From Evidence to an Auditable Recommendation
A dose recommendation is more useful when a reviewer can follow how the evidence supports it, where uncertainty remains, and which assumptions connect the two. A list of favorable results is not yet a justification: it may omit the comparator, the clinically meaningful threshold, the safety constraint, or the reason a higher dose was not selected. A traceable argument makes those links explicit without claiming more certainty than the data provide.
The dose-response evidence, safety constraints, treatment duration, and dose-reduction evidence considered in earlier material can each contribute to the rationale. The task here is to connect them in a consistent account, not to re-estimate every model or repeat every analysis. The proposed dose should be identifiable, the evidence should be interpretable in context, and limitations should be stated alongside the claims they qualify.
A practical argument usually answers five questions. What dose and regimen are proposed? What efficacy evidence supports it, relative to what comparator and on what endpoint? What uncertainty accompanies the estimates? What safety findings or constraints limit the choice? What assumptions and remaining risks could change the conclusion? These questions do not prescribe one statistical model. They provide a check that the decision can be reconstructed from the evidence presented.
Name the dose, regimen, intended population, and treatment context. Avoid referring only to “the selected dose.”
Identify the study, endpoint, time point, comparator, estimate, and units that support the efficacy and safety claims.
Explain what the estimates do and do not establish, and how safety findings or predefined criteria affect the choice.
Separate observed findings from extrapolations, state assumptions that matter, and identify unresolved risks and how they will be monitored.
What Makes a Claim Traceable?
A reviewer should be able to move from a conclusion back to the evidence that supports it. For example, “100 mg is selected because efficacy is maintained with an acceptable safety profile” is too compressed to audit. “At Week 12, the estimated improvement relative to placebo at 100 mg was 5.0 points on the prespecified symptom scale (95% confidence interval 3.24 to 6.76); the estimated improvement at 150 mg was 5.4 points, while adverse-event incidence was higher at 150 mg” is more informative. The latter still requires study context and a clear account of what “acceptable” means, but it exposes the relevant comparisons.
An evidence record can be kept as a table during analysis even if the final rationale is written as prose. The table should preserve the distinction between an observation, an interpretation, and a decision. For example, an observed event count is not itself a safety conclusion; interpreting it requires its denominator, follow-up, population, and uncertainty. A recommendation is not a direct output of an estimate; it also reflects clinical criteria and assumptions.
| Argument element | What to record | Traceability check |
|---|---|---|
| Recommendation | Dose, regimen, population, and treatment context | Is the proposed use unambiguous? |
| Efficacy evidence | Endpoint, time point, comparator, estimate, units, and interval | Can the reader identify what the estimate measures? |
| Safety evidence | Event definition, counts and denominators, exposure context, and uncertainty | Is the safety claim proportional to the amount of evidence? |
| Decision criterion | Relevant clinical or development threshold and its source | Is the criterion stated rather than implied? |
| Assumptions and risks | Key assumptions, extrapolations, and unresolved uncertainties | Could a changed assumption alter the recommendation? |
A confidence interval summarizes sampling uncertainty under the specified analysis and its assumptions. It does not cover every source of uncertainty: for example, an interval may not account for an inappropriate endpoint, biased missing-data assumptions, or limited generalizability to a different population. A traceable argument should therefore name both the reported statistical uncertainty and material uncertainties not represented by the interval.
Worked Examples
Worked Example: Connecting Efficacy Estimates to a Dose Choice
Hypothetical data. A randomized, double-blind, 12-week trial compares placebo with three daily doses in adults with a chronic condition. The prespecified endpoint is improvement on a symptom scale, in points, with larger positive values indicating greater improvement. Model-adjusted mean improvements are 2.0 points for placebo, 5.8 for 50 mg, 7.0 for 100 mg, and 7.4 for 150 mg. The adjusted dose-versus-placebo contrasts and their standard errors are reported as follows: 50 mg, 3.8 points (SE 0.85); 100 mg, 5.0 points (SE 0.90); and 150 mg, 5.4 points (SE 0.95). Assume the analysis plan supports normal-based 95% confidence intervals for these contrasts.
