Worked Interpretation of Renal and Hepatic Covariate Models

From Fitted Coefficients to Patient Predictions

A fitted renal and hepatic covariate model is useful only when its coefficients can be translated into predictions for clinically interpretable profiles. A coefficient on a logarithmic scale is not, by itself, a percentage change; a power coefficient is not a fixed change in clearance per unit of eGFR. The interpretation depends on the model form, the reference values, and the units in which each covariate is recorded.

This tutorial uses a hypothetical fitted model to work through those translations. The calculations illustrate how eGFR, bilirubin, and albumin contribute to a predicted typical clearance, and why a prediction that can be calculated algebraically is not necessarily supported by the data.

Definition: A typical clearance prediction is the clearance predicted by the population covariate model for specified covariate values, with individual random effects set to their typical value of zero. It is not a guarantee of an individual patient’s clearance, and it does not include uncertainty in the fitted coefficients unless that uncertainty is calculated separately.

A Hypothetical Fitted Model

Suppose a population PK analysis reports the following hypothetical model for typical clearance:

CL_i^{\mathrm{typ}} = 8\left(\frac{\mathrm{eGFR}_i}{90}\right)^{0.65} \exp\left[-0.10(\mathrm{BIL}_i-1)+0.20(\mathrm{ALB}_i-4)\right]\ \mathrm{L/h}.

Here eGFR is measured in mL/min/1.73 m², total bilirubin (BIL) in mg/dL, and albumin (ALB) in g/dL. The reference profile is eGFR 90 mL/min/1.73 m², bilirubin 1 mg/dL, and albumin 4 g/dL; the model predicts 8 L/h at that profile. The coefficients and values are illustrative, not estimates from a real study.

The model assumes a power relationship for eGFR and an additive linear relationship on the log-clearance scale for bilirubin and albumin. In particular, it contains no bilirubin-by-albumin interaction and no interaction between renal and hepatic covariates. Thus, each covariate contributes a multiplicative factor to clearance, and the factors multiply.

Interpretation of coefficients: The eGFR exponent 0.65 is an elasticity: holding bilirubin and albumin fixed, a 1% increase in eGFR corresponds approximately to a 0.65% increase in typical clearance locally. A 1 mg/dL increase in bilirubin multiplies typical clearance by \(\exp(-0.10)\), about 0.9048. A 1 g/dL increase in albumin multiplies it by \(\exp(0.20)\), about 1.221. These are model-based associations conditional on the other included covariates, not causal effects.

The exact percentage interpretation of an exponential coefficient comes from the ratio, not from treating the coefficient itself as a percentage. For bilirubin, the ratio is \(0.9048\), so a 1 mg/dL increase corresponds to a predicted decrease of about \(100(1-0.9048)=9.52\%\), holding the other covariates fixed. For albumin, the ratio is \(1.221\), or about a 22.1% increase per 1 g/dL. These changes are multiplicative; they are not fixed additions in L/h.

Worked Patient Predictions

Worked Example: Lower eGFR and Higher Bilirubin

For a hypothetical patient with eGFR 45 mL/min/1.73 m², bilirubin 2 mg/dL, and albumin 4 g/dL, calculate typical clearance and compare it with the reference profile.

Substitute the values into the fitted model:

CL_i^{\mathrm{typ}} = 8\left(\frac{45}{90}\right)^{0.65} \exp[-0.10(2-1)+0.20(4-4)]\ \mathrm{L/h}.

The eGFR factor is \(0.5^{0.65}=\exp[0.65\log(0.5)]\approx0.6373\). The hepatic factor is \(\exp(-0.10)\approx0.9048\). Therefore:

CL_i^{\mathrm{typ}} = 8(0.6373)(0.9048) \approx 4.61\ \mathrm{L/h}.

The ratio to the reference prediction is \(0.5^{0.65}\exp(-0.10)\approx0.5766\). The model therefore predicts typical clearance about 42.3% lower than at the reference profile, with bilirubin and albumin changes taken into account. This is a comparison of model predictions, not evidence that changing either covariate would cause the predicted change.

