: its larger treatment groups can reduce treatment-mean variability. Correlation describes how the contrasts vary together, whereas their variances describe their individual precision. Those are related but distinct features of a design.

Comparing Allocation Patterns

The formula helps compare design patterns, provided the shared-control and common-variance assumptions remain appropriate. Equal ratios produce a compact calculation; unequal ratios require both ratios rather than an average ratio. The table summarizes these distinctions.

Allocation pattern	Ratios	Correlation	What to keep in mind
Equal group sizes	\(r_1=r_2=1\)	\(0.5000\)	A benchmark, not a universal value for shared-control contrasts.
Equal, treatment-heavy allocation	\(r_1=r_2=r\)	\(r/(1+r)\)	Increasing both ratios increases the correlation.
Unequal treatment ratios	\(r_1\ne r_2\)	\(\sqrt{r_1r_2/[(1+r_1)(1+r_2)]}\)	Calculate using each comparison's own ratio.

An alternative to reasoning in ratios is to calculate directly from the three group sizes. That route can be a useful arithmetic check, as in the unequal-ratio example. The ratio form is usually clearer for comparing allocation plans because it separates the relative allocation from the overall scale. Neither calculation replaces checking the assumptions: the formula is for two treatment-control contrasts that use the same control observations, under independent groups and a common positive variance.

Diagnostics, Pitfalls, and Limitations

Before applying the formula, check that each comparison uses the same control group and that \(n_0,n_1,n_2\) are positive. Confirm that the ratios are defined as treatment size divided by control size; reversing that convention changes the numbers entered into the formula. Under the stated model, a calculated correlation must lie between zero and one. A negative result or a value above one signals an algebra, data-entry, or model-setup error.

The common variance must be positive. If \(\sigma^2=0\), all observations are constant under this model, both contrasts have zero variance, and their correlation is undefined—not zero and not the value obtained by cancelling \(\sigma^2\) in the formula. Correlation is defined only when both variables have positive variance.

Pitfall: treating a larger correlation as lower precision. That conclusion does not follow. Correlation measures co-movement; contrast variance and standard error measure individual uncertainty.
Pitfall: using the absolute group sizes as if they alone determine correlation. Under this model, multiplying all sizes by the same positive constant leaves the correlation unchanged; the treatment-to-control ratios determine it.
Pitfall: using one ratio for both contrasts when allocations differ. Compute \(r_1=n_1/n_0\) and \(r_2=n_2/n_0\) separately, then substitute both.

The formula is specific to the stated shared-control structure and common-variance model. If group variances differ, the equal-variance expression need not hold. Also, knowing the correlation alone does not supply the means, standard errors, power, or the covariance structure needed for every analysis. It is one design feature to interpret alongside those quantities.

Key takeaway: With independent groups and common positive variance, the correlation between two treatment-minus-shared-control sample mean contrasts is \(\sqrt{r_1r_2/[(1+r_1)(1+r_2)]}\), where each \(r_i\) is that treatment's sample size divided by the control sample size. Increasing either ratio while holding the other fixed increases the correlation.

Check Your Understanding

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

A hypothetical design has \(n_0=40\), \(n_1=80\), and \(n_2=80\). What are the ratios and the contrast correlation?
If both treatment-to-control ratios are \(0.5\), calculate the correlation.
For fixed \(r_2>0\), does increasing \(r_1\) increase, decrease, or leave unchanged the correlation? Give a mathematical reason.
Why is the correlation undefined if the common variance is zero?
Two designs have the same treatment-to-control ratios, but one has every group four times as large. Do they have the same contrast correlation under this model? Does that imply equal contrast variances?

Solutions. (1) \(r_1=80/40=2\) and \(r_2=2\). The correlation is \(\sqrt{2\cdot2/[(1+2)(1+2)]}=2/3=0.6667\). (2) \(\sqrt{0.5\cdot0.5/[(1.5)(1.5)]}=\sqrt{0.25/2.25}=1/3=0.3333\). (3) It increases, because the derivative of the log correlation with respect to \(r_1\) is \(1/[2r_1(1+r_1)]>0\). (4) Zero common variance makes each contrast variance zero, and correlation divides by the product of contrast standard deviations; that denominator is zero. (5) Yes. The ratios, and therefore the correlation formula, are unchanged. No: multiplying the sample sizes changes the contrast variances, so equal correlation does not imply equal precision.