Case Study in Reporting Terminal Data Below Quantification

Why Below-Quantification Terminal Data Need Careful Reporting

Terminal concentrations often approach an assay’s lower limit of quantification (LLOQ). Consequently, a substantial share of late samples may be reported as below quantification (BLQ). Those records do not provide an exact concentration, but they do provide information: the concentration is below a known threshold. Treating that information as if it were an exact zero, a fixed fraction of the LLOQ, or an ordinary missing value can alter the apparent terminal decline and the parameters derived from it.

This hypothetical case focuses on how to report that problem and its consequences. It does not imply that one handling choice is automatically appropriate for every analysis. The report should identify the primary handling approach, state its assumptions, show how conclusions change under reasonable alternatives, and explain which interpretations are supported by the data.

Definition: A concentration is below quantification when the assay cannot quantify it reliably below a specified LLOQ. A BLQ record is therefore an inequality observation, \(C<LLOQ\), rather than an exact concentration or proof that \(C=0\). The terminal-phase rate constant, denoted here by \(\lambda_z\), describes the rate of decline in the terminal part of a concentration-time profile. Under a log-linear terminal decline, the corresponding terminal half-life is \(t_{1/2,z}=\log(2)/\lambda_z\).

Case Data and Primary Handling Approach

Suppose a hypothetical intravenous study has 40 participants, each with eight scheduled terminal-phase samples at 24, 36, 48, 60, 72, 84, 96, and 120 hours after dosing. The assay LLOQ is 0.50 mg/L. There are 320 scheduled records in this terminal window, of which 112 are BLQ, or 35.0%. The counts by scheduled time are 2, 6, 12, 18, 20, 20, 18, and 16 BLQ records, respectively. This increasing prevalence at later times is consistent with concentrations approaching the assay limit, although it does not by itself establish that the model describes the terminal decline correctly.

For the primary analysis, suppose the population PK model uses the quantifiable concentrations as measured values and represents each BLQ record as a censored observation below 0.50 mg/L. For illustration, assume lognormal residual error: conditional on a participant’s model-predicted concentration, the natural logarithm of a measured concentration differs from the natural logarithm of that prediction by a normally distributed residual with standard deviation \(\sigma\). The model must also account for relevant between-participant variability and the study’s dosing and sampling structure.

Under this assumption, a BLQ observation contributes the probability that the model-generated concentration falls below the LLOQ. If the conditional log concentration has mean \(\mu_i\) and standard deviation \(\sigma\), that probability is:

P(C_i<LLOQ)=\Phi\left(\frac{\log(LLOQ)-\mu_i}{\sigma}\right)

Here, \(\Phi\) is the cumulative distribution function of a standard normal random variable. The expression follows by taking logarithms: \(C_i<LLOQ\) is equivalent to \(\log(C_i)<\log(LLOQ)\); standardizing the normally distributed log concentration gives the stated probability. For population inference, the model must also handle the distribution of participant-level effects, rather than treating the conditional probability for one participant as the entire population likelihood.

This approach relies on the LLOQ being correctly recorded and on BLQ meaning that the true assay result lies below that threshold, not that the value is missing for an unrelated reason. It also relies on an adequate structural model, an appropriate residual-error model, and sufficient information in the observed data to estimate the parameters. The terminal samples must genuinely inform the terminal portion of the profile; a late scheduled time alone does not guarantee that it does.

Worked Example: Calculating a Terminal Rate From Two Quantifiable Samples

For one hypothetical participant, suppose the concentrations at 24 and 36 hours are 2.00 and 0.580 mg/L, respectively. Both exceed the LLOQ of 0.50 mg/L. If these observations lie in a log-linear terminal phase, the two-point estimate of the decline rate is the log concentration ratio divided by the time interval:

\lambda_z=\frac{\log(C_{24}/C_{36})}{36-24} =\frac{\log(2.00/0.580)}{12} =\frac{1.23787436}{12} =0.1031562\ \text{h}^{-1}

The two-point terminal half-life is:

t_{1/2,z}=\frac{\log(2)}{\lambda_z} =\frac{0.69314718}{0.1031562} =6.719\ \text{h}

This calculation illustrates the interpretation of a terminal slope; it is not a population estimate based on all participants. Its assumptions are that both points are quantifiable, belong to the terminal phase, and are not dominated by measurement error or a transient departure from log-linearity. Two observations cannot establish those conditions on their own, so the calculation should not replace assessment of the full concentration-time data and the population model.

