Within-Subject Reference Variance in Replicate Data

Why the Source of Variability Matters

Repeated reference administrations provide information that a single reference observation per subject cannot: they allow variability between subjects to be distinguished from variability within a subject across reference occasions. That distinction matters in reference-scaled bioequivalence because the relevant variability is specific to the reference treatment, not simply the spread of all observed concentrations or exposure values.

The endpoint is commonly analyzed on a logarithmic scale. The variance discussed here is therefore a variance of log-transformed endpoint values, with units of squared log units. This tutorial focuses on identifying that reference-specific variance from repeated reference observations. It does not convert the variance to a within-subject coefficient of variation or apply a bioequivalence decision rule.

A Model That Separates the Sources

Let \(Y_{ijR}\) denote the log-scale endpoint for subject \(i\) on reference occasion \(j\). A simple repeated-reference model is

Here, \( \mu_R \) is the population mean response under reference, \(b_i\) is subject \(i\)’s deviation from that mean, and \( \varepsilon_{ijR} \) is the occasion-specific deviation for that subject under reference. Assume \(b_i\) has mean zero and variance \( \sigma_B^2 \), and that the occasion deviations have mean zero and variance \( \sigma_{WR}^2 \). The subject effects and occasion deviations are independent; occasion deviations from distinct reference administrations are also independent and have the same variance.

The subscript \(B\) marks between-subject variability: some subjects tend to have higher responses than others. The subscript \(WR\) marks within-subject reference variability: the same subject’s responses can differ across reference occasions. These components have different meanings even though both contribute to the observed spread.

If the study has period effects, the model should include them, for example as fixed period terms. In the equations and simple calculation below, \(Y_{ijR}\) is understood to have been adjusted for modeled period effects. When the design gives each subject the same two reference periods, a common period difference shifts paired differences by a constant; centering those differences removes that shift. With differing period patterns or subject-specific period effects, an appropriate model adjustment is needed rather than assuming the raw differences are comparable.

Why Repetition Identifies the Within-Subject Component

Under the model, a single reference observation varies both because subjects differ and because an individual’s response varies from one reference occasion to another. Repeated observations on the same subject share the same \(b_i\), so the covariance between two adjusted reference responses from that subject is

The variance of one response is \( \sigma_B^2+\sigma_{WR}^2 \). Subtracting the two responses for the same subject removes \(b_i\), because that subject effect appears in both:

Independence and equal variance of the two occasion errors then give the key identity:

This explains the information supplied by replication: the repeated-reference difference does not contain the stable between-subject effect. If \(d_i=Y_{i1R}-Y_{i2R}\) for each of \(n\) subjects, and the assumptions above hold, the sample variance \(s_d^2\) estimates the variance of those differences. Thus \(s_d^2/2\) estimates the reference-specific within-subject variance.

Under independent normally distributed occasion errors, \(d_i\) are normally distributed with variance \(2\sigma_{WR}^2\). Consequently, \(s_d^2/2\) is unbiased for \( \sigma_{WR}^2 \), and \(s_d^2\) has \(n-1\) degrees of freedom when estimating the difference variance. Without normality, the variance identity still follows from the stated second-moment assumptions, but exact small-sample distributional claims need not hold.

Distinguishing Marginal, Between-Subject, and Pooled Variability

The marginal variance of a single reference response is

It is therefore generally incorrect to treat the variance across all reference observations as the within-subject reference variance. That overall spread includes both sources. Conversely, the variability among subject averages is influenced by between-subject variation and by the amount of within-subject noise averaged into each subject’s mean. The level of analysis must match the variance being discussed.

A pooled residual variance from a model containing both test and reference observations is also not automatically \( \sigma_{WR}^2 \). A model may impose one common residual variance across treatments, in which case that pooled quantity estimates a common variance only if the common-variance assumption is reasonable. If residual variability differs by treatment, pooling combines treatment-specific variability and does not isolate reference variability.

Diagnostics, Alternatives, and Limitations

The calculation is simple, but its interpretation depends on the design and model. Useful checks include verifying that each included subject has the intended repeated reference observations; confirming that the endpoint, transformation, and units match across occasions; examining paired differences for outliers or unusual patterns; and checking whether differences depend on period, sequence, or other design features. A strong period pattern, carryover, or correlation between occasion errors can make the simple \(s_d^2/2\) estimate misleading.

The paired-difference calculation is a direct approach for complete pairs under the stated assumptions. A mixed-effects model is a more flexible alternative when there are additional periods, incomplete observations, or modeled period and sequence effects. Such a model can estimate a reference-specific residual component while accounting for the design, but its interpretation depends on the model specification. A model that imposes one residual variance across treatments does not become reference-specific merely because reference observations are included.

The estimate can also be imprecise when few subjects have repeated reference observations. Sampling variation may make a variance estimate small or large relative to the underlying value; an estimate of zero in a fitted variance-components model, for example, does not prove that the true variability is exactly zero. Normality is particularly relevant for exact variance-based inference in small samples, and independence is essential to the factor-of-two identity. If repeated occasion errors have covariance \(c\), then their difference variance is \(2\sigma_{WR}^2-2c\), not \(2\sigma_{WR}^2\).

Common Mistake

Wrong reasoning: “The standard deviation of all observed reference values is the reference within-subject standard deviation because all the values are reference measurements.” This ignores the fact that subjects can have different stable response levels.

Correction: All-reference data can still contain both between-subject and within-subject variation. Repeated observations identify the within-subject component by comparing reference responses within the same subject, under a model that accounts for the study design. Likewise, a pooled treatment residual is reference-specific only if its model and assumptions justify that interpretation.