Why Check Proportionality Before Extrapolating?
A dose recommendation is often based on simulated exposure at several candidate doses. If exposure is proportional to dose, an exposure observed at one dose can be scaled to estimate the dose needed for a different exposure. But that arithmetic is only reliable over a range where proportionality is a reasonable description of the model and its input. If the model predicts nonlinear scaling, or if the target dose lies outside the dose range supported by the model and its data, a simple proportional extrapolation can be misleading.
The key question is not simply whether simulated exposure rises as dose rises. It is whether it rises by the amount proportional scaling predicts, for the exposure metric and regimen being considered. This assessment should be made over the proposed dose range, not inferred from a single pair of doses or from a visually smooth curve.
Here \(Y(D)\) can be an individual’s AUC or Cmax, or a population summary of one of these metrics. Dose must be defined consistently: for example, as the amount per administration when comparing the same dosing interval and schedule. If the interval, route, formulation, or number of administrations changes, the comparison is between regimens, not a clean test of dose proportionality.
Why Linear Pharmacokinetics Predict Proportional Exposure
Under linear pharmacokinetics, concentration responds linearly to the input: multiplying the input by a constant multiplies the concentration-time profile by that same constant. For a fixed regimen with dose-proportional input, increasing dose from \(D\) to \(cD\) therefore multiplies every concentration at the corresponding time by \(c\). This result assumes the same kinetic parameters and the same timing and definition of the exposure metric.
For AUC, integrating the scaled concentration-time profile multiplies the integral by \(c\). For the true Cmax, taking the maximum of the scaled profile also multiplies the maximum by \(c\). Thus, under linear pharmacokinetics with dose-proportional input and a fixed regimen, both AUC and Cmax are dose proportional. A different scaling pattern requires some relevant condition to change—for example, nonlinear pharmacokinetics, dose-dependent absorption or input, or different measurement conditions.
The same reasoning applies within a simulated population when each virtual subject’s parameters and regimen are held fixed across doses. If a subject’s AUC doubles when dose doubles, the population mean AUC also doubles, and scale-equivariant summaries such as the median scale as well. Comparing different independently generated populations can obscure this pattern with Monte Carlo variation, which is one reason paired simulations are useful.
Practical Ways to Assess Simulated Proportionality
Two complementary summaries are useful. First, calculate dose-normalized exposure, \(Y(D)/D\), at each dose. A roughly constant sequence supports proportionality; a systematic increase or decrease suggests a departure. Second, compare exposures as ratios. Relative to a reference dose \(D_{\mathrm{ref}}\), define the proportionality ratio
If \(R(D)=1\), the observed or simulated exposure ratio equals the dose ratio. Values below 1 mean exposure rises less than proportionally relative to the reference; values above 1 mean it rises more than proportionally. A tolerance around 1 should be chosen for the decision at hand. For example, a team might regard ratios from 0.90 to 1.10 as acceptably close for a preliminary assessment. That is a practical criterion, not a universal biological boundary.
A second summary fits a power relationship \(Y(D)=aD^\beta\), where \(a\) is a positive scale constant and \(\beta\) describes scaling. Taking logarithms gives a straight-line relationship:
The proportional case is \(\beta=1\). A slope below 1 indicates less-than-proportional scaling on average over the fitted range; a slope above 1 indicates more-than-proportional scaling. This summary compresses the dose trend into one number, so it should not replace inspection of dose-specific ratios: a single slope can conceal a bend or a change in scaling near the upper end of the range.
For Monte Carlo simulations, use the same virtual subjects or parameter draws at every candidate dose when feasible. This paired design makes dose contrasts less noisy because each subject serves as their own comparison. Evaluate the metric consistently for each subject and dose, then inspect individual dose-normalized values or paired ratios as well as population summaries. Consider Monte Carlo error when interpreting small deviations: a small apparent departure may reflect finite simulation size, while a consistent trend across doses and subjects is more concerning. A statistical failure to distinguish a slope from 1 is not, by itself, evidence that deviations are small enough for the dosing decision.
Worked Example: Dose-Normalized AUC Across Four Doses
Suppose a hypothetical simulation uses the same fixed regimen and paired virtual subjects at doses of 100, 200, 400, and 800 mg. The population mean AUC values are 9.80, 19.7, 39.5, and 79.1 mg·h/L. Assess proportionality using the 100 mg dose as reference.