Worked Example: A Severe Covariate Profile Outside the Data Range

Consider a second hypothetical patient with eGFR 15 mL/min/1.73 m², bilirubin 6 mg/dL, and albumin 1.5 g/dL. Calculate the model prediction, then assess what the number can support. Suppose the hypothetical estimation data included eGFR values from 45 to 120, bilirubin from 0.5 to 2.0, and albumin from 3.0 to 4.5.

The renal factor is \((15/90)^{0.65}=(1/6)^{0.65}\approx0.3120\). The exponent for the hepatic factor is \(-0.10(6-1)+0.20(1.5-4)=-0.50-0.50=-1.00\), so that factor is \(\exp(-1)\approx0.3679\). Substitution gives:

CL_i^{\mathrm{typ}} = 8(0.3120)(0.3679) \approx 0.918\ \mathrm{L/h}.

The arithmetic is valid for the stated equation. However, this profile is outside the stated data ranges for all three covariates. The prediction is therefore an extrapolation: it relies on the power and exponential relationships continuing in a region where the hypothetical data provide no direct support. It should not be presented as a well-validated prediction for such a patient merely because the model returns a numerical value.

Worked Example: A Jointly Supported Covariate Contrast

Suppose the hypothetical data contain adequate observations near both a reference profile (eGFR 90, bilirubin 1, albumin 4) and a second profile (eGFR 60, bilirubin 1.5, albumin 3.5). Assume these profiles are represented jointly, rather than each value appearing only in unrelated combinations. Calculate the predicted contrast.

The ratio of predictions is obtained by dividing the model at the second profile by the model at the reference profile. The common baseline of 8 L/h cancels:

\frac{CL_{\mathrm{second}}^{\mathrm{typ}}}{CL_{\mathrm{reference}}^{\mathrm{typ}}} = \left(\frac{60}{90}\right)^{0.65} \exp[-0.10(1.5-1)+0.20(3.5-4)].

The renal factor is \((2/3)^{0.65}=\exp[0.65\log(2/3)]\approx0.7683\). The hepatic exponent is \(-0.05-0.10=-0.15\), giving a hepatic factor of \(\exp(-0.15)\approx0.8607\). The combined ratio is approximately \(0.7683(0.8607)=0.6613\). Thus, the model predicts about 33.9% lower typical clearance for the second profile than for the reference profile. Its absolute prediction is \(8(0.6613)\approx5.29\) L/h.

Because the example stipulates adequate observations near both joint profiles, this is an interpolation-supported contrast in the hypothetical setting. Being inside each covariate’s separate minimum-to-maximum range would not, by itself, establish joint support: the data might contain few or no individuals with that particular combination.

Why the Ratio Calculation Works

A ratio is often clearer than comparing two separately rounded clearance predictions. To derive it, write the model for a profile \(A\) and a profile \(B\), then divide. The baseline clearance cancels, and the remaining terms show exactly which covariate changes contribute:

\frac{CL_B^{\mathrm{typ}}}{CL_A^{\mathrm{typ}}} = \left(\frac{\mathrm{eGFR}_B}{\mathrm{eGFR}_A}\right)^{0.65} \exp\left[-0.10(\mathrm{BIL}_B-\mathrm{BIL}_A) +0.20(\mathrm{ALB}_B-\mathrm{ALB}_A)\right].

This follows because \((\mathrm{eGFR}_B/90)^{0.65}/(\mathrm{eGFR}_A/90)^{0.65}=(\mathrm{eGFR}_B/\mathrm{eGFR}_A)^{0.65}\), and because the difference between the two exponential exponents cancels the common reference terms. The ratio is dimensionless. If it is \(r\), the percentage change from profile \(A\) to profile \(B\) is \(100(r-1)\%\); a negative result indicates a decrease.

The calculation assumes that the fitted model is appropriate for both profiles and that the same model coefficients apply to each. It also holds the model’s other covariates fixed. If a model contains interactions, dividing predictions still works, but the ratio must retain the interaction terms; the simple expression above would no longer be complete.

