Upload
lmunet
View
0
Download
0
Embed Size (px)
Citation preview
Selection of Variables for Discriminant Analysis of Human Crania for
Ancestry Determination
Adam Kolatorowicz
B.S., Northern Illinois University, 2002
A Thesis Submitted in Partial Fulfillment of the Requirements for the Degree of Master of
Science in Human Biology in the Graduate School of the University of Indianapolis
January 2006
Copyright © 2006 Adam Kolatorowicz. All Rights Reserved.
i
FORM B
Accepted by the faculty of the Graduate School, University of Indianapolis, in the partial
fulfillment of the requirements for the Master of Science degree in
HUMAN BIOLOGY
__________________________________
Thesis Advisor – Dr. Stephen P. Nawrocki, D.A.B.F.A
__________________________________
Reader – Dr. John H. Langdon
Date:
ii
ACKNOWLEDGEMENTS
I would like to thank Noel D. Justice at the Indiana University Glenn A. Black Laboratory of
Archaeology for allowing me access to the Angel Mounds skeletal material on which I refined
my measurement technique.
Thanks go to Dr. Della C. Cook at the Indiana University Anthropology Department
Osteological Collection for allowing me access to the Arkansaw Mounds material.
Additional thanks go to Lyman Jellema at the Cleveland Museum of Natural History for
allowing me access to the Hamann-Todd Osteological Collection and for providing me with the
Mitutoyo direct input tool and foot pedal. Without those instruments I surely would not have
completed data collection on time.
Last, but not least, I would like to extend much gratitude to Dr. Stephen P. Nawrocki for letting
me borrow the very expensive instrumentation to complete the study and for helping me develop
the proper measuring technique.
This project was funded in part by:
The Connective Tissue Graduate Student Research Fund
iii
ABSTRACT
Forensic anthropologists use the computer program FORDISC 2.0 (FD2) as an analytical
tool for the determination of ancestry of unknown individuals. There are an almost endless
number of measurements that can be taken on the human skeleton, yet FORDISC includes only
78 measurements for its analysis. In particular, the program will only utilize up to 24
measurements of the cranium. These 24 cranial variables are used because they require simple,
relatively inexpensive instruments that most biological anthropology laboratories have
(spreading and sliding calipers). Also, individuals with a basic knowledge of the anatomical
landmarks can take the measurements with relative ease. Unconventional measurements of the
cranium require unusual, costly instruments (such as the radiometer and coordinate caliper) and
are more difficult to take. This study will examine which measurements of the human cranium
provide the greatest classificatory power when constructing discriminant function formulae for
the determination of ancestry and will answer the question of whether the use of variables that
require more time, training, and equipment are worth the effort.
Sixty five cranial measurement were taken on 155 adult human crania from three
different ancestral groups: (1) African American (n = 50), (2) European American (n = 50),
and (3) Coyotero Apache (n = 55). The 65 measurements were broken up into four subsets for
statistical analysis: (1) FD2 (1996), (2) Howells (1973), (3) Gill (1984), and (4) All
Measurements. A predictive discriminant analysis with a forward stepwise methodology of
p = 0.05 to enter and p = 0.15 to remove was run using the computer software package SPSS
13.0. The analysis produced 4 sets of discriminant function formulae. The classificatory power
of each set of formulae was determined by comparing the hit-rate estimation (the percent
correctly classified) of each of the subsets. First, the resubstitution rate was compared to the
iv
leave-one-out (LOO) rate for each subset and then both rates were compared across all subsets.
The FD2 subset had a resubstitution rate of 90.3% and LOO rate of 85.8%. The Howells subset
had a resubstitution rate of 92.9% and a LOO rate of 90.3%. The Gill subset had a resubstitution
rate of 63.2% and a LOO rate of 61.9%. Finally, the All Measurements subset had a
resubstitution rate of 95.5% and a LOO rate of 93.5%. The non-standard measurements of the
All Measurements subset performed the best and the standard FD2 measurements performed
third best. Non-standard measurements incorporated in the All Measurements formulae included
frontal subtense, mid-orbital breadth, bistephanic breadth, bimaxillary breadth, and molar
alveolar radius.