Reporting the Population Estimate and Its Sensitivity

In this hypothetical case, the censored-likelihood population model estimates a typical terminal-phase rate constant of 0.0950 h⁻¹, corresponding to a terminal half-life of 7.30 hours. Suppose a participant-level bootstrap gives a 95% interval of 0.0820 to 0.110 h⁻¹ for the terminal rate. Because the half-life is a decreasing function of the rate, the corresponding interval is approximately 6.30 to 8.45 hours. The interval describes uncertainty under the fitted model and bootstrap procedure; it is not a probability statement that the fixed parameter lies in this particular interval.

The report should make the BLQ handling choice and the sensitivity findings visible. For example, suppose the following hypothetical refits use the same structural model and data, changing only the handling of BLQ records:

Handling of terminal BLQ records	Estimated \(\lambda_z\) (h⁻¹)	Derived \(t_{1/2,z}\) (h)	Estimated clearance (L/h)
Censored likelihood; primary analysis	0.0950	7.30	4.80
Substitute LLOQ/2 for each BLQ record	0.0810	8.56	4.86
Omit BLQ records	0.112	6.19	4.73

These are stipulated hypothetical model outputs, not estimates calculated directly from the BLQ counts. The half-lives are derived from the listed rate constants. For example, \(0.693147/0.0810=8.56\) hours, while \(0.693147/0.112=6.19\) hours. Relative to the primary half-life, the substitution estimate is about 17.3% longer and the omission estimate about 15.2% shorter. In this example, clearance changes less than the terminal half-life across the handling choices. That is a finding to report, not a general rule: the pattern depends on the model, the data, and which observations identify each parameter.

Worked Example: Quantifying the Extent of BLQ Records

The terminal window contains 40 participants with eight scheduled records each, so the total is \(40\times8=320\). There are 112 BLQ records. The fraction that is BLQ is:

\frac{112}{320}=0.350=35.0\%

A useful report gives this overall proportion and shows its distribution over time. Here the BLQ count rises from 2 at 24 hours to 20 at both 72 and 84 hours, then remains high at later times. Reporting only “35% of observations were BLQ” would hide the concentration of censoring in the terminal tail. Reporting only the late-time counts would hide the overall scale of the issue.

Worked Example: Interpreting Sensitivity in the Derived Half-Life

The primary terminal rate is 0.0950 h⁻¹. Under LLOQ/2 substitution, it is 0.0810 h⁻¹. The primary and sensitivity half-lives are:

t_{1/2,z,\text{primary}}=\frac{0.693147}{0.0950}=7.296\ \text{h}

t_{1/2,z,\text{substitution}}=\frac{0.693147}{0.0810}=8.557\ \text{h}

The change is \(8.557-7.296=1.261\) hours, or \(1.261/7.296\times100=17.3\%\) longer than the primary estimate. Because half-life is inversely related to the terminal rate, the lower rate produces the longer half-life. This comparison shows why the terminal parameter interpretation should be qualified when it is sensitive to BLQ handling.

What the Terminal Estimate Does and Does Not Mean

The terminal half-life summarizes the slow decline in the terminal portion of the modeled concentration-time profile. It is not automatically the half-life of every earlier phase, and in a multicompartment model the terminal rate can reflect distribution as well as elimination. In particular, \(\lambda_z\) should not be treated as synonymous with clearance. A terminal rate multiplied by a volume yields clearance only under a model and volume definition for which that relationship is justified. The report should therefore name the fitted or derived quantity and avoid inferring clearance from the terminal slope alone.

