Correlated Random Effects in Saturable Elimination

Why the Pairing of Vmax and Km Matters

In a Michaelis–Menten model, the elimination rate for an individual at concentration $C$ is $V_{\max}C/(K_m+C)$. A larger $V_{\max}$ raises the rate, whereas a larger $K_m$ lowers it at a fixed concentration. The distribution of rates across individuals therefore depends not only on how much each parameter varies, but also on whether individuals with high $V_{\max}$ tend to have high or low $K_m$.

This tutorial concerns that pairing at a fixed concentration. It does not treat the distribution of rates as a substitute for a population-average concentration-time prediction: nonlinear models require care when averaging, which is a separate question. The key point here is that a correlation between parameter effects can alter the spread and shape of individual elimination rates even when the marginal distributions of $V_{\max}$ and $K_m$ are unchanged.

Definition: For individual $i$, write the log-scale random effects as $\eta_{V,i}$ and $\eta_{K,i}$, and define $$ V_{\max,i}=V_{\max,\mathrm{typ}}\exp(\eta_{V,i}), \qquad K_{m,i}=K_{m,\mathrm{typ}}\exp(\eta_{K,i}). $$ Here $V_{\max,\mathrm{typ}}$ and $K_{m,\mathrm{typ}}$ are typical population parameters. A common assumption is that the vector $(\eta_{V,i},\eta_{K,i})$ has a bivariate normal distribution with mean zero and covariance matrix $$ \Omega = \begin{pmatrix} \omega_V^2 & \omega_{VK}\\ \omega_{VK} & \omega_K^2 \end{pmatrix}. $$ The random-effect correlation is $\rho_{VK}=\omega_{VK}/(\omega_V\omega_K)$, when both variances are positive. It describes association between the two log-scale effects, not correlation between observed concentrations.

From Parameter Correlation to Rate Variation

At a specified concentration $C>0$, define the individual elimination rate $R_i(C)=V_{\max,i}C/(K_{m,i}+C)$. A useful way to understand the effect of the random effects is to examine the log rate near the typical parameter values. Let $q=K_{m,\mathrm{typ}}/(K_{m,\mathrm{typ}}+C)$. Differentiating the log rate with respect to the two log-scale effects gives

\left.\frac{\partial \log R_i(C)}{\partial \eta_{V,i}}\right|_{\eta=0}=1, \qquad \left.\frac{\partial \log R_i(C)}{\partial \eta_{K,i}}\right|_{\eta=0}=-q.

Thus, for modest random-effect variation, a first-order expansion around the typical values is

\log R_i(C)\ \approx\ \log R_{\mathrm{typ}}(C)+\eta_{V,i}-q\eta_{K,i}, \qquad R_{\mathrm{typ}}(C)=\frac{V_{\max,\mathrm{typ}}C}{K_{m,\mathrm{typ}}+C}.

The variance of this linear combination follows from the variance formula for two correlated random variables:

\operatorname{Var}\{\log R_i(C)\} \ \approx\ \omega_V^2+q^2\omega_K^2-2q\omega_{VK} = \omega_V^2+q^2\omega_K^2-2q\rho_{VK}\omega_V\omega_K.

The minus sign is important. Since the local effect of $\eta_K$ on log rate is negative, positive covariance makes the two contributions tend to offset: individuals with higher $V_{\max}$ also tend to have higher $K_m$. Negative covariance tends to pair higher $V_{\max}$ with lower $K_m$, reinforcing the rate differences. This statement concerns local rate variability at a fixed $C$; it is not a claim that correlation has the same effect on every population summary.

Worked Example: Positive Versus Negative Correlation

Suppose $V_{\max,\mathrm{typ}}=100$ mg/h and $K_{m,\mathrm{typ}}=10$ mg/L. At $C=10$ mg/L, the typical elimination rate is $100(10)/(10+10)=50$ mg/h, and $q=0.5$. Assume the log-scale standard deviations are $\omega_V=0.30$ and $\omega_K=0.40$.