The dose-normalized values are \(9.80/100=0.0980\), \(19.7/200=0.0985\), \(39.5/400=0.09875\), and \(79.1/800=0.098875\) (mg·h/L)/mg. They are close to one another.
For 800 mg, the dose ratio is \(800/100=8\), while the AUC ratio is \(79.1/9.80=8.071\). Thus, \(R(800)=8.071/8=1.0089\). The simulated AUC at 800 mg is about 0.89% higher than exact proportional scaling from 100 mg predicts.
Using a prespecified practical tolerance of 0.90 to 1.10, this comparison is consistent with proportionality. The intermediate dose-normalized values also show no meaningful trend. This supports proportional scaling within these simulated conditions and this dose range; it does not establish that the same pattern holds beyond 800 mg or under a different regimen.
Recognizing Nonlinear Scaling and Its Consequences
Nonlinear scaling can arise when processes governing absorption, distribution, metabolism, or elimination change with concentration or dose. For example, saturable elimination can make exposure rise more than proportionally as dose increases, whereas saturation of absorption can make it rise less than proportionally. These are possible explanations, not conclusions established by a dose-response pattern alone. The simulated model structure and the evidence supporting its parameterization determine whether a mechanism is credible.
A useful diagnostic is to examine dose-normalized exposure by dose and the paired ratios between adjacent dose levels. A declining \(Y(D)/D\) suggests less-than-proportional scaling; an increasing value suggests more-than-proportional scaling. Also inspect the actual exposure distributions and individual patterns: a population mean trend can mask subgroups or a small number of subjects with markedly different scaling. If the trend changes across the range, report that pattern rather than relying only on an average log-log slope.
Worked Example: Less-Than-Proportional AUC at Higher Doses
In a hypothetical fixed-regimen simulation, mean AUC is 10, 19, 35, and 58 mg·h/L at doses of 100, 200, 400, and 800 mg. Calculate dose-normalized AUC and the endpoint log-log slope, then interpret the pattern.
The dose-normalized values are \(10/100=0.100\), \(19/200=0.0950\), \(35/400=0.0875\), and \(58/800=0.0725\) (mg·h/L)/mg. The steady decline indicates less-than-proportional exposure scaling across the range.
The 800-to-100 mg dose ratio is 8, whereas the AUC ratio is \(58/10=5.80\). The proportionality ratio is \(5.80/8=0.725\), meaning the endpoint AUC ratio is 27.5% below the proportional prediction.
The endpoint power-model slope is \( \beta=\log(58/10)/\log(800/100)=\log(5.80)/\log(8)=1.7579/2.0794=0.8454 \). This is an average description of the endpoints, not proof that the response follows one exact power law at every dose. Both summaries indicate that linear dose scaling would overpredict AUC at the high dose in this simulation. Any dose recommendation should use the simulated nonlinear relationship rather than a proportional calculation, and should be qualified by the model’s support at those doses.
Exposure metrics should also be considered separately. Under the fixed, linear conditions stated above, AUC and true Cmax both scale with dose. If simulated AUC appears proportional but Cmax does not, first check whether the input profile, absorption assumptions, regimen, timing, or peak-measurement procedure changes with dose. A change in sampling times can alter a measured maximum even when the underlying concentration-time curve scales exactly. It is not correct to attribute different Cmax scaling to absorption timing alone when the input and measurement conditions are fixed and dose proportional.
Worked Example: Distinguishing a Regimen Change From Nonlinearity
Consider two hypothetical simulations. In Simulation A, the same oral formulation and dosing interval are used at 100 and 200 mg; mean AUC is 20 and 40 mg·h/L, and true Cmax is 3.0 and 6.0 mg/L. In Simulation B, 100 mg is given every 12 hours and 200 mg every 24 hours; the total daily dose is the same. What can be concluded?
In Simulation A, the dose ratio is \(200/100=2\), the AUC ratio is \(40/20=2\), and the Cmax ratio is \(6.0/3.0=2\). Both metrics are proportional for these two doses under the fixed regimen. This pair is compatible with linear scaling but does not establish proportionality across a wider range.
Simulation B changes the dosing interval, so its exposure comparison cannot isolate the effect of dose size. Equal total daily doses do not guarantee equal Cmax, trough concentration, or time course. The results may still be useful for comparing those two regimens, but they are not a dose-proportionality test. To test dose proportionality, hold the interval and other regimen features fixed while changing dose.