The formulae provided the best separation of the Apache group from the other two
groups. Stepwise analysis showed that the use of more variables is not necessarily better, as not
all of the variables were included in the final formulae. Only 12 of the 24 FD2 measurements,
12 of the 57 Howells measurements, 4 of the 6 Gill measurements, and 15 of the 65 All
Measurements were used. Results show that the non-standard measurements can be useful for
determining the ancestry of unknown human crania. These measurements could be especially
useful for incomplete crania. It is suggested that biological anthropology laboratories purchase
radiometers and coordinate calipers to record data that would be missed with spreading and
sliding calipers. Standard measurements can be combined with non-standard measurements to
produce more powerful discriminant function formulae for the determination of ancestry.
v
TABLE OF CONTENTS
Signature Page . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . i
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .ii
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
Chapter 1: Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Chapter 2: Multiple Discriminant Function Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Chapter 3: Standard and Non-Standard Measurements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
Chapter 4: Materials and Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Chapter 5: Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Chapter 6: Discussion and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Literature Cited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Appendix A: Cranial Measurement Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Appendix B: Craniometrics Recording Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Appendix C: Casewise Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
1
CHAPTER 1: INTRODUCTION
The term “ancestry” as used by biological anthropologists refers to the population
affiliation, based on geographical location, of an individual. The determination of the ancestry
of human skeletal remains is a key step in the course of identifying unknown individuals. The
predicted ancestry, sex, age, and stature form a biological profile of the decedent which, in turn,
helps with the identification process by narrowing down the number of potential matches.
The biological profile is also used by anthropologists to understand human variation in
historic and prehistoric populations. Bioarcheologists analyze the skeletons of long-deceased
individuals to better understand the history of human populations, including migration patterns,
demographic structure, and the effects of disease. Creating an accurate biological profile,
whether for recent or ancient remains, is the starting point of all future analysis. Therefore, it is
imperative to refine existing methods for determining ancestry and sex and for estimating age
and stature in order to improve the accuracy of the biological profile, and, by extension,
conclusions based on those profiles.
In this thesis the term ancestry is used instead of the term race. Race is an outmoded
term that implies discrete, typological, and thus artificial subdivisions of the human species
based on their overt physical and behavioral characteristics (Nawrocki, 1993). Ancestry, on the
hand, refers to the broad geographical origin of a group of humans and emphasizes ancestral-
descendant relationships between groups and the microevolutionary forces that produce human
variation.
Forensic anthropologists correctly note that certain morphological traits are known to be
present at higher frequencies in certain ancestral populations, and the data suggest that clusters of
traits may occur at higher frequencies in specific groups. However, any specific trait may not
2
occur exclusively in one particular group, and no trait or assemblage of traits can perfectly
distinguish between all members of ancestral groups. In other words, the presence of a trait or a
cluster of traits may suggest, but not confirm, group affiliation. If one were to examine the
global population one would find that many morphological traits are clinally distributed across
the globe with no discrete boundaries between groups of people. In some ways the population of
the United States is an exception because of its unique structure and history.
Groups from all over the world have recently migrated to the United States. In their
native lands, morphological traits can be clinally distributed and grade gradually into
neighboring groups. Migrating subsets of these ancestral groups have largely tended to maintain
their cultural and biological identities in North America, creating non-clinal distributions of traits
and increasing the appearance of biological discreteness relative to their native continents.
Ancestral categories as used in the American medicolegal setting are descriptive tools used to
identify unknown individuals and to communicate with law enforcement agencies. Forensic
anthropologists predict facial, hair, and skin characteristics from skeletal morphology. This
prediction of ancestral appearance can only be accomplished when the anthropologist is familiar
with the variation that exists within populations that are likely to be found in the area that he/she
is working.
Skeletal Ancestry Determination
Determining the ancestry of a decedent from their skeletal remains is perhaps the most
difficult aspect of the biological profile to resolve (Church, 1995). Variables used to determine
ancestry can be placed into one of two categories: (1) non-metric variables or (2) metric
variables. Non-metric variables are traits of the skeleton that are discontinuous, discrete, or
quasi-continuous, such as orbital shape or prognathism curvature. Metric variables rely on
3
continuous measurements of the skeleton, such as maximum cranial breadth or orbital height.
One type of variable is not superior to the other; rather, “. . . the chances of being right in a racial
identification depends largely on the observer’s experience” (Stewart, 1948:318). Nonetheless,
there has long been debate within biological anthropology as to which variables yield the best
results. Classically, non-metric variables were seen as a subjective, less scientific means of
determining ancestry compared to metric variables, which were considered more objective and
scientific. Stewart (1948) notes that measuring bones is a “mechanical procedure,” but that does
not mean that a possibility for error does not exist. Carpenter (1976) examined the utility of non-
metric traits versus metric traits to determine ancestry from the skull, at a time when non-metric
traits dominated the field. He observed both discrete traits and took measurements of skulls from
the Terry Collection and concluded that craniometrics are an excellent way to determine
ancestry, with 86 - 91% correctly identified. Krogman and İşcan (1986) state that on average, 85
- 90% of individuals are correctly identified by using discrete traits and 80 - 88% of individuals
are identified correctly using metric methods. Today, both metric and non-metric traits are used
to estimate one’s ancestry in forensic anthropology (Krogman and İşcan, 1986; Reichs, 1986;
Gill and Rhine, 1990; Church, 1995).
