Why Stress-Test a Human Prediction?
A human prediction is conditional: it follows from a model together with assumptions about biology, measurements, and how information transfers from preclinical species to people. A single predicted value can hide that dependence. Stress testing makes it visible by recalculating the prediction under defensible alternatives and asking whether the resulting change matters for the decision at hand.
The purpose is not to find the most extreme number imaginable. Each altered assumption needs a plausible range grounded in data, biological knowledge, or a clearly stated expert rationale. A useful stress test therefore answers two questions: how much does the prediction change across credible conditions, and does that change affect a decision? It does not, by itself, assign probabilities to those conditions or say which assumption is most responsible for uncertainty.
A stress test is conditional on its chosen model, assumptions, and ranges. Its scenario results are not confidence limits, credible intervals, or probabilities. For example, if a prediction spans 2 to 7 L/h across a range of allometric exponents, that range means only that those specified exponents produce those predictions under the model. It does not mean that the true human clearance has a particular probability of lying within that range.
Designing a Defensible Stress Test
Start with the prediction and the decision it informs. Specify the predicted quantity and its units, the reference case, and what size or kind of change would be material. Then list assumptions that could plausibly affect that quantity. Examples include an allometric exponent, a species-to-human scaling relationship, an uncertain physiological covariate, or a structural choice about which biological process limits the prediction.
For each assumption, distinguish its evidence-based range from a deliberately broader challenge scenario. A range may be informed by uncertainty in preclinical estimates, relevant biological measurements, or the range of values compatible with the model’s development data. Record the source and rationale. A range chosen solely because it produces an interesting result is not a defensible stress test.
Next decide how assumptions will be varied. Changing one assumption at a time shows how the prediction responds when the rest of the reference case is held fixed. Joint scenarios can be more realistic when assumptions are linked, but the combinations must be biologically coherent. For instance, pairing two values that each occur in isolation may create an implausible joint scenario if the underlying biology makes them dependent. Earlier material on sampling correlated inputs addresses dependence in probabilistic analyses; in a stress test, the practical requirement is to avoid presenting an implausible combination as a credible scenario.
Finally, calculate the prediction for every specified case, compare it with the reference, and interpret the difference in context. If the range crosses a decision threshold, describe which assumptions and conditions produce the crossing. If it does not, say so only within the tested ranges and model; it is not proof that the prediction is robust to every possible failure.
Name the quantity, units, reference prediction, and decision criterion before looking at scenario results.
For every varied assumption, state its plausible values and the evidence or rationale supporting them.
Vary assumptions singly for diagnosis or jointly when a justified biological scenario requires it; do not combine incompatible extremes.
Show the model calculation, compare results with the reference and decision criterion, and state the limits of the tested cases.
Worked Example: Allometric Exponent and Human Clearance
Worked Example: Allometric Exponent and Human Clearance
Suppose a hypothetical rat weighing 0.25 kg has measured clearance of 0.080 L/h. A simple allometric model predicts human clearance at 70 kg as \(CL_H=CL_R(W_H/W_R)^b\), where \(b\) is the scaling exponent. Use \(b=0.75\) as the reference value, and stress-test the prediction over a biologically justified range from 0.60 to 0.80, informed here by plausible alternative scaling fits to the available preclinical evidence. The range is an assumption range, not a confidence interval.
The body-weight ratio is \(70/0.25=280\). The reference prediction is:
At the lower exponent, \(280^{0.60}\approx29.40\), so:
At the upper exponent, \(280^{0.80}\approx90.72\), so:
Thus the tested exponent range produces predictions from about 2.35 to 7.26 L/h, compared with 5.48 L/h at the reference exponent. This is a substantial change. Whether it is material depends on the use of the prediction: if a dosing decision changes over that clearance range, the exponent deserves explicit attention. The calculation assumes that the same allometric relationship applies across this species and body-size range, that the rat clearance value is suitable, and that other determinants of clearance do not invalidate this simple model. The arithmetic cannot validate those assumptions.
Reference Cases, Scenarios, and Decision Materiality
One-at-a-time stress tests are easy to interpret: each result isolates the consequence of changing a specified assumption while holding the others at reference values. They are useful for exposing a strong dependence, but they can miss interactions. If two assumptions jointly determine the prediction, the effect of changing both may be different from the sum of changing each separately. A scenario analysis instead specifies complete, internally consistent cases—for example, a lower-clearance case tied to a particular physiological profile and a higher-clearance case tied to another profile.
