Why Positivity Matters
Log-scale analysis is common for pharmacokinetic endpoints such as AUC and Cmax because ratios of exposures are often more meaningful than differences. But the natural logarithm, \(\ln(x)\), is defined for real \(x\) only when \(x>0\). Thus every endpoint value to be transformed must be strictly positive. This is a mathematical requirement, not a preference that can be relaxed by a statistical model.
A zero or negative value should therefore trigger an investigation before analysis. It is not, by itself, permission to omit a subject, replace the value, or change the endpoint. The value might reflect a data-entry or processing problem, a code for missing data, an assay reporting convention, or a genuine observation that requires a different analysis. Those possibilities have different scientific meanings.
The positivity condition applies to the values being logged, not to every quantity in a statistical model. For example, a treatment contrast on the log scale may be negative; that contrast is a difference between logarithms, not a log of a nonpositive endpoint. Separately, assumptions about the model for log-transformed outcomes—such as the distribution of residuals—must be assessed as appropriate. Those assumptions do not make \(\ln(0)\) or \(\ln(x)\) for \(x<0\) valid.
First Determine What the Value Represents
For conventional exposure endpoints, AUC and Cmax calculated from nonnegative concentrations are themselves nonnegative. A negative value for either endpoint is therefore a warning that the recorded quantity may not be the intended endpoint, or that an input, unit conversion, baseline adjustment, or calculation has altered its meaning. It should be traced to the source records and derivation rather than passed directly to a logarithm.
A zero is less immediately diagnostic. It may be a genuine recorded result under a particular measurement or reporting convention, or it may stand for “not measured,” “below quantification,” or another status. A reported zero does not necessarily establish that the underlying exposure was exactly zero. For instance, concentrations below an assay’s lower limit of quantification (LLOQ) may be represented in a data set using a convention that yields zero. The resulting AUC or Cmax then reflects that convention as well as the measurements; it does not automatically reveal the unobserved concentration.
A careful review distinguishes at least four situations:
- Coding or data-management issue: zero or a negative number was used as a missing-value code, a placeholder, or a default. Recover the correct value or missingness status from the source.
- Calculation or unit issue: a derivation, import, sign, or unit conversion is incorrect. Correct and document the derivation, then verify the endpoint from its inputs.
- Assay or reporting convention: the recorded value represents a below-quantification result or another measurement status. Review the assay and data-handling rules; do not infer an exact value merely from the code.
- Genuine nonpositive result: the endpoint as defined is genuinely zero, or the recorded quantity is a different endpoint that can be negative. A standard log analysis of that value is not available; the scientific question and analysis strategy need review.
The distinction matters for the estimand: the quantity the analysis is intended to compare. If the observed value is missing, the problem is missing data. If it is censored by an assay limit, the observation contains limited information about the underlying value. If zero is a genuine endpoint value, it is part of the endpoint’s distribution. Treating all three cases as interchangeable changes the scientific interpretation.
Worked Example: Auditing a Zero Cmax
In a hypothetical crossover data set, the recorded Cmax values for four subjects after one formulation are 840, 0, 760, and 920 ng/mL. The analysis plan calls for natural logarithms of positive Cmax values. The second value cannot be transformed as recorded: \(\ln(0)\) is undefined.
A data review finds that the zero was inserted when a concentration-profile file could not be imported. The original records contain a valid profile, and reprocessing under the documented endpoint calculation gives Cmax = 680 ng/mL for that subject. The corrected value is positive, so its transformed value is
The zero was a processing placeholder, not a measured zero exposure. Replacing it with the verified result is justified because the source data establish that value and the derivation is corrected. It would not be justified to replace it with an arbitrary small positive number simply to make the logarithm computable. The remaining positive Cmax values can be logged, subject to the analysis plan and the model’s other assumptions.
Why Adding a Constant Changes the Question
A tempting workaround is to add a positive constant \(c\) to every value and analyze \(\ln(x+c)\). This creates a defined logarithm when \(x=0\), provided \(c>0\), but it does not preserve the original log-scale comparison. The transformed treatment contrast is now based on \((x+c)\), not \(x\).
