Documenting Below-Quantification Data Handling

Why Below-Quantification Records Need Explicit Reporting

A concentration reported as below the assay’s lower limit of quantification is not the same as a measured concentration of zero. It indicates that the assay did not quantify the value at or above its stated threshold. How an analysis handles that record can affect the data used, the model likelihood, and potentially the resulting parameter estimates. A report should therefore let a reader determine both what happened to these records and whether the main findings depend on that decision.

This is a documentation problem as well as a modeling problem. A statement such as “below-quantification values were handled appropriately” is not reproducible: it does not say which records were affected, which rule was applied, or whether the results changed under a reasonable alternative. The report should connect the assay threshold, the record-level rule, the analysis population, and the sensitivity results.

The LLOQ may differ by analyte, assay, or sample-specific laboratory conditions. State the applicable threshold and units, including any variation, rather than assuming that readers can infer it from a table. Distinguish below-quantification results from missing samples, rejected assay runs, and values outside the assay’s upper range: these have different meanings and may receive different processing.

What a Reproducible Description Contains

The methods and results should jointly answer four questions: what qualified as below quantification, what rule was applied, which records and participants were affected, and how much the results changed under a relevant alternative. If a rule differed by analysis, analyte, phase, or sample type, describe those differences explicitly.

• Define the threshold and classification. Give the LLOQ and units, identify the assay or analyte it applies to, and explain how a reported “<LLOQ” result was identified. If thresholds varied, say how each record was assigned its threshold.
• State the analysis rule. Say whether the record was excluded, replaced (and with what value or rule), or included as censored information in the likelihood. Specify whether the rule applied to every below-quantification record or only to a defined subset, such as predose samples.
• Describe the analysis records affected. Report the number of below-quantification records and the total number of relevant concentration records. Also report the number of participants with at least one affected record. If records were excluded, state how many records and participants were removed from the analysis; if they were retained as censored, say so.
• Show meaningful breakdowns. Counts by analyte, study phase, or predose versus postdose status can reveal where the records occur. Use denominators for each subgroup, not just percentages, and state whether categories are mutually exclusive.
• Report a sensitivity analysis. Identify the alternative rule, confirm what else was held constant, and compare relevant estimates or conclusions with the primary analysis. Include enough numerical detail to assess the size and direction of differences.

A censored-likelihood approach uses the fact that a result was below a threshold without pretending that its exact concentration was observed. In a simple representation, if the latent assay-scale result is \(Y\), a censored record contributes the probability of falling below the applicable LLOQ, conditional on the model, rather than a density evaluated at an invented numeric value:

The precise likelihood contribution depends on the specified observation and residual-error model. A report describing this approach should name the method used and its implementation, not imply that all censored-likelihood methods are identical. By contrast, replacing a below-quantification record with a constant such as LLOQ/2 treats that assigned value as if it were numerical input under the analysis; it is a different rule and should be reported as such.

Counting Records and Participants

A record count and a participant count answer different questions. One participant may contribute many below-quantification samples, so the number of affected participants is usually smaller than the number of affected records. State both where possible. Make clear whether “affected participants” means anyone with at least one such result, and whether the reported denominator includes all enrolled participants or only participants in the analysis population.

For an exclusion rule, describe its consequences at both levels. If below-quantification records are removed but the participants remain in the analysis through other observations, say so. If a participant is removed because no usable concentration records remain, report that separately. These distinctions prevent a reader from mistaking record exclusion for participant exclusion.

A useful percentage has an explicit numerator and denominator. For example, “96 of 1,200 concentration records (8.00%)” is interpretable; “8% were below quantification” is incomplete if the relevant record set is unclear. For subgroup summaries, define the subgroup and its denominator. If categories overlap, explain that their counts may not sum to the total.

Designing and Interpreting Sensitivity Results

A sensitivity analysis asks whether the main results change under a specified alternative handling choice. It does not, by itself, establish which rule is correct. To make the comparison interpretable, state the primary rule and the alternative, identify whether the model and other data-processing choices were held constant, and report the estimates or conclusions compared. If other choices also changed, readers cannot attribute the difference specifically to below-quantification handling.

For an estimate \(\theta\), a simple relative difference from the primary analysis can be reported as

This calculation describes the estimate’s change relative to the primary value; it is not a test of statistical significance and does not provide an uncertainty interval. State the units for the estimates and report enough digits to make the calculation transparent. For parameters where a percentage difference is misleading or the primary estimate is near zero, give the absolute difference and units instead.

Alternatives are not interchangeable. Excluding below-quantification records, substituting a fixed value, and retaining them as censored observations encode different information and assumptions. A sensitivity analysis is most informative when the alternative is plausible for the application and its purpose is stated. For example, a comparison with a substitution rule can show how dependent estimates are on assigning numeric values; it does not make that substitution equivalent to censoring.

Diagnostics, Limitations, and Common Pitfalls

Before interpreting a sensitivity result, check that the record counts reconcile and that each rule was implemented as described. Compare the number of affected records and participants between analyses, confirm that any removed participants or records are identified, and verify that the compared estimates come from the intended analysis versions. If the alternative analysis does not converge or produces unusable estimates, report that outcome rather than silently omitting it. A failed sensitivity analysis does not demonstrate robustness.

Also examine where the below-quantification records occur. A high share in a particular sampling phase, analyte, or group may matter even when the overall percentage is low. A report should not imply that a small overall fraction guarantees negligible impact. Conversely, a change in an estimate under an alternative rule is evidence of sensitivity to that choice, not proof that the primary analysis is biased. The interpretation depends on the model, assay, data pattern, and intended use.

• “Below quantification means zero.” This confuses failure to quantify with a measured zero. Report the censoring threshold and the analysis rule instead.
• “The values were included.” This leaves unclear whether they were retained as censored observations, assigned numeric replacements, or handled some other way. Name the exact treatment.
• Reporting only a percentage. Without a numerator, denominator, and record definition, readers cannot assess or reproduce the count. Give counts and denominators, preferably with relevant strata.
• Reporting only affected records. This hides how broadly they are distributed across participants. Report the number of participants with at least one affected record as well.
• Calling a sensitivity result proof of robustness. A similar estimate under one alternative shows limited change under that particular comparison; it does not establish insensitivity to every plausible handling choice.
• Changing several analysis decisions at once. If the model, data set, and below-quantification rule all change, the source of any difference is unclear. Isolate the rule when feasible, or disclose all simultaneous changes.