The materiality criterion should be set in terms of use, not chosen after seeing the output. A 20% change may be unimportant for a broad planning estimate but consequential if it changes a dose category or crosses a safety boundary. Conversely, a large numerical difference may not alter the decision. Report both the prediction scale and the decision implication, and state the threshold or operational rule used.
Worked Example: Stress-Testing a Dose Projection
Consider a hypothetical linear pharmacokinetic model in which average steady-state exposure is approximated by \(AUC=F\times D/CL\). Here \(D\) is dose over the interval, \(F\) is bioavailability, and \(CL\) is clearance. Suppose the target \(AUC\) is 10 mg·h/L. The reference assumptions are \(CL=5\) L/h and \(F=0.50\). The dose needed to reach the target is obtained by rearranging the equation:
Suppose evidence supports testing \(CL\) from 4 to 7 L/h and \(F\) from 0.40 to 0.60. These are hypothetical ranges for this example. The independent lower-dose corner gives:
and the independent higher-dose corner gives:
The required dose across these corners is 66.7 to 175 mg, versus 100 mg at the reference assumptions. This calculation assumes linear exposure-dose proportionality, that the stated bioavailability and clearance apply to the same dosing conditions, and that the target is expressed on the same AUC basis. It also assumes the corner combinations are plausible together. If evidence indicates that low clearance is associated with high bioavailability, the simple independent corners may not represent realistic joint cases; construct scenarios consistent with that relationship instead.
If a hypothetical development rule selects between 100 mg and 150 mg, a prediction of 175 mg could change the selected dose. The stress test therefore identifies a decision-sensitive region, but it does not establish that 175 mg is likely or recommend that dose. It shows what follows if the specified assumptions and model apply.
Testing a Biological Translation Assumption
Stress testing need not be limited to numerical parameters in a scaling equation. It can examine a biological translation assumption by representing alternatives explicitly. The model still has to be defined before the alternatives are compared, and the range must be credible rather than an arbitrary search for extremes.
Worked Example: Human Potency Relative to an Animal Estimate
Suppose a hypothetical exposure-response model predicts response as \(E=E_{\max}C/(EC_{50}+C)\). Let \(E_{\max}=100\%\), predicted human concentration \(C=4\) mg/L, and the animal estimate of \(EC_{50}\) be 2 mg/L. Define \(q\) as the human-to-animal potency scaling factor, so that human \(EC_{50}=2q\) mg/L. Assume that assay and biological evidence make \(q=0.75\), 1.0, and 2.0 defensible scenarios. These values are not assigned probabilities.
For \(q=0.75\), human \(EC_{50}=1.5\) mg/L and:
For the reference \(q=1\), human \(EC_{50}=2\) mg/L and \(E=100\%\times4/(2+4)=66.7\%\). For \(q=2\), human \(EC_{50}=4\) mg/L and \(E=100\%\times4/(4+4)=50.0\%\). Thus the tested translation assumptions shift the predicted response from 50.0% to 72.7%.
If the hypothetical decision threshold is a predicted response of at least 60%, the exact boundary can be found from:
The prediction meets the threshold for \(q\) up to about 1.33, but not for larger values. This result makes the decision dependence clear. It does not show which \(q\) is correct, and it relies on the Emax model, the predicted concentration, and the assumed scaling of \(EC_{50}\). If the model form or the exposure-response relationship is unsuitable, varying \(q\) alone cannot repair that failure.
Justifying the Comparison and Diagnosing Failure
A useful comparison keeps a traceable chain from assumption to prediction. For a model \(g\) with input assumptions \(\theta\), write the prediction as \(\hat{Y}=g(\theta)\). For reference assumptions \(\theta_0\), the reference is \(\hat{Y}_0=g(\theta_0)\). A scenario \(\theta_j\) gives \(\hat{Y}_j=g(\theta_j)\), and its difference from the reference is \(\Delta_j=\hat{Y}_j-\hat{Y}_0\). For positive predictions, a relative change can also be reported as \(100(\hat{Y}_j-\hat{Y}_0)/\hat{Y}_0\%\). These are descriptive contrasts between specified model calculations, not estimates of sampling error.