For two positive values \(T\) and \(R\), the original log difference is \(\ln(T)-\ln(R)=\ln(T/R)\), which compares their ratio. After shifting, the difference is \(\ln(T+c)-\ln(R+c)=\ln((T+c)/(R+c))\). In general, this is not the original log ratio. The shift also depends on the units used to record the endpoint: a numerical constant of 1 has different sizes if the endpoint is recorded in different units.
The change can be seen by considering what happens as \(c\) grows. For fixed positive \(T\) and \(R\), \((T+c)/(R+c)\) approaches 1 as \(c\) becomes large. Thus the shifted log difference approaches zero, even though the original ratio \(T/R\) has not changed. Adding a constant can therefore alter the magnitude of the estimated contrast and, consequently, its interpretation.
Worked Example: A Shifted Ratio Is a Different Ratio
Suppose hypothetical endpoint values are \(T=12\) and \(R=10\) mg·h/L. Without a shift, the ratio and log difference are
Now add \(c=5\) mg·h/L to both values. The ratio of shifted values is \(17/15\), and the log difference becomes
With the larger shift \(c=100\) mg·h/L, the corresponding calculations are
The original values did not change, but the shifted contrast moved closer to zero as the added constant increased. The shifted analysis is not estimating the same ratio as the unshifted analysis. Selecting \(c\) after seeing the data does not solve the underlying endpoint problem; it defines a different analysis whose scientific justification would need to be established in advance.
Zeros, Assay Limits, and Negative Values
When a zero is connected to an assay limit, the appropriate response depends on how the endpoint was defined and how the study handled measurements below quantification. A result below the LLOQ is not automatically an exact zero, and replacing it with a fixed fraction of the LLOQ is not a universal remedy. Such substitution imposes a value that was not measured and can affect derived endpoints and treatment contrasts. Review the prespecified data-handling rules, the concentration-time records, and the sensitivity of the intended analysis to the convention used.
If all recorded concentrations for a profile are below the LLOQ, a reported Cmax or AUC of zero may be the result of the reporting and calculation rules. The log transformation remains undefined for that reported value. The correct next step is not to take the logarithm of a substitute chosen for convenience, but to establish what the record means and whether the planned log-scale estimand can be supported by the data.
Negative values require particular care. A conventional AUC or Cmax derived from nonnegative concentrations should not be negative. Check whether the file contains a signed change from baseline, a corrected signal, a difference between quantities, or an error in units or calculation. A signed change may be a legitimate endpoint in its own right, but it is not the same quantity as a positive exposure endpoint, and a log transformation of that signed change is not defined when it is nonpositive.
Worked Example: A Zero Created by Below-LLOQ Records
In a hypothetical profile, the assay LLOQ is 0.50 ng/mL and every scheduled concentration is reported as below the LLOQ. The data-processing convention records each of these concentrations as 0 ng/mL. Under a calculation that uses those recorded values, the resulting Cmax and AUC are both 0. Neither can be logged: \(\ln(0)\) is undefined.
Suppose an analyst proposes replacing each zero endpoint with 0.25 ng/mL, half the LLOQ. For Cmax this would give \(\ln(0.25)\approx -1.386\), with the concentration unit understood as ng/mL. But the data do not show that the true Cmax equals 0.25 ng/mL. The proposed number is a convention, not an observed endpoint, and it does not establish that the transformed value represents the intended exposure comparison.
The analyst should instead verify how below-LLOQ results were to be handled, whether the reported zero is an assay status or a true endpoint value, and whether the planned analysis remains appropriate. Any alternative handling must be scientifically justified and specified transparently; it must not be selected merely because it permits a logarithm.
Analysis Choices and Diagnostics
Once the source and meaning of a nonpositive value are established, possible responses differ. Correct a verified data or derivation error and retain an audit trail. If the value is a missingness code, represent missingness according to the data rules rather than pretending it is a measured zero. If it is tied to a quantification limit, use a handling or modeling strategy appropriate to that measurement process and the prespecified objective. If it is a genuine zero or a signed endpoint, reconsider whether a log-scale analysis answers the scientific question.
