Why Build an Assumption and Uncertainty Register?
A development decision can depend on more than estimated model parameters. It may also depend on assumptions about the target population, dose implementation, endpoint definition, external evidence, or how a model will be used. If these assumptions remain implicit, a decision-maker may see a precise prediction without seeing the conditions on which it rests.
An assumption and uncertainty register is a structured record of decision-relevant assumptions, the evidence supporting them, plausible alternatives, the predictions they may affect, and a plan for communicating their influence. Its purpose is traceability: a reader should be able to follow an assumption from its basis, through the model or analysis, to the decision it could inform. A register organizes the work; it does not, by itself, estimate probabilities, quantify predictive uncertainty, or establish that an assumption is correct.
This is different from listing every simplification in a model. A computational convention may be worth documenting, but it need not be a register priority if it cannot plausibly affect the decision. Conversely, an assumption supported by substantial evidence may still belong in the register if an alternative would change a decision-relevant prediction. The earlier material on distinguishing variability from decision uncertainty is useful here: variation among individuals is not itself the same thing as uncertainty about which assumptions or model inputs are appropriate.
What a Useful Entry Records
A register should be specific enough that another analyst can understand and review each entry without guessing what was intended. A practical entry identifies the decision context, states one assumption in testable language, describes its evidence and limitations, records at least one plausible alternative when appropriate, and names the prediction or decision criterion that could be affected. It also states how the issue will be communicated and who is responsible for review.
A concise traceability chain can be written as:
The arrows are substantive, not decorative. If the assumption has no credible path to a decision-relevant prediction, its priority may be low. If the prediction is named but the model quantity connecting it to the assumption is missing, the entry is not yet sufficiently traceable. A clear entry also separates what is observed from what is inferred: evidence is the source and its findings; the assumption is the proposition adopted for analysis.
| Register field | What to record |
|---|---|
| Decision and scope | The decision, population, analysis, and time point to which the entry applies. |
| Assumption | A concise statement of what is taken to hold for the analysis. |
| Evidence and limitations | Source, relevance to the decision context, and important gaps or differences. |
| Plausible alternative | A credible competing proposition, not an arbitrary extreme. |
| Affected prediction | The model quantity, predicted outcome, or decision criterion with a traceable link to the assumption. |
| Communication and ownership | How the issue will be described, by whom, and when it should be reviewed. |
Evidence descriptions should be concrete. “Supported by prior data” is less useful than naming the data source, the population and regimen it covers, and the mismatch that limits its relevance. Likewise, “could affect efficacy” is less useful than identifying the predicted endpoint and how that endpoint enters the decision rule. The register need not contain an exhaustive model specification; it should contain enough information to explain the decision-relevant chain and let a reviewer locate the underlying analysis.
Worked Example: Registering an Extrapolation Assumption
Hypothetical case. A team is considering whether to advance a regimen for a new patient population. Its current model predicts a 58% response rate, while the prespecified decision criterion is a predicted response rate of at least 55%. The model uses an assumption that the exposure-response relationship estimated in adults applies in the new population after accounting for body size.
State the assumption. “After the planned body-size adjustment, the adult exposure-response relationship is applicable to the target population.” This is more reviewable than “the model extrapolates well,” because it names what is being carried across populations.
Record evidence and its limitation. Suppose the hypothetical evidence consists of adult exposure-response data and a small target-population pharmacokinetic study. The former informs the relationship between exposure and response in adults; the latter informs exposure in the new population, but does not independently establish that the exposure-response relationship is unchanged. The entry should identify both sources and that limitation rather than presenting the combined evidence as direct validation.
Name a plausible alternative and the affected prediction. A plausible alternative is that the relationship differs in the target population. Suppose a prespecified alternative model fit gives a predicted response rate of 51%. The comparison is \(58\%-51\%=7\) percentage points. Against the criterion, the base prediction exceeds the threshold by \(58\%-55\%=3\) percentage points, whereas the alternative prediction falls below it by \(55\%-51\%=4\) percentage points. These are conditional predictions under two assumptions, not probabilities that either assumption is true.
Complete the communication plan. Record that the assumption could change whether the predicted response meets the criterion, identify the model owner responsible for updating the evidence review, and specify that both predictions and the assumption behind each will be shown to the decision group. The entry now connects evidence, an alternative, a prediction, and the decision criterion without claiming to quantify the probability of the decision changing.
Constructing and Reviewing the Register
The most reliable way to build a register is to begin with the decision rather than with an unfiltered list of model limitations. First state the decision and its criterion. Then work backward from the predictions used by that criterion: identify the assumptions needed to generate or interpret those predictions, and record the evidence and alternatives for each. This keeps attention on assumptions that could matter to the stated decision.
Name the decision, target population, analysis version, and decision criterion so that each entry has a defined scope.
State what the analysis takes to be true. Avoid broad labels such as “extrapolation” unless the specific proposition is also stated.
Identify the source and its relevance, then record a credible alternative that would be considered if the assumption were not adopted.
Name the affected output and decision criterion, then state how the dependence will be explained, who will review it, and when.
The register should distinguish the uncertainty about an assumption from the consequence if an alternative applies. The first concerns how well the evidence supports the adopted proposition; the second concerns the prediction or decision that could be affected. Neither should be silently substituted for the other. A weak evidence base does not automatically imply a large decision consequence, and a potentially large consequence does not establish that the alternative is likely.
