Building a Comparability Matrix for Conflicting Values

Why a Comparability Matrix Is Useful

A parameter dossier may contain several reported values for what appears to be the same PBPK input. Yet a shared label does not guarantee that the values describe the same quantity under the same conditions. For example, two values called “unbound fraction” may come from different biological matrices or assays. Placing them side by side without recording those distinctions can make a difference look like a contradiction, or conceal a meaningful difference between measurements.

A comparability matrix makes the basis for each comparison explicit. It organizes measurements and their metadata so a modeler can decide whether two values address the same endpoint, whether their measurement contexts align, and what uncertainty or missing information qualifies the comparison. Its purpose is to assess comparability, not to determine which value should ultimately be used or to explain why published measurements differ.

Definition: A measurement record is a reported result together with the information needed to interpret and trace it, including its source, measured endpoint, assay or method, population or material, experimental conditions, units, uncertainty, and any reported qualifiers. A comparability matrix is a structured table that records these attributes for multiple measurements and supports explicit pairwise judgments about whether the measurements can be compared directly, only conditionally, or not on the available information.

Build the Matrix Around the Measurement

Use one row for each distinct measurement record, not one row for each article. A single publication can report multiple endpoints, experimental conditions, or estimates; these should not be collapsed into one row if they represent different records. Keep the reported value and qualifiers intact, following the earlier material on extracting values and recording uncertainty.

A useful matrix includes fields such as these:

Traceability: citation or source identifier, table or figure, and any link to the original measurement.
Parameter and endpoint: the parameter label and the operational quantity measured, including how it was defined or calculated.
Population or material: species, donor group, tissue, matrix, preparation, or other relevant characteristics.
Assay or method: measurement technique, analytical procedure, and material preparation where reported.
Conditions: values and settings that may define the measurement, such as temperature, pH, incubation time, concentration range, or binding-protein concentration.
Reported result: the original value, units, qualifiers, and whether it is an observation, summary, or transformed quantity.
Uncertainty: the reported type of uncertainty and its numerical value, if available. Do not relabel a standard deviation as a confidence interval or assume an unreported uncertainty is zero.
Comparability notes: which fields align, differ, or are unknown, plus the resulting pairwise status and its rationale.

Keep source facts separate from your assessment. For example, “temperature not reported” is a fact about the available record; “temperature may affect comparability” is an assessment. This distinction makes the matrix auditable and prevents an assumption from silently becoming reported evidence.

Key distinction: A pair of values is directly comparable when the records support that both address the same operational endpoint and their relevant measurement contexts are aligned, with units made commensurate. They are conditionally comparable when a specified difference or missing detail could affect interpretation, so the comparison requires an explicit caveat. They are not comparable on the available information when the endpoint differs, a material mismatch prevents a meaningful comparison, or essential metadata are absent. These labels describe the comparison, not the quality of either source.

A Pairwise Check, Not a Single Score

Comparability is multidimensional. Two records may match on species and endpoint but differ in assay; another pair may use the same assay but measure different endpoints. A single numerical score can hide which dimension matters, so retain the dimension-by-dimension evidence and state the pairwise judgment in words.

A compact notation can help organize the review. For records \(i\) and \(j\), let \(c_{ij,k}\) denote the comparison of metadata dimension \(k\), recorded as aligned, different, or unknown. Let \(S_{ij}\) be the overall status: direct, conditional, or not comparable on the available information. The matrix is not computed by averaging the \(c_{ij,k}\) values. Instead, the endpoint and the dimensions relevant to interpreting it determine the status.

S_{ij} = \begin{cases} \text{direct}, & \text{same endpoint and relevant contexts align;}\\ \text{conditional}, & \text{endpoint aligns but a relevant difference or gap remains;}\\ \text{not comparable}, & \text{endpoint or essential basis for comparison does not align.} \end{cases}

This is a decision framework, not a statistical test. In particular, overlapping uncertainty intervals do not establish that two measurements are comparable: intervals describe uncertainty in reported estimates, not whether the estimates refer to the same endpoint or conditions. Conversely, a difference in reported values alone does not establish that the records are incomparable.