Alternatives are not interchangeable. An analysis of raw differences targets additive contrasts rather than ratios. A model for censored observations can represent information that values lie below a limit, but it requires a suitable censoring model and adequate data. A two-part approach can separate the probability of a zero from the distribution of positive values when that distinction is scientifically meaningful. Each option has assumptions and may target a different estimand; none should be chosen solely to bypass the undefined logarithm.
The following checks help identify when log-scale analysis is going wrong:
- Tabulate nonpositive values by subject, treatment, period, and endpoint. A cluster in one treatment or period may indicate a process difference worth investigating.
- Compare analysis records with source concentration-time data, assay status flags, units, and endpoint derivation code. Verify the record’s provenance rather than relying only on a summary value.
- Check whether zero or negative numbers are documented codes, and whether those codes have been converted consistently. A numeric placeholder can otherwise be mistaken for a measured outcome.
- Confirm that the endpoint definition itself is positive-valued. A baseline-adjusted or signed quantity may not be a conventional AUC or Cmax.
- After correcting an error or applying a justified rule, verify the number of usable observations and the resulting analysis population. Do not silently remove observations that cannot be logged.
If the model is fit only after nonpositive observations are dropped, the analysis may describe a selected subset rather than the originally intended population. If these observations occur more often under one treatment, dropping them can distort the comparison especially strongly. The missingness or zero-generating mechanism is part of the scientific problem, not merely a computational inconvenience.
Common Mistakes and Limitations
A second mistake is treating every zero as a true absence of exposure. A zero can instead encode missingness or a below-quantification result. Conversely, assuming every zero is a data error can erase a meaningful feature of the endpoint. The value must be interpreted using its source and the measurement process.
A third mistake is deleting a subject with a nonpositive endpoint and reporting the remaining log-scale analysis as if nothing changed. Deletion changes the analyzed records and may change the target population. Any exclusion must follow a scientifically and procedurally justified rule, with its consequences assessed and reported.
Finally, positivity is necessary but not sufficient. Having only positive values makes logarithms mathematically available; it does not by itself establish that the selected model is appropriate or that the analysis answers the intended question. A log transformation cannot repair an incorrectly defined endpoint, an undocumented assay convention, or a flawed data derivation.
A Practical Decision Sequence
A disciplined sequence helps preserve both the data and the estimand:
Identify every endpoint value \(x\leq 0\) before applying the logarithm; do not let software silently discard or alter it.
Check source records, assay flags, units, endpoint definitions, and calculation steps to determine whether it is a code, an error, a censored measurement, or a genuine value.
Correct verified errors, handle missingness or assay limits according to a defined strategy, or reconsider the estimand and model if the nonpositive value is genuine.
Record the rule applied, verify the resulting data and analysis population, and assess whether the method still supports the intended scientific interpretation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- Why can a negative treatment contrast be valid on the log scale even though a negative endpoint cannot be logged?
- A recorded AUC is zero because missing profile data were filled with zeros by a software import. What should be checked before analysis?
- For \(T=20\) and \(R=10\), compare the original ratio with the ratio after adding \(c=10\). What does the comparison show?
- Why does replacing every below-LLOQ endpoint with half the LLOQ not establish that the true endpoint equals that value?
Solutions. (1) A contrast such as \(\ln(T)-\ln(R)\) subtracts logarithms of positive endpoints; the resulting difference can be negative when \(T<R\). It is not the logarithm of a negative endpoint. (2) Trace the import problem to source records, recover or correctly classify the missing data, and verify the endpoint derivation and documented handling rules. The placeholder zero is not a measured AUC. (3) The original ratio is \(20/10=2\). After the shift it is \((20+10)/(10+10)=30/20=1.5\). The shifted ratio is different, so the shift changes the comparison. (4) A below-LLOQ result indicates a measurement limitation, not an exact value of half the LLOQ. Substitution imposes an unmeasured value and needs scientific justification; it is not a mathematical fix that preserves the original estimand.