Why Effect Sizes Are Easy to Misread
A covariate effect can be numerically correct and still be described incorrectly. A 25% higher clearance does not imply 25% lower AUC; a model prediction for an unusually high covariate value may be far beyond the data that informed the model; and a fitted association does not, by itself, establish that changing the covariate would change drug disposition. Each error arises from moving too quickly from a model parameter to a clinical interpretation.
The earlier material on interpreting clearance effects and translating clearance effects into AUC changes provides the relevant model relationships. Here the focus is on preserving those relationships when describing effect sizes, selecting contrasts, and discussing what the data can support. The numerical examples are hypothetical.
Clearance Percentages and AUC Percentages Are Not Symmetric
For a linear pharmacokinetic model, the area under the concentration–time curve (AUC) over a dosing interval is \(AUC=F\cdot Dose/CL\), where \(F\) is bioavailability and \(CL\) is clearance. For intravenous administration, \(F=1\). If dose and bioavailability are unchanged between two covariate values, AUC is inversely proportional to clearance. This interpretation assumes linear pharmacokinetics over the relevant range, comparable AUC definitions, and no covariate-related change in bioavailability.
Thus, if clearance changes by a multiplicative factor \(r_{CL}\), AUC changes by \(1/r_{CL}\), not by subtracting the same percentage. Percent changes use the starting value as their denominator, so reversing a ratio generally does not preserve the percentage.
If \(CL\) increases by \(q\%\), then \(r_{CL}=1+q/100\), and the AUC percent change is \(100\{1/(1+q/100)-1\}\%\). If clearance decreases by \(q\%\), then \(r_{CL}=1-q/100\), and the AUC percent change is \(100\{1/(1-q/100)-1\}\%\). The two results are not equal in magnitude.
Worked Example: Converting a Clearance Increase to an AUC Change
In a hypothetical comparison at the same intravenous dose, a model predicts \(CL_0=8.00\) L/h in one group and \(CL_1=10.0\) L/h in another. Assume linear pharmacokinetics and complete AUC measurement.
The clearance ratio is \(r_{CL}=10.0/8.00=1.25\), a 25.0% increase. With dose fixed, the AUC ratio is the reciprocal:
The relative AUC change is \((0.800-1)\times100\%=-20.0\%\). For example, if the reference AUC were 100 mg·h/L, the comparison AUC would be \(100\times0.800=80.0\) mg·h/L, a decrease of 20.0 mg·h/L. The conclusion is a 25.0% higher clearance and, under the stated assumptions, a 20.0% lower AUC—not a 25.0% lower AUC.
The reciprocal relationship also works in the other direction, but the percentage asymmetry remains. A 20% decrease in clearance gives \(r_{CL}=0.800\) and \(r_{AUC}=1/0.800=1.25\): a 25% increase in AUC. These percentages describe ratios relative to different reference values.
Deriving a Covariate Contrast and Respecting Its Range
For a continuous covariate \(x\), a power model can express clearance relative to a reference value \(x_{\mathrm{ref}}\):
Here \(CL_{\mathrm{ref}}\) is clearance at \(x_{\mathrm{ref}}\), and \(\theta\) is the covariate exponent. The model implies a ratio for any two positive covariate values. Dividing the prediction at \(x_1\) by the prediction at \(x_0\) cancels the reference clearance and reference covariate:
This derivation explains why the covariate contrast must be stated: the exponent alone is not the percent effect for every possible comparison. The ratio depends on the selected \(x_1/x_0\). A mathematically computable prediction is not automatically a well-supported prediction. If \(x_1\) lies outside the range represented in the data, the result is an extrapolation whose reliability depends on the model continuing to describe the relationship there.
Worked Example: Recognizing an Extrapolated Prediction
Suppose, hypothetically, that body weight ranged from 50 to 100 kg in the data used to fit a power model, with \(x_{\mathrm{ref}}=70\) kg and estimated exponent \(\theta=0.75\). Compare the model predictions at 100 kg and 140 kg with the prediction at 70 kg.
For 100 kg, the ratio is \(CL(100)/CL(70)=(100/70)^{0.75}\). Since \(\ln(100/70)\approx0.3567\), the ratio is \(\exp(0.75\times0.3567)=\exp(0.2675)\approx1.307\). The model predicts approximately 30.7% higher clearance at 100 kg than at 70 kg. Both values are within the stated observed range.
For 140 kg, \(CL(140)/CL(70)=(140/70)^{0.75}=2^{0.75}\approx1.682\), or approximately 68.2% higher clearance. But 140 kg is outside the observed range, so this is an extrapolation. The calculation follows from the fitted equation; it does not demonstrate that the same power relationship holds up to 140 kg. A report should identify that limitation rather than presenting the 68.2% as equally data-supported as the within-range contrast.
A useful check is to inspect the covariate distribution and the support for the particular contrast, rather than only the model equation. A prediction at an extreme value may be informed by very few observations even when that value technically falls within the minimum-to-maximum range. Sparse tails, gaps in the distribution, and a small number of subjects in a category can all make an apparently ordinary contrast uncertain.
