Why Scenario Analysis Is Useful With Sparse Evidence
When preclinical evidence is limited, a single human prediction can conceal how much depends on assumptions. One animal species may identify a plausible clearance pathway, for example, without establishing whether that pathway has the same contribution in humans. A numerical prediction may still be useful, but only if the assumptions that produce it are visible and alternatives are considered explicitly.
Scenario analysis makes those alternatives concrete. It asks: what human predictions follow under different biologically plausible, internally coherent descriptions of the translation problem? The scenarios are not automatically statistical samples, confidence limits, or equally likely explanations. They are structured conditional cases. Their value is to expose dependence on uncertain biology and to help determine what evidence would distinguish the cases.
A scenario is not simply a label such as “low,” “typical,” or “high.” It must state what biological circumstances the label represents and how those circumstances enter the calculation. Nor is it a substitute for evidence: the assumptions should be linked to observations, established biology, or a clearly identified knowledge gap. The earlier material on structural uncertainty and model discrepancy provides relevant context, but the practical emphasis here is on organizing and calculating coherent alternatives when the evidence does not select one uniquely.
Constructing Scenarios Before Calculating
Start by separating three things that are often mixed together: the measurements, the model inputs estimated from those measurements, and assumptions that cannot be identified from the available data. A sparse data set may estimate an animal clearance reasonably well while leaving human enzyme contribution uncertain. The animal estimate is not itself an alternative scenario; alternative scenarios arise from different plausible ways to translate that estimate to humans.
A useful scenario set should meet four conditions. First, each case has a biological rationale. Second, assumptions within a case are mutually compatible. Third, the cases cover consequential alternatives rather than merely changing a number without explanation. Fourth, shared assumptions remain shared, so that differences in predictions can be traced to the assumptions that actually vary.
For example, if a compound is cleared by a pathway observed in one animal species, plausible cases might represent (1) similar human pathway contribution, (2) lower human contribution with limited compensation by other pathways, or (3) lower contribution but substantial compensation by another human pathway. These are coherent hypotheses about pathway contributions. Combining the lowest clearance with the highest bioavailability merely because both are individually possible may instead create an implausible corner case. The biological story should determine which inputs vary together.
A prediction also requires a defined quantity and conditions. In the worked examples below, the quantity is an oral dose that would produce a specified AUC under linear pharmacokinetics. The formula is used as a transparent calculation device; it does not establish that linearity, the target AUC, or any scenario assumption is true in humans.
Worked Scenario Predictions
Worked Example: Three Coherent Human Exposure Scenarios
Hypothetical data and assumptions. Suppose sparse preclinical and in vitro evidence supports a reference human clearance estimate of 20 L/h if the dominant pathway is conserved. It does not determine whether human clearance is lower because that pathway contributes less, or higher because another pathway compensates. Oral bioavailability is also uncertain. For illustration, a translational team specifies three coherent cases: conserved pathway and moderate absorption (\(CL=20\) L/h, \(F=0.40\)); lower pathway contribution with greater absorption (\(CL=12\) L/h, \(F=0.60\)); and compensatory human clearance with lower absorption (\(CL=32\) L/h, \(F=0.25\)). These are hypothetical scenario inputs, not estimates established by the sparse data. The common target is \(AUC^*=10\) mg·h/L.
The dose calculation is applied separately within each case. In the conserved-pathway case, \(Dose = (10\ \mathrm{mg\cdot h/L})(20\ \mathrm{L/h})/0.40 = 500\ \mathrm{mg}\). The units reduce to mg because \((\mathrm{mg\cdot h/L})(\mathrm{L/h})=\mathrm{mg}\).
In the lower-clearance, greater-absorption case, \(Dose = (10)(12)/0.60 = 200\) mg. In the compensatory-clearance, lower-absorption case, \(Dose = (10)(32)/0.25 = 1{,}280\) mg. The resulting scenario-specific predictions are therefore 200, 500, and 1,280 mg for the same target AUC.
The range from 200 to 1,280 mg is a scenario range, conditional on the three cases chosen. It is not a 95% interval, and it does not imply that doses inside the range are equally plausible or that values outside it are impossible. The large difference is informative: both clearance and bioavailability assumptions materially affect the dose calculation. It also makes clear why reporting only the 500 mg reference case would hide important dependence on assumptions.