Measurements of a cranium may be used to give a numerical description of an already
recognizable discrete trait. In other words, it is possible to convert features that anthropologists
have traditionally classified as discrete traits into metric traits by measuring the features in one or
more dimensions. Mastoid size is one example. Conversely, some metric traits, such as nasal
breadth and alveolar prognathism, are sometimes treated as discrete traits.
Craniometric Assessment of Ancestry
Out of all the components of the human skeleton, the traits of the skull and face are
4
the most often used and most effective in ancestry determination (Gill, 1986, 1998; Novotný et
al., 1993). Large, geographically-defined populations share similar features of the cranium. It is
the slight variation in the size and shape of these features that allows the anthropologist to
determine ancestral group affiliation (Sauer, 1992; Brace, 1995; Kennedy, 1995).
A feature on one individual will be smaller or larger than the same feature on another
individual; therefore, single measurements describe the size of a feature. It is then easy to
compare sizes of traits between two individuals using simple, univariate statistical methods. If
two measurements are taken of a feature, the second measurement is said to provide a description
of its shape. Shape is the proportion of the two measurements, which will make a feature on one
cranium “look” different from the same feature on another. If more measurements are taken then
one can be more detailed in one’s description of the skull under study. In other words, shape can
be more accurately described. Analyzing multiple variables at the same time requires the use of
multivariate statistical methods. Although having the same foundation as univariate methods,
multivariate methods require the use of advanced statistical computer software.
Some researches have called for the refinement of craniometric techniques as well as
metric techniques in general (Brues, 1990; Church, 1995; Gill, 1998). In the early 1900s, there
was little to no standardization among measurement devices. Instruments were often custom-
made by the researcher and different institutions/laboratories used different instrumentation and
techniques to collect data from the skeleton. It was very difficult to duplicate methodology and
compare results of previous experiments with later studies. Efforts by researchers such as
Hrdlička (1920), Martin (1928), and Howells (1937) helped to standardize osteometric data
collection by providing definitions of measurements and landmarks, which included a
description of the placement of instrumentation on the cranium. In fact, many researchers still
refer to Martin for measurement definitions (Howells, 1973; Krogman and İşcan, 1986; Bass,
5
1995; Buikstra and Ubelaker, 1994; Moore-Jansen et al., 1994). As there are several references
(Hrdlička, 1920; Martin, 1928; Howells, 1937, 1973; Buikstra and Ubelaker, 1994; Moore-
Jansen et al., 1994; Bass, 1995) available today that define measurements of the skeleton, it has
become easier to test the utility of earlier techniques on modern populations. Most of the error in
recording measurements of the cranium no longer lies with the instrumentation but instead
involves interobserver error.
Currently, most biological anthropology laboratories have basic measuring devices as
part of their collection. Instruments such as spreading and sliding calipers are relatively
inexpensive and are simple to learn how to use. A cursory knowledge of anatomical landmark
location will allow the researcher to take measurements of the human cranium with virtual ease.
Using only the spreading and sliding calipers limits the researcher to the amount of data that can
be collected from the cranium. Within forensic anthropology, there is a standard set of 24
measurements that correspond to the measurements on the computer program FORDISC (Jantz
and Ousley, 1993; Ousley and Jantz, 1996), which practitioners turn to in their osteological
analysis. The creators of FORDISC have found the 24 measurements to be the most effective
when using discriminant function analysis to determine the sex and ancestry of an unknown
cranium. There are, however, other measurement devices that allow the researcher to more fully
describe the shape of the cranium by recording different types of measurements. Instruments
such as the coordinate caliper and radiometer are more expensive and require more training to
manipulate properly. The landmarks that define the measurements taken by these instruments
are more difficult to locate, but the instruments can measure features that sliding and/or
spreading calipers cannot.
6
Purpose and Hypothesis
The purpose of this study is to determine which measurements of the human cranium
provide the greatest classificatory power when constructing discriminant function formulae for
the determination of ancestry. The study will identify the most effective combinations of
standard and non-standard measurements and will answer the question of whether the use of
variables that require more time, training, and equipment are worth the effort. Perhaps biological
anthropology laboratories should invest in other instrumentation besides spreading and sliding
calipers. It is hypothesized that a combination of standard and non-standard measurements will
provide higher classificatory power over only standard measurements. The study will test the
null hypothesis that adding non-standard measurements to standard measurements will not
increase the classificatory power of a linear discrimination function.