Worked Example: Recording a Dose-Implementation Assumption
Hypothetical case. A development team predicts a 72% probability of meeting a clinical exposure target. The prediction assumes that 90% of intended doses are administered on schedule. The assumed administration rate is based on a feasibility study in a different setting. The team has also examined an alternative model input of 80%, which gives a predicted probability of 68%.
The register entry should not say simply “adherence may be lower.” It should state: “For the target setting, 90% of intended doses are administered on schedule.” The evidence field identifies the feasibility study and notes the setting difference. The alternative field records 80% as an examined alternative and explains why it is plausible, rather than presenting it as an established fact.
The affected prediction is the probability of meeting the clinical exposure target. The difference between the two conditional predictions is \(72\%-68\%=4\) percentage points. This arithmetic describes the difference in model outputs under the recorded inputs; it does not show that the administration rate is uncertain by 10 percentage points, nor that the prediction itself has a 4-percentage-point confidence interval.
For communication, the team can state that the 72% prediction depends on an administration-rate assumption informed by evidence from another setting, and that the examined 80% alternative gives 68%. The register should also assign an owner to assess whether more relevant implementation evidence is available. It should not turn the two outputs into a probability distribution or claim that one is more likely without a separate evidential basis.
Consistency, Duplication, and Prioritization
A register is only as useful as its internal consistency. One assumption may affect several predictions, and several assumptions may contribute to the same prediction. Entries should preserve those relationships rather than imply that each assumption operates in isolation. If two entries describe the same underlying uncertainty in different language, link or consolidate them. Otherwise, a reader may count one issue twice when interpreting the register or planning follow-up.
This concern is especially important when evidence sources overlap. For example, two reports may use the same underlying patients, or one estimate may already incorporate an adjustment that another entry describes separately. The register should note the shared source or dependency. Recording two labels does not create two independent pieces of evidence. Where the dependence cannot yet be resolved, state that limitation rather than treating the entries as separate confirmation.
Prioritization should consider both decision relevance and the quality and applicability of evidence. A team may flag an entry for review because its alternative could change whether a criterion is met, because the evidence is poorly matched to the target setting, or both. Such flags are prompts for discussion, not calibrated measures of probability, impact, or value. If a numerical rank or score is used for operational reasons, its construction and limitations need to be stated; a score should not be presented as a statistical result unless it has a defined statistical interpretation.
Worked Example: Handling Overlapping Evidence
Hypothetical case. Two draft entries concern exposure in a target population. Entry A says that the body-size adjustment is adequate, citing a 24-person pharmacokinetic study. Entry B says that predicted target-population exposure is supported by that same study. A reviewer finds that both entries cite the same 24 participants and both lead to the same predicted exposure distribution used in the decision analysis.
Check the evidence count. The evidence source is one study with \(n=24\) participants, not two independent studies. Counting the source once gives 24 participants; adding the two draft entries does not give \(24+24=48\) participants. The two entries may represent distinct propositions, but they do not provide independent data.
Resolve the register structure. Keep separate entries only if the propositions and their consequences are genuinely distinct: for example, whether the adjustment form is appropriate and whether the study population represents the intended population. In that case, cross-reference the shared source and state which feature of the study informs each proposition. If both entries make the same proposition and point to the same prediction, merge them to avoid duplication.
Plan communication. State that the same study informs both considerations and avoid language suggesting independent corroboration. This preserves the evidential provenance and helps decision-makers understand what is known without overstating the amount of supporting evidence.
Common Failure Modes and Limits
- Listing vague concerns. “Model uncertainty” does not specify an assumption. State the proposition, evidence, and decision-relevant link.
- Treating an alternative as equally supported. Listing a plausible alternative does not imply that it has the same evidential weight as the adopted assumption. Describe the evidence for each accurately.
- Using evidence labels without provenance. “External data support this” is not enough. Name what the data cover and how the context differs.
- Confusing a changed prediction with a probability of being wrong. A difference between two conditional predictions is not the probability that one assumption is false, and it is not automatically an uncertainty interval.
- Recording every technical detail at equal priority. An exhaustive issue list can hide the assumptions connected to the decision. Use the stated decision and prediction chain to focus the register.
- Leaving the register static. New evidence, a changed population, or a revised decision criterion can make an old entry incomplete. Version the register and record when and why entries are updated.
A register is not a substitute for model diagnostics, a quantitative uncertainty analysis, or a justified decision rule. It does not establish that all important assumptions have been found, and it cannot repair weak evidence. Its contribution is to make assumptions and their decision pathways visible, reviewable, and communicable. Quantifying how parameter uncertainty propagates to a clinical decision requires additional analysis beyond the register itself.
Check Your Understanding
Answer each question using the distinction between an assumption, its evidence, an alternative, and an affected prediction.
- A model uses data from adults to predict an outcome in adolescents. What should an assumption-register entry say beyond “extrapolation uncertainty”?
- A prediction is 64% under one assumption and 59% under a plausible alternative. What is the numerical difference, and what does that difference not establish?
- Two register entries cite the same study. What should be checked before describing them as two sources of support?
- Why can an assumption with strong evidence still merit an entry?
Solutions. (1) State the specific proposition carried from adults to adolescents, identify the adult evidence and its relevance limits, name a plausible alternative, and link both to the affected prediction and decision criterion. (2) The difference is \(64\%-59\%=5\) percentage points. It does not establish the probability that either assumption is true, nor does it automatically quantify uncertainty in the prediction. (3) Check whether the entries represent distinct propositions and whether the study data overlap; the same participants or analysis must not be presented as independent evidence. (4) Evidence strength and decision relevance are different: even a well-supported assumption may matter if a credible alternative could affect an important prediction or criterion.