Missing and Misclassified Covariate Records

Why Covariate Record Quality Matters

A population PK covariate model relates a recorded characteristic to a population parameter, such as clearance. That relationship is only as meaningful as the records used to define the characteristic. A blank genotype field is not the same as a reference genotype, and a co-medication listed without reliable dates may not establish whether the patient was exposed at the time of a concentration measurement. Treating either record as if it were known can alter both the estimated covariate effect and the interpretation of the model.

Three situations need to be kept distinct. A covariate can be missing, meaning its value was not recorded or cannot be determined from available data. It can be misclassified, meaning a value was recorded but differs from the relevant true or intended value. Or it can be measured correctly but coded in a way that does not represent the scientific question, such as marking a patient as exposed throughout a study based on a medication history that does not include dates. These problems require different checks.

Put Missingness Into the Model Explicitly

For illustration, consider a binary covariate \(Z_i\), such as a recorded genotype carrier status, and a clearance model:

Here \(CL_{\mathrm{pop}}\) is the typical clearance for the reference covariate value, \(\beta\) is the log clearance ratio for \(Z_i=1\) versus \(Z_i=0\), and \(\eta_i\) is an individual random effect. The model presumes that the covariate is correctly recorded and available for each patient included in this expression. If a missing value is silently coded as \(Z_i=0\), the model asserts that the patient belongs to the reference group. That is not a neutral choice.

One explicit missing-category parameterization defines \(Z_i^*=Z_i^{\mathrm{rec}}\) for observed records and sets \(Z_i^*=0\) for missing records. It then includes both \(Z_i^*\) and \(M_i\):

Under this coding, \(\beta\) compares recorded covariate groups among patients with observed records, conditional on the model, while \(\gamma\) represents the modeled difference between patients with missing records and observed reference-group patients. It is an association with record status, not an estimate of the effect of the unknown covariate values. If missingness is related to the unobserved value or to PK outcomes in ways the model does not capture, adding \(M_i\) does not remove that bias.

The interpretation also depends on the chosen reference and coding. For a categorical variable with several recorded levels, a missing category can be represented with an indicator alongside the indicators for observed levels. For a time-varying co-medication, \(M_i\) may be insufficient: whether the medication status is missing can vary by occasion, so the missingness and exposure indicators may need to be defined at the occasion level.

Missingness Assumptions and Alternatives

A useful analysis asks why values are missing and what information is available about that process. Missing completely at random means the probability a value is missing does not depend on observed or unobserved data. Missing at random means that, conditional on observed information included in the analysis, missingness does not depend on the missing value itself. Missing not at random means missingness still depends on the unobserved value after conditioning on the observed information. These are assumptions about the data-generation process; a table of observed records cannot prove which one holds.

Complete-record analysis excludes patients with missing values for the covariate being analyzed. It is straightforward, but can lose precision and can be biased if inclusion in the complete-record subset is related to the outcome or covariate after conditioning on the model. A missing-category indicator retains those patients and can make record status visible in the fitted model, but does not recover their covariate values. Multiple imputation instead draws plausible missing values from a specified imputation model, repeats the PK analysis across completed data sets, and accounts for both within-imputation and between-imputation uncertainty. Its validity depends on a suitable imputation model and defensible assumptions about missingness; in PK applications, that model should preserve relevant covariate relationships, time structure, and uncertainty rather than treating one filled-in value as known.

For records that may be misclassified, alternatives include checking source records, restricting to high-confidence classifications, and fitting a model that explicitly represents classification uncertainty when sensitivity and specificity information is available. A simple coding sensitivity analysis is often a practical first step, but it is not a substitute for validation data when the error rate itself matters.

How Misclassification Changes an Estimated Effect

If a binary covariate is recorded incorrectly, some patients are assigned to the wrong group. Under a simple setting in which errors are independent of the PK outcome conditional on the true group, mixing the groups often pulls their observed difference toward each other. The extent depends on the true group frequencies and the sensitivity and specificity of classification. This familiar direction is not guaranteed in a real population PK analysis: differential errors, confounding, nonlinear model structure, and uneven subgroup representation can attenuate, exaggerate, or reverse an estimated association.

A short mixture calculation shows the mechanism. Suppose the true carrier prevalence is \(p\), sensitivity is \(Se=P(Z^{\mathrm{rec}}=1\mid Z=1)\), and specificity is \(Sp=P(Z^{\mathrm{rec}}=0\mid Z=0)\). The fraction of recorded carriers who are truly carriers is

The denominator includes both correctly recorded carriers and non-carriers incorrectly recorded as carriers. Thus, the recorded carrier group is a mixture unless classification is perfect. The recorded non-carrier group is also a mixture whenever sensitivity is below 1.

Alternative Coding Checks and Diagnostics

A coding check asks whether reasonable alternative representations of the same records materially change the estimated PK relationship. The alternatives should be motivated by actual ambiguities, not selected after looking for the most favorable result. Examples include different defensible mappings for an uncertain race or ethnicity label, a genotype carrier indicator versus a dosage representation when genotype calls support it, or a co-medication status based on exposure at the PK occasion versus a broader history window. For sex records, verify the source, the meaning of the recorded field, and how unknown or unavailable values are handled rather than silently assigning them to a reference category.

Useful diagnostics include tabulating missingness and category counts by study site, visit, subgroup, and relevant PK sampling period; checking whether missing records cluster in particular calendar periods or data sources; and comparing covariate distributions among included and excluded patients. For co-medications, inspect start and stop dates against PK sampling times. For genotypes, examine call quality and discrepancies across sources where available. These checks can identify patterns consistent with record-process differences, but cannot establish the true missingness mechanism or error rate by themselves.

Compare estimates and uncertainty under prespecified approaches, such as complete records, explicit missing-category coding, and a justified imputation analysis. Examine whether the population parameter estimates and individual predictions change in scientifically important ways, not merely whether a significance threshold is crossed. Large changes signal dependence on record assumptions and should temper interpretation. Similar estimates provide reassurance only about the alternatives examined, not proof that all records are correct.

Common Pitfall and Practical Limits

Wrong reasoning: “I added a missingness indicator, so patients with missing genotype data are accounted for and the genotype effect is unbiased.” This confuses a flag for missingness with knowledge of the missing genotype. The indicator can estimate a difference associated with being missing, but it does not identify which genotype those patients carry. Correcting bias requires assumptions and information sufficient for the chosen method, and those assumptions should be stated.

A second pitfall is treating a sensitivity analysis as a search for a preferred answer. Alternative codings are informative when they correspond to credible interpretations of records and are reported transparently. They cannot diagnose every data error, and sparse subgroups may make estimates unstable even when coding is correct. In a nonlinear mixed-effects model, conclusions also depend on the structural and random-effects assumptions; record-quality checks do not validate those other assumptions.

A practical sequence is to define the scientific covariate and time window, distinguish missing from uncertain or misclassified records, summarize record patterns, choose an analysis that matches plausible assumptions, and compare prespecified alternatives. Report the number of records affected and describe how the estimates change. This makes clear which conclusions are supported by the observed data and which rely on unverifiable coding or missingness assumptions.