Standardize Covariate Units and Reference Values

Why Covariate Units Need a Consistent Convention

A continuous covariate such as body weight, age, or a laboratory measurement may arrive in more than one unit or format. A recorded value of 70 could mean 70 kg, 70 lb, or a value encoded in a different scale. If those values are treated as though they share one unit, summaries and later model inputs can be misleading. In population pharmacokinetics, a unit mistake can also make a covariate appear implausible or change its numerical scale substantially.

The task is not to make every covariate numerically similar. It is to put measurements of the same quantity into a declared, consistent unit and, when useful, express them relative to a declared reference value. Unit conversion, centering, and reference scaling are distinct operations. Keeping them distinct makes the transformation auditable and helps prevent a convenient numerical convention from being mistaken for a correction to the source data.

Conversion, Centering, and Scaling

Let \(x\) be a measured covariate, with units \(U\). A unit conversion to units \(V\) has the form \(x_V=cx_U\), where \(c\) has units \(V/U\). The numerical factor must correspond to the actual units and quantity being converted. For example, a factor for converting pounds to kilograms is not applicable to a laboratory concentration.

After choosing a common unit, a reference value \(x_{\mathrm{ref}}\) can be used in either of two common ways. The centered value \(x-x_{\mathrm{ref}}\) retains the covariate's units. The ratio \(x/x_{\mathrm{ref}}\) is dimensionless when numerator and denominator refer to the same quantity in the same units. A log-ratio, \(\log(x/x_{\mathrm{ref}})\), is also dimensionless, but it is a different transformation and should only be used when that convention is intended.

A reference value is a declared point of comparison, not automatically the sample mean, a normal value, or a clinical threshold. For example, 70 kg may be selected as a convenient weight reference. The choice should be recorded with the analysis convention and used consistently. If a reference value is expressed in a different unit from the covariate, convert one before subtracting or dividing.

A Short Justification: Transformations Must Preserve Meaning

A consistent change of units does not change the underlying measurement. If a value in unit \(U\) is converted to unit \(V\) using \(x_V=cx_U\), then a reference value must be converted by the same factor. Consequently, the relative value is unchanged:

This cancellation is why ratios are unit-invariant only when both values have been converted consistently. Mixing a value in pounds with a reference in kilograms does not produce a valid ratio.

Centering also has a precise effect when a quantity is written as a linear expression. If \(\eta=\theta_0+\theta_1x\) and \(x_c=x-x_{\mathrm{ref}}\), substitution gives \(x=x_c+x_{\mathrm{ref}}\), so

Thus the expression can be represented with the centered covariate if the intercept is adjusted as shown; the slope is unchanged. This is an algebraic identity, not a claim that centering corrects measurement errors or improves every analysis. The reference value and units still need to be stated so that the transformed quantity can be interpreted.

Harmonizing Laboratory Units

Laboratory covariates require particular care because the same analyte may be reported in different units by different laboratories or source files. The unit belongs to the measurement, not merely to the variable name. A column called “creatinine” is not sufficient evidence that every row uses the same unit. Use source metadata, laboratory documentation, or an explicit data dictionary to establish the unit for each record or source segment before conversion.

For serum creatinine, a commonly used conversion is \(1\text{ mg/dL}=88.4\text{ µmol/L}\). This factor is specific to creatinine. It must not be reused for a different analyte simply because that analyte is also measured as a concentration. The converted data should have one declared unit, and quality checks should compare source values with converted values using the recorded source units.

Incompatible Formats and Scale Checks

Sometimes records express the same covariate using different formats rather than a clearly labelled unit column. Height might appear as “1.72 m,” “172 cm,” or “68 in.” A numeric field may also have been imported as text, or decimal punctuation may differ by locale. These cases require two separate questions: what quantity and unit does the source represent, and how should it be parsed without losing its original representation?

Use known format rules and source documentation to interpret the value; do not infer units from the number alone. In particular, an unexplained factor of 10, 100, or 1,000 is a reason to investigate scale, prefixes, and decimal conventions, not permission to multiply or divide until the values look typical. A plausibility screen can identify records for review, but it cannot establish the correct unit.

Practical Quality-Control Sequence

A reliable workflow separates evidence, transformation, and review. The following sequence is useful for continuous covariates whether the final convention is a common unit alone or a common unit plus a reference value.

Useful diagnostics include counts by source unit, minimum and maximum values before and after conversion, and summaries stratified by laboratory or file. A sudden shift in range at a source boundary can indicate mixed units or a changed encoding. Also check whether conversion changed the ordering of values: a positive conversion factor should preserve it. If a positive measurement becomes negative, or if an expected unit conversion produces an implausible magnitude, review the factor, parsing, and source assignment.

These checks flag inconsistencies; they do not identify their cause automatically. An extreme value may be genuine, as established in the earlier material on distinguishing data errors from genuine extreme concentrations. For covariates as well, do not delete or alter a value solely because it is unusual. Investigate it against the source record and document any correction under the established data-correction process.

Choosing a Reference and Avoiding Common Errors

A reference value should be selected for a clear, stable purpose. It may be a clinically familiar value, a convenient round value, or a value chosen to make a transformed covariate easy to interpret. Whatever the reason, it must be stated with units. A reference value does not change the raw measurement, and it does not make different source units interchangeable. First harmonize units; then apply the reference convention.

The main alternatives serve different purposes. Keeping the covariate in its common physical unit preserves direct interpretation, such as kilograms or years. Centering expresses distance from a reference in the original unit. Dividing by a reference gives a dimensionless ratio and can make the comparison to that reference explicit. These are not interchangeable labels: for instance, a centered value of 10 kg and a ratio of 1.10 encode different numerical quantities.

A common mistake is to treat an unexpectedly large or small number as proof of a unit error and rescale it without checking the source. The correction is to verify the unit and encoding from documentation, apply only a justified conversion factor, and keep the original measurement available. Another mistake is to convert the covariate but not the reference value, producing an invalid centered difference or ratio. Convert both to the same unit before applying the reference operation.