Chapter 2 provides a description of predictive multiple discriminant function analysis and
its early and later applications to biological anthropology research. Chapter 3 discusses the
different types of measurements used in craniometry as well as special instrumentation. Chapter
4 details the study sample that was used along with the statistical methodology employed. A
description of the results of the statistical analysis is provided in Chapter 5, followed by
conclusions reached from the study in Chapter 6.
7
CHAPTER 2: MULTIPLE DISCRIMINANT FUNCTION ANALYSIS
Multiple discriminant function analysis is a type of multivariate statistical technique
which reduces large data sets to explain the relationship between two types of variables. It
involves both grouping variables (the groups) and response variables (the observed
characteristics). There are two major types of multiple discriminant function analyses: (1)
descriptive (DDA) and (2) predictive (PDA). In DDA, the response variables are viewed as
outcome variables and the grouping variables are viewed as the explanatory variables. In other
words, the response variables are used to separate the groups. DDA is very similar to a
multivariate analysis of variance in that both methods analyze the effect of more than one
response variable on more than one grouping variable. In PDA, the response variables act as
predictor variables and the grouping variables are viewed as outcome variables. In other words,
the response variables are used to predict the grouping variables, or, the roles of the two
variables are reversed compared to DDA. PDA is very similar to multiple regression analysis in
that both methods are used to predict a dependent grouping variable from independent response
variables. As this study examines the applications of PDA; DDA will be discussed no further.
Huberty (1994) provides a detailed explanation of both types of analysis.
It is essential to understand the theoretical and mathematical foundations of PDA before
one can apply the method to a research model. Basic PDA theory is explained below, followed
by a discussion of its most common uses within biological anthropology. Next, a history of PDA
beginning with its inception in statistical analysis will be reviewed along with its development in
the first half of the 20th
century. The later applications of PDA coinciding with the rise of
computer technology will then be addressed.
8
Predictive Discriminant Analysis
PDA is used to predict group membership of unknown objects. The object could be a
living person, an ancient artifact, a car, or a human skull. Applying a PDA to these objects might
tell the researcher if the person has risk factors for a disease, what the artifact was made for, what
make and model the car is, or the geographical origin of the skull. The prediction is based on a
suite of characteristics (response variables) that are measured or observed from objects of known
membership from different groups (grouping variables). The characteristics are used to construct
a model of the average group member. The unknown object is then compared to the average
model of each group. The group to which the unknown object is the closest becomes the
predicted group membership.
When constructing a PDA, the grouping variables must be defined prior to the analysis.
When predicting the grouping variable from the response variables of an unknown object, one
must be certain that it does in fact belong to one of the grouping variables. A PDA will assign
the object to one of the groups regardless of whether or not it actually belongs to one of those
groups.
In its simplest terms, PDA maximizes the ratio of variation between groups to the
variation within a group (Howells, 1973; McLachlan, 1992; Huberty, 1994; Kachigan, 1991;
Afifi et al., 2004). In other words, PDA provides the greatest separation possible of the groups
under study. The response variables are used to create variation and correlation matrices, which,
after undergoing algebraic transformation, are converted into simple, algebraic linear
discriminant functions (LDF) (Fisher, 1936) that look like the following:
f1x1 + f2x2 + f3x3 + . . . . + fnxn + c = L
9
where f is the discriminant coefficient determined by the matrices, x is the response variable, c is
a constant, and L is the function score. The number of LDFs that are produced by the PDA is
determined by the number of grouping variables and it can be described by the following
formula:
n = k – 1
where n is the number of LDFs and k is the number of grouping variables. In other words, there
will be one fewer LDF than there are groups. The score of the unknown object is then compared
to each grouping variables’ average score, or group centroid. All scores are placed in
multidimensional space and the distances between them are calculated. This distance is referred
to as the squared Mahalanobis distance or D2. It is important to differentiate D
2 from Euclidean
distances.
Euclidean distance is the straight line distance between two points (p and q) in space. In
multiple dimensions the Euclidean distance is
√(∑I = 1N (pi - qi)²)
where i is the dimension, N is the number of dimensions, and pi (or qi) is the coordinate of p (or
q) in dimension i. D2 values take into account the variation between grouping variables, so it can
be said that they are distances scaled by the statistical variation of each point (Rao, 1952) or “a
measure of the actual magnitude of divergence between the . . . groups under comparison”
(Mahalanobis et al., 1949:237). In multiple dimensions the Mahalanobis distance is
(pi – qi)’si-1
(pi – qi)
where si-1
is the inverse covariance matrix. The larger the D2 distance of an object from a group
10
centroid, the less likely the object belongs to that particular group. The smaller the D2 distance
of an object from a group centroid, the more likely the object belongs to that particular group.
