Geometric Interpretation of Contrast Correlation

From Shared Control to Geometry

When two treatments are each compared with the same control, the estimated contrasts share one source of sampling error: the error in the control mean. Algebraically, this shared component creates covariance between the contrasts. Geometrically, it appears as a shared coordinate in their error vectors. Under the standard equal-variance sampling model, the correlation between the contrast errors is the cosine of the angle between those vectors.

The geometry is a way to see the covariance calculation, not a different statistical model. It also helps distinguish the shared part from the treatment-specific parts: increasing the contribution of the control coordinate relative to the other coordinates makes the contrast vectors point more nearly in the same direction.

Model and Contrast-Error Vectors

Suppose there is one control group, labelled \(0\), and treatment groups \(1,\ldots,m\). In group \(j\), let \(\bar{x}_j\) estimate the population mean \(\mu_j\), and define its mean error as \(E_j=\bar{x}_j-\mu_j\). Assume that individual observations are independent within each group and independent across groups, and that every observation has the same finite variance \(\sigma^2\). These assumptions give independent group-mean errors with variances \(v_j=\sigma^2/n_j\), where \(n_j\) is the sample size in group \(j\).

The estimated treatment-control contrast is \(\bar{x}_i-\bar{x}_0\). Its error around the population contrast \(\mu_i-\mu_0\) is therefore \(D_i=E_i-E_0\). To express that error as an ordinary dot product with coordinates of unit variance, define \(Z_j=E_j/\sqrt{v_j}\). The \(Z_j\) have mean zero, variance one, and are uncorrelated with one another. Independence follows from the sampling assumptions; uncorrelatedness and unit variances are what the covariance calculation below requires.

For example, with two treatments the vectors have coordinates \(a_1=(-\sqrt{v_0},\sqrt{v_1},0)\) and \(a_2=(-\sqrt{v_0},0,\sqrt{v_2})\), in the order control, treatment 1, treatment 2. Their control coordinates have the same sign because both contrasts subtract the same control mean. Their treatment coordinates lie on different axes.

Why Correlation Is a Cosine

For vectors \(a_i\) and \(a_k\), the dot product is the sum of products of corresponding coordinates. Since the coordinates of \(Z\) have variance one and are uncorrelated, the covariance of the two contrast errors is exactly that dot product. The squared length of each vector is the variance of its contrast error.

Here \(\theta_{ik}\) is the angle between the two nonzero vectors. The last equality is the definition of the cosine of the angle between vectors. In this shared-control setting, the only nonzero coordinate they share is the control coordinate, so \(a_i^\mathsf{T}a_k=v_0\) for distinct treatments \(i\) and \(k\). Their squared lengths are \(v_0+v_i\) and \(v_0+v_k\). Thus the geometric expression reproduces the usual correlation formula:

The second expression uses the common-variance assumption, which cancels \(\sigma^2\). This angle interpretation concerns the coefficient vectors in whitened error space. It is not the angle between the observed group means, nor does it depend on the particular numerical values of the observed contrasts.

Allocation Changes the Angle

With unequal sample sizes, the group-mean errors have unequal variances, so the vector coordinates have unequal lengths. A larger control variance lengthens the shared coordinate in both vectors. Larger treatment variances lengthen their separate coordinates. The cosine formula shows how these changes alter the angle without requiring a drawing.

What the Geometry Reveals—and What It Does Not

The angle gives an immediate visual interpretation of the correlation: a small angle corresponds to a cosine near one, a right angle to zero, and an obtuse angle to a negative cosine. For two treatment-versus-control contrasts with the sign convention used here, the dot product is \(v_0>0\); consequently their angle is acute and their correlation is positive. In this particular setting, reducing the relative contribution of the shared control error makes the vectors less aligned.

The argument is a short linear-algebra justification, not a normal-theory result. It uses finite variances and uncorrelated standardized group-mean errors; it does not require the errors to have a normal distribution. The geometric calculation describes covariance and correlation. It does not, by itself, establish the joint probability distribution of the contrast statistics.

In practice, the geometry is useful in three ways. First, it makes the source of dependence visible: the control coordinate is shared, whereas each treatment coordinate is specific to one contrast. Second, it provides a check on calculations. The dot product must reproduce the covariance, and each vector's squared length must reproduce its contrast variance. Third, it shows why marginal standard errors alone do not specify the relationship between two contrasts: their directions, not just their lengths, matter.

Assumptions, Diagnostics, and Alternatives

Before using this representation, check that the group means have the variances assigned to their coordinates and that errors from different groups are uncorrelated. Independence of observations within each group is one standard way to justify \(\operatorname{Var}(\bar{x}_j)=\sigma^2/n_j\); independence across groups justifies zero covariance between distinct group-mean errors. If observations within a group are correlated, the sample mean need not have variance \(\sigma^2/n_j\), and the coordinates based on that formula are not justified. The stated geometry also uses a common individual variance. With unequal variances, a more general covariance calculation is needed rather than this equal-variance version.

A useful diagnostic is to compute the covariance directly from the contrast definitions, then compare it with the vector dot product. Also check that each vector's squared length equals the corresponding contrast variance and that the resulting correlation lies between \(-1\) and \(1\). A mismatch can signal a sign error, a missing variance contribution, or an incorrectly specified sampling variance. A correlation outside that range is impossible for a valid covariance and variance calculation.

The main alternative is to calculate covariance and variance algebraically, then divide covariance by the product of standard deviations. That route is compact and often easiest for a single pair of contrasts. The vector method is a complementary representation: it turns the same calculation into a dot product and makes shared and contrast-specific error components visible. Neither method repairs a wrong sampling model. In particular, the simple vectors above are limited to treatment-control contrasts under the stated equal-variance assumptions; they do not cover every contrast structure or every covariance pattern.