Defining Analysis Populations and Data-Handling Decisions

Why Population and Data-Handling Rules Need to Be Explicit

A pharmacometric result depends not only on the model and estimation method, but also on which participants and observations contribute to the analysis. If those choices are made after examining model results, apparently technical decisions can change the evidence being analyzed. A clear analysis plan therefore defines populations and data-handling rules in advance, in terms that can be applied consistently.

These rules are related, but they are not the same as dataset construction. A population definition states which participants are eligible for a specified analysis. A data-handling rule states what to do with particular records or values, such as a missing covariate or a concentration below the assay’s quantification limit. Dataset construction is the later process of assembling and transforming records to implement those decisions. A participant can belong to a population even when some of their observations are missing or unusable for a particular model component.

There may be several populations because different objectives require different eligibility criteria. For example, a safety summary may include every participant who received at least one dose, while a pharmacokinetic analysis may require at least one valid post-dose concentration. These populations can overlap without being identical. The plan should give each population a clear name, purpose, and operational definition, rather than relying on labels such as “evaluable” without explaining what they mean.

Defining Populations With Reproducible Criteria

A useful population definition identifies the unit being included, the relevant study conditions, and any objective criterion for membership. It should also say how the criterion is evaluated. “Participants with usable PK data” leaves room for judgment; “participants who received at least one dose and have at least one valid post-dose concentration with a recorded sampling time” is more reproducible. If a valid result below the quantification limit counts as a concentration record, say so explicitly.

Population definitions should be tied to the analysis objective, not to whether a participant’s data produce a convenient fit or a result that agrees with expectations. Exclusions should be based on prespecified, observable facts. If a participant’s data are incomplete, decide whether the participant remains in the population and which available records can still contribute. Do not assume that one unusable observation makes every observation from that participant unusable.

Prespecifying Observation-Level Data Handling

The plan should identify important kinds of missingness and state how each will be handled for the intended analysis. A missing observation is not a measured value of zero. Nor does a missing value have one universal treatment: the appropriate rule depends on which variable is missing, why it is missing, the model, and the analysis objective. State whether a record is omitted from a particular likelihood contribution, whether a participant remains eligible, and what information is retained for review. Avoid silently replacing missing values with a typical value unless that method is justified and prespecified.

Below-quantification measurements require a separate rule. A result below the lower limit of quantification (LLOQ) is not necessarily zero or absent; it indicates that the measurement process did not provide a quantifiable value below that limit. Possible approaches include treating the result as censored in a likelihood, excluding it under a stated rule, or using another justified method supported by the model and assay. The plan should identify the selected approach and how it applies to observations before or after dosing, if those situations differ. It should not switch approaches just because one produces a more favorable fit.

The reason for prespecifying this rule is not just administrative. Omitting below-quantification observations can alter the distribution of the retained measurements. For example, if latent measurements \(X\) follow a normal distribution with mean \(\mu\) and standard deviation \(\sigma\), and only values at or above a threshold \(L\) are retained, the retained values are drawn from a truncated distribution. Let \(a=(L-\mu)/\sigma\), and let \(\phi\) and \(\Phi\) denote the standard normal density and distribution functions. Standardizing and integrating the truncated density gives:

The inequality holds because \(\sigma>0\), \(\phi(a)>0\), and \(1-\Phi(a)>0\). This result shows why simply dropping low observations can raise the average of the retained values under these assumptions. It does not establish that every analysis that omits such values will have a particular bias; the direction and size of an effect depend on the model and data. It does establish that the handling rule can affect the information used for estimation.

Protocol Deviations and Exclusion Rules

A protocol deviation is an event in study conduct that differs from the protocol. Its existence should be recorded, but the label alone does not determine whether the participant or an observation is excluded. The plan should distinguish deviations that affect population membership from deviations that affect only particular records or their interpretation. Examples that may matter to a pharmacometric analysis include an incorrect dose, a dose-time discrepancy, a sample collected outside its permitted window, or a prohibited concomitant treatment. The plan needs criteria relevant to the actual objectives; a generic list is not a substitute for an operational rule.

For each potentially consequential deviation, specify the threshold or decision rule, who or what determines that it occurred, and the analysis consequence. For example, a deviation might exclude a particular concentration from a primary analysis while leaving the participant in the PK population. Alternatively, a prespecified population might exclude participants meeting a defined major-deviation criterion. If a decision requires clinical or data review, specify the review process and document the decision before examining its effect on the model result. A deviation should not be declared “major” solely because removing the affected data improves fit.

Checks, Alternatives, and Limitations

Before fitting the model, check that the written criteria can be applied to the available information. Count participants at each population step and reconcile transitions: enrolled, dosed, meeting the analysis-population criteria, and excluded under each stated reason. Check observation counts separately, by key variable and relevant sampling period. Investigate duplicate records, impossible or inconsistent times, unexpected missingness, and values marked below quantification but lacking an LLOQ. These checks identify implementation problems; they do not authorize ad hoc changes to the planned rules.

There is no single best data-handling approach for every setting. Censored-likelihood methods use the information that a value lies below a threshold, but require a suitable likelihood and reliable quantification limits. Exclusion is simpler but discards information and can select the retained observations. Substitution with a fixed value is easy to implement but can distort distributions and create artificial repeated values. For missing data, complete-record analysis, likelihood-based use of available observations, and imputation make different assumptions; the chosen method must be compatible with the analysis objective and model. State the approach and its rationale rather than presenting any option as universally correct.

Population criteria also limit the scope of inference. A PK population requiring post-dose samples does not represent every enrolled participant, and a population restricted by dose receipt does not describe those who were never dosed. When exclusions or missing records are substantial, the resulting analysis may describe a selected subset. Documenting that limitation is necessary even when the rules were followed exactly.

Common Pitfall

A common mistake is to remove every participant with any missing value, below-quantification result, or protocol deviation and call the remainder “evaluable.” This combines three distinct decisions without stating a rationale. The correction is to apply participant-level population criteria separately from observation-level rules, and to define in advance whether a specific deviation affects a participant, a record, or neither. Another mistake is to treat a below-quantification result as zero; the correction is to apply a justified, prespecified method that reflects what the assay result actually says.