Precision Of Variance Components

Variability Is Not Precision

In a population PK model, a between-subject variability estimate describes how much individuals differ according to the model. A precision estimate describes how uncertain we are about that estimated amount of variability. These are different quantities: a model can estimate substantial variability imprecisely, or estimate a small amount of variability with relatively little uncertainty. Reporting an omega or sigma estimate without a measure of its precision can obscure that distinction.

Use the following simple model to fix the terminology. Let \(Y_{ij}\) denote the observation for individual \(i\) at observation \(j\), and let \(f_{ij}(\theta,\eta_i)\) denote the model prediction, depending on population parameters \(\theta\) and individual random effects \(\eta_i\). For one random effect and additive residual error, suppose

Here \(\omega\) is the between-subject standard deviation of the random effect, and \(\omega^2\) is its variance. The residual standard deviation is \(\sigma\), and the residual variance is \(\sigma^2\). The random effects are assumed independent across individuals, and the residual errors are assumed independent across observations and independent of the random effects, conditional on the model. These assumptions are illustrative: actual models may use multiple random effects, correlated random effects, or a different residual-error structure. The meaning and units of \(\omega\) and \(\sigma\) depend on the scale on which the model defines them. For example, an omega on log clearance has units of log clearance; it is not a concentration-scale standard deviation.

Some software reports a variance component directly, while other summaries report its square root. State which scale is being reported. If \(\hat{\omega}=0.30\), for example, then the estimated random-effect variance is \(\hat{\omega}^2=0.09\). The value 0.30 is not itself a variance. Precision must also be attached to the reported scale: an SE for \(\hat{\omega}\) is not an SE for \(\hat{\omega}^2\).

Standard Errors and Intervals for Omega and Sigma

For an estimate \(\hat{\psi}\) of a model parameter \(\psi\), its standard error (SE) summarizes the estimated sampling variability of the estimator under repeated datasets and the assumed model. In likelihood-based fitting, a common approximation obtains the covariance matrix of the parameter estimates from the inverse of the observed information matrix, the negative curvature of the log likelihood near its maximum. The square root of the relevant diagonal element is the approximate SE. This approximation is most dependable when the likelihood is well behaved near an interior estimate and the model is adequately identified.

If the fitted parameterization estimates \(\omega\) and \(\sigma\) directly, a basic Wald interval is the estimate plus or minus 1.96 SE for an approximate 95% interval. This uses an approximate normal sampling distribution. Because standard deviations cannot be negative, an interval that extends below zero is a warning that the direct-scale approximation may be unsuitable; it is not a scientifically plausible negative variability estimate.

An approximate confidence interval has a repeated-sampling interpretation: if the procedure and its assumptions were applied to many comparable datasets, about 95% of the resulting intervals would cover the fixed parameter. It is not a statement that the parameter has a 95% probability of lying in this particular interval.

Changing Between Variance and Standard-Deviation Scales

A variance estimate may be available even when the desired report is a standard deviation, or vice versa. Uncertainty does not transform by simply taking the square root of an SE. For a differentiable transformation \(g(\psi)\), the first-order delta method approximates

For a variance \(v=\omega^2\), the standard deviation is \(g(v)=\sqrt{v}\), so \(g'(v)=1/(2\sqrt{v})\). Therefore, away from zero,

This is a local approximation. Near zero, the derivative becomes large and the square-root transformation is strongly nonlinear, so symmetric normal approximations may be particularly unreliable.

The delta-method relationship also explains why an SE must be labelled by parameter scale. An SE of 0.060 for \(\omega^2\) is not an SE of 0.060 for \(\omega\). Confusing these changes both the numerical uncertainty and its units.

Boundary Effects and Alternative Uncertainty Summaries

Variance components are constrained to be nonnegative. When the estimated \(\omega^2\) or \(\sigma^2\) is near zero, the likelihood may be asymmetric, flat in one direction, or maximized at the boundary. The usual normal approximation behind a Wald interval can then perform poorly. An SE alone may also fail to reveal that a likelihood is nearly flat or that estimates from repeated fits frequently land on the boundary.

Three common approaches have different strengths and limitations:

A profile likelihood is formed by fixing the parameter of interest at candidate values and maximizing the likelihood over the remaining parameters at each value. Its shape can show asymmetry or a plateau that a local SE hides. Likelihood-ratio cutoffs commonly used for interior parameters do not automatically apply to a variance component at its boundary: the usual regularity assumptions can fail. A bootstrap can provide another view, but its simulations must reproduce the study design and model, and failed fits or boundary estimates must be reported rather than silently discarded.

Diagnostics, Limitations, and Common Mistakes

Precision estimates depend on both the data and the model. A small number of individuals, few observations per individual, a narrow range of sampling times, or heavy censoring can leave a variance component weakly informed. Between-subject and residual variation may also be difficult to distinguish: when individual profiles are sparse, multiple combinations of omega and sigma can explain similar patterns. In such cases, the fitted point estimates can look plausible while their uncertainty is large or sensitive to modeling choices.

Useful checks include examining whether the information matrix is positive definite and the covariance estimate is finite; checking for unusually large SEs relative to the estimate; and inspecting profile likelihoods or bootstrap distributions when estimates approach zero or uncertainty is visibly asymmetric. Compare fits under scientifically reasonable residual-error assumptions and inspect whether the variance-component estimates change materially. Convergence messages alone do not establish that the model has estimated the variance components precisely. A model can converge to a boundary or to a shallow likelihood maximum.

A frequent mistake is to say, “The estimated omega is 0.30, so between-subject variability is precisely 0.30.” The correction is that 0.30 is a point estimate of the random-effect SD; precision requires an uncertainty summary, such as an SE and an appropriate interval, and the scale and units must be stated. Another mistake is to compare an SE for \(\omega^2\) directly with an estimate of \(\omega\). Those are different parameters on different scales, so convert consistently before interpreting them.

No uncertainty summary repairs a misspecified model. The assumptions about random effects, residual errors, the observation process, and the likelihood still matter. An apparently narrow interval conditional on a model does not account automatically for uncertainty about model structure, data quality, or which residual-error form is scientifically appropriate.