After LDFs are created, one must test their efficacy on other objects to determine how
well the formulae perform outside of the original sample. The performance of the LDFs is
reported as a percent of correct classification, or hit-rate, of a test sample of objects with known
group membership.
Some researchers state that PDA’s should be used with caution (Klepinger and Giles,
1998). One must be careful of how one uses the functions because they are not easy to interpret
as they do not correspond to something that we can easily identify in the objects. LDFs are used
only to describe the relationships among groups (Howells, 1969; Kowalski, 1972; DiBennardo,
1986). The linear discriminant function score is only a number that abstractly represents an
ancestral group. Applying the technique to anthropological questions of group relationships and
individual identification transforms ancestry and sex into statistical abstractions of trait
complexes (Gill, 1995).
Uses in Anthropology. Within biological anthropology, there are two common uses of
predictive discriminant analysis: (1) evolutionary relationships within hominid studies and (2)
predicting sex and/or ancestry of unknown individuals (Feldesman, 1997). As early as 1948,
Rao suggested using PDA and geographical classification to trace the evolution of species.
Since then, anthropologists have been using PDA to construct evolutionary trees and outline the
history of humankind from australopithecines (Ashton et al., 1957) to the Bushmen-Hottentot
groups of Africa (Rightmire, 1970) to Australia and Oceania (Pietrusewsky, 1990) to Plains
Indians (Key, 1994). Howells (1972, 1973) has completed broader studies using PDA to
understand the variation of cranial morphology across the entire globe as well as the
relationships among different groups. Still, the methodology used to discover evolutionary ties
11
must be used with caution because the sample sizes are small, the number of populations is
unknown, and the degree to which the variables and groups covary is unknown (van Vark, 1994).
Researchers have also used PDA to determine the sex and ancestry of unknown skeletal remains
from historic contexts (Jantz and Owsley, 1994) and modern, forensic contexts (DiBennardo and
Taylor, 1983). PDA plays a particularly important role in identification when morphological
traits are not clear (Novotný et al., 1993), as when remains are damaged or when non-metric
traits are inconclusive. All in all, PDA has been a valuable tool for anthropologists asking
questions about group relationships and individual identification.
Early Applications
The use of PDA to aid in identifying unknown skeletons has its roots in the first half of
the 20th
century. At the height of the Eugenics movement, statisticians were at the forefront in
developing new mathematical classificatory methods as they implemented biometric techniques
to study and classify humans. Journals of the day such as Biometrika were inundated with
articles involving the measurement of human crania. Researchers scoured the globe looking for
crania to measure, including everything from Oxford undergraduates and Royal Engineers
(Benington and Pearson, 1911), Congo and Gabon Africans (Benington and Pearson, 1912),
Bushmen and Hottentots of South Africa (Broom, 1923), dynastic Egyptians (Pearson and
Davin, 1924), Eastern Islanders (von Bonin, 1931), Kenyans (Kitson, 1931), Southern, Eastern,
and Northern Asians (Woo and Morant, 1932), New Britains from Indonesia (von Bonin, 1936),
and Native Americans (von Bonin and Morant, 1938). It seemed as though researchers had an
obsession with craniometrics as they measured skulls to describe and classify groups of humans.
Pearson and Davin (1924) boldly stated, “In vulgar estimation the craniologist is still something
of a body-snatcher.” This was a comment on the fact that researchers were taking skeletons back
12
to their laboratories without necessarily having the permission to do so. At this point,
biometricians were only comparing means and deriving correlations of measurements of the
cranium.
R.A. Fisher developed the first discriminant function formulae in the early 1930’s by
using matrix algebra. This new approach allowed the analyst to maximize the ratio of the
between sample variance to the within sample variance. The result of this procedure is known
today as Fisher’s linear discriminant function (LDF). Fisher’s colleague, Barnard (1935), was
the first to publish the results of the application of PDA to biometry when he examined variation
in skull shape in Egyptian populations. Barnard examined four series of predynastic and
dynastic Egyptian crania to find temporal differences. Although he did not specifically label the
technique as discriminant analysis, he aimed to maximize the difference between the series
relative to the variance within a series. Barnard goes so far as to suggest using the method as a
supplement to anatomical sexing of the skeleton. Fisher (1936) was the first to use the term
discriminant functions and introduced it as a method to separate groups based on multiple
measurements. He used measurements of three species of iris as an example to describe how the
technique worked. Fisher (1938) provided a more detailed statistical explanation of PDA in a
later publication. Thus far, PDA had been limited to studying three or fewer groups at a time.