Worked Example: Aligning Units Before Comparing Microsomal Clearance

Suppose two hypothetical records report intrinsic clearance in human liver microsomes. Record A reports \(42\ \mu\text{L}/\text{min}/\text{mg protein}\). Record B reports \(0.050\ \text{mL}/\text{min}/\text{mg protein}\). Both specify the same operational endpoint, species, microsomal preparation, assay protocol, incubation conditions, and summary statistic. Record B reports an uncertainty of \(0.006\ \text{mL}/\text{min}/\text{mg protein}\); Record A reports no uncertainty estimate.

First convert Record B to the units of Record A. Since \(1\ \text{mL}=1000\ \mu\text{L}\),

0.050\ \frac{\text{mL}}{\text{min}\cdot\text{mg}} \times 1000\ \frac{\mu\text{L}}{\text{mL}} = 50\ \frac{\mu\text{L}}{\text{min}\cdot\text{mg}}.

Its reported uncertainty converts in the same way: \(0.006\times1000=6.0\ \mu\text{L}/\text{min}/\text{mg protein}\). The matrix therefore records \(42\) and \(50\ \mu\text{L}/\text{min}/\text{mg protein}\), while retaining that uncertainty is available only for Record B.

Dimension	Record A	Record B	Assessment
Endpoint	Microsomal intrinsic clearance	Microsomal intrinsic clearance	Aligned
Species and preparation	Human liver microsomes	Human liver microsomes	Aligned
Units and value	42 \(\mu\text{L}/\text{min}/\text{mg}\)	50 \(\mu\text{L}/\text{min}/\text{mg}\)	Commensurate after conversion
Uncertainty	Not reported	0.006 mL/min/mg	Incomplete for pairwise uncertainty assessment

Because the example stipulates that the endpoint and relevant assay conditions align, the values are directly comparable as measurements. The comparison should still note that their uncertainty reporting differs. The arithmetic difference is \(50-42=8\ \mu\text{L}/\text{min}/\text{mg protein}\), but that difference alone neither establishes a meaningful conflict nor identifies its cause.

Apply the Matrix to Assays and Populations

The comparison should follow the endpoint. An assay method matters when it defines or may affect how that endpoint was measured; a population or biological matrix matters when it changes what the measurement represents. Do not mark every difference as disqualifying. Record the difference first, then explain whether it prevents direct comparison, makes the comparison conditional, or is immaterial to the stated endpoint.

Worked Example: Comparing Unbound Fractions From Different Assays

Consider two hypothetical estimates of plasma unbound fraction, \(f_u\), both reported as dimensionless fractions for pooled adult human plasma at \(37^\circ\text{C}\). Equilibrium dialysis gives \(0.018\), while ultrafiltration gives \(0.026\). Both records identify plasma as the matrix, but the reports do not establish that the assay procedures, equilibration conditions, or nonspecific-binding checks are equivalent.

The values have the same stated endpoint and units, so no numerical unit conversion is needed:

$$ 0.026 - 0.018 = 0.008. $$

The matrix records a difference of \(0.008\) in fraction units, the matching stated matrix and temperature, and a method difference with relevant procedural details not shown to be equivalent. The pair is therefore conditionally comparable, not directly comparable on the available documentation. The assay distinction must remain visible alongside the values. The matrix does not establish that either estimate is incorrect or explain the observed difference.

This conclusion depends on the stated record. If a later metadata check showed a different plasma matrix or a different operational definition of the endpoint, the pairwise status would need to be reassessed. A comparability judgment is only as specific as its documented evidence.

