Why Model Hepatic Biomarkers Continuously?
Albumin, bilirubin, and the international normalized ratio (INR) are commonly available clinical measurements, but they describe different aspects of hepatic status and are not interchangeable measures of drug-metabolizing capacity. Albumin reflects, among other things, hepatic synthesis; bilirubin reflects production, uptake, conjugation, and excretion; INR reflects coagulation and can also be influenced by vitamin K status, anticoagulant treatment, and other factors. Their values may help explain variability in a pharmacokinetic parameter, but a covariate relationship is an empirical model, not proof that the biomarker itself causes a change in drug clearance.
A continuous model retains differences among individuals who would otherwise be grouped into broad categories. The model must still choose how a difference in a biomarker translates into a difference in a PK parameter. A straight-line change, a constant change on the log scale, and a constant proportional change are distinct assumptions. The data and scientific context should guide that choice; no one form is automatically appropriate for every biomarker or drug.
Three Common Covariate Forms
Let \(P_i\) denote a positive PK parameter for individual \(i\), and let \(x_i\) be one biomarker. The following forms express different relationships between \(x_i\) and \(P_i\). Each is usually considered alongside the same structural PK model, random-effects structure, and residual-error model, so comparisons focus on the covariate form rather than unrelated changes to the analysis.
Log-linear form: \(P_i=P_{\mathrm{ref}}\exp\{\gamma(x_i-x_{\mathrm{ref}})\}\), where \(\gamma\) is the change in \(\log(P_i)\) per one-unit increase in \(x\).
Power form: \(P_i=P_{\mathrm{ref}}(x_i/x_{\mathrm{ref}})^\theta\), where \(\theta\) is an elasticity: a 1% relative increase in \(x\) corresponds approximately to a \(\theta\)% relative change in \(P\), locally.
The linear form changes the parameter by a fixed absolute amount per biomarker unit. Its predictions can become zero or negative when extrapolated far enough, which is unsuitable for a parameter that must be positive. The log-linear form ensures \(P_i>0\) and gives a constant multiplicative change in \(P\) for each fixed additive change in \(x\). The power form also ensures positivity when \(x_i>0\), and gives a constant multiplicative response to a fixed proportional change in \(x\).
The reference value affects the interpretation and numerical scale of the model, but not the fitted relationship when the parameterization is otherwise equivalent. For albumin, a reference such as 40 g/L can make the intercept clinically legible. Bilirubin may be measured in mg/dL or μmol/L; INR is dimensionless. The selected units and reference values must be stated because the numerical coefficient changes when units or scaling change.
Worked Example: A Linear Albumin Effect
Suppose a hypothetical population PK analysis estimates clearance as a linear function of albumin:
where albumin is measured in g/L, 40 g/L is the reference, and the coefficient means a decrease of 0.150 L/h in predicted clearance per 1 g/L increase in albumin. At albumin 30 g/L:
At albumin 50 g/L, the prediction is \(10.0-0.150(10)=8.50\) L/h. These calculations describe the fitted association conditional on the model; they do not establish a causal effect of albumin. The form is only credible across a range in which predictions remain physiologically and empirically plausible. For example, extrapolating the equation to sufficiently high albumin would eventually predict negative clearance, a clear warning against unconstrained extrapolation.
Worked Example: A Power Effect of Bilirubin
Consider the hypothetical model
with bilirubin in mg/dL and the reference value 1.0 mg/dL. At 4.0 mg/dL, the ratio to reference is 4, so:
A doubling of bilirubin multiplies predicted clearance by \(2^{-0.25}=0.8409\), a decrease of about 15.9%, regardless of the starting bilirubin value within the model's range. The negative exponent represents an inverse fitted association. It does not establish that bilirubin itself is the mechanism that reduces clearance. Because the power model requires positive covariate values, a zero or nonpositive value cannot be inserted; the data definition and measurement scale must be checked before using this form.
Log-Linear Effects and Coefficient Interpretation
In a log-linear model, the parameter changes by a constant proportion for a fixed additive biomarker difference. If \(x\) increases by \(\Delta x\), then the ratio of the new prediction to the original prediction is \(\exp(\gamma\Delta x)\). Thus, \(\gamma\) itself is not generally the exact percentage change for a one-unit increase; the exact percentage change is \(100\{\exp(\gamma)-1\}\%\). This distinction matters when the coefficient is not close to zero.
Worked Example: An INR Log-Linear Effect
Suppose a hypothetical model relates clearance to INR by
The predicted clearance at INR 1.50 is:
For a one-unit increase in INR, the model's predicted clearance ratio is \(\exp(-0.300)=0.7408\), corresponding to an exact decrease of \(100(0.7408-1)=-25.92\%\). That is a model-based association, not a direct interpretation of INR as a measure of metabolic activity. The model also predicts clearance at INR 1.00 to be exactly the reference value, 9.00 L/h.
