Why Parameter Shifts Need a Profile-Based Interpretation
A below-quantification-limit (BQL) record says that the concentration is below the assay’s lower limit of quantification (LLOQ); it does not report a measured concentration equal to zero, LLOQ/2, or LLOQ. Substituting one of those values creates a numerical point that can influence a fitted concentration-time profile. When many late samples are BQL, those artificial points can have particular leverage on the estimated terminal decline.
The direction of the effect is not fixed. A substituted value may be higher or lower than the unobserved concentration that generated the BQL record. A value above that concentration tends to pull a declining profile upward at late times and flatten its fitted terminal slope. A value below it tends to pull the profile downward and steepen the slope. The resulting changes in clearance and volume also depend on how the model links those parameters to the profile.
To make the links transparent, the calculations below use a one-compartment intravenous bolus model with first-order elimination. After a bolus dose, its concentration profile is \(C(t)=A\exp(-kt)\), where \(A\) is the concentration at time zero and \(k\) is the terminal elimination rate constant. With dose \(D\), volume \(V=D/A\), and clearance \(CL=kV\). These simple relationships are useful for interpreting fitted profiles, but they do not apply unchanged to every dosing route or structural model.
How a Substituted Point Changes a Log-Linear Fit
Taking logarithms of the one-compartment profile gives \(\log C(t)=\log A-kt\). Thus, in a regression of log concentration on time, the intercept estimates \(\log A\) and the slope estimates \(-k\). Substitution changes the log concentration at a BQL time, and therefore can change both the fitted slope and intercept.
The following short derivation shows the effect of changing just one point in an ordinary least-squares log-linear fit. Suppose the times are \(t_i\), and one value at \(t_j\) is changed by \(\delta=\log(C_{\mathrm{sub}})-\log(C_{\mathrm{old}})\). The change to the fitted slope is the point’s centered time multiplied by its change in log concentration, divided by the sum of squared centered times:
This follows directly from the least-squares slope formula, \(\hat{\beta}_1=\sum_i(t_i-\bar{t})y_i/\sum_i(t_i-\bar{t})^2\), because changing only \(y_j\) by \(\delta\) changes the numerator by \((t_j-\bar{t})\delta\) and leaves the denominator unchanged. For a late point, \(t_j-\bar{t}\) is positive. A substitution above the old value has \(\delta>0\), so it makes the slope less negative and the estimated \(k=-\hat{\beta}_1\) smaller. A substitution below the old value has \(\delta<0\), so it makes the slope more negative and the estimated \(k\) larger.
This derivation describes a particular log-linear fit, not a universal formula for a nonlinear mixed-effects fit. In an NLME model, the change can also depend on the observation model, the other records, individual random effects, and how population parameters are estimated. The calculation is a diagnostic aid for understanding a profile, not a replacement for fitting the specified model.
Worked Example: One Late BQL Point Shifts Slope, Intercept, and PK Parameters
Hypothetical data. Consider a 100 mg intravenous bolus with a concentration profile \(C(t)=10\exp(-0.200t)\) mg/L, where \(t\) is in hours. The corresponding parameters are \(A=10.0\) mg/L, \(V=100/10=10.0\) L, and \(CL=0.200\times10.0=2.00\) L/h. Suppose concentrations are recorded at 0, 4, 8, 12, and 16 h, and the LLOQ is 0.500 mg/L. The profile gives approximately 10.00, 4.493, 2.019, 0.9072, and 0.4076 mg/L. The 16 h record is therefore BQL.
For this illustration, fit an ordinary least-squares line to log concentration versus time, using the exact profile values at the first four times and a substituted value at 16 h. This is a deliberately simple calculation: it assumes equal-weight errors on the log scale and does not represent a full NLME fit. The five times have \(\bar{t}=8\) h and \(\sum_i(t_i-\bar{t})^2=160\ \mathrm{h}^2\). At the last time, the slope changes by \((16-8)\delta/160=\delta/20\).
Substitute the LLOQ. The actual profile value at 16 h is 0.4076 mg/L, while the substituted value is 0.500 mg/L. Hence \(\delta=\log(0.500/0.4076)=0.2043\), and the fitted slope becomes \(-0.200+0.2043/20=-0.1898\ \mathrm{h}^{-1}\). For this point, the intercept changes by \(\delta[1/5-\bar{t}(16-\bar{t})/160]= -0.2\delta=-0.04085\). Thus \(\hat{A}=10\exp(-0.04085)=9.600\) mg/L, \(\hat{V}=100/9.600=10.42\) L, and \(\widehat{CL}=0.1898\times10.42=1.977\) L/h. In this example, LLOQ substitution flattens the fitted terminal slope, increases the implied volume, and slightly decreases the implied clearance.
