Why Clearance–Volume Correlation Matters
In a one-compartment model with first-order elimination, an individual’s elimination rate depends on the ratio of clearance to volume, \(CL_i/V_i\). The half-life depends on the reciprocal ratio, \(V_i/CL_i\). Consequently, knowing how clearance and volume vary separately is not enough to determine how much elimination rates or half-lives vary: the way their individual deviations move together also matters.
This tutorial considers between-subject random effects on clearance and volume. Its focus is the interpretation of their correlation, not how to estimate every aspect of a mixed-effects model. The central result is that positive correlation between the log-scale effects tends to reduce variability in the clearance-to-volume ratio, whereas negative correlation tends to increase it. The correlation does not, by itself, specify any one individual’s clearance, volume, or concentration profile.
Model and Assumptions
Here, \(\rho\) is the correlation between the two log-scale random effects, not a statement that every subject with above-typical clearance must also have above-typical volume. A positive value describes a tendency for the effects to deviate in the same direction; a negative value describes a tendency to deviate in opposite directions. Individual subjects can depart from that tendency.
The derivations below assume the one-compartment, first-order elimination relationship \(k_i=CL_i/V_i\), with half-life \(t_{1/2,i}=\log(2)/k_i\). The normal random-effect assumption makes the difference of the two effects normally distributed. The formula for the variance of that difference follows from their variances and covariance. These assumptions define the scope of the conclusions; a different structural model or random-effect distribution may require a different interpretation.
Deriving the Effect on Elimination Rate and Half-Life
Substitution of the individual parameter equations into the clearance-to-volume ratio gives
The typical elimination rate is \(CL/V\). On the log scale, the subject-specific departure from that typical rate is the contrast \(\eta_{CL,i}-\eta_{V,i}\). Its variance is obtained by expanding the variance of a difference:
Call this variance \(s_k^2\). Since the two effects are jointly normal, their difference is normal, so \(\log(k_i/(CL/V))\) is normal with mean zero and variance \(s_k^2\). The half-life has the opposite contrast:
The log-scale half-life departure is therefore also normal with variance \(s_k^2\). The rate and half-life departures have opposite signs for each subject: a larger-than-typical rate corresponds to a shorter-than-typical half-life.
If \(\rho=0\), the log-scale effects are uncorrelated and the variance is the sum of their variances. If \(\rho=1\), it is \((\omega_{CL}-\omega_V)^2\); it is zero only when the two standard deviations are equal. If \(\rho=-1\), it is \((\omega_{CL}+\omega_V)^2\), the largest possible value for these fixed marginal standard deviations. These boundary cases clarify why “positive correlation cancels variability” is only approximate: cancellation is complete only under the stated equal-standard-deviation condition.
Worked Example: Comparing Half-Life Variability at Two Correlations
Suppose a hypothetical one-compartment model has typical \(CL=5.00\) L/h and \(V=50.0\) L, with \(\omega_{CL}=0.200\) and \(\omega_V=0.300\). Compare \(\rho=0.750\) with \(\rho=-0.750\). The positive parameters and the bivariate normal random-effect assumption satisfy the conditions above.
The typical elimination rate is \(CL/V=5.00/50.0=0.100\) h\(^{-1}\), and the typical half-life is
For \(\rho=0.750\), the variance of the log-rate contrast is
For \(\rho=-0.750\), it is
Thus the log-scale standard deviations are \(0.200\) and \(\sqrt{0.220}=0.4690\), respectively. The negative correlation produces substantially greater dispersion in individual rates and half-lives, even though the marginal random-effect variances have not changed.
Because the half-life is lognormal under these assumptions, its median is \(6.931\) h in either case. Its arithmetic mean is the median multiplied by \(\exp(s_k^2/2)\). For \(\rho=0.750\), this is \(6.9314718\exp(0.0400/2)=7.071\) h. For \(\rho=-0.750\), it is \(6.9314718\exp(0.220/2)=7.737\) h. These are population arithmetic means under the model, not the half-life of a typical subject.
What the Correlation Says About Concentration Profiles
For an instantaneous intravenous bolus dose \(D\) in a one-compartment model, the individual concentration is
Volume affects the initial concentration through \(D/V_i\), while the ratio \(CL_i/V_i\) determines the exponential decline rate. Correlation between random effects can therefore affect the distribution of decline rates. It does not imply that every subject has the same initial concentration or prescribe the full concentration curve. In particular, two subjects can have the same clearance-to-volume ratio, and thus the same half-life, but different volumes and different initial concentrations.
Worked Example: Same Half-Life, Different Concentration Scale
Consider a hypothetical 100 mg bolus dose with typical \(CL=5.00\) L/h and \(V=50.0\) L. The typical concentration at \(t=10.0\) h is \(2.00\exp(-0.100\times10.0)=0.7358\) mg/L. Now consider a subject with \(\eta_{CL}=\eta_V=\log(1.25)=0.2231\). These paired effects are one possible individual realization; they are not implied for every subject by a positive population correlation.
For this subject, \(CL_i=5.00(1.25)=6.25\) L/h and \(V_i=50.0(1.25)=62.5\) L. Therefore \(k_i=6.25/62.5=0.100\) h\(^{-1}\), and the half-life remains \(6.931\) h. The initial concentration is \(100/62.5=1.600\) mg/L, so at 10.0 h,
The subject has the same decline rate as the typical profile but a lower concentration at every time because the larger volume lowers the initial concentration. This illustrates why interpreting the correlation through half-life alone does not fully characterize concentration profiles.