Before treating these outputs as meaningful, check the assumptions used in the calculation. The dose-AUC relation presumes linearity, and the scenarios must refer to the same target AUC and compatible definitions of clearance and bioavailability. The arithmetic alone cannot validate those conditions. In practice, check whether the assumed dose range is plausibly linear, whether the bioavailability values have a biological rationale, and whether clearance and bioavailability combinations describe coherent human physiology.
Scenario Spread Is Not Statistical Uncertainty
Statistical uncertainty concerns variation in an estimated quantity conditional on a specified model and its assumptions—for example, uncertainty in clearance estimated from noisy observations under one scenario. Scenario spread instead reflects differences in predictions when assumptions or model descriptions change. Sparse evidence can produce both at once. Keeping them separate prevents an attractive but incorrect interpretation of a range of scenario outputs as though it were a calibrated prediction interval.
Suppose, for a given scenario, a model supplies a distribution for the prediction because parameters have been estimated with uncertainty. That distribution describes uncertainty conditional on the scenario. If the scenario itself is uncertain, an overall probability distribution can be constructed only after specifying how scenarios are weighted, such as with justified probabilities. Without such weights, the scenarios can be compared, but their outputs should not be pooled as if each had a known probability.
Worked Example: Conditional Intervals Within Scenarios
Continue the hypothetical dose example. Suppose a statistical analysis within each scenario represents uncertainty in the scenario-specific dose prediction as lognormal, with the scenario point prediction as the median and log-scale standard deviation \(0.15\). This is an additional illustrative assumption; it does not describe uncertainty about which scenario is correct. For a lognormal distribution with median \(m\) and log-scale standard deviation \(s_{\log}\), the central 95% interval is \(m \times [\exp(-1.96s_{\log}),\exp(1.96s_{\log})]\).
Here \(1.96(0.15)=0.294\), so the multiplicative limits are \(\exp(-0.294)=0.745\) and \(\exp(0.294)=1.342\). For the 200 mg scenario, the conditional interval is \(200 \times [0.745,1.342]=[149,268]\) mg. For the 500 mg scenario it is \([373,671]\) mg, and for the 1,280 mg scenario it is \([954,1{,}718]\) mg, with rounding to three significant figures.
Each interval describes uncertainty conditional on its own scenario and the stated lognormal assumption. The intervals do not assign probabilities to the scenarios. In particular, the gap between 671 mg and 954 mg does not establish that doses in that gap are impossible; it arises because these are separate conditional distributions, not one continuous interval covering all assumptions.
Why the Variance Decomposition Matters
When defensible probabilities for scenarios are available, the distinction can be expressed precisely. Let \(Y\) be a prediction and \(S\) identify the scenario. If \(S\) has specified probabilities and the conditional distributions of \(Y\) have finite variance, then \(\mathrm{Var}(Y)=E[\mathrm{Var}(Y\mid S)]+\mathrm{Var}(E[Y\mid S])\). The first term is the average within-scenario variance; the second is the variance among scenario-specific means. This identity follows by writing \(Y-E[Y]=(Y-E[Y\mid S])+(E[Y\mid S]-E[Y])\), squaring, and taking expectations. The cross-product has expectation zero because \(E[Y-E[Y\mid S]\mid S]=0\). Thus the two components have distinct meanings and together give total variance under the specified scenario probabilities.
The identity does not tell analysts what scenario probabilities to use. If the evidence does not justify weights, reporting a variance from arbitrarily weighted cases gives a numerical appearance of precision without a defensible probability interpretation.
Worked Example: Separating Within-Scenario and Between-Scenario Variance
For a purely illustrative calculation, suppose the three dose predictions 200, 500, and 1,280 mg are assigned scenario probabilities 0.50, 0.30, and 0.20. Also suppose the conditional arithmetic standard deviation in each scenario is 20% of its scenario mean. These probabilities and conditional standard deviations are stipulated for the example; they are not inferred from the sparse preclinical evidence.