In the case above, the smaller change in estimated clearance than in terminal half-life suggests that the reported clearance estimate is less sensitive to the chosen BLQ handling in these refits. It does not demonstrate that clearance is insensitive in all analyses, nor does it remove uncertainty in the terminal half-life. If a clinical or dosing conclusion depends on the terminal tail, the sensitivity in that tail deserves particular attention even when another parameter is comparatively stable.

Diagnostics, Alternatives, and Failure Modes

Before interpreting the comparison, check whether the assumptions that support it are plausible. Review observed and predicted concentrations over time, the proportion BLQ by scheduled time and participant, and whether the model predicts a plausible fraction below the LLOQ. Examine whether the censored records cluster by assay batch, site, or participant in a way the model does not represent. Check parameter precision, convergence, and whether a small number of participants or quantifiable late samples dominate the terminal estimate. A model can converge and still provide weakly identified terminal parameters.

The main alternatives answer different questions and have different drawbacks:

Approach	What it does	Important limitation
Censored likelihood	Uses a BLQ record as evidence that concentration is below the LLOQ.	Depends on the model and error distribution adequately describing concentrations near the threshold.
Omit BLQ records	Fits the model using quantifiable concentrations only.	Discards information and can select a systematically different subset of the terminal data.
Substitute a fixed value such as LLOQ/2	Assigns an exact concentration to each BLQ record.	Creates values that were not measured and can distort the terminal slope and its uncertainty.

The sensitivity refits are informative when they are presented as alternative handling scenarios, not as equally valid measurements of the hidden concentrations. For instance, an LLOQ/2 fit does not reveal that the true BLQ values were half the LLOQ. It shows how the fitted results change under that imposed convention.

A common mistake is to say that BLQ concentrations equal zero and then include them as exact zeros in a log-linear regression. A logarithm of zero is undefined, and, more fundamentally, the assay did not establish that the concentration was zero. Another mistake is to omit BLQ values without explaining that choice, then interpret the resulting terminal slope as if all scheduled observations had been used. The correction is to state what each record contributes to the analysis, distinguish censored information from exact measurement, and report how a justified sensitivity analysis affects the conclusion.

Censored-likelihood handling also has limitations. A misspecified structural model or residual distribution can yield misleading probabilities below the LLOQ. If few quantifiable observations remain in the terminal region, the data may not identify the terminal rate precisely even when censored records are included. The report should acknowledge these limitations and avoid claiming that the primary method eliminates uncertainty or model dependence.

Key takeaway: Report how many terminal records are BLQ, where they occur, how the primary model uses them, and how alternative handling changes the terminal rate and half-life. Interpret those estimates in light of model assumptions and distinguish terminal decline from clearance.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

In the hypothetical data set, what fraction of terminal records is BLQ, and why is the time pattern also important?
What information does a censored-likelihood model use from a record reported as \(C<LLOQ\)?
If \(\lambda_z=0.100\) h⁻¹, what is the corresponding terminal half-life?
Why does an LLOQ/2 sensitivity analysis not establish that the true BLQ concentrations equal LLOQ/2?
Why should a terminal rate constant not automatically be reported as clearance?

Solutions. (1) The fraction is \(112/320=0.350\), or 35.0%. The timing matters because BLQ records are concentrated at later times, where they can affect interpretation of the terminal tail. (2) It uses the event that the concentration is below the known LLOQ, contributing the model probability of that event rather than an exact concentration. (3) \(t_{1/2,z}=0.693147/0.100=6.93\) hours. (4) LLOQ/2 is an imposed value for a sensitivity refit, not a measured value or a fact established about the censored concentrations. (5) The terminal rate describes decline in the terminal phase; its relationship to clearance depends on the structural model and volume definition, and in a multicompartment model the terminal phase can reflect distribution as well as elimination.