J. theor. Biol. (1992) 158, 109-128
Detecting and Displaying Size Bimodality: Kurtosis, Skewness
and Bimodalizable Distributions
TOMASZ WYSZOMIRSKI
Department of Phytosociology and Plant Ecology, Institute of Botany,
Warsaw University, Al. Ujazdowskie 4, PL-00-478 Warsaw, Poland
(Received on 25 July 1991, Accepted on 5 March 1992)
Bimodality of size distributions is often found both in plant and animal populations,
but any widely accepted measure of it does not exist. Kurtosis coefficient g2 can be,
with some reservations, considered to be a measure of bimodality in symmetric
distributions. Bimodality-generating mechanisms usually make distributions platy-
kurtic. Strongly asymmetric distributions are ieptokurtic and also may remain as
such when an additional mode emerges. This disqualifies kurtosis as a bimodality
measure. The logarithmic transformation, which is often used to make distributions
less asymmetric, may create bimodality. This leads to the concept of bimodalizable
and platykurtizable distributions, i.e. distributions becoming bimodal or platykurtic,
respectively, after transformation. The log-transformation is appropriate only in
some cases. In this paper, the Box-Cox (BC) transformation to symmetry is proposed
as a basis for assessing bimodalizability. Bimodalizable and platykurtizable distribu-
tions are defined as distributions becoming bimodai or platykurtic, respectively,
after the symmetrizing BC-transformation. Further transformation by the normal
cumulative distribution function allows us to present platykurtizability graphically.
Kurtosis of BC-transformed data indicates the operation of bimodality-promoting
mechanisms much better than the common kurtosis coefficient. Properties of the
proposed procedure are illustrated by its application to distribution mixtures.
Examples of its behaviour are also presented for data drawn from simulation of
growth and competition in even-aged plant populations.
1. Introduction
Size frequency distributions are usually far from normality both in natural and
experimental populations. Mass distributions are often positively skew (see e.g.
Uchmafiski, 1985). Apart from asymmetry, bimodality of size distributions is found
in many cases in plant (Ford, 1975; Mohler et al., 1978; Rabinowitz, 1979) as well
as in animal populations (see Huston & DeAngelis, 1987 for a review).
The study of bimodality is strongly restricted by the lack of its measure. Usually,
analyses were based mainly on the visual inspection of histograms. This is a rather
unreliable method, as the appearance of a histogram depends markedly on the
number of class intervals used. Histograms can only reveal gross distribution changes
(Hutchings, 1975). Existing statistical procedures designed for the analysis of bimod-
ality are oriented mainly at the resolution of empirical distribution into two or more
components (Everitt & Hand, 1981). Such methods are dependent upon
the assumptions about the form of component distributions. This imposes serious
109
0022-5193/92/170109 + 20 $08.00/0 © 1992 Academic Press Limited