In the 1940’s and early 1950’s, Rao refined discriminant techniques by examining the
uses of multiple measurements in biological classification and explored its limits within
anthropometry (1946, 1948, 1949, 1952). Mahalanobis and colleagues (1949) undertook an
anthropometric study of colossal proportions as they measured and classified the inbreeding
castes of men in India, which pushed the threshold for the number of variables and groups that
could be included in a discriminant analysis. It was in this study that D2 values were used as a
measure of generalized distance between groups. The use of D2 allowed researchers to better
13
understand group variation.
The process of creating linear discriminant functions was time consuming and all done by
hand, which limited the size of the datasets that could be used. Early uses of PDA lacked
refinement and standardization, which made researchers shy away from metric techniques and
use the more traditional discrete traits for identification purposes (Rightmire, 1976). It was not
until the latter half of the 1900’s that the fields of statistics and anthropology saw a refinement of
these techniques, coinciding with the rise of computer technology.
Later Applications
Anthropologists found the application of PDA especially useful within a medico-legal
context, specifically when called upon to identify recent human skeletal remains. One of the first
anthropological studies that utilized predictive discriminant analysis at the beginning of the
computer era was that of Thieme and Schull (1957). They took seven measurements of the
postcranial skeletons of modern European American and African American specimens from the
Terry Collection. Discriminant functions were calculated and Thieme and Schull found that they
could correctly identify the sex of 98% of the specimens. Richman and colleagues (1979) tested
the method on different European American and African American specimens in the Terry
collection as well as on new specimens from the Howard University Medical School Collection.
They were able to correctly classify 91% of the sample, concluding that the Thieme and Schull
method is useful for sex determination. The original study paved the way for other
anthropologists to use predictive discriminant analysis as an applied technique.
Giles and Elliot (1962, 1963) developed discriminant functions for ancestry and sex
determination using modern European American and African American skeletons from the Terry
and Hamman-Todd Osteological Collections. They also included a Native American sample
14
from the Indian Knoll site in Illinois, dating to 3450 B.C. The authors used 8 measurements of
the cranium to produce formulae for male and female specimens. The only reason given for
selecting those variables was that they required simple measurement devices. Approximately
83% of the males and 88% of the females were classified correctly. The Giles and Elliot method
would be used for the next thirty years by anthropologists employing metric analyses of sex and
ancestry. However, in the years following the publication of the method, other researchers found
that the formulae did not produce the same results with populations other than those used in the
original study.
Birkby (1966) was the first to examine the effectiveness of the Giles and Elliot method.
He applied the formulae to a group of Native American and Labrador Eskimo crania. Birkby
found that the Native American crania were misclassified as European American and African
American 40% of the time. The female crania were misidentified 50% of the time. Birkby
concluded that the Indian Knoll sample used to represent Native Americans was not, in fact,
representative of the Native Americans at all, as shown by the low hit-rates. Giles (1966)
responded to Birkby’s study by commenting on the invalid nature of Birkby’s conclusions
stemming from his misunderstanding of discriminant analysis. Later tests of the Giles and Elliot
method provided evidence that supported Birkby’s conclusions. The method works well when
predicting sex but is not as accurate when predicting ancestry.
Snow and associates (1979) tested the Giles and Elliot method on a sample of forensic
cases from Oklahoma and surrounding areas. Sex was correctly identified for 88% of the crania,
not unlike the original study. However, only 71% of the crania were correctly identified for
ancestry. Specifically, only 14% of the Native American crania were correctly identified.
Researchers began to notice that the formulae were ineffective when applied to different Native
American groups. Fisher and Gill (1990) examined a series of Northwest Plains Native
15
Americans and found that the method correctly identified the ancestry of only 26% of the males
and 38% of the females. The sex of 100% of the males and 75% of the females was correctly
identified. Fisher and Gill remind readers that the one downfall of discriminant analysis is when
one applies the technique to populations not used in its construction. Ayers and colleagues
(1990) tested the method on a sample of forensic cases from the Forensic Data Bank at the
University of Tennessee. A total of 85% of European American males, 83% of European
American females, 48% of African American males, 90% of African American females, 11% of
Native American males, and 50% of Native American females were correctly classified. The
authors state that the sample used to test the method contributes to its inefficacy. Dissatisfied
with the results of the Giles and Elliot method, anthropologists began to develop their own
discriminant functions specific to their geographical location and tried to refine and test the
established methodology.
