Coefficient of variability of a logit model
In pmxcode: Create Pharmacometric Models

knitr::opts_chunk$set(echo = TRUE)

General derivation

Let's assume a model that includes a parameter $P$ which needs to be constrained between 0 and 1, while being associated with interindividual variability (IIV). The typical value ($TVP$) of the parameter $P$ can be estimated by $\theta_P$, such that $\theta_P \in ]a, b[$.

We will show below that the magnitude of variability associated with the individual parameter values $P_i$ can be calculated as a function of $\theta_P$, $a$, $b$, and the standard deviations of the individual variability estimates ($\eta_i\in \mathbb{R}$):

$$ \begin{array}{lll} \text{ if a > 0 and b > 0}, & \%CV(P_i) & \approx 100 \times \dfrac{ (\theta_P - a) \times (b-\theta_P) }{\theta_P\times(b - a)} \times SD(\eta_i) \ \text{ if a = 0 and b > 0}, & \%CV(P_i) & \approx 100 \times ( 1 - \dfrac{\theta_P}{b} ) \times SD(\eta_i)\ \text{ if a = 0 and b = 1}, & \%CV(P_i) & \approx 100 \times ( 1 - \theta_P ) \times SD(\eta_i) \ \end{array} $$

The logit-transformed value of $P$ can be defined as follows :

$LP = ln \left(\dfrac{\frac{\theta_P - a}{b-a}}{1-\frac{\theta_P - a}{b-a}}\right)$

$\forall \: \theta_P \in ]a, b[$, $LP \in \mathbb{R}$.

We can add an IIV term ($\eta_i\in \mathbb{R}$) prior to applying the back-transformation to the linear scale.

$$ \begin{array}{ll} LP_i &= ln \left(\dfrac{\frac{\theta_P - a}{b-a}}{1-\frac{\theta_P - a}{b-a}}\right) + \eta_i\ \ P_i &= a + (b-a) \times \dfrac{e^{LP_i}}{1+e^{LP_i}} \end{array} $$

The back-transformation enforces that $P \in ]a, b[$ and $P_i \in ]a, b[$.

Calculation of the magnitude of IIV

$$ \begin{array}{ll} P_i &= a + (b -a) \times \dfrac{e^{ln \left(\dfrac{\frac{\theta_P - a}{b-a}}{1-\frac{\theta_P - a}{b-a}}\right) + \eta_i}}{1+e^{ln \left(\dfrac{\frac{\theta_P - a}{b-a}}{1-\frac{\theta_P - a}{b-a}}\right) + \eta_i}}\ &= a + (b -a) \times \dfrac{e^{ln \left(\dfrac{\frac{\theta_P - a}{b-a}}{1-\frac{\theta_P - a}{b-a}}\right)} \times e^{\eta_i}}{1+e^{ln \left(\dfrac{\frac{\theta_P - a}{b-a}}{1-\frac{\theta_P - a}{b-a}}\right)} \times e^{\eta_i}}\ &= a + (b -a) \times \dfrac{\dfrac{\frac{\theta_P - a}{b-a}}{1-\frac{\theta_P - a}{b-a}} \times e^{\eta_i}}{1+\dfrac{\frac{\theta_P - a}{b-a}}{1-\frac{\theta_P - a}{b-a}} \times e^{\eta_i}}\ &= a + (b -a) \times \dfrac{ \frac{\theta_P - a}{b-a} \times e^{\eta_i}}{1-\frac{\theta_P - a}{b-a} + \frac{\theta_P - a}{b-a} \times e^{\eta_i}}\ &= a + (b -a) \times \dfrac{ (\theta_P - a) \times e^{\eta_i}}{(b-a)-(\theta_P - a) + (\theta_P - a) \times e^{\eta_i}}\ &= a + (b -a) \times \dfrac{ (\theta_P - a) \times e^{\eta_i}}{(b-\theta_P) + (\theta_P - a) \times e^{\eta_i}}\ \end{array} $$

The magnitude of variability is expressed as the \% coefficient of variation: $\%CV(P_i) = 100 \times \dfrac{SD(P_i)}{mean(P_i)}$

The mean of $P_i$ is the typical value of $P$, thus $mean(P_i) = \theta_p$

We calculate the variance of $P_i$ which is the square of the standard deviation of $P_i$:

$$ \begin{array}{ll} var(P_i) &= var \left( a + (b -a) \times \dfrac{ (\theta_P - a) \times e^{\eta_i}}{(b-\theta_P) + (\theta_P - a) \times e^{\eta_i}} \right)\ &= (b - a)^2 \times (\theta_P - a)^2 \times var \left( \dfrac{ e^{\eta_i}}{(b-\theta_P) + (\theta_P - a) \times e^{\eta_i}} \right) \end{array} $$

We use the first-order Taylor series expansion to approximate the variance of $P_i$.

