1. Define the result and data range before choosing a model
State the relative-potency estimand, response and concentration transformations, reference assignment, intended reportable range, uncertainty requirement, and whether the design targets a bounded central region or the broader sigmoid.
The model must represent that prespecified question. Selecting doses or a model after seeing which combination produces the preferred potency changes the analysis and hides selection uncertainty.
2. Match the retained response design to the model question
The retained teaching fixture spans 1 to 160 ng/mL across the lower response, transition, and upper response for both reference and test. Reference and test midpoints are 20.000 and 25.000 ng/mL. Because the intended result uses the entire prespecified sigmoid design and development evidence supports a symmetric shared shape, the selected routine model is common-shape 4PL.
A parallel-line candidate would require a separately predefined transformed interval where both preparations are adequately linear and support a common slope. A 5PL candidate would require repeatable asymmetry and stable estimation of its extra parameter. Neither condition is created merely by selecting a convenient subset or a more flexible fit after seeing this plate.
| Candidate | Question applied to the fixture | Worked disposition |
|---|---|---|
| Parallel-line log-log | Was an approximately linear transformed interval predefined and supported? | Not selected for this full-sigmoid design |
| Common-shape 4PL | Do both preparations support a shared symmetric sigmoid shape over 1–160 ng/mL? | Selected development model |
| Common-shape 5PL | Is repeatable asymmetry supported and the fifth parameter stable? | Not justified by the symmetric teaching fixture |
The intended range determines which relationship is tested
Prespecified central transformed range
Question: do predefined transformed responses support linearity and a common slope?
Full designed sigmoid range
Question: do all designed concentrations support the method-defined sigmoid relationship?
- Parallel-line view
- Uses four named central levels only to explain a prespecified linear-range question.
- Common-shape 4PL view
- Uses all eight retained levels because this fixture was designed for the broader sigmoid workflow.
3. Compare candidates on decision-relevant evidence
During development, fit prespecified candidates and retain every result. Evaluate whether the designed doses support the required region and parameters, whether comparability assumptions hold, how residuals behave, whether uncertainty is stable, and whether known-sample recovery and repeated-run results improve.
| Dimension | Parallel-line | Common-shape 4PL | Common-shape 5PL |
|---|---|---|---|
| Designed response region | Predefined approximately linear transformed interval | Supported lower, transition, and upper sigmoid regions | Sigmoid with repeatable supported asymmetry |
| Comparison assumption | Common transformed-dose slope | Method-defined shared nonlinear shape parameters | Method-defined shared asymmetric shape parameters |
| Primary diagnostics | Linearity, parallelism, residuals, range | Coverage, independent shape evidence, residuals, parameters | 4PL evidence plus stability of asymmetry |
| Primary risk | Selecting the linear range after seeing results | Unsupported plateaus or imposed shared shape | Weakly identified extra parameter |
4. Lock one exact calculation for routine use
Approve the included doses, transformations, equation, restrictions, weighting, bounds, optimizer and convergence rule, uncertainty method, comparability diagnostic, suitability criteria, units, and rounding. Give that complete calculation a method version.
Routine analysis applies the approved version without searching across parallel-line, 4PL, and 5PL results. A model or range change creates a new method version and requires impact assessment and appropriate validation evidence.
5. Hand the selected model to the controlled workflow
The routine workflow maps reference and test observations, applies the locked fit, evaluates comparability, calculates potency and uncertainty, applies suitability, and reports the result with the model identity and retained evidence.
Report transformed variables, included doses, parameters and restrictions, variance and weighting, estimates and uncertainty, diagnostics, exclusions, reference assignment, units, criterion outcomes, software version, and configuration fingerprint.
Limits and current Provenarium support
A model is not better because it is newer or uses more data. Parallel-line, 4PL, and 5PL answer different prespecified questions and require their own expected-result, boundary, and failure tests.
Provenarium supports three exact supplier-verified quantitative relative-potency contracts: parallel-line log-log, common-shape 4PL, and common-shape 5PL. Model choice remains a controlled development decision; routine execution does not compare alternatives or switch after a failed fit.
Frequently asked questions
Is 4PL always more modern and therefore better?
No. Model choice depends on the analytical response, design, assumptions, intended result, and evidence—not chronology.
Does parallelism mean the plotted curves look parallel?
Visual inspection is useful but not sufficient. Define the statistical or model-based comparability assessment and its acceptance rule.
Which relative-potency models does Provenarium currently support?
Provenarium supports three exact supplier-verified contracts: parallel-line log-log, common-shape 4PL, and common-shape 5PL. A method fixes one before routine execution.