Using discriminant functions constructed from forensic cases from the New York region,
members of the Metropolitan Forensic Anthropology Team have correctly identified the sex of
82% of forensic cases (Taylor et al., 1984; Zugibe et al., 1985). Scholars outside of the United
States noticed that techniques developed on “white” and “black” populations did not work as
well on non-North American populations. Townsend and associates (1982) used a series of
Australian Aboriginal crania and obtained a success rate of 80%. At an extraordinary rate of
97% correct classification, Song and colleagues (1992) identified the sex of Chinese crania.
Other researchers have developed their own formulae for the identification of South African
“blacks” and “whites”, with an average correct identification of 86% (Steyn and İşcan, 1998;
İşcan and Steyn, 1999; Patriquin et al., 2002).
PDA and Large Databases. Most of the early work on predictive discriminant analysis
centered on discriminating between only two or three populations. As computer technology
16
advanced and more data could be analyzed at one time, anthropologists began to ask much larger
questions concerning relationships among populations from across the globe. Enormous
databases of osteometric data from thousands of crania and multiple groups were constructed and
analyzed. This work was pioneered by W.W. Howells in his landmark monograph Cranial
Variation in Man (1973), a culmination of decades of research in craniometrics. He took 57
measurements of crania from 17 world populations to discover the relationships between the
groups. For years to come, this study would serve as a model for all others, whether for
evolutionary studies or individual identification.
The use of PDA with large databases in anthropology, particularly within a forensic
context, was revolutionized by FORDISC, a personal computer forensic discriminant function
analysis program created by researchers at the University of Tennessee at Knoxville (UTK)
(Jantz and Ousley, 1993; Ousley and Jantz, 1996). This program, available for purchase, allows
the user to create personalized discriminant functions to aid in the identification of unknown
individuals. The functions are produced using the UTK Forensic Database, a collection of 11
populations comprised of ≈1200 individuals born after 1900. The ancestry and sex of most of
the individuals is known. Also included are data from 28 populations measured by Howells
(1973). Cranial and postcranial metric data from an individual specimen are entered into the
program and the user is able to select which ancestry/sex groups to compare their unknown to.
The output of the program includes probabilities of group affiliation. However, the analysis is
based on only 24 cranial measurements and 44 postcranial measurements, chosen because they
are among the easiest and quickest measurements to take. This ease of use is actually one of the
program’s downfalls in that it becomes attractive to those with little experience in skeletal
analysis. The program can produce fast results, which if not interpreted correctly can lead to
erroneous conclusions. Also, observer error of the magnitude of only ± 1 mm can create a
17
misclassification of the specimen (Zambrano et al., 2005).
Summary
Multiple predictive discriminant function analysis was created in the early 1900’s and has
since then been the primary metric method for the prediction of sex and ancestry from the human
skeleton. It has been shown that PDA can correctly identify specimens anywhere from 80-97%
of the time. Understanding the theoretical and mathematical foundations of PDA allows one to
more critically examine the results of a discriminant analysis. It is important to apply the LDF
only to populations that are similar to the ones on which they were constructed. Using a LDF on
a different population will result in low hit-rates. Researchers have developed functions that are
specific to a geographical location, producing much higher hit-rates. Over the past seventy years
scholars have refined the technique to become an invaluable tool for the biological
anthropologist when non-metric methods are not conclusive.
18
CHAPTER 3: STANDARD AND NON-STANDARD MEASUREMENTS
The measurements used in a predictive discriminant analysis are rarely accompanied by
an explanation as to why they were chosen. If an explanation is provided, it oftentimes identifies
the measurements as “standard”, “easily taken”, or “restricted because of time.” For the
purposes of this study, “standard” measurements are defined as the twenty-four cranial
measurements used in a FORDISC analysis, which require the use of conventional
instrumentation and are also included in osteological data collection manuals (Buikstra and
Ubelaker, 1994; Moore-Jansen et al., 1994; Bass, 1995). (See Appendix A, 1-24 for a list and
definition of the FORDISC measurements). “Non-standard” measurements are those not
included in the FORDISC analysis and are less frequently used because they require special
instruments to record (Howells, 1973).
In this chapter, a description of the major categories and types of measurements is
provided. Specific measurements and instrumentation as used in past and current research is
described. The discussion then turns to a survey of the literature to determine which
measurements are currently being used by researchers. Finally, I look at how researchers are
selecting the measurements for inclusion in predictive discriminant analyses and how many they
are choosing.