Why the Power Form Is Linear on Log Scales
The power form is often useful when a proportional change in a biomarker is a more natural comparison than an additive change. Its mathematical form also clarifies what is being estimated. Taking logarithms of both sides gives:
Consequently, \(\theta\) is the slope relating log parameter to log covariate. For a small relative change \(\Delta x/x\), the corresponding relative parameter change is approximately \(\Delta P/P\approx\theta(\Delta x/x)\). The approximation follows by differentiating the log form: \(d\log(P)=\theta\,d\log(x)\). It is local; for a finite change, the exact ratio remains \((x_{\mathrm{new}}/x_{\mathrm{old}})^\theta\).
Assessing Whether the Form Matches the Data
A covariate form should be evaluated within the population PK model, not by visually connecting a few raw concentration measurements to biomarker values. Concentrations also depend on dose, sampling time, disposition structure, and other modeled sources of variability. The relevant question is whether the candidate covariate relationship improves the description or prediction of the PK parameter, given the model and available data.
A practical comparison fits candidate forms using the same observations and otherwise comparable model assumptions. Examine parameter estimates and uncertainty, the likelihood-based fit, and diagnostic plots. The Akaike information criterion, \(AIC=-2\ell+2k\), combines the maximized log-likelihood \(\ell\) with a penalty for \(k\) estimated parameters; lower AIC indicates better relative support among the compared models, not proof that a model is true. AIC comparisons require likelihoods fitted to the same data under comparable likelihood definitions.
Useful checks include plots of individual or population residuals against the biomarker, predicted PK parameter or concentration against the biomarker, and observed versus predicted concentrations over the covariate range. Look for systematic curvature, changing spread, influential observations, sparse coverage at an endpoint, and implausible extrapolations. A trend in residuals can indicate that the chosen form misses structure, but it can also reflect an inadequate PK model, measurement error, or another covariate pattern. It does not by itself identify the cause.
| Form | What stays constant | Important condition or risk |
|---|---|---|
| Linear | Absolute change in parameter per covariate unit | May predict nonpositive values outside the observed range |
| Log-linear | Multiplicative parameter ratio for a fixed additive covariate change | Effect depends on the unit used for the covariate |
| Power | Multiplicative parameter ratio for a fixed proportional covariate change | Covariate must be positive; effects are defined relative to scale |
Name the PK parameter, biomarker, units, reference value, and observed range. Confirm that the proposed covariate values are defined for the selected form.
Keep the rest of the model and analysis data comparable so that differences in fit can be attributed to the covariate relationship as far as possible.
Use likelihood-based comparison together with residual and prediction diagnostics. Check predicted parameter values across the full observed range, not just at the reference.
Describe the estimated association, its uncertainty, the biomarker range represented, and limitations. Avoid claims about effects beyond the data or mechanisms not identified by the model.
Assumptions, Limitations, and Common Pitfalls
These forms assume the selected biomarker is measured and coded consistently, the proposed relationship is an adequate approximation over the modeled range, and the broader PK model can separate the covariate pattern from other sources of variation sufficiently for estimation. The reference value should be meaningful and within or near the observed data; a reference far outside the data can make the intercept difficult to interpret. A fitted trend may be unstable when only a narrow biomarker range is represented.
Albumin, bilirubin, and INR should not be treated as interchangeable indicators or automatically combined into a single linear score. Their clinical determinants differ, and a measured association with a PK parameter need not be causal. Likewise, a biomarker's biological meaning does not dictate a particular mathematical form. A power relationship may be convenient for a positive, right-skewed covariate, but convenience alone does not validate it.
A common mistake is to choose a power model because the parameter must be positive, then interpret its exponent as an absolute change per one-unit increase in the biomarker. The correction is that positivity is guaranteed by the form when the covariate is positive, while the exponent describes a relative-scale relationship: it is the slope on log-parameter versus log-covariate scales. Another pitfall is choosing the model with the smallest AIC and treating that as definitive. AIC is a relative comparison; residual diagnostics, uncertainty, clinical plausibility, and the covariate range still matter.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- A power model has \(\theta=0.40\). What parameter ratio does the model predict when the positive covariate doubles?
- In a log-linear model, \(\gamma=0.10\) per 1 g/L increase in albumin. What is the exact percent change in the modeled parameter for a 1 g/L increase?
- Why is a power model not directly defined at bilirubin equal to zero?
- A linear clearance model predicts a negative value at an albumin concentration far above all observed values. What should be concluded?
Solutions. (1) The ratio is \(2^{0.40}=\exp(0.40\log 2)=1.320\), so the modeled parameter is 32.0% higher. (2) The exact change is \(100\{\exp(0.10)-1\}=10.52\%\), not exactly 10%. (3) The form uses \((x_i/x_{\mathrm{ref}})^\theta\); at zero, the log-scale representation is undefined, and the model's relative-change interpretation requires a positive covariate. (4) The prediction is an extrapolation warning: the linear form is not credible there for a positive parameter. Do not interpret the negative value biologically; reassess the supported range and candidate model forms.