Handling Outlying Log Exposures Without Automatic Deletion

Why an Extreme Log Exposure Needs Investigation

A log-transformed exposure value can look far from the rest of the data and still be correct. It may reflect real subject-to-subject variation, a genuine treatment response, or a data-processing problem. These possibilities have different implications: an erroneous value may need correction, whereas a valid but extreme value generally remains part of the analysis. The purpose of examining an outlying log exposure is therefore to establish what happened to the observation and how much it matters—not to find a statistical rule that permits its removal.

The log scale is used for exposure measures such as AUC and Cmax because multiplicative differences on the original scale become additive differences after logging. This tutorial concerns what to do when an observation on that analysis scale appears unusual. It does not treat extremeness by itself as evidence of invalidity, nor does it replace the earlier material on residual diagnostics and influential subjects. Here the emphasis is the observation-level investigation and a transparent assessment of sensitivity.

Model and Assumptions Before Diagnosing

In a crossover analysis, let \(Y_{ijk}\) denote the natural logarithm of an exposure endpoint for subject \(j\) in sequence \(i\) and period \(k\). A typical model includes sequence, subject within sequence, period, and treatment effects. In a mixed-model formulation, subject effects account for the repeated observations on a subject; the fixed effects represent the design and treatment contrasts of interest. The exact model must match the design and the estimand. A residual is the observed log exposure minus the value predicted by that fitted model.

Residual diagnostics are informative only relative to a stated model. The usual model-based interpretation relies on an appropriate mean structure, an appropriate covariance structure for repeated observations, and residual variation that is adequately represented by the model. Normality is often assessed for model residuals when small-sample inference relies on a normal-error model. An observation can look extreme because the model omitted a relevant design feature, not because the measurement is erroneous. Conversely, a plausible-looking residual does not prove that every data entry is correct.

The investigation should start with the observation’s provenance. Confirm the subject, treatment, period, endpoint, units, and original source record. Check the exposure calculation, the log base, rounding, and any conversions between units. For instance, values expressed in two units differing by a factor of ten differ by \(\ln(10)\) on the natural-log scale. A conversion mismatch can therefore create a striking log-scale discrepancy while leaving the underlying measurement intact.

What Residual Extremeness Can and Cannot Show

A raw log exposure is not a model residual. A value may be high in the sample but fit the model well after accounting for treatment, period, sequence, and subject structure. Conversely, a moderately high value may have a large residual if the model predicts a much lower value for that design cell. Residual plots can reveal patterns such as an observation far from its fitted value, a variance that changes with fitted value, or a missing design pattern. A normal quantile plot can help identify tail departures, but it cannot identify the cause of an unusual observation.

Numerical cutoffs are screening aids rather than validity tests. A standardized or studentized residual above a chosen threshold can prompt a closer data check, but the chance of seeing at least one large residual rises as the number of observations grows. Repeatedly examining many outcomes, treatments, periods, and diagnostic displays also increases the opportunity to find an apparently unusual point by chance. Do not interpret a residual threshold as a calibrated probability that the observation is erroneous.

A residual’s size also does not by itself measure its impact on the treatment estimate. Impact depends on the model, the observation’s position in the design, and the way the estimate is calculated. The earlier material on influential subjects treats influence diagnostics in depth; in this setting, the key follow-up is to quantify how the treatment estimate and uncertainty change under clearly specified analyses.

Sensitivity Analysis and Its Interpretation

A useful sensitivity analysis shows how conclusions depend on a defensible analysis choice. The primary analysis should follow the prespecified model and include valid observations. A comparison that omits an observation can be informative about dependence on that observation, but unless independent evidence establishes an error or a prespecified rule applies, the omission is a sensitivity analysis—not a replacement primary analysis.

For a simple mean of \(n\) subject-level contrasts, the estimate after leaving out contrast \(d_i\) is related to the full-data mean by

The second equation follows by subtracting \(\bar{d}\) from the first and simplifying. It shows why a value far from the mean can shift an estimate, especially when \(n\) is small. In a crossover model with additional terms, the same simple expression need not hold; refit the specified model to calculate the effect of an omission rather than applying this formula to a more complex estimator.

Correct Errors, Preserve Valid Observations

A documented data-processing error calls for correction according to the source record and the analysis plan. That is different from deleting an observation because it is extreme. Preserve an audit trail: record what was checked, the evidence for a correction, the original and corrected values, and how the analysis changed. If the source cannot resolve a discrepancy, state that uncertainty and show how the unresolved value affects results where feasible.

Alternatives, Failure Modes, and Limits

The main alternatives are not simply “keep” versus “delete.” First, correct a confirmed recording or processing error using documented evidence. Second, retain a valid observation in the primary model and report sensitivity analyses, such as a refitted model omitting the flagged value for diagnostic comparison. Third, if the model assumptions are inadequate, consider a scientifically justified alternative model or a robust analysis, while making clear that it changes the analysis target or assumptions. A robust method is not a license to discard inconvenient observations, and a different model should not be selected solely because it yields a preferred conclusion.

Common failure modes include checking only the raw exposure rather than the model residual; using a cutoff as an automatic deletion rule; changing a value without documenting its source; and reporting only the analysis that gives the desired result. Other warning signs are a diagnostic pattern shared by several observations, a residual that changes substantially when the design model is corrected, or conclusions that depend strongly on one valid observation. Such findings call for careful reporting and review of the model and data, not concealment of the sensitivity.

A leave-one-out analysis is not a new estimate of what the population treatment effect “really” is when the omitted observation is valid. It answers a narrower question: how much does this fitted analysis depend on that observation? Sensitivity comparisons can be unstable in small samples, and neither a residual plot nor a robust fit proves that a value is correct. The source record, protocol, analysis plan, design, and model assumptions remain central to the decision.