For the 100 mg contrast, the interval is the estimate plus or minus 1.96 standard errors:
Rounded to two decimal places, this is 3.24 to 6.76 points. Suppose the prespecified clinically meaningful improvement relative to placebo is 3 points. The point estimate at 100 mg exceeds that threshold, and the lower confidence limit is also above it. Under the analysis assumptions, this supports a clinically meaningful effect at this dose; it does not establish that every patient benefits or eliminate uncertainty about the endpoint and model.
The 150 mg estimate is 5.4 points, only 0.4 points greater than the 100 mg estimate. That difference alone is not evidence that 150 mg provides a meaningful advantage. A dose-to-dose contrast and its uncertainty, estimated with the appropriate covariance from the same trial, would be needed to characterize that comparison. A traceable rationale would therefore report the evidence for both doses, avoid describing the 150 mg estimate as definitively superior, and connect the choice between them to the safety evidence and decision criteria.
Worked Example: Showing Safety Evidence and Its Uncertainty
Hypothetical data. In the same imagined trial, an adverse event of special interest occurs in 5 of 200 placebo recipients, 10 of 200 recipients of 100 mg, and 18 of 200 recipients of 150 mg. The observed proportions are calculated as the event count divided by the number of recipients:
For a simple illustration of uncertainty in each proportion, use a 95% Wilson interval. For \(x\) events among \(n\) recipients, let \(\hat{p}=x/n\), \(z=1.96\), and define the interval center \(C\) and half-width \(H\) by:
For 100 mg, \(\hat{p}=0.050\) and \(n=200\). Substitution gives \(C\approx0.05848\) and \(H\approx0.03110\), so the interval is approximately \(0.02738\) to \(0.08958\), or 2.74% to 8.96%. For 150 mg, \(\hat{p}=0.090\); the same calculation gives \(C\approx0.09773\) and \(H\approx0.04004\), or approximately 5.77% to 13.78%. These intervals describe uncertainty in each event proportion, not a confidence interval for the difference between doses.
Suppose, for illustration, the development team has specified that a dose with an observed incidence above 8% requires additional justification. The 150 mg observed incidence is 9%, whereas the 100 mg incidence is 5%. The 150 mg estimate triggers the stated review criterion, but its interval also shows that the true incidence is not precisely determined. The 100 mg interval extends above 8%, so the point estimate alone cannot establish that the underlying incidence is below that level. The rationale should therefore report the criterion, counts, denominators, uncertainty, and any further safety review; it should not call 100 mg “proven safe” or treat 150 mg as having a known fixed risk.
Worked Example: Writing the Rationale With Assumptions Visible
Hypothetical conclusion using the examples above. A concise, traceable paragraph could read: “We propose 100 mg once daily for adults represented by the randomized 12-week trial. At Week 12, the adjusted improvement relative to placebo was 5.0 points on the symptom scale (95% confidence interval 3.24 to 6.76), exceeding the prespecified 3-point threshold in the point estimate and interval. The 150 mg point estimate was 5.4 points, only 0.4 points greater; the available summary does not establish that this increment is clinically meaningful. The adverse event of special interest occurred in 10/200 recipients of 100 mg (5.0%; 95% Wilson interval 2.74% to 8.96%) and 18/200 recipients of 150 mg (9.0%; 5.77% to 13.78%). The 100 mg estimate is below the illustrative 8% review criterion, but its interval does not exclude incidence above that criterion. This recommendation assumes that the trial population, endpoint, 12-week exposure, and analysis are relevant to the intended use. Longer-term safety and effects in populations not adequately represented remain residual risks and require continued assessment.”
Each sentence has a role: the first specifies the proposal; the next two provide comparative efficacy evidence; the safety sentence gives event counts and uncertainty; and the final sentences identify an assumption and residual risks. The paragraph does not claim that 100 mg is risk-free, that 150 mg is definitively more efficacious, or that the trial proves effects beyond its observed duration and population.
