Covariate Balance and Population Information

Why the Joint Covariate Distribution Matters

A population design can include many subjects and still provide weak information about covariate effects. The issue is not only how many subjects have each covariate value, but also how the covariates occur together. If two covariates tend to move together, their effects may be difficult to distinguish. If their joint distribution is well spread, the design can provide more distinct information about each effect.

This tutorial studies that issue with two covariates coded as \(-1\) or \(+1\). A deliberately simple local model makes the allocation effect visible without requiring a full pharmacometric model. The calculations concern information for a typical value and two covariate effects; they do not imply that covariate balance alone guarantees a good design for every population parameter.

The example model assumes that the mean is linear in the coded covariates, the residual variance is the same for all subjects, and subject responses are independent. The covariates are treated as fixed design values. For the numerical examples, take \(\sigma^2=4\), so each subject contributes one quarter of the outer product of its design row to the information matrix. This is a transparent special case of a local information calculation: in a more general model, subject-level contributions can differ, but the same design question remains—how do their covariate sensitivities combine?

Balanced Allocation: An Orthogonal Reference

Suppose 40 hypothetical subjects are allocated equally to the four possible covariate patterns. With 10 subjects in each cell, both covariates have mean zero, and their cross-product sum is zero. Their squared-value sums are each 40 because every coded value has square 1.

This reference design is balanced both marginally and jointly. Each covariate has equal numbers at its two levels, and every combination occurs equally often. The resulting columns of \(X\) are orthogonal: no column can be represented by a linear combination of the others. That property explains the diagonal information matrix, rather than balance being a guarantee that every possible model parameter is estimated precisely.

Centering Separates the Typical Value From the Effects

When covariate means are not zero, the intercept \(\theta_0\) refers to the response at covariate values zero, which may not describe a typical subject. Define the sample covariate means as \(\bar{x}_1\) and \(\bar{x}_2\), and center each covariate by subtracting its mean. The corresponding typical value is \(\alpha=\theta_0+\beta_1\bar{x}_1+\beta_2\bar{x}_2\). The slopes are unchanged; the intercept is now the mean-model value at the observed covariate means.

The centered information matrix has a useful form. Write \(S_{jk}=\sum_i(x_{ij}-\bar{x}_j)(x_{ik}-\bar{x}_k)\). The intercept column of the centered design matrix is a column of ones, while each centered covariate column sums to zero. Therefore its cross-product with the intercept column is zero, and

This is a reparameterization of the same model, not a change to the observations or an increase in information. It makes the typical value orthogonal to the effects in this parameterization. The two effects can still be correlated with each other if \(S_{12}\ne 0\). In addition, the information for the effects after accounting for the typical value is determined by the centered cross-products, not simply by the uncentered covariate sums of squares.

Imbalance Can Couple the Covariate Effects

Consider another hypothetical allocation: 20 subjects have pattern \((+1,+1)\), 10 have \((+1,-1)\), and 10 have \((-1,+1)\); there are no subjects with \((-1,-1)\). Each covariate has mean \(0.5\). The covariates are not independent in this allocation, even though their uncentered cross-product sum happens to be zero. Centering is essential for assessing their association around the observed means.

The arithmetic also illustrates why it is unsafe to infer balance from an uncentered cross-product alone. In this allocation the raw cross-product is zero, but both covariate means are nonzero. Subtracting \(n\bar{x}_1\bar{x}_2=10\) gives a negative centered cross-product. The centered quantities describe variation around the actual population represented by the design.

Marginal Balance Does Not Ensure Joint Balance

Equal numbers at each level of each covariate are not sufficient to ensure that the effects are distinguishable. A design can be perfectly balanced on both margins but place subjects only in two diagonal cells. In that case the covariates move together exactly: observing one reveals the other. The model can then estimate only a combination of the two effects.

Diagnostics, Alternatives, and Limits

For a two-covariate design, inspect the joint cell counts as well as the marginal counts. In a local information calculation, also inspect the centered cross-products and the associated correlation. A large absolute correlation indicates that the two covariate columns point in similar or opposite directions, so separating their effects may be difficult. A zero or nearly zero determinant of the effect-information block signals non-identifiability or weak information in at least one direction. These checks should be performed after confirming that covariates were coded and centered as intended.

There are several useful design alternatives, each answering a different question:

• Marginal balancing makes the individual covariate distributions comparable across levels, but it does not control dependence between covariates. The diagonal-cell example shows its limitation.
• Balanced factorial allocation spreads subjects across all joint patterns. In the equal-weight, two-level example, equal allocation makes the covariate columns orthogonal and gives equal information for the two effects.
• Model-based allocation comparison evaluates the information matrix under the intended model and compares the precision or other chosen design criterion. This can accommodate unequal subject contributions, but the result depends on those modeled contributions and assumptions.

Balanced allocation is not automatically optimal when covariate patterns have different costs, are rare in the target population, or yield different response sensitivities. In a pharmacometric setting, a subject's contribution may depend on the covariate values through the model and may also depend on their planned observations. The equal-\(\sigma^2\) calculation here isolates the geometry of the covariates; it is not a substitute for evaluating the relevant population information under the planned model and sampling design.

A common mistake is to reason, “Both covariates have half their subjects at each level, so both effects must be well estimated.” That conclusion uses marginal counts but ignores the joint patterns. If only \((+1,+1)\) and \((-1,-1)\) occur, the effects are inseparable despite perfectly balanced margins. Check whether all relevant directions in the design are supported; centering changes the meaning of the intercept but cannot create information absent from the joint allocation.