Association Is Not a Causal Effect
In a population PK model, a covariate coefficient typically describes how the modeled parameter varies with the covariate, conditional on the model structure and included covariates. This is an association in the analyzed data. Calling it causal makes a stronger claim: that intervening to change the covariate would change the parameter, with other relevant causes appropriately accounted for.
For example, a covariate may be associated with clearance because it is related to another factor that affects clearance, because groups differ in concomitant treatments or disease severity, or because of selection into the study. A well-fitting covariate model does not alone rule out these possibilities. Causal interpretation requires a suitable study design and assumptions about confounding, measurement, and the relation between the covariate and outcome. A fitted effect should therefore be described as an estimated association unless the evidence and assumptions justify a causal claim.
Worked Example: Interpreting a Categorical Contrast Without Claiming Causation
In a hypothetical observational analysis, estimated clearance is 6.00 L/h for a reference category and 4.32 L/h for a second category. Assume the model comparison is otherwise appropriate. The clearance ratio is \(4.32/6.00=0.720\), or 28.0% lower clearance in the second category relative to the reference.
If dose and bioavailability are the same and pharmacokinetics are linear, the implied AUC ratio is \(1/0.720=1.389\), or approximately 38.9% higher AUC. This is the model-based exposure contrast under those assumptions.
The data description does not establish that category membership caused the clearance difference. If the categories also differ in disease severity or co-medication, those factors could contribute to the observed association. A defensible statement is: “The model estimated 28.0% lower clearance in the second category; under equal dose and bioavailability with linear pharmacokinetics, this corresponds to approximately 38.9% higher AUC.” It would not be justified from these results alone to say that changing a person’s category would cause the AUC increase.
Checks, Alternatives, and Failure Modes
Before interpreting a covariate effect, check what is being compared, how the effect is scaled, and whether the comparison is supported by the data. The following distinctions help prevent a numerical result from acquiring a meaning it does not have.
| Interpretation choice | What it communicates | Common risk |
|---|---|---|
| Ratio or relative percent change | Proportional difference for a stated contrast | Confusing the ratio for one parameter with the reciprocal ratio for AUC |
| Absolute difference | Change in parameter units, such as L/h | Assuming the same absolute difference has the same importance at every baseline value |
| Within-range contrast | Model comparison among covariate values represented in the data | Overlooking sparse support near an endpoint |
| Extrapolated contrast | Model prediction beyond the observed covariate support | Presenting a mathematical prediction as an empirically verified effect |
| Model association | Conditional relationship estimated from the analyzed data | Describing the association as a causal effect without adequate justification |
Diagnostics for interpretation are not limited to a model’s numerical output. Check the covariate range and distribution, the number of observations supporting each part of the contrast, and whether the assumed functional form is plausible over that range. For an exposure translation, verify that the dose and bioavailability assumptions apply and that linear pharmacokinetics are reasonable. If bioavailability differs by covariate, or if exposure is not inversely proportional to clearance in the relevant setting, the simple reciprocal calculation need not hold.
Alternative summaries can be useful, but each answers a different question. A ratio is suited to multiplicative effects; an absolute difference retains the parameter’s units. A contrast between two specified covariate values is more interpretable than quoting an exponent without saying what values it compares. Neither a different scale nor a narrower contrast removes confounding or makes extrapolation reliable. These are limitations of the evidence and model, not merely wording problems.
Check Your Understanding
For each item, identify the correct interpretation. Solutions follow each question.
- A hypothetical model predicts clearance to increase from 5.00 to 6.00 L/h. At the same dose and bioavailability, with linear pharmacokinetics, what is the AUC percent change?
- A power model has \(\theta=0.50\). What clearance ratio does it predict for \(x_1=80\) compared with \(x_0=20\)? If the first value is outside the observed range, what additional qualification is needed?
- A covariate is associated with lower clearance in an observational model. Does this alone show that changing the covariate would lower clearance? Explain.
- Why is “20% lower clearance means 20% higher AUC” incorrect when dose and bioavailability are fixed and pharmacokinetics are linear?
Solutions. 1. The clearance ratio is \(6.00/5.00=1.20\), so the AUC ratio is \(1/1.20=0.8333\). The AUC change is \((0.8333-1)\times100\%=-16.7\%\), approximately. 2. The clearance ratio is \((80/20)^{0.50}=\sqrt{4}=2.00\), a predicted doubling. If 80 is outside the observed covariate range, this is an extrapolation; the model’s relationship at that value is not directly supported by the observed range. 3. No. The result is an association conditional on the model and included covariates. Confounding or other design and data limitations may explain some or all of it; causal interpretation needs additional justification. 4. A 20% decrease gives a clearance ratio of 0.800, whose reciprocal is an AUC ratio of \(1/0.800=1.25\). AUC therefore increases by 25%, not 20%, because the percentage is calculated relative to the original AUC.