$$ \begin{array}{ll} var(P_i) & \approx (b - a)^2 \times (\theta_P - a)^2 \times var \left( \dfrac{ e^{\eta_i}}{(b-\theta_P) + (\theta_P - a) \times e^{\eta_i}}\Big\rvert_{\eta_i=0} + \dfrac{ e^{\eta_i} \times \left( (b-\theta_P) + (\theta_P - a) \times e^{\eta_i} \right) - e^{\eta_i} \times \left( (\theta_P - a) \times e^{\eta_i} \right) }{\left( (b-\theta_P) + (\theta_P - a) \times e^{\eta_i}\right)^2}\Big\rvert_{\eta_i=0} \times \eta_i \right)\ & \approx (b - a)^2 \times (\theta_P - a)^2 \times var \left( \dfrac{1}{b-\theta_P + \theta_P - a} + \dfrac{ b-\theta_P + \theta_P - a - \theta_P + a }{(b-\theta_P + \theta_P - a)^2} \times \eta_i \right)\ & \approx (b - a)^2 \times (\theta_P - a)^2 \times var \left( \dfrac{1}{b - a} \right) \times var \left( \dfrac{ b-\theta_P }{(b - a)^2} \times \eta_i \right)\ & \approx (b - a)^2 \times (\theta_P - a)^2 \times \dfrac{ (b-\theta_P)^2 }{(b - a)^4} \times var(\eta_i)\ & \approx \dfrac{ (\theta_P - a)^2 \times (b-\theta_P)^2 }{(b - a)^2} \times var(\eta_i)\ \text{Therefore:}&\ SD(P_i) & \approx \dfrac{ (\theta_P - a) \times (b-\theta_P) }{b - a} \times SD(\eta_i) \end{array}\ $$

Therefore:

$$ \%CV(P_i) \approx 100 \times \dfrac{ (\theta_P - a) \times (b-\theta_P) }{\theta_P\times(b - a)} \times SD(\eta_i) $$

Special cases:

a = 0: $\%CV(P_i) \approx 100 \times ( 1 - \frac{\theta_P}{b} ) \times SD(\eta_i)$
a = 0 and b = 1: $\%CV(P_i) \approx 100 \times ( 1 - \theta_P ) \times SD(\eta_i)$

Alternative derivations for parameters bound between 0 and 1

Let's assume a model that includes a parameter $P$ which needs to be constrained between 0 and 1, while being associated with IIV and covariate effects. The typical value ($TVP$) of the parameter $P$ can be estimated by $\theta_P$, such that $\theta_P \in ]0, 1[$.

The logit-transformed value of $P$ can be defined as follows :

$LP = ln \left(\dfrac{\theta_P}{1-\theta_P}\right)$

$\forall \: \theta_P \in ]0, 1[$, $LP \in \mathbb{R}$.

We can add an IIV term ($\eta_i\in \mathbb{R}$) prior to applying the back-transformation to the linear scale.

$$ \begin{array}{ll} LP_i &= ln \left(\dfrac{\theta_P}{1-\theta_P}\right) + \eta_i\ \ P_i &= \dfrac{e^{LP_i}}{1+e^{LP_i}} \end{array} $$

The back-transformation enforces that $P \in ]0, 1[$ and $P_i \in ]0, 1[$.

Calculation of the magnitude of IIV

To calculate the magnitude of variability in $P_i$, we re-arrange its expression assuming $\kappa = \dfrac{\theta_P}{1-\theta_P}$:

$$ P_i = \dfrac{e^{ln(\kappa) + \eta_i}}{1+e^{ln(\kappa) + \eta_i}} $$

The magnitude of variability is expressed as the \% coefficient of variation: $\%CV(P_i) = 100 \times \dfrac{SD(P_i)}{mean(P_i)}$

The mean of $P_i$ is the typical value of $P$, thus $mean(P_i) = \dfrac{\kappa}{1+\kappa}$ ( that simplifies to $\theta_p$)

We calculate the variance of $P_i$ which is the square of the standard deviation of $P_i$:

$$ \begin{array}{ll} var(P_i) &= var \left( \dfrac{\kappa \times e^{\eta_i}}{1+\kappa \times e^{\eta_i}}\right)\ &= \kappa^2 \times var \left( \dfrac{ e^{\eta_i}}{1+\kappa \times e^{\eta_i}}\right) \end{array} $$

We use the first-order Taylor series expansion to approximate the variance of $P_i$.

$$ \begin{array}{ll} var(P_i) & \approx \kappa^2 \times var \left( \dfrac{e^{\eta_i}}{1+\kappa \times e^{\eta_i}}\Big\rvert_{\eta_i=0} + \dfrac{ e^{\eta_i} \times \left(1+\kappa \times e^{\eta_i}\right) - e^{\eta_i} \times \left(\kappa \times e^{\eta_i}\right) }{\left(1+\kappa \times e^{\eta_i}\right)^2}\Big\rvert_{\eta_i=0} \times \eta_i \right)\ & \approx \kappa^2 \times var \left( \dfrac{1}{1+\kappa \times 1} + \dfrac{ 1 \times \left(1+\kappa \times 1\right) - 1 \times \left(\kappa \times 1\right) }{\left(1+\kappa \times 1\right)^2} \times \eta_i \right) \ & \approx \kappa^2 \times var \left( \dfrac{1}{1+\kappa} \right) \times var \left( \dfrac{\eta_i}{\left(1+\kappa \right)^2} \right)\ & \approx \kappa^2 \times 1 \times \dfrac{var \left(\eta_i\right)}{\left(1+\kappa \right)^4} \ \text{Therefore:}&\ SD(P_i) & \approx \kappa \times \dfrac{SD(\eta_i)}{\left(1+\kappa\right)^2} \end{array}\ $$

Therefore:

$$ \begin{array}{ll} \%CV(P_i) & \approx 100 \times \kappa \times \dfrac{SD(\eta_i)}{\left(1+\kappa\right)^2} \times \dfrac{1+\kappa}{\kappa}\ & \approx 100 \times \dfrac{SD(\eta_i)}{1+\kappa}\ & \approx 100 \times \dfrac{SD(\eta_i)}{1+\dfrac{\theta_P}{1-\theta_P}}\ & \approx 100 \times \dfrac{SD(\eta_i)}{\dfrac{1-\theta_P + \theta_P}{1-\theta_P}}\ & \approx 100 \times (1-\theta_P) \times SD(\eta_i)\ \end{array} $$