With $\rho_{VK}=0.75$, the covariance is $0.75(0.30)(0.40)=0.0900$. The first-order approximate variance of log rate is $$ 0.30^2+0.5^2(0.40^2)-2(0.5)(0.0900) =0.0900+0.0400-0.0900=0.0400. $$ The approximate standard deviation of log rate is $\sqrt{0.0400}=0.200$.

With $\rho_{VK}=-0.75$, the covariance is $-0.0900$, giving $$ 0.0900+0.0400-2(0.5)(-0.0900)=0.220, $$ and an approximate log-rate standard deviation of $\sqrt{0.220}=0.469$. Under this local approximation, negative correlation produces much greater rate dispersion at 10 mg/L than positive correlation, despite identical marginal random-effect variances. These are approximate log-scale spreads, not exact rate intervals.

Worked Parameter Pairings

The derivative calculation explains the mechanism locally. A direct calculation with hypothetical individual parameter values makes the pairing more concrete. The following example uses the same four values for each random effect in both populations, changing only which $K_m$ effect is paired with each $V_{\max}$ effect.

Worked Example: Same Marginal Effects, Different Pairings

Take $V_{\max,\mathrm{typ}}=100$ mg/h, $K_{m,\mathrm{typ}}=10$ mg/L, and $C=10$ mg/L. Use four hypothetical values for each effect: $-0.4,-0.1,0.1,0.4$. For any paired values $a=\eta_V$ and $b=\eta_K$, the rate is $$ R(C)=\frac{100e^a(10)}{10e^b+10} =\frac{100e^a}{e^b+1}\ \text{mg/h}. $$

For positive pairing, set $b=a$. The four rates are $100e^a/(1+e^a)$, giving approximately 40.13, 47.50, 52.50, and 59.87 mg/h. Their mean in this four-person illustration is $(40.13+47.50+52.50+59.87)/4=50.00$ mg/h.

For negative pairing, pair the same $a$ values with $b=0.4,0.1,-0.1,-0.4$, respectively. Substitution gives approximately 26.90, 42.98, 58.02, and 89.32 mg/h. The mean is $54.31$ mg/h. The negative pairing makes the rates more spread out in this small set because high $V_{\max}$ is paired with low $K_m$, and low $V_{\max}$ with high $K_m$.

This constructed four-person comparison illustrates the effect of pairing; its means are not general predictions of how correlation changes the population mean rate. In other parameter distributions or at other concentrations, the exact mean-rate change must be evaluated from the model rather than inferred from the correlation sign alone.

Concentration Determines the Strength of the Effect

The coefficient $q=K_{m,\mathrm{typ}}/(K_{m,\mathrm{typ}}+C)$ ranges between zero and one. At concentrations well below typical $K_m$, $q$ is near one: the rate is locally sensitive to both parameters, so their covariance can matter substantially. At concentrations well above typical $K_m$, $q$ is near zero: the rate approaches $V_{\max}$, so variation in $K_m$ and its covariance with $V_{\max}$ have little local influence on rate variation.

Worked Example: Correlation at Low and High Concentrations

Suppose $\omega_V=0.30$, $\omega_K=0.40$, and $\rho_{VK}=0.50$, so $\omega_{VK}=0.0600$. Use $K_{m,\mathrm{typ}}=10$ mg/L. At $C=10$ mg/L, $q=0.5$, and the approximate log-rate variance is $$ 0.30^2+0.5^2(0.40^2)-2(0.5)(0.0600) =0.0900+0.0400-0.0600=0.0700. $$

At $C=90$ mg/L, which is nine times typical $K_m$, $q=10/(10+90)=0.1$. The corresponding variance is $$ 0.0900+0.1^2(0.1600)-2(0.1)(0.0600) =0.0900+0.0016-0.0120=0.0796. $$ The covariance contribution has decreased from $-0.0600$ to $-0.0120$. At high concentration the rate is locally governed mainly by $V_{\max}$; the correlation with $K_m$ has less influence. Both values are first-order approximations and rely on the assumed log-scale effects and modest variability.

