Variance Components for Reference-Scaled Analysis

Why Separate the Sources of Variation?

A log-transformed pharmacokinetic endpoint can vary because subjects differ from one another, because the same subject's measurements vary across occasions, or because of both. For reference-scaled analysis, these sources are not interchangeable: the target is the within-subject variability under the reference treatment, not the total spread of measurements across all subjects. A model that keeps the sources distinct makes clear which part of observed variation informs that target.

The earlier material on within-subject reference variance and reference-replicate contrasts establishes why repeated reference observations carry information about this quantity. Here the emphasis is the model: how its terms represent subject, treatment, and residual variation, and how its covariance structure identifies the reference-specific variance.

A Mixed-Effects Model for a Replicate Crossover

Let \(Y_{ijk}\) be the log-transformed endpoint for subject \(j\) in sequence \(i\), observed in period \(k\). Let \(d_{ijk}\) indicate the treatment received in that period, either reference \(R\) or test \(T\). A conventional model is

Y_{ijk} = \mu + \mathrm{Sequence}_i + \mathrm{Subject}_{j(i)} + \mathrm{Period}_k + \mathrm{Treatment}_{d_{ijk}} + \epsilon_{ijk}.

The sequence, period, and treatment terms are fixed effects: they describe systematic mean differences represented in the design. Subject is a random effect, with subjects nested within sequence. In the model below, subjects vary around the overall mean with variance \(\sigma_S^2\). Residual variability depends on the treatment received:

\mathrm{Subject}_{j(i)} \sim N(0,\sigma_S^2), \qquad \epsilon_{ijk} \sim N(0,\sigma^2_{W,d_{ijk}}).

Thus, \(\sigma^2_{W,R}\) is the reference-specific within-subject residual variance, and \(\sigma^2_{W,T}\) is the corresponding test-treatment residual variance. The model assumes subject effects are independent of residuals, and residuals from distinct periods are independent conditional on the modeled effects. Fixed-effect coding must make the mean terms identifiable, for example by using appropriate reference levels or constraints.

Definition: In this model, the reference-specific within-subject variance is \(\sigma^2_{W,R}\), the variance of a reference observation's residual after accounting for the fixed effects and the subject's random effect. Its square root, \(\sigma_{W,R}\), is the reference within-subject standard deviation on the log scale. The subject variance \(\sigma_S^2\) describes between-subject variation and is not part of \(\sigma^2_{W,R}\).

The residual term represents within-subject variation only under the assumptions just stated. If a subject has an additional treatment-specific random response, or if residuals are correlated across periods, this simple model may not separate those features correctly. The interpretation of \(\sigma^2_{W,R}\) therefore depends on the model as well as on the data.

What the Covariance Structure Identifies

For an observation on treatment \(d\), the random subject effect and residual are independent. Conditional on the fixed effects, the observation variance is therefore the sum of the two component variances. For two different periods on the same subject, the shared subject effect induces covariance; independent residuals do not.

\operatorname{Var}(Y_{ijk}) = \sigma_S^2 + \sigma^2_{W,d_{ijk}}, \qquad \operatorname{Cov}(Y_{ijk},Y_{ijk'}) = \sigma_S^2 \quad (k \ne k').

In particular, two reference observations from the same subject each have variance \(\sigma_S^2+\sigma^2_{W,R}\), while their covariance is \(\sigma_S^2\). Their difference removes the shared subject effect. After adjusting for any systematic period difference, its variance is

\operatorname{Var}(Y_{R,1}-Y_{R,2}) = (\sigma_S^2+\sigma^2_{W,R})+ (\sigma_S^2+\sigma^2_{W,R})-2\sigma_S^2 =2\sigma^2_{W,R}.

This identity is the key justification for using replicated reference observations to identify reference-specific within-subject variability. A single reference observation per subject does not supply a within-subject reference contrast. Across subjects, its observed spread combines subject and residual variation; without further design information or assumptions, that total alone does not separate the two components.

Worked Example: Separating Variance and Covariance

Suppose a hypothetical fitted model for a log endpoint gives \(\sigma_S^2=0.0900\), \(\sigma^2_{W,R}=0.0400\), and \(\sigma^2_{W,T}=0.0625\), in squared log units. Calculate the reference observation variance, the covariance between two reference observations on the same subject, and the variance of their period-adjusted difference.