When a Target-Based Extrapolation Is Unreliable
If proportionality is supported over a relevant range, a target-based dose calculation can be a useful approximation. For a reference dose \(D_{\mathrm{ref}}\) producing exposure \(Y_{\mathrm{ref}}\), the proportional estimate for target \(Y_{\mathrm{target}}\) is
This calculation assumes that the target metric scales proportionally at the estimated dose. It is especially vulnerable when the estimate lies well beyond the dose range supported by the model and its data. Simulating a dose outside that range can show what the model predicts, but simulation alone does not validate the model’s extrapolation. The prediction may depend strongly on structural assumptions or parameter values that were poorly informed at high doses.
Worked Example: A Proportional Estimate Beyond the Supported Range
Suppose a hypothetical model is supported by data and evaluation only from 50 to 200 mg under a fixed regimen. At 200 mg, the simulated mean AUC is 40 mg·h/L. A target mean AUC of 100 mg·h/L is proposed. Calculate the proportional dose estimate and assess whether it is reliable.
The proportional calculation gives \(D_{\mathrm{target}}=200\times(100/40)=200\times2.5=500\) mg. The estimate is 2.5 times the highest dose in the supported range and is not an interpolation within 50 to 200 mg.
The arithmetic is correct under proportionality, but the data-supported range does not establish proportionality through 500 mg. The 500 mg value should therefore be treated as an extrapolative estimate, not as a reliable dose recommendation on this evidence alone. A defensible assessment would require model predictions at relevant doses plus justification that the model structure and input assumptions remain credible there; if that support is absent, the result should be reported as uncertain or not supported.
Common Pitfalls and a Practical Assessment Sequence
- Using only one dose pair. Two points define a ratio but cannot reveal curvature. Inspect all candidate doses, particularly the upper end relevant to extrapolation.
- Treating a nonsignificant slope difference as proof of proportionality. Failure to detect a difference from 1 can reflect limited simulation precision. Judge whether observed deviations are acceptably small for the decision, using a stated practical tolerance.
- Changing more than dose. If interval, route, formulation, or input timing changes, differences cannot be attributed to dose scaling alone. Analyze the comparison as a regimen comparison instead.
- Assuming model simulation validates extrapolation. A model can produce a numerical prediction at any supplied dose, but numerical output does not demonstrate empirical support or reliability at that dose.
- Generalizing from one exposure metric. Assess the metric tied to the decision. Under fixed linear conditions AUC and true Cmax are both proportional, but dose-dependent input or nonlinear behavior may affect metrics differently.
Specify the exposure metric, dose definition, regimen, and dose range. Keep non-dose conditions consistent when testing proportionality.
Use paired virtual subjects when feasible; inspect dose-normalized exposure, proportionality ratios, and the dose trend rather than relying on a single point.
State a tolerance appropriate to the decision, consider Monte Carlo error, and note systematic departures or curvature.
Compare the target-attaining dose with the model-supported range. If it lies beyond that range, label the estimate as extrapolative and explain what additional support is needed.
The central distinction is between model prediction and support for that prediction. Dose-normalized exposure and log-log slopes describe what the simulated model implies; they do not establish that the model is accurate outside the range where its assumptions are credible. Proportionality is therefore a useful check before extrapolation, not a substitute for evaluating the range and conditions under which the model can be trusted.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- At 50 mg, a hypothetical simulation predicts a mean AUC of 12 mg·h/L; at 100 mg it predicts 24 mg·h/L under the same regimen. What is the proportionality ratio at 100 mg relative to 50 mg?
- Why is a slope of 1 in a log exposure-versus-log dose model associated with dose proportionality?
- A team changes both dose and dosing interval between two simulations. Can the resulting exposure ratio be interpreted as a dose-proportionality assessment? Explain.
- A proportional calculation gives a target dose of 600 mg, but the model is supported only through 300 mg. What is the key limitation of the estimate?
Solutions. (1) The dose ratio is \(100/50=2\), and the AUC ratio is \(24/12=2\). Thus \(R(100)=2/2=1\), exactly proportional for this pair. (2) In \(Y(D)=aD^\beta\), a one-unit increase in \(\log D\) changes \(\log Y\) by \(\beta\); when \(\beta=1\), exposure changes in the same proportion as dose, so \(Y(D)/D=a\) is constant. (3) No. The interval changed as well as dose, so the comparison is between regimens and cannot isolate the effect of dose size. (4) The estimate lies beyond the model-supported range. The calculation assumes proportionality through 600 mg, but support only through 300 mg does not establish that assumption; the estimate is extrapolative and should not be presented as reliable without additional justification.