Why the Rule Matters
A baseline covariate is a subject characteristic defined at or before the relevant baseline assessment and treated as fixed for that analysis. Examples include baseline body weight, age, or a baseline category such as genotype group. If a baseline value is missing, the rule used to fill it changes the analysis-ready covariate data. It can also change which subjects appear to share a category or value in a population pharmacokinetic model.
There is no universally correct default replacement. A population-level rule assigns a value calculated across subjects; a subject-level rule uses information specific to that subject; and a reference-category rule assigns a missing categorical value to a specified reference group. Each rule answers a different practical question, so the rule should be chosen for the covariate, justified, and specified before model fitting. The classification of missingness helps frame that choice, but does not by itself establish that any one replacement is unbiased.
Specify the Rule Before Constructing the Dataset
For a baseline covariate \(X_i\) measured on subject \(i\), let \(M_i=1\) indicate that the analysis value is missing and \(M_i=0\) indicate that it is observed. A replacement rule defines the value to place in the analysis covariate column when \(M_i=1\). Preserve the original value and a missingness or imputation flag separately; the analysis value should not erase the distinction between recorded and imputed information.
A population-level rule might use the observed median for a continuous covariate or the most frequent category for a categorical covariate. The statistic should be calculated from the prespecified eligible population and according to a consistent unit and data-cleaning policy. For example, do not calculate a body-weight median until units have been harmonized and implausible records have been investigated. The choice of mean, median, mode, or a fixed value also needs a reason: the mean is sensitive to extreme values, while a median or mode has a different interpretation and does not recover an individual’s unobserved value.
A subject-level rule is appropriate only when the subject’s own information supports the assigned value. Examples include using a verified duplicate measurement from the same baseline assessment, or deriving age at baseline from a documented date of birth and assessment date under the study’s age convention. A value from another person, a population summary, or an unsupported clinical guess is not subject-level evidence. If sources conflict, resolve the discrepancy using the data-governance process rather than selecting the convenient value.
A reference-category rule is specific to categorical covariates. It places missing values in the reference category; it does not establish that a subject belongs to that category. In a model using indicator variables, the reference group is represented by the baseline level of the factor coding. Assigning missing observations to it therefore makes them indistinguishable in that covariate from subjects whose reference-category membership was actually observed, unless a separate missingness flag or other modeling strategy preserves that distinction.
Worked Example: Population Median for a Continuous Covariate
Hypothetical data. Baseline body weights, in kg, are 72, 84, missing, 96, and 108. The prespecified rule is to replace a missing weight with the median of the observed eligible subjects. Assume all weights are already in the same units and have passed the study’s data checks.
Sort the observed values: 72, 84, 96, 108. With four observations, the median is the average of the middle two:
The analysis-ready values are therefore 72, 84, 90, 96, and 108 kg. The third value is an imputation, not a measured weight. A suitable data structure keeps its raw value missing, records the analysis value as 90 kg, and marks that subject’s weight as imputed.
This rule is simple and reproducible, but it assumes that assigning the observed-population median is an acceptable approximation for the analysis purpose. It does not use the third subject’s individual information and does not represent the uncertainty about that person’s actual weight.
How the Rules Change Model Inputs
Covariate values may enter a population PK model through a continuous relationship or a categorical effect. For example, a continuous covariate could be centered around a reference value, while a categorical covariate could be represented by indicator variables. Imputation changes those model inputs directly. A subject assigned a population median may receive a centered value of zero if that median is the centering value; a subject assigned the reference category receives zero on every non-reference indicator. These are coding consequences, not evidence that the subject truly has the reference value.
A simple mean-replacement calculation illustrates one limitation. Suppose \(n_o\) subjects have observed values with mean \(\bar{x}_o\), and \(n_m\) subjects have missing values. Replacing every missing value by \(\bar{x}_o\) leaves the combined mean equal to \(\bar{x}_o\), because the imputed values add no deviation from that mean. The sum of squared deviations about the combined mean comes only from the observed subjects:
Thus, the imputed subjects contribute zero squared deviation. Using the population-style variance with divisor \(n_o+n_m\), the variance after replacement is the observed sum of squares divided by \(n_o+n_m\), rather than by \(n_o\). When \(n_m>0\), this is smaller than the observed-data variance calculated with divisor \(n_o\), provided that sum of squares is positive. The precise sample-variance comparison depends on the divisor convention, but the added values still create a point mass at the mean and do not add genuine information about between-subject variability.
Worked Example: Subject-Level Reconstruction From Baseline Records
Hypothetical data. A subject’s baseline age is missing in the covariate file. The source record contains a date of birth of 12 April 1990 and a baseline assessment date of 1 March 2024. The prespecified convention is age in completed years on the assessment date. Assume both dates are verified and the convention is applicable.