Types of Measurements
Hursh (1976) classifies cranial measurements into three categories: (1) box
measurements, (2) sutural measurements, and (3) extreme curvature measurements. Box
measurements are those that measure extremes of the skull. There are no specific, fixed points to
identify in order to take a box measurement. Maximum cranial breadth, “the maximum width of
19
the skull perpendicular to the mid-sagittal plane wherever it is located,” (Howells, 1973:172) is
an example of a box measurement. Sutural measurements are distances between two points, one
of which is defined by a suture or similar feature. An example of a sutural measurement is nasal
height, “the direct distance from nasion to nasospinale” (Howells, 1973:175). Extreme curvature
measurements are those based on regions of maximum change in curve of a surface. The frontal
subtense, “the maximum subtense, at the highest point on the convexity of the frontal bone in the
midplane, to the nasion-bregma chord,” (Howells, 1973:181) is one example of an extreme
curvature measurement. Of the three categories, Hursh states that extreme curvature
measurements are of the highest informational value because they can be used to predict the
shape of other parts of the cranium. He also suggests the use of coordinate points because more
information can be gathered from a single coordinate point than from other measurements.
There are numerous types of measurements that one can take on the human skull to
describe its size and shape, most of which fall into one of ten categories: lengths, widths,
breadths, heights, chords, arcs, subtenses, fractions, radii, and angles. The first six are
measurements between two landmarks (or other defined points) on the skull. Lengths are taken
from anterior to posterior, such as cranial base length, “the direct distance from nasion to
basion” (Howells, 1973:171). Breadths are taken from the left to right, such as minimum cranial
breadth, “the breadth across the sphenoid at the base of the temporal fossa, at the infratemporal
crests” (Howells, 1973:173) Heights are taken from superior to inferior (e.g., upper facial
height, “the direct distance from nasion to prosthion” (Howells, 1973:174)). The one width
measurement, mastoid width, is taken through the transverse axis of the base of the mastoid
process, wherever it may lie. Chords are a measure of direct distance and almost always refer to
measurements of the vault in which the shortest distance between two points on a curved surface
is recorded. An example is the occipital chord, “the direct distance from lambda to opisthion”
20
(Howells, 1973:182). An arc is a curved line or segment of a circle as in the bones of the vault
(e.g., frontal arc, the distance from nasion to bregma along the curvature of the vault). The
remaining four measurements are not measurements between two defined landmarks of the skull.
A subtense is a perpendicular measurement from the chord to the outer surface of bone,
describing how far the bone projects from a given plane (e.g., a magnitude of curvature). An
example of a subtense measurement is the parietal subtense, “the maximum subtense, at the
highest point on the convexity of the parietal bones in the midplane, to the bregma-lambda
chord” (Howells, 1973:182). A fraction is a measure of how far along the chord the subtense lies
such as the frontal fraction, “the distance along the nasion-bregma chord, recorded from nasion,
at which the frontal subtense falls” (Howells, 1973:181). Figure 3.1 shows how chords,
subtenses, fractions, and arcs are measured on a curved surface. All radial measurements are
taken from the transmeatal axis, a line that passes between the two external auditory meati. The
perpendicular distance from the transmeatal axis to a specified landmark is recorded as in the
zygomaxillare radius, “the perpendicular to the transmeatal axis from the left zygomaxillare
anterior” (Howells, 1973:184). Angles are not measurements taken directly from the skull;
rather, they are calculated from two measurements that share a common landmark. For example,
nasion angle is the angle at nasion whose sides are cranial base length and upper facial height.
___ ___ ___
FIGURE 3.1. Measurements of the Cranium. AC = chord, BD = subtense, AB = fraction,
and ADC = arc.
C B A
D
21
Specific Measurements and Instrumentation
Sliding and spreading calipers are the standard instruments found in most biological
anthropology laboratories. Special instruments are required to record non-standard
measurements such as subtenses, fractions, and radii. For subtenses and fractions one uses a
coordinate caliper or simometer, which is a modified sliding caliper with an extra (third) arm.
The simometer was first described in 1882 by de Mérejkowsky (Woo and Morant, 1934). The
extra arm moves up and down in a plane perpendicular to the other arms and is used to measure
subtenses. The third arm can also be moved anywhere horizontally between the two main arms.
A fraction is taken from the fixed arm of the caliper to the third arm. The only difference
between a simometer and a coordinate caliper is that the points of the two main arms face
towards each other in a simometer while those of the coordinate caliper come straight down.
Manufacturers sometimes combine both into one instrument (Figure 3.2a).
An even more specialized instrument is used to record radii. The radiometer (Figure
3.2b) has three arms just like a coordinate caliper, but only the third, middle arm is used to
record measurements. Two of the arms in the horizontal plane face each other and are inserted
into the external auditory meati to lock the radiometer in the transmeatal axis. At this point the
device is free to rotate 360° about the axis while the third arm in the vertical