The comparison is meaningful only if the model is being used within its applicable domain and the varied values are defined consistently. Check units, parameter bounds, input transformations, and the numerical implementation. Verify that the reference result is reproduced before changing assumptions. If predictions jump discontinuously, become biologically impossible, or change because a parameter has silently switched units, diagnose the model or coding before interpreting the result.
Other warning signs include a chosen range unsupported by evidence, an extreme combination that cannot occur biologically, or a scenario that changes several assumptions without explaining why they belong together. A prediction may also be insensitive to the tested assumptions but highly sensitive to an omitted structural feature. A narrow output range is therefore not general proof of predictive reliability. Where a scenario takes the model outside the data or biological domain that supports it, label that extrapolation clearly instead of presenting it as an equally credible prediction.
How Stress Testing Differs From Alternatives
One-at-a-time sensitivity analysis changes one input around a reference point and is helpful for seeing direction and scale, but does not by itself represent joint uncertainty or interactions. Global sensitivity analysis evaluates variation across multiple uncertain inputs, usually under specified distributions or sampling schemes, and can provide a systematic assessment of influence. Stress testing has a different emphasis: it asks whether predictions remain acceptable under selected, justified conditions, especially conditions that could change a decision. A stress test can complement global analysis, but its scenario range should not be mistaken for a probability distribution or an attribution of uncertainty.
Model averaging addresses uncertainty across alternative models by combining their predictions according to assigned weights. It is useful when several models are considered plausible, but it answers a different question from recalculating a particular prediction under specified assumption changes. Likewise, a prediction interval summarizes predictive uncertainty under a specified model and uncertainty framework. Scenario bounds are not a substitute for such an interval. Choose the approach that matches the question, and be explicit if several approaches are used together.
Common Pitfall and Key Takeaway
- Wrong reasoning: “The stress-test range is 2 to 7 L/h, so there is a 95% probability human clearance is between 2 and 7 L/h.” This assigns a probability that the scenario calculations do not provide.
- Correction: “Under the specified model, the tested exponent values produced predictions from 2 to 7 L/h. These are conditional scenario results, not a probability interval.”
- Wrong reasoning: “I tested the largest possible values of every input, so this is a realistic worst case.” Arbitrary simultaneous extremes can be biologically incompatible.
- Correction: Define a worst-case scenario only when its assumptions and their joint occurrence are defensible; otherwise report separate or coherent scenarios.
Check Your Understanding
Try each problem, then compare your reasoning with the solutions.
- A human prediction is 12 units at a reference parameter value. Values justified by available evidence produce predictions from 9 to 15 units. What can and cannot be concluded from this range?
Solution: Under the stated model and tested values, predictions range from 9 to 15 units. The range is not a 95% interval and does not assign probabilities to values within it. Whether the range matters depends on the decision criterion and the units and context of the predicted quantity. - In the allometric example, what is the human clearance prediction at \(b=0.60\), given rat clearance 0.080 L/h and a body-weight ratio of 280?
Solution: \(280^{0.60}\approx29.40\), so \(CL_H=0.080\times29.40=2.352\) L/h, or about 2.35 L/h. - For \(AUC=F D/CL\), calculate the dose needed for \(AUC=10\) mg·h/L, \(CL=6\) L/h, and \(F=0.50\).
Solution: Rearrange to \(D=AUC\times CL/F\). Substitution gives \(D=10\times6/0.50=120\) mg. This result assumes the linear exposure-dose relationship and stated values apply. - In the potency example, does \(q=1.5\) meet a 60% response threshold? Use the derived boundary.
Solution: The threshold is met when \(q\leq1.333\). Since \(1.5>1.333\), it is not met. Direct calculation gives \(EC_{50}=2(1.5)=3\) mg/L and \(E=100\%\times4/(3+4)=57.1\%\), also below 60%. - Why might a stress test that varies clearance and bioavailability independently be misleading?
Solution: If the assumptions are biologically related, independent extreme combinations may not be plausible. A defensible analysis should preserve the known relationship or define coherent joint scenarios; otherwise the calculated corners should not be described as realistic cases.