The subject had not yet reached the 2024 birthday on the assessment date, so the completed age is:
The subject-level analysis value is 33 years. Unlike a population mean or median, this value is obtained from records specific to the subject and follows a defined calculation. The date inputs, calculation convention, resulting value, and reconstruction flag should all be retained for auditability.
This is defensible only if the dates are reliable and the age convention is prespecified. If the date of birth is uncertain, the calculation does not make it certain; the missing value should not be filled as though the derivation were verified.
Worked Example: Reference-Category Assignment
Hypothetical data. A baseline genotype covariate has categories A and B, with A prespecified as the reference category. Three subjects have observed genotypes A, B, and missing. Use indicator coding in which \(I(B)=1\) for category B and \(I(B)=0\) for category A.
Before a missing-value rule is applied, the indicator is 0 for the observed A subject and 1 for the observed B subject; it is undefined for the missing subject. Under the prespecified reference-category rule, assign the missing subject to A. The analysis-ready indicators are then 0, 1, and 0.
The model treats the missing subject’s genotype contribution as identical to category A for this covariate. It does not learn that the genotype was A. If that distinction matters, retain a missingness flag or choose a different strategy; a reference assignment alone cannot preserve it.
Choose by Meaning, Not Convenience
Before applying a rule, state what its assigned value means and what evidence supports it. A useful decision sequence is:
Verify that the covariate is defined at baseline for this analysis, and that the missing entry is not a coding error, an unharmonized unit, or a recorded value stored in another source.
If a verified duplicate measurement or prespecified deterministic derivation exists, assess whether it directly supports the missing value. Do not substitute an unsupported proxy.
For a population summary, identify the eligible observations and summary statistic. For categorical reference assignment, define the reference category and acknowledge that missing subjects are being grouped with it.
Keep raw and analysis values distinguishable, flag imputed records, and compare covariate distributions and missingness counts before and after construction.
These approaches differ from alternatives such as leaving the covariate missing, excluding subjects from an analysis that requires it, or using a model-based missing-data method. Those alternatives also require assumptions and may change the usable population or uncertainty represented in the analysis. A simple replacement is not automatically preferable because it produces a complete covariate column. If missingness is substantial, related to subject characteristics, or influential for the covariate effect, consider whether a single deterministic value is adequate for the scientific and modeling objective.
Diagnostics, Failure Modes, and Documentation
Review the completed dataset rather than assuming that a rule worked because the program ran. Check the number and proportion of imputed values by covariate and, where relevant, by study site or other prespecified groups. Compare observed and analysis-ready distributions, including the number of unique values and the frequency of ties at a population summary. For categorical variables, tabulate observed categories separately from imputed reference assignments. Confirm that the reference coding in the model matches the documented rule.
- Do not confuse replacement with recovery. A population median is a common assigned value, not a reconstructed measurement. Labeling it as observed overstates the data.
- Do not treat missing as the reference without stating the consequence. The rule combines unknown membership with known reference membership in the covariate coding.
- Do not derive a value from unreliable inputs. A subject-level calculation is only as valid as its source records and prespecified convention.
- Do not use one statistic without checking its suitability. An extreme value can distort a mean, and a mode can be unstable when category counts are similar.
- Do not let imputation hide missingness patterns. Retain an indicator or separate audit trail so analysts can identify which entries were filled.
A simple rule can be useful for a small amount of missing baseline information when its assumptions are credible and its effect is transparent. It is limited because a single substituted value does not express uncertainty about what was missing, can alter the distribution, and may affect estimated covariate relationships. The choice should be documented with the variable definition, rule, eligible source records, calculation, units or category coding, and flags used to identify imputed entries.
Check Your Understanding
Answer each question, then compare with the solutions below.
- Observed baseline values are 10, 14, and 18 mg/L, with one value missing. Under median replacement, what value is assigned?
- Why is an age calculated from a subject’s documented birth date and baseline date subject-level reconstruction rather than population-level imputation?
- A missing categorical value is assigned to the model reference category. What distinction is lost in the covariate coding?
- Give one diagnostic that can reveal the effect of population mean replacement on a continuous covariate.
- Why should an imputed value not overwrite the raw missing value?
Solutions. (1) The median of 10, 14, and 18 mg/L is 14 mg/L, so the assigned value is 14 mg/L. (2) The value is calculated from verified records specific to that subject under a stated convention, rather than from a summary across subjects. (3) The coding no longer distinguishes subjects whose category is known to be the reference from subjects whose category was missing but assigned to it. (4) Compare observed and analysis-ready distributions, including variance, unique-value counts, or the frequency of ties at the imputed mean; mean replacement creates a pile-up at that value. (5) Retaining the raw missing value preserves provenance and allows the analysis value to be identified as imputed rather than observed.