Substitute LLOQ/2. Now the assigned value is 0.250 mg/L, below the profile’s 0.4076 mg/L concentration. Then \(\delta=\log(0.250/0.4076)=-0.4889\). The fitted slope becomes \(-0.200-0.4889/20=-0.2244\ \mathrm{h}^{-1}\), and the intercept changes by \(-0.2\delta=0.09778\). Therefore, \(\hat{A}=10\exp(0.09778)=11.03\) mg/L, \(\hat{V}=100/11.03=9.068\) L, and \(\widehat{CL}=0.2244\times9.068=2.035\) L/h. Here LLOQ/2 steepens the fitted slope, decreases the implied volume, and slightly increases the implied clearance.
The opposite shifts arise because the two assigned values fall on opposite sides of the profile’s value at 16 h. The BQL record itself does not reveal that value; the numbers only illustrate how different substitutions can affect a particular assumed profile.
Reading the Direction of the Shift
For the one-compartment IV-bolus model, \(k\) is the negative terminal slope and \(A\) determines the volume through \(V=D/A\). Therefore, a substitution can change clearance through both \(k\) and \(V\), since \(CL=kV\). The clearance shift is not necessarily in the same direction or proportion as the slope shift: an increase in \(k\), for example, can be partly offset by a decrease in \(V\).
The following comparison uses two points to make the profile effect easy to see. It does not imply that two-point estimation is a suitable general NLME procedure. For a declining profile, whether a substitution flattens or steepens the slope depends on the value assigned relative to the unobserved concentration, not simply on whether the rule is called “LLOQ” or “LLOQ/2.”
Worked Example: The Same BQL Record Can Imply Different Apparent Declines
Hypothetical profile. A concentration of 0.800 mg/L is quantified at 12 h. A candidate profile with terminal rate \(0.120\ \mathrm{h}^{-1}\) predicts, 12 h later, \(0.800\exp(-0.120\times12)=0.1895\) mg/L. If the LLOQ is 0.200 mg/L, that predicted concentration is BQL.
If the BQL result at 24 h is replaced by the LLOQ, the two-point apparent rate is \(\log(0.800/0.200)/12=\log(4)/12=0.1155\ \mathrm{h}^{-1}\). That is flatter than the candidate rate of \(0.120\ \mathrm{h}^{-1}\), because 0.200 mg/L is above the candidate profile’s 0.1895 mg/L. If the same result is replaced by LLOQ/2, the apparent rate is \(\log(0.800/0.100)/12=\log(8)/12=0.1733\ \mathrm{h}^{-1}\). This is steeper, because 0.100 mg/L is below 0.1895 mg/L.
As an additional parameter translation, suppose a hypothetical fit after substitution reports \(k=0.1182\ \mathrm{h}^{-1}\) and \(V=1.40\) L. The implied clearance is \(CL=kV=0.1182\times1.40=0.1655\) L/h. This multiplication translates the reported fit into the model’s clearance parameter; it does not show that substitution alone caused that estimate. For comparison, \(0.120\times1.40=0.1680\) L/h if the rate were \(0.120\ \mathrm{h}^{-1}\) at the same volume.
Population Parameters and Assumptions
An individual profile calculation helps interpret a fitted curve, but NLME estimates describe a population model as well as individual deviations from it. When many late observations are BQL, a substitution can alter the individual profile contributions and thereby influence population estimates. The direction and magnitude depend on how many records are substituted, their timing, the spread of concentrations, the residual-error model, and the relationship between individual and population parameters.
Keep the assumptions visible when using parameter relationships. The equations \(V=D/A\) and \(CL=kV\) in the examples presume an IV bolus, a one-compartment model, first-order elimination, and a post-dose profile governed by that model. For an oral model, the early profile may reflect absorption as well as disposition; for a multi-compartment model, a terminal slope need not equal \(CL/V\). In those settings, reading a volume or clearance shift directly from a terminal slope and intercept can be misleading.