The weighted mean is \(E[Y]=(0.50)(200)+(0.30)(500)+(0.20)(1{,}280)=506\) mg. The between-scenario variance is \((0.50)(200-506)^2+(0.30)(500-506)^2+(0.20)(1{,}280-506)^2=166{,}644\) mg\(^2\).
The within-scenario variances are \(40^2=1{,}600\), \(100^2=10{,}000\), and \(256^2=65{,}536\) mg\(^2\). Their weighted average is \((0.50)(1{,}600)+(0.30)(10{,}000)+(0.20)(65{,}536)=16{,}907.2\) mg\(^2\). Therefore, total variance is \(166{,}644+16{,}907.2=183{,}551.2\) mg\(^2\), and the total standard deviation is \(\sqrt{183{,}551.2}=428\) mg. Under these stipulated weights and distributions, variation among scenario means is the larger variance component. This numerical conclusion applies only to the assumptions in this example.
Diagnostics, Alternatives, and Limitations
Scenario analysis is most useful when the scenario set is inspected as carefully as the calculations. Warning signs include a scenario with no biological rationale, an assumption that conflicts with observed pathway evidence, an extreme combination created by varying inputs independently when they are biologically linked, or an omitted case that could change the decision. Predictions should also be checked for dimensional consistency, plausible magnitude, and sensitivity to the assumptions that distinguish the scenarios. If all scenarios produce nearly identical outputs, verify that the uncertain assumptions were actually propagated into the calculation.
The main alternatives answer somewhat different questions. A single reference-case prediction is concise but conceals dependence on assumptions. A probabilistic predictive analysis represents uncertainty with distributions and can combine sources of uncertainty, but requires defensible distributions and dependence assumptions. Model averaging can combine alternative models when weights are supportable. Scenario analysis is particularly useful when plausible alternatives can be described but the evidence is insufficient to assign credible probabilities. It can also be used alongside a probabilistic analysis: quantify parameter uncertainty within cases, then show how predictions change across cases.
Scenario analysis has limitations. The selected cases may miss a relevant biological mechanism; a narrow scenario range can therefore be falsely reassuring. Conversely, an intentionally broad range may be uninformative if it combines incompatible assumptions. Scenario spread is sensitive to which cases are included, and it does not establish which case is most likely. Where scenario-specific statistical intervals overlap, that overlap does not resolve uncertainty about the biology. The next evidence-gathering step should be chosen by asking which measurement or experiment would distinguish consequential cases, rather than by treating the current scenario range as a final answer.
A Practical Sequence
Specify the human quantity, target, population or conditions, and units before selecting assumptions.
Record what the preclinical observations estimate and what remains unmeasured or unidentified for humans.
Describe the biological interpretation of each case and retain dependencies among assumptions that should vary together.
Use the same prediction definition in every case; verify units, model conditions, and arithmetic.
Label conditional statistical uncertainty and scenario spread distinctly, and state whether any scenario weights are justified.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- In a hypothetical scenario, \(CL=15\) L/h, \(F=0.50\), and the target AUC is 8 mg·h/L. What dose follows from the stated linear-PK model?
- Two scenarios give predictions of 300 mg and 900 mg. Can their difference alone be described as statistical uncertainty? Why or why not?
- Within one scenario, the conditional dose median is 400 mg and the log-scale standard deviation is 0.10. Using the stated central 95% lognormal interval formula, calculate the interval.
- What additional information is needed to combine scenario-specific distributions into an overall probability distribution?
Solutions. (1) \(Dose=(8\ \mathrm{mg\cdot h/L})(15\ \mathrm{L/h})/0.50=240\) mg. This calculation is conditional on the linear-PK model and stated scenario inputs. (2) No. The 600 mg difference is scenario spread: it compares conditional predictions under different assumptions, and supplies neither a sampling distribution nor a coverage probability. (3) The multiplicative limits are \(\exp(-1.96(0.10))=0.822\) and \(\exp(1.96(0.10))=1.217\), so the interval is \(400\times[0.822,1.217]=[329,487]\) mg, approximately. It is conditional on the scenario and lognormal assumption. (4) Defensible probabilities or weights for the scenarios, along with their conditional predictive distributions, are required. Without justified weights, compare the scenario-specific results rather than presenting a weighted overall probability distribution.