Why Selecting the Terminal Slope Requires Judgment
After a concentration-time profile reaches its peak, concentrations often decline. The late decline can sometimes be represented by a straight line after taking logarithms of concentration. The slope of that line is used to estimate the terminal rate constant, commonly denoted \(\lambda_z\). The word terminal matters: a straight line fitted to an earlier decline may describe a distribution phase, ongoing absorption, or a transition between phases rather than the late-time behavior of the profile.
Selecting the terminal points is therefore a modeling decision, not merely a request to fit a line. The candidate observations should form a biologically plausible late segment, have a reasonably linear pattern on the log scale, and yield a slope that is not unduly sensitive to small changes in the selected range. The fit statistics help assess a candidate; they do not decide the pharmacokinetic interpretation by themselves.
Model and Estimation
For one concentration-time profile, let \(C_i\) be the measured concentration at time \(t_i\). A candidate terminal segment is modeled as
Here \(\beta_0\) is the intercept, \(\beta_1\) is the slope, and \(\varepsilon_i\) is the deviation of an observation from the fitted line. The regression is fitted only to selected, positive, quantifiable concentrations. The concentration-time decline must be plausibly in its terminal phase, and a straight-line approximation on the log scale must be reasonable across the chosen times.
For the usual least-squares interpretation and its standard regression diagnostics, assume the log concentrations follow a linear mean trend over the candidate segment, the errors have mean zero and are independent with approximately constant variance, and (for conventional small-sample inference) are approximately normal. These assumptions concern the fitted segment; they do not establish that the segment is truly terminal. Concentration measurements from the same individual also should not be treated as independent subjects in a population analysis. In standard noncompartmental analysis, the slope is typically estimated separately for each individual profile.
The least-squares slope can be calculated from the selected times and log concentrations:
This uses natural logarithms. If common logarithms are used instead, the fitted slope must be multiplied by \(\ln(10)\) to obtain the natural-log slope; \(\lambda_z\) is still the negative of that natural-log slope. Mixing log bases without conversion gives an incorrect rate constant.
Why the Regression Slope Represents a Rate Constant
In a period where concentration declines exponentially with time, write the concentration as \(C(t)=C_0e^{-\lambda_z t}\), where \(C_0\) is the concentration implied at the start of that exponential segment. Taking natural logarithms gives a linear equation:
Thus the intercept is \(\ln C_0\), and the slope is \(-\lambda_z\). In a regression with measurement deviations, the estimated slope \(\hat{\beta}_1\) estimates this underlying slope, so \(\hat{\lambda}_z=-\hat{\beta}_1\). This connection depends on the selected observations representing the same exponential terminal decline. A negative slope by itself is not enough.
Worked Example: Estimating a Rate Constant From a Log-Linear Decline
Suppose a hypothetical profile has concentrations of 20, 10, 5, and 2.5 mg/L at 8, 12, 16, and 20 hours. Assume, for this calculation, that these are positive, quantifiable observations from a plausible terminal segment.
The natural logs are approximately 2.99573, 2.30259, 1.60944, and 0.916291. Each four-hour step reduces concentration by one half, so the log concentration drops by \(\ln(2)=0.693147\) in four hours. The slope is therefore
For example, the exponential model implies that the concentration halves over \(\ln(2)/\hat{\lambda}_z=0.693147/0.173287=4.000\) hours. The calculation gives a rate estimate, but the regular halving pattern alone does not prove that these observations are terminal; their timing and the rest of the profile must also be considered.
Assessing Candidate Terminal Points
Start by identifying plausible late observations from the profile, rather than automatically starting at a fixed time or including every post-peak value. Candidate points should be ordered in time, quantifiable, and part of a declining phase. Fit a line to a candidate segment, then examine the residuals \(e_i=\ln(C_i)-(\hat{\beta}_0+\hat{\beta}_1t_i)\). Residuals should generally fluctuate around zero without a systematic curve, run, or abrupt shift. A clear pattern suggests that a single straight line is not describing the candidate segment adequately.
Next, compare scientifically plausible alternative segments. Ask whether adding an earlier point changes the slope substantially, whether omitting the last point changes it substantially, and whether the residual pattern improves or worsens. A stable slope across reasonable choices supports the robustness of the estimate. Instability can signal a phase transition, curvature, a highly influential observation, or too few useful sampling times.
The coefficient of determination \(R^2\) summarizes how much variation in the observed log concentrations is accounted for by the fitted line. It can be useful when comparing candidate fits of the same profile, but it is not a test of whether the selected observations belong to the terminal phase. A short segment can have a very high \(R^2\), and a late point with measurement error may reduce it even when the overall terminal pattern is plausible. Use \(R^2\) alongside residuals, slope stability, quantifiability, sampling context, and pharmacokinetic judgment.
Use the full profile and sampling context to distinguish a possible terminal decline from earlier post-peak phases. Do not log-transform zero or nonpositive values.
Regress natural-log concentration on time using a reasonable set of late, quantifiable observations, and calculate \(\hat{\lambda}_z=-\hat{\beta}_1\).
Look for structure in residuals and compare slopes from reasonable alternative point sets. A change in slope or a patterned residual sequence requires investigation.
Retain a segment whose log-linear pattern and stability are defensible in context. Record the included times and the reasons for excluding plausible alternatives.