Support, Diagnostics, and Alternative Forms

A fitted equation can be evaluated at any positive eGFR and any recorded bilirubin or albumin value, but data support is a separate question. Examine the observed ranges, the joint distribution of the covariates, and whether there are observations near the profile or contrast of interest. Collinearity can make individual coefficient interpretations unstable even when combined predictions in a well-sampled region are more stable. Measurement error in organ-function covariates can also weaken or distort estimated associations.

Useful checks include plots of observed versus predicted clearance, residuals against each covariate, and residuals against clinically meaningful combinations of covariates. Look for systematic trends, curvature, clusters, or increasing spread that the model has not captured. Inspect whether a small number of extreme profiles drive the fit. A monotone-looking fitted curve is not proof that the chosen functional form is adequate, especially at the edges of the observed data.

The power model for eGFR is one possible continuous form. A model linear in eGFR on the log-clearance scale would impose a different pattern: equal absolute eGFR increments, rather than equal proportional increments, would have the same log-scale contribution. Categories can make certain contrasts easy to describe, but discard within-category information and impose abrupt boundaries. More flexible curves can represent curvature but require enough informative observations and make interpretation more involved. The appropriate form depends on the scientific question, data support, and diagnostics; no form makes extrapolation automatically reliable.

A further limitation is the distinction between typical and individual predictions. A population model may include a random effect for individual clearance, but setting that random effect to zero gives a typical prediction, not the actual clearance of a particular patient. Individual prediction also depends on the random-effects distribution, relevant patient data, and uncertainty in model parameters. The calculations here do not provide a prediction interval or account for parameter uncertainty.

Common Mistakes

Reading an exponential coefficient as a percentage. A bilirubin coefficient of \(-0.10\) is not exactly a 10% decrease. The ratio is \(\exp(-0.10)\approx0.9048\), a decrease of about 9.52% per 1 mg/dL, conditional on the other covariates.
Reporting a prediction as a patient-specific fact. The equation gives a typical model prediction for the stated profile. It does not establish that an individual’s clearance equals that value.
Assuming in-range values guarantee support. Separate marginal ranges do not show that a joint combination is represented. Check the covariate combinations actually observed.
Calling an extrapolation a validated estimate. A computed value outside the data-supported region depends on unverified continuation of the model form.
Rounding intermediate factors too early. Keep adequate precision in the ratio and factors, then round the final clearance to a sensible number of significant figures.

Key takeaway: Interpret coefficients through the fitted model’s scale and reference values, calculate contrasts as ratios when possible, and judge predictions against joint data support. A precise calculation does not remove uncertainty about the model or make extrapolation reliable.

Check Your Understanding

Use the hypothetical model in this tutorial. Solutions are included after the questions.

Holding eGFR and albumin fixed, what is the predicted clearance ratio for a 1 mg/dL increase in bilirubin, and what percentage change does it represent?
Holding bilirubin and albumin fixed, calculate the eGFR-only clearance ratio for eGFR 60 versus 90. What percentage decrease does this imply?
Explain why a prediction at eGFR 60, bilirubin 1.5, and albumin 3.5 might be unsupported even if each value lies within its observed marginal range.
For the severe profile in the second worked example, identify one reason the calculated clearance should be treated cautiously.

Solutions. (1) The ratio is \(\exp(-0.10)\approx0.9048\), so predicted typical clearance is about 9.52% lower per 1 mg/dL increase, with other covariates fixed. (2) The ratio is \((60/90)^{0.65}=(2/3)^{0.65}=\exp[0.65\log(2/3)]\approx0.7683\); the prediction is about \(100(1-0.7683)=23.2\%\) lower. (3) The data may not contain observations near that joint combination, even though each value separately falls within its observed range; sparse joint support makes the prediction model-dependent. (4) The profile lies outside the hypothetical observed ranges, so the value is an extrapolation and depends on the power and exponential relationships holding beyond the data.