Reference observation variance. The model gives \(0.0900+0.0400=0.1300\) squared log units.

Same-subject reference covariance. The shared random subject effect contributes \(0.0900\) squared log units. The residuals are assumed independent across periods, so they add no covariance.

Reference difference variance. By the contrast identity, \(2(0.0400)=0.0800\) squared log units. The standard deviation of this difference is \(\sqrt{0.0800}=0.2828\) log units.

The test observation variance would instead be \(0.0900+0.0625=0.1525\) squared log units. The larger test observation variance does not change the reference-specific target: it remains \(\sigma^2_{W,R}=0.0400\) squared log units.

Deriving a Reference-Contrast Estimate

The mixed-effects model supplies a general framework, but the same model identity yields a simple moment-based illustration. Suppose each of \(m\) subjects has two reference observations, and let \(D_j\) be the difference after adjusting for the expected period effect. Under the model, \(E(D_j)=0\) and \(\operatorname{Var}(D_j)=2\sigma^2_{W,R}\). If the adjusted differences are independent across subjects, their sample variance \(s_D^2\) estimates \(2\sigma^2_{W,R}\); dividing by two estimates the reference variance.

\widehat{\sigma^2_{W,R}}_{\text{moment}} = \frac{s_D^2}{2}, \qquad s_D^2 = \frac{\sum_{j=1}^{m}(D_j-\bar{D})^2}{m-1}.

This derivation uses the ordinary sample variance of the adjusted contrasts. Since the contrasts have mean zero under the model, \(E(s_D^2)=\operatorname{Var}(D_j)=2\sigma^2_{W,R}\). Dividing by two consequently gives an unbiased moment estimate under these conditions. This is a justification of the contrast relationship, not a claim that this simple calculation is the best estimator for every replicate design or data pattern.

Worked Example: A Moment Estimate From Replicated References

In a hypothetical study, five subjects have two reference observations. After accounting for their period effects, suppose their reference differences \(D_j\) are \(0.10,-0.20,0.30,-0.10,-0.10\) log units. Estimate the reference-specific within-subject variance using the moment calculation.

The mean difference is \(\bar{D}=(0.10-0.20+0.30-0.10-0.10)/5=0.00\) log units. The sum of squared deviations is \(0.10^2+(-0.20)^2+0.30^2+(-0.10)^2+(-0.10)^2=0.1600\) squared log units.

Thus \(s_D^2=0.1600/(5-1)=0.0400\) squared log units, and \(\widehat{\sigma^2_{W,R}}_{\text{moment}}=0.0400/2=0.0200\) squared log units. The corresponding estimated within-subject standard deviation is \(\sqrt{0.0200}=0.1414\) log units.

This calculation assumes the differences have been adjusted for systematic period effects and that subjects' differences are independent with a common variance. Five subjects provide little precision; the estimate illustrates the decomposition, not a regulatory conclusion.

Design Requirements, Diagnostics, and Alternatives

The reference-specific variance is identified by replicated reference exposure within subjects, together with a design that allows treatment, period, and sequence effects to be distinguished. In a full replicate design, both treatments are repeated within subjects. In a partial replicate design, the reference is repeated while the test may be observed once. Either can provide repeated reference information, but the available treatment-specific replication differs. With only one reference observation per subject, the reference within-subject variance is not directly identified from reference contrasts.

Before interpreting a fitted component, check that the design actually contains the relevant repeats and that the model matches the assignment structure. Useful diagnostics and warning signs include:

Residual structure: Plot residuals by treatment, period, and sequence. Large systematic patterns suggest that fixed mean effects may be missing or incorrectly represented.
Unequal spread: Compare residual spread under reference and test. A single common residual variance can obscure treatment-specific variability when the spreads differ.
Within-subject dependence: Examine replicate contrasts and residual patterns across periods. Serial correlation or a subject-by-treatment response pattern can violate the independent-residual model.
Boundary estimates: A fitted variance component of zero can indicate weak information, limited replication, or a genuinely small component. It does not prove that the underlying variance is exactly zero.
Design imbalance or missing observations: These may reduce information or make some components difficult to distinguish. Check that each relevant treatment and period pattern is represented before relying on an estimate.