Worked Example: Comparing Candidate Point Sets
Consider hypothetical log concentrations (natural-log concentration, with concentration expressed in mg/L) at 4, 6, 8, 10, and 12 hours: 2.90, 2.00, 1.50, 1.00, and 0.50. All five are positive and quantifiable. The late four points from 6 to 12 hours form an exact straight-line pattern in these displayed values: each two-hour increase in time corresponds to a decrease of 0.50 log units.
For the candidate using 6, 8, 10, and 12 hours, \(\bar{t}=9\), \(\overline{\ln C}=1.25\), and
The resulting rate estimate is \(\hat{\lambda}_z=0.2500\ \mathrm{h}^{-1}\), and the residuals for these four displayed points are all zero. The earlier candidate point at 4 hours has log concentration 2.90. The late-line prediction at 4 hours is \(1.25+(-0.2500)(4-9)=2.50\), giving a residual of \(2.90-2.50=0.40\) log units. This earlier point does not follow the late segment and may belong to an earlier phase.
If all five points are fitted together, \(\bar{t}=8\), \(\overline{\ln C}=1.58\), and the slope is \(-11.6/40=-0.2900\ \mathrm{h}^{-1}\), yielding \(\hat{\lambda}_z=0.2900\ \mathrm{h}^{-1}\). The earlier point changes the estimate by \(0.0400\ \mathrm{h}^{-1}\), or 16% relative to 0.2500. The five-point fit still has a high \(R^2\) (approximately 0.981), but its residuals show a change in pattern around the added earlier observation. The late four-point segment is more consistent with a stable terminal slope in this hypothetical profile. The choice is supported by residuals and slope stability, not by \(R^2\) alone.
Worked Example: Checking Sensitivity to a Late Observation
A second hypothetical profile has natural-log concentrations 1.00, 0.62, 0.20, and \(-0.23\) at 10, 12, 14, and 16 hours, respectively. These correspond to positive concentrations of approximately 2.718, 1.859, 1.221, and 0.794 mg/L. Suppose the four measurements are quantifiable and the profile is declining.
For all four points, \(\bar{t}=13\), \(\overline{\ln C}=0.3975\), and the slope numerator is \(-4.11\), while the denominator is \(20\). Thus
Now omit the 10-hour observation and refit the three later points. Their mean time is 14 hours, their mean log concentration is \(0.196667\), the slope numerator is \(-1.70000\), and the denominator is 8. Therefore \(\hat{\beta}_1=-1.70000/8=-0.2125\ \mathrm{h}^{-1}\), giving \(\hat{\lambda}_z=0.2125\ \mathrm{h}^{-1}\). The two estimates differ by approximately \((0.2125-0.2055)/0.2125=3.3\%\) relative to the three-point estimate. This is a much smaller slope difference, though the analyst should still investigate the observations, sampling, and phase interpretation rather than report a rate as robust simply because the four-point regression looks linear. With only three points in the alternative fit, residual diagnostics also have little power to reveal a pattern.
Failure Modes, Alternatives, and Limits
Several practical problems can make a terminal-slope estimate misleading. If too few late samples are collected, there may be no defensible segment to fit. If concentrations approach or fall below the quantification limit, the remaining observations may not support reliable log-scale regression; a below-quantification result is not a zero concentration that can be logged or substituted casually. A single unusual point can exert substantial influence, especially in a short segment. Curvature or a change in residual pattern can indicate that the chosen times span more than one kinetic phase.
The fitted terminal slope describes the observed concentration-time profile over the selected segment. It is not automatically a pure elimination constant for the underlying drug. After extravascular dosing, for example, slow absorption can govern the late decline; this is often called flip-flop behavior. In that case the observed terminal slope may reflect absorption rather than elimination. Likewise, a multicompartment profile can have an earlier distribution decline followed by a different terminal decline. The data and design determine what can be inferred.
A reasonable alternative to a single log-linear segment is a pharmacokinetic compartmental model that represents multiple processes explicitly. That approach requires a specified structural model and additional assumptions; it is not a justification for fitting a visually convenient line. If the available profile cannot support a stable terminal segment, report that limitation rather than forcing a rate estimate. The introductory material on AUC from zero to the last quantifiable time provides the context for the observed profile; the terminal rate estimate is a distinct modeling step.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions below.
- A regression of natural-log concentration on time has slope \(-0.12\ \mathrm{h}^{-1}\). What is \(\hat{\lambda}_z\), and what are its units?
- Why is a high \(R^2\) insufficient evidence that a candidate segment is terminal?
- What does a systematic curve in residuals suggest about a single straight-line model for the selected points?
- Why should a below-quantification concentration not be replaced by zero and included in a log-linear regression?
- A four-point fit gives \(\hat{\lambda}_z=0.18\ \mathrm{h}^{-1}\), while a plausible three-point late fit gives \(0.27\ \mathrm{h}^{-1}\). What should the analyst examine before selecting a value?
Solutions. (1) \(\hat{\lambda}_z=-(-0.12)=0.12\ \mathrm{h}^{-1}\), the negative of the fitted slope. (2) \(R^2\) describes how well a line accounts for variation among the included points; it does not show that those points belong to the terminal process. (3) A curve suggests that the log concentrations do not follow a single linear mean trend across that segment, perhaps because of curvature or a phase transition. (4) The natural logarithm of zero is undefined, and a below-quantification result does not provide an observed exact concentration suitable for ordinary log-linear regression. (5) Examine the residuals, the included sampling times, data quality and quantifiability, and whether the slope change reflects a phase transition or an influential observation; assess whether any candidate has a defensible terminal interpretation rather than choosing by fit statistic alone.