Assumptions, Diagnostics, and Alternatives

The interpretation above assumes that individual parameters are positive and are represented by exponentiated log-scale effects; that the random effects have mean zero, finite variances, and the specified covariance; and that individuals’ random-effect vectors follow the population distribution independently. It also assumes that $C$ is held fixed while comparing individual rates. If concentration changes over time, the value of $q$, and therefore the local impact of covariance, changes too. The first-order variance formula additionally relies on a local approximation; it is most useful when random-effect variation is not so large that curvature dominates.

The covariance matrix must be valid. In the two-effect case, this requires $\omega_V^2\geq 0$, $\omega_K^2\geq 0$, and $\omega_{VK}^2\leq\omega_V^2\omega_K^2$, equivalently $-1\leq\rho_{VK}\leq1$ when both standard deviations are positive. A fitted correlation near either boundary can signal that the data provide weak information about the joint variability, rather than compelling evidence for an extreme biological relationship.

Useful checks include examining whether the estimated covariance matrix is positive definite, whether uncertainty in the covariance is large, and whether individual parameter estimates show a pattern consistent with the proposed pairing. Sparse concentration coverage can make $V_{\max}$ and $K_m$ difficult to distinguish: when observations occupy a narrow concentration range, apparent correlation may partly reflect weak identification of the two parameters. Individual empirical Bayes estimates can also be pulled toward the population mean, so a scatterplot of those estimates alone is not a definitive diagnostic of the population correlation.

A model with independent random effects fixes $\omega_{VK}=0$. It is simpler and may be appropriate when the data do not support estimating covariance, but it excludes a potentially important pattern of paired individual differences. Estimating an unrestricted covariance allows the data to express that pairing, at the cost of requiring enough information to estimate another parameter reliably. A shared latent effect that influences both parameters is another possible biological structure, but it imposes a different model and interpretation; it is not equivalent to freely estimating a correlation unless its assumptions imply the same covariance structure.

Common mistake: “Positive correlation must increase elimination rates because high $V_{\max}$ goes with high $K_m$.” This reasoning considers only the numerator. At fixed concentration, higher $K_m$ reduces the rate and can offset the higher $V_{\max}$. In the local variance formula, the covariance enters as $-2q\omega_{VK}$, so positive covariance reduces approximate rate variation when $q>0$. The exact rate distribution remains nonlinear, and its mean or other summaries should not be inferred from that variance result alone.

Key takeaway: Correlated random effects describe how individual $V_{\max}$ and $K_m$ values are paired. At a fixed concentration, positive correlation tends locally to offset their opposing effects on elimination rate, while negative correlation tends to reinforce them. The strength of this effect depends on concentration and on the validity of the random-effects model.

Check Your Understanding

Answer each question, then compare with the solutions below.

At a fixed concentration, why does positive covariance between log $V_{\max}$ and log $K_m$ have a negative contribution to the first-order log-rate variance?
If $C$ is much greater than typical $K_m$, what happens to $q$, and which parameter increasingly dominates the rate?
For $\omega_V=0.20$, $\omega_K=0.30$, $\rho_{VK}=-0.50$, and $q=0.5$, calculate the approximate log-rate variance.
State one condition that a valid two-effect covariance matrix must satisfy.
Why is an observed correlation among individual empirical Bayes estimates not, by itself, conclusive evidence about the population random-effect correlation?

Solutions. (1) The local log-rate effects are $\eta_V-q\eta_K$, so the covariance term in their variance is $-2q\omega_{VK}$; for $C>0$, $q>0$, making the contribution negative when covariance is positive. (2) $q=K_{m,\mathrm{typ}}/(K_{m,\mathrm{typ}}+C)$ approaches zero, and the rate approaches $V_{\max}$. (3) The covariance is $(-0.50)(0.20)(0.30)=-0.0300$. The variance is $0.20^2+0.5^2(0.30^2)-2(0.5)(-0.0300)=0.0400+0.0225+0.0300=0.0925$. (4) For example, $\omega_{VK}^2\leq\omega_V^2\omega_K^2$, or equivalently $|\rho_{VK}|\leq1$ when both variances are positive. (5) Such estimates can be shrunk toward the population means, and sparse concentration information can make the two individual parameters difficult to distinguish; the scatterplot alone does not account for those limitations.