A direct reference-replicate contrast is transparent when subjects have comparable, complete replicate data and period effects can be handled appropriately. A mixed-effects model is more flexible for representing the fixed design effects and the shared subject covariance, and can use data that are not perfectly balanced. Its flexibility depends on its assumptions: a sophisticated fit does not compensate for a design without reference replication or for an incorrectly specified covariance structure. Neither approach makes the subject variance itself a substitute for reference within-subject variance.

Worked Example: Spotting a Non-Identifiable Target

A hypothetical two-period study gives each subject reference treatment \(R\) in period 1 and test treatment \(T\) in period 2. Each subject has only one observation on each treatment. A colleague proposes estimating \(\sigma^2_{W,R}\) as the across-subject variance of the period-1 reference observations. Is that justified by the model?

No. For a reference observation, the model variance across subjects is \(\sigma_S^2+\sigma^2_{W,R}\). The observed variance therefore combines the between-subject component with the reference residual component; it is not, by itself, an estimate of \(\sigma^2_{W,R}\).

There is also no within-subject reference pair: each subject contributes only one \(R\) observation. The reference contrast with variance \(2\sigma^2_{W,R}\) cannot be formed. Moreover, treatment is tied to period in this design, so treatment and period mean effects cannot be separated without additional assumptions. The proposed across-subject calculation consequently does not identify the reference-specific within-subject variance.

The correction is to use a replicate design that repeats reference within subjects and to model the period and treatment effects according to the design. Even then, the variance estimate relies on the residual-independence and random-subject assumptions rather than following from replication alone.

Common Pitfall and Practical Limits

A frequent mistake is to call the variance of all reference observations “the within-subject reference variance.” In the stated model, the across-subject variance of a single reference measurement is \(\sigma_S^2+\sigma^2_{W,R}\), whereas the within-subject reference variance is only \(\sigma^2_{W,R}\). The correction is to use the replicate covariance or contrast structure: the shared subject effect cancels in a within-subject reference difference.

The model also simplifies reality. It assumes a common reference residual variance across subjects and periods, independent residuals conditional on subject and fixed effects, and no extra subject-specific treatment response beyond the random intercept. If these assumptions are implausible, the fitted reference component may not represent the intended variability. Model diagnostics, design review, and sensitivity to reasonable alternative covariance assumptions are therefore part of interpretation, not optional decoration.

Key takeaway: In a random-intercept crossover model with treatment-specific residual variances, \(\sigma^2_{W,R}\) is identified by replicated reference observations because the subject effect is shared and cancels in within-subject reference contrasts. It is distinct from both total reference observation variance and between-subject variance.

Check Your Understanding

Answer each question, then compare with the short solutions below.

If \(\sigma_S^2=0.16\) and \(\sigma^2_{W,R}=0.05\), what are the variance of one reference observation and the variance of a pairwise reference difference?
Why does the covariance of two reference observations from the same subject equal \(\sigma_S^2\) under the model?
What design feature is required to estimate the reference-specific within-subject variance directly from reference data?
Suppose the across-subject variance of reference observations is \(0.21\) squared log units. Can this value alone be reported as \(\sigma^2_{W,R}\)? Explain.
Name one assumption whose failure could make the model's residual variance a misleading description of within-subject variability.

Solutions. (1) The single-observation variance is \(0.16+0.05=0.21\) squared log units; the reference-difference variance is \(2(0.05)=0.10\) squared log units. (2) Both observations share the same subject random effect, while independent period residuals have zero covariance; hence their covariance is the subject variance. (3) Subjects must have replicated reference observations, so a within-subject reference contrast is available. (4) No. Under the model, that across-subject variance combines \(\sigma_S^2\) and \(\sigma^2_{W,R}\); the components cannot be separated from this value alone. (5) For example, correlated residuals across periods, a subject-by-treatment random response, or treatment-dependent spread represented incorrectly by one common residual variance could mislead the interpretation.