Computational Statistics & Data Analysis 1 (1983) 239-255
North-Holland
239
On an integrated approach to
member selection and parameter
estimation for Pearson distributions
Rudolph S. PARRISH
Computer Sciences Corporation (c/o U.S. EPA, Athens, GA)
Received February 1983
Revised September 1983
Abstract: Empirical data often are summarized by selecting and fitting an appropriate member from
a wide class of frequency distributions. A common technique used for selection of a density
function and estimation of its parameters relies upon the method of moments. In the case of
Pearson distributions, it is possible to adopt a loss function approach and thereby avoid some of the
problems associated with the moments method. Discussed here is a general technique that can be
utilized for selecting and fitting, concurrently, a Pearson distribution on the basis of any of several
loss functions. Except for Pearson distribution evaluation routines, which are available, only
commonly available numerical minimization methods are required. In order to facilitate implemen-
tation of this procedure, analytical derivatives of Pearson density functions are provided. In
addition, the technique is illustrated for cases of maximum likelihood, minimum chi-square,
truncated Pearson distributions, estimation of Pearson priors, and an application from chemistry.
Keywords and phrases: Estimation of distribution parameters, Loss functions, Pearson distributions.
I. Introduction
Whenever empirical data are to be summarized using a single distribution
function, it is often convenient to fit a distribution which has been selected from
an extensive family of empirical distributions. Two such families in common use
are the Pearson system and the Johnson system [5]. Both of these consist of
four-parameter distributions of various types. The Pearson system, for example,
contains as members the normal, gamma, beta, and Student's t, among others.
For both the Johnson system and the Pearson system, a member is selected for
use in a given application usuaUy on the basis of sample estimates of the third and
fourth standardized moments. Indeed, the selection procedure frequently is sum-
marized geometrically by partitioning a subset of the skewness-kurtosis plane into
regions corresponding to specific distribution types from the particular family (see
Figure 1). However, as pointed out by Slifker and Shapiro [13], there are
shortcomings to this procedure. Specifically they mention that the variances of
0167-9473/83/$3.00 © 1983, Elsevier Science Publishers B.V. (North-Holland)