Exponential Random Effects for Positive Pharmacokinetic Parameters

Why Use an Exponential Random Effect?

Clearance and volume are positive quantities. A model for individual differences should therefore avoid assigning them negative values. The exponential random-effects parameterization does this while allowing an individual parameter to be either above or below its population typical value.

In the hierarchical model, the population typical value is a fixed-effect quantity, while an individual’s random effect describes that individual’s deviation. For a positive parameter, write the relationship as a multiplicative adjustment: the random effect changes the typical value by a factor rather than by adding an amount in the parameter’s units.

The exponential transformation is defined for every real value of \(\eta_i\), and \(\exp(\eta_i)>0\). Since \(\theta>0\), their product is also positive. Thus this parameterization guarantees \(P_i>0\) without needing to restrict \(\eta_i\) to a particular range. The random effect itself has no units: it operates on a log scale, whereas \(P_i\) and \(\theta\) retain the parameter’s units.

A common model assumption is that individual random effects have mean zero and follow a normal distribution, such as \(\eta_i\sim N(0,\omega^2)\), where \(\omega^2\) is the random-effect variance. Under this assumption, \(\eta_i\) is normal on the log scale and \(P_i\) is lognormally distributed on the parameter scale. The exponential form guarantees positivity even if a different distribution for \(\eta_i\) is used. Normality is an additional distributional assumption, not the reason positivity holds.

From Eta to a Multiplicative Change

Taking logarithms of the parameter equation exposes the role of the random effect:

Equivalently, the ratio \(P_i/\theta\) is \(\exp(\eta_i)\). This gives a direct interpretation: a positive eta makes the individual parameter larger than the typical value, a negative eta makes it smaller, and eta equal to zero leaves it unchanged. The numerical size of an eta should be interpreted through the exponential, not as a percentage by itself.

For instance, \(\eta_i=0.20\) does not mean that the parameter is exactly 20% higher. Its factor is \(\exp(0.20)\), approximately 1.221, so it corresponds to an increase of about 22.1%. Negative and positive eta values of equal magnitude also do not produce equal-sized percentage changes: \(\exp(-a)=1/\exp(a)\), rather than \(1-a\) in general.

What the Parameterization Assumes

The equation has a few distinct requirements and modeling choices. Keeping them separate helps prevent over-interpreting the formula.

• Positive typical value: \(\theta\) must be greater than zero for the parameter to be positive under this form.
• One random effect per modeled individual parameter: \(\eta_i\) is the individual-level deviation for that parameter. In a standard subject-level model, the same individual effect applies across that subject’s observations rather than changing independently at every measurement.
• A distribution for population modeling: to specify a conventional random-effects model, the distribution of eta must be stated. A common choice is \(\eta_i\sim N(0,\omega^2)\). This assumption describes variation across individuals; it does not say that each realized eta must equal zero.
• Meaningful parameter scale: the units belong to \(\theta\) and \(P_i\), not eta. The ratio \(P_i/\theta\) is dimensionless, as required for its logarithm.

A mean-zero normal eta is centered at zero on the log scale. That condition does not, by itself, say that the arithmetic average of the individual parameters equals \(\theta\). The interpretation of typical value, median, and arithmetic mean under lognormality requires care; the key point here is to use \(\theta\) as the reference value in the stated parameterization, not to silently relabel it as an arithmetic mean.

Short Derivation and Practical Checks

The multiplicative interpretation follows directly from the model, without an approximation. Start with \(P_i=\theta\exp(\eta_i)\), divide both sides by the positive \(\theta\), and obtain

Subtracting one and multiplying by 100 expresses the relative difference from the typical value as a percentage:

The positivity result is equally direct: \(\theta>0\), and \(\exp(\eta_i)>0\) for every real \(\eta_i\), so \(P_i=\theta\exp(\eta_i)>0\). No assumption that eta is small is needed for either conclusion.

When calculating or reviewing an individual parameter, check that the typical value has the correct units and is positive, eta is dimensionless, the exponential is applied to eta, and the result has the original parameter’s units. If a percentage change is reported, compute it from the factor as \(100\{\exp(\eta_i)-1\}\%\), not by multiplying eta by 100.

Common Mistakes, Alternatives, and Limitations

• Adding eta directly. The expression \(P_i=\theta+\eta_i\) does not use the exponential parameterization, and it mixes quantities with incompatible units if eta is dimensionless. Use \(P_i=\theta\exp(\eta_i)\) when the intended effect is multiplicative.
• Calling eta a percentage. Eta is a log-scale deviation. Convert it using \(\exp(\eta_i)\) before describing a factor or percentage change.
• Assuming equal positive and negative percentages. Eta values \(+a\) and \(-a\) give reciprocal factors, not percentage changes with equal magnitudes.
• Treating positivity as evidence of a good fit. The form prevents negative parameter predictions, but it does not establish that the chosen pharmacokinetic model, distribution of eta, or fitted variability is appropriate.

An additive random effect on the parameter scale is the main alternative parameterization. It describes deviations in the parameter’s original units and does not automatically guarantee positivity. The exponential form instead describes proportional changes and guarantees positivity, but it imposes a log-scale structure; it may be unsuitable if the scientific variability is naturally additive or if the log-scale distribution poorly represents the data. The choice should reflect the parameter and modeling assumptions, not be made solely because one equation is convenient.

In model diagnostics, investigate implausibly large or small individual parameter estimates, sensitivity to the assumed eta distribution, and patterns in residuals or individual fits that the model does not capture. Such findings do not alone identify the cause: they may reflect a poor structural model, sparse information for individuals, influential observations, or a mismatch in the random-effects specification. The exponential form enforces a mathematical constraint, not a guarantee that estimates are physiologically plausible or well identified.