University of Groningen
Calculation of the Coefficient of Variation of Log-Normally Distributed Parameter Values
Proost, Johannes H.
Published in:
Clinical Pharmacokinetics
DOI:
10.1007/s40262-019-00760-6
IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to cite from
it. Please check the document version below.
Document Version
Publisher's PDF, also known as Version of record
Publication date:
2019
Link to publication in University of Groningen/UMCG research database
Citation for published version (APA):
Proost, J. H. (2019). Calculation of the Coefficient of Variation of Log-Normally Distributed Parameter
Values. Clinical Pharmacokinetics, 58(9), 1101-1102. [Link]
Copyright
Other than for strictly personal use, it is not permitted to download or to forward/distribute the text or part of it without the consent of the
author(s) and/or copyright holder(s), unless the work is under an open content license (like Creative Commons).
The publication may also be distributed here under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license.
More information can be found on the University of Groningen website: [Link]
amendment.
Take-down policy
If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately
and investigate your claim.
Downloaded from the University of Groningen/UMCG research database (Pure): [Link] For technical reasons the
number of authors shown on this cover page is limited to 10 maximum.
Download date: 04-04-2024
Clinical Pharmacokinetics (2019) 58:1101–1102
[Link]
EDITORIAL
Calculation of the Coefficient of Variation of Log‑Normally Distributed
Parameter Values
Johannes H. Proost1,2
Published online: 30 April 2019
© Springer Nature Switzerland AG 2019
In population pharmacokinetics/pharmacodynamics (PK/ It is well-recognized (e.g. Elassaiss-Schaap and Heis-
PD), individual parameters are usually assumed to be log- terkamp [2]), that, in the case of low variability, the CV
normally distributed (rather than normally distributed) calculated using Eq. 2 is a good approximation for the CV
among the individuals in a population. Equations for mean calculated using Eq. 1. For low values of σ, the CV from
and variance in the log-domain, as well as in the untrans- Eqs. 1 and 2 is approaching asymptotically, as can be shown
formed domain, can be found in many statistical textbooks mathematically. For higher values of σ, Eq. 1 provides higher
and internet sites (e.g. Wikipedia [1]). values for CV than Eq. 2. For example, for σ = 0.5, Eq. 1
For ease of interpretation, the interindividual variability yields a CV of 0.533.
is often reported as a coefficient of variation (CV), rather For this reason, the use of Eq. 2 to calculate the CV as
than the variance of the log-normal distribution. The CV in a measure of variability has been criticized [2]. However,
the normal, untransformed domain is obtained from Eq. 1: this difference is due to the upper tail of the log-normal
distribution. The mean, standard deviation, and coefficient
(1)
√
CV = exp(𝜎 2 ) − 1, of variation in the normal domain are ‘inflated’ by the con-
where σ2 is the variance in the log-domain (to keep as close tribution of high values in the upper tail of the distribution,
as possible to general statistical literature, the symbol for and this effect increases with increasing σ. This is reflected
variance is σ2, rather than ω2 as used in NONMEM for inter- in higher values of CV calculated using Eq. 1. Therefore, it
individual variance). If desired, the CV may be expressed as may be questioned whether the CV according to Eq. 1 is a
a percentage by multiplying by 100 (%CV). useful and meaningful measure of variability if the underly-
Equation 1 is used for the conversion of variances ing values are log-normally distributed. In addition, it may
(obtained from a population analysis program, e.g. NON- be questioned whether it is correct to state that Eq. 2 under-
MEM) to the CV. Alternatively, the CV may be obtained estimates the CV [2], as may be concluded from the follow-
from Eq. 2: ing arguments:
CV = 𝜎. (2)
1. σ is the appropriate measure of variability if the distri-
A survey of recent pharmacokinetic research papers pub-
bution is assumed to be log-normal, in contrast to the
lished in several relevant journals in the field showed that
CV according to Eq. 1, since the latter is a ‘contradictio
both equations are used in practice. In several papers, the
in terminis’; it calculates the CV for the untransformed
CV is reported as a measure of interindividual variability,
data, whereas it is assumed that the data are log-nor-
but the method for calculation of the CV is not mentioned.
mally distributed.