Interpreting the Population Distribution
Since \(k_i\) and \(t_{1/2,i}\) are exponentials of normal contrasts, they are lognormally distributed under the stated assumptions. The typical value \(CL/V\), and likewise \(\log(2)V/CL\), is the median of the corresponding individual distribution when the contrast has mean zero. It is not generally the arithmetic mean. For a lognormal variable with median \(m\) and log-scale variance \(s^2\), its arithmetic mean is \(m\exp(s^2/2)\). Thus, correlation changes both the spread and the arithmetic mean of the individual rate and half-life distributions, while leaving their medians fixed when the marginal variances and typical parameters are held fixed.
Worked Example: Probability of a Half-Life Above a Threshold
Use the hypothetical values from the first example. Calculate the model-implied probability that an individual half-life exceeds 8.00 h for each correlation. The threshold is positive, and the log half-life is normal with median 6.931 h and log-scale standard deviation \(s_k\), so the normal probability calculation is applicable.
The event \(t_{1/2,i}>8.00\) is equivalent to the log-scale contrast exceeding \(\log(8.00/6.9314718)=0.1434\). With \(\rho=0.750\), \(s_k=0.200\), so the standardized value is \(z=0.1434/0.200=0.717\). Hence \(P(t_{1/2,i}>8.00)=1-\Phi(0.717)\approx0.237\).
With \(\rho=-0.750\), \(s_k=0.4690\), so \(z=0.1434/0.4690=0.306\). Hence \(P(t_{1/2,i}>8.00)=1-\Phi(0.306)\approx0.380\). Under the respective model assumptions, a larger fraction of the population has a half-life above 8.00 h in the negative-correlation case. These are model-implied population proportions, not probabilities that a particular subject’s unknown half-life exceeds the threshold after observing data.
Estimation, Diagnostics, and Alternatives
The interpretation above is conditional on the fitted values of the two marginal variances and their correlation. A correlation estimate should therefore be read alongside uncertainty in those quantities and the data’s ability to distinguish their effects. Information about the earlier material on identifiability of correlated random effects is relevant here: a numerical estimate of \(\rho\) is not automatically a precise or well-supported estimate.
In a one-compartment concentration profile, volume can influence the scale near the dose, while clearance and volume together determine the decline rate. If observations are sparse, concentrated in a limited time range, or noisy, the data may provide weak information to separate these influences. Useful checks include inspecting whether the estimated covariance matrix is valid, whether the correlation estimate is near its allowed boundary, whether its uncertainty is large, and whether the fitted model reproduces the observed concentration patterns across time. Implausible or unstable estimates call for scrutiny of the data, model structure, and parameterization rather than a literal biological interpretation.
The main alternatives answer different modeling questions:
| Approach | What it assumes or represents | Implication for the rate contrast |
|---|---|---|
| Correlated random effects | Estimates \(\rho\) along with the two marginal variances. | Variance is \(\omega_{CL}^2+\omega_V^2-2\rho\omega_{CL}\omega_V\). |
| Independent random effects | Sets \(\rho=0\). | Rate-contrast variance is the sum of the two marginal variances. |
| One random effect omitted or fixed | Does not allow both parameters to vary independently across subjects. | Rate variability is attributed only to the effects retained in the model. |
Setting \(\rho=0\) can be a useful simpler model when the data do not support estimating a covariance, but it is an assumption, not evidence that the true effects are uncorrelated. Conversely, estimating a correlation adds flexibility but may be difficult with limited information. Omitting a random effect changes the scientific variability represented by the model and should not be treated as an automatic numerical fix.
Common Pitfalls and Limitations
- Reversing the sign. The rate uses \(\eta_{CL}-\eta_V\), so its variance has the term \(-2\rho\omega_{CL}\omega_V\). The half-life uses the opposite contrast, but its variance has the same covariance term.
- Assuming positive correlation means no rate variability. It reduces the variance for fixed marginal variances, but complete cancellation requires \(\rho=1\) and equal random-effect standard deviations.
- Treating a population tendency as an individual rule. A positive \(\rho\) does not require each individual’s two effects to have the same sign or size.
- Confusing typical and arithmetic mean values. With nonzero log-scale variance, the median and arithmetic mean half-life differ; report which summary is being used.
- Equating half-life with the whole profile. Half-life describes the decline rate in this model, but volume also affects the concentration scale after a bolus dose.
The formulas rely on a particular structural relationship and a bivariate normal distribution on the log scale. They describe model-based between-subject variability; they do not establish a causal relationship between clearance and volume. Nor does a fitted correlation alone show that a biological mechanism links them. Such interpretations require evidence beyond the covariance parameter.
Check Your Understanding
Answer each question using the stated model and assumptions. Brief solutions follow the questions.
- If \(\omega_{CL}=0.25\), \(\omega_V=0.20\), and \(\rho=0\), what is the variance of the log-rate contrast?
- With the same marginal standard deviations, would changing \(\rho\) from \(0.5\) to \(-0.5\) increase or decrease the log-half-life variance? Why?
- Why can two subjects have the same half-life but different concentrations after the same bolus dose?
- For a lognormal half-life distribution, is the typical half-life generally its arithmetic mean? State the condition under which they coincide.
Solutions. (1) \(0.25^2+0.20^2=0.0625+0.0400=0.1025\). (2) It increases: the term \(-2\rho\omega_{CL}\omega_V\) becomes larger when \(\rho\) decreases, so the variance increases. (3) Equal \(CL_i/V_i\) gives equal half-lives, but different \(V_i\) gives different initial concentrations \(D/V_i\). (4) Not generally. If the log-scale variance is zero, the distribution has no random-effect spread and its median and arithmetic mean coincide.