Check the Endpoint Before the Number

A shared parameter name can conceal distinct measured quantities. Before comparing numbers, state the endpoint in operational terms: what was measured, on what material, and how the reported quantity was obtained. If records use different endpoints, the right action is to preserve them as distinct records rather than force them into one numerical comparison.

Worked Example: Detecting an Endpoint Mismatch

Two hypothetical reports provide values associated with a drug’s acidity. Record A reports an aqueous \(pK_a\) of \(7.4\) at \(25^\circ\text{C}\). Record B reports an apparent \(pK_a\) of \(7.8\) in a specified water–cosolvent mixture at \(25^\circ\text{C}\). The numerical difference is \(7.8-7.4=0.4\), but the records do not describe the same measurement medium or explicitly establish the same operational endpoint.

The matrix should retain the reported labels, values, temperature, and media. It should not treat the difference of \(0.4\) as a direct comparison of like-for-like aqueous measurements. On the information given, the records are not directly comparable; the endpoint and medium differ. Whether a transformation could relate the measurements would require additional information and is outside this comparison. Keeping the original values and qualifiers prevents an apparent numerical match from erasing the distinction.

Review the Matrix for Failure Modes

A matrix is useful only if its entries preserve the evidence and its status labels are justified. Review it for omissions that could make a comparison appear stronger or weaker than the records support.

Parameter-name matching: treating identical labels as proof of a shared endpoint. Correct this by writing the operational endpoint for each record.
Unconverted units: interpreting numerical magnitude before checking unit definitions and scales. Convert explicitly and retain the original reported units in the traceable record.
Missing metadata treated as a match: assuming an unreported condition equals the condition in another record. Mark it unknown, then assess whether the gap matters.
Uncertainty used as a comparability test: concluding that overlapping intervals make endpoints or methods equivalent. Assess endpoint and context separately from uncertainty.
Premature explanation: using a matrix to claim why values disagree. The matrix identifies documented similarities, differences, and gaps; explaining disagreement is a separate task.
Premature selection: treating “directly comparable” as “preferred input.” Comparability does not by itself rank sources or decide which value belongs in a model.

The main alternative to a structured matrix is a prose summary or a table containing values alone. Prose can provide context but makes systematic pairwise review difficult; a values-only table is compact but can suppress assay, population, conditions, or uncertainty. A matrix makes those attributes auditable, although it takes more effort to populate and cannot compensate for absent source information. Use only columns that help interpret the records, while retaining the details needed to trace each measurement.

Key takeaway: Compare records in a fixed order: verify the endpoint, check population or material and relevant conditions, reconcile units, inspect method and uncertainty reporting, then state a pairwise status with its rationale. Leave the cause of disagreement and the choice of a final model input unresolved at this stage.

Check Your Understanding

For each question, identify the relevant matrix entries and justify the comparison status.

Two records report the same endpoint and assay in the same species and conditions. One value is \(30\ \mu\text{L}/\text{min}/\text{mg}\); the other is \(0.040\ \text{mL}/\text{min}/\text{mg}\). What is the converted value, and what can the matrix conclude?
Two unbound-fraction records give values in the same units for the same stated plasma population, but one uses a different assay and the reports omit details needed to establish equivalent conditions. What status is appropriate?
One record reports an aqueous endpoint; another reports an apparent endpoint in a mixed solvent. Is the numerical difference alone a direct comparison?
Both records have overlapping reported uncertainty intervals, but one measures a different endpoint. Does overlap establish comparability?

Solutions. (1) \(0.040\times1000=40\ \mu\text{L}/\text{min}/\text{mg}\). The stated units are commensurate after conversion; given the stipulated matching endpoint and contexts, the measurements are directly comparable, while any difference in uncertainty reporting should be noted. (2) Conditional comparability: the endpoint and stated population align, but the method and missing condition details qualify the comparison. (3) No. The endpoints or measurement media differ, so they are not directly comparable on the information given; retain both records and their labels. (4) No. Interval overlap concerns reported uncertainty and cannot make different endpoints comparable.