A second assumption is that the substituted numerical values are treated by the fit as if they were observed concentrations. That is precisely why the fit can be distorted: the BQL status establishes an inequality relative to the LLOQ, whereas substitution imposes an exact value. The earlier material on censored observation models distinguishes such records from quantified values. A likelihood that uses the censoring information does not assert that a BQL concentration equals a chosen substitute.
Diagnostics, Alternatives, and Limitations
To diagnose a possible substitution-driven shift, compare the fitted curves with the original observation pattern and mark the times and values assigned to BQL records. In a terminal run of BQL observations, look for a fitted tail that bends toward the common substituted value or appears flatter or steeper than is compatible with the quantified concentrations. Then ask whether the estimated \(k\), \(V\), and \(CL\) move together in a way consistent with the model equations. A change in clearance cannot be interpreted from the slope alone when the fitted volume also changes.
A useful check is to calculate the concentration predicted by a candidate fitted profile at each BQL time and compare it with the assigned value and the LLOQ. This does not reveal the true unobserved concentration; it shows whether the substitution lies above or below the candidate profile and therefore which direction it may pull a simple fit. Also confirm that every quantified record remains above the LLOQ, that BQL records have not been inadvertently treated as measured zeros, and that reported parameter units are consistent.
Substitution is easy to implement, but it discards the distinction between an interval statement and an exact measurement. By contrast, the M3 approach uses the probability of a result falling below the LLOQ, given the model and observation error. That distinction is especially relevant when late BQL values are numerous. Neither a simple profile calculation nor a parameter-shift comparison proves that one fitted model is correct; the result remains conditional on the structural and observation models used.
These calculations also have limits. The true concentrations corresponding to BQL records are unknown, so a profile-based comparison cannot establish the exact bias in an estimate. Parameter shifts can be small, large, or change direction across sampling designs and models. A stable-looking estimate under one substitution is not evidence that the assigned concentrations were observed, and a shift in an estimate does not by itself establish which alternative estimate is closer to the truth.
Worked Example: Interpreting a Reported Shift Without Overclaiming
Hypothetical fit comparison. A one-compartment IV-bolus analysis reports \(k=0.120\ \mathrm{h}^{-1}\) and \(V=1.40\) L under one handling rule, and \(k=0.1182\ \mathrm{h}^{-1}\) with the same reported volume under a substitution rule. The clearance values are \(0.120\times1.40=0.1680\) L/h and \(0.1182\times1.40=0.1655\) L/h. The difference is \(0.1655-0.1680=-0.0025\) L/h, approximately a 1.50% decrease relative to 0.1680 L/h.
The supported interpretation is that, in this hypothetical comparison and under the stated model, the second fit has a slightly lower estimated terminal rate and implied clearance. It is not justified to conclude from this arithmetic alone that the underlying population clearance is lower, or that substitution must produce a decrease in other data sets. To assess the shift, inspect where the substituted values lie relative to the fitted profiles and how much the volume and other model components change.
Check Your Understanding
Try each question, then check the brief solution immediately below it.
- A late substituted concentration is above the candidate profile’s predicted concentration. For a log-linear terminal fit, which way does the fitted rate \(k\) tend to move?
- For a one-compartment IV-bolus model, a fitted intercept corresponds to \(A=8.00\) mg/L after a 96 mg dose. Calculate \(V\).
- At a fixed volume of 2.00 L, calculate the clearance corresponding to \(k=0.150\ \mathrm{h}^{-1}\). What is the clearance if the fitted rate instead becomes \(0.180\ \mathrm{h}^{-1}\) and volume stays fixed?
- Why can LLOQ/2 substitution steepen a terminal slope even though the record is BQL?
- Name two assumptions needed before using \(V=D/A\) and \(CL=kV\) to interpret a fitted terminal profile.
Solutions. (1) The assigned value is above the candidate profile, so it raises the late log concentration; the fitted slope becomes less negative and \(k\) tends to decrease. (2) \(V=D/A=96/8.00=12.0\) L. (3) Initially, \(CL=kV=0.150\times2.00=0.300\) L/h. At \(k=0.180\ \mathrm{h}^{-1}\), \(CL=0.180\times2.00=0.360\) L/h. (4) BQL means below LLOQ, not equal to a particular unobserved concentration. If LLOQ/2 is below the profile’s predicted concentration at that time, substituting it pulls the late point downward and can make the decline steeper. (5) Examples include an IV-bolus dose, a one-compartment model, first-order elimination, and a post-dose profile governed by that model. The parameter relationships should not be transferred automatically to other structural models or dosing routes.