Justifying the Level of Detail
A concise argument is not an argument with missing links. One useful test is to ask whether a skeptical reviewer could answer three questions from the record: what evidence supports each key claim, what assumptions are required to use that evidence for the proposed dose, and what evidence might weaken the recommendation? If the answer to any question depends on an unstated premise, the rationale should either document that premise or narrow the claim.
For instance, a dose tested for 12 weeks does not by itself establish an acceptable profile over a year. Applying the estimate to longer treatment requires an explicit basis, such as longer-term data or a stated extrapolation, and a description of remaining uncertainty. Likewise, a trial estimate for its enrolled population does not automatically establish the same response in an underrepresented subgroup. These are not reasons to omit the proposed dose; they are reasons to describe the reach of the evidence accurately.
The appropriate comparison depends on the question. A placebo comparison can establish the estimated benefit relative to placebo under the study design, but it may not answer whether a higher dose adds worthwhile benefit over the proposed dose. A dose-to-dose comparison addresses that question, provided its estimate and uncertainty are reported. Similarly, observed safety proportions describe event frequency in the studied groups; a prespecified exposure-adjusted analysis may be more appropriate when follow-up or exposure differs materially. The rationale should use the comparison that matches the claim rather than substituting one comparison for another.
A model-based estimate also depends on the model and analysis assumptions. Relevant items may include the endpoint definition, model specification, handling of missing observations, treatment adherence, and whether the estimate is interpolated within observed doses or extrapolated beyond them. State the assumptions that materially affect the conclusion, not every technical detail in the analysis plan. Where a conclusion changes under a plausible alternative assumption, describe that sensitivity and avoid presenting the conclusion as unconditional.
Diagnostics and Common Failure Modes
A rationale is becoming difficult to audit when its key estimates lack a comparator, units, time point, or population; when “acceptable safety” appears without a stated basis; or when a model prediction is presented as if it were directly observed. A further warning sign is an apparent contradiction between the evidence and the recommendation, such as selecting a dose because of a small estimated efficacy increment while omitting a dose-related safety concern. The explanation should show how the evidence and criteria resolve the trade-off, or acknowledge that the decision remains uncertain.
Common mistake: “The 100 mg dose is safe because only 5% of participants had the event, below the 8% limit.” This reasoning treats an observed proportion as the known underlying risk and ignores the interval, study duration, and population. The correction is to report the count and denominator, quantify uncertainty with an appropriate method, define what the 8% criterion means, and state that the observed estimate is below the criterion while uncertainty remains about the underlying incidence.
Other frequent errors include presenting a dose-response trend without showing the underlying doses and estimates; implying a causal effect from a nonrandomized association without discussing its limitations; and using “no statistically significant difference” as proof that two doses are equivalent. A traceable rationale reports the relevant estimates and their uncertainty, limits claims to what the design supports, and identifies evidence gaps that matter to the decision.
Check Your Understanding
For each question, try to identify the evidence-to-claim link before reading the solution.
- A report says, “The higher dose was better.” Name three details needed to make that claim traceable.
- In a hypothetical trial, 8 of 160 recipients experienced an event. What is the observed proportion, and why is it not enough by itself to conclude that the dose is safe?
- A 12-week study supports a dose in its enrolled population. What must be addressed before claiming the same benefit and safety profile for a different population over one year?
- Why does a placebo comparison not necessarily establish that a higher dose offers worthwhile benefit over a proposed lower dose?
Solutions. (1) State at least the endpoint and units, the doses being compared, the comparator or contrast, the time point, the population, and the estimate with uncertainty; three or more of these are needed to interpret the claim. (2) The observed proportion is \(8/160=0.050\), or 5.0%. The underlying risk is uncertain, and interpretation also depends on the event definition, exposure duration, population, and appropriate safety criterion. (3) State the population and duration extrapolations, support them with relevant evidence or explicit assumptions, address uncertainty and differences in exposure or baseline risk, and identify remaining risks and monitoring. (4) The placebo contrast estimates benefit relative to placebo; the incremental benefit of the higher versus lower dose requires a direct dose-to-dose contrast and its uncertainty, interpreted against a relevant clinical criterion.