2. Eq. 1 calculates the CV in the untransformed domain.
However, in the population analysis, it is assumed that
* Johannes H. Proost the distribution of individual parameter values is log-
[Link]@[Link] normal. Therefore, it may be questioned why one should
1 consider the CV from Eq. 1 as ‘the true measure’ of
Department of Pharmacokinetics, Toxicology and Targeting,
Groningen Research Institute of Pharmacy, University variability [2]. In the untransformed domain, the distri-
of Groningen, Groningen, The Netherlands bution of the individual values will be highly skewed,
2
Department of Anesthesiology, University Medical and the CV cannot be used, for example, to estimate
Center Groningen, University of Groningen, Groningen, confidence intervals and prediction intervals, to perform
The Netherlands
Vol.:(0123456789)
1102 J. H. Proost
Monte Carlo simulations, or to use as prior informa- pharmacometrics [1]. The classical statistical approach is
tion in Bayesian forecasting. If the population analysis to express all statistical parameters in the normal domain,
is performed assuming a log-normal distribution, the which is in conflict with the use of Eq. 2. However, in phar-
same should be assumed using the statistical results, i.e. macometrics, usually a log-normal distribution is assumed,
mean, standard deviation and CV, in the above-men- and, consequently, the statistical parameters used in further
tioned applications. To this purpose, Eq. 2 provides the steps (see above under point 2), should be expressed in the
appropriate measure of variability in the log-domain. log-domain. The present paper provides a simple extension
In contrast, Eq. 1 provides a correct value of CV in the of the definition of CV in the log-domain, which is more
normal domain, but this value cannot be used for further appropriate in data analysis where a log-normal distribution
calculations. So, why do we need it? is assumed.
3. The geometric mean value, as obtained in the log- In conclusion, the CV according to Eq. 2 is the appropri-
domain, and back-transformed into the normal domain, ate measure of variability if the distribution is assumed to
will be quite different from that of the mean of the be log-normal. It is recommended not to use Eq. 1 to cal-
untransformed values. For example, for σ = 0.5, and culate CV. Finally, it should be stressed that the method for
geometric mean exp(μ) = 100, the mean value in the calculation of CV should be explicitly described in reports
untransformed domain is 113.3. If the CV is calculated and articles.
using Eq. 1, the latter value should be used as the mean
value, but this is not clear when reporting only the CV.
4. If the CV is calculated using Eq. 1 to estimate predic- Compliance with Ethical Standards
tion intervals, to perform Monte Carlo simulations, or
to use as prior information in Bayesian forecasting, the Conflicts of interest Johannes H. Proost has no conflicts of interest to
declare regarding this manuscript.
value σ needs to be calculated (see above). If the CV is
calculated using Eq. 2, no conversion is required. Funding No funding was received in relation to this manuscript.
5. The prediction interval around the geometric mean
can be easily obtained from σ. Defining a factor (a)
[a = exp(1.96 ⋅ 𝜎)], the 95% prediction interval is from
exp(μ)/a to exp(μ) × a. For example, for exp(μ) = 100 and
References
σ = 0.5, a = 2.66 and the prediction interval is 38–266. 1. Wikipedia. Log-normal distribution. https://[Link][Link]/
This makes more sense than CV = 0.533 (or 53.3%) as wiki/Log-normal_distribution. Accessed 3 Feb 2019.
obtained from Eq. 1. 2. Elassaiss-Schaap J, Heisterkamp SH. Variability as constant coef-
ficient of variation: can we right two decades in error? Abstracts
of the Annual Meeting of the Population Approach Group in
There are some arguments against the use of Eq. 2. Most Europe. 2009. [Link]-meeti[Link]/?abstract=1508. Accessed
importantly, this approach does not seem to be mentioned 3 Feb 2019 (Abstract 1508, ISSN 1871-6032).
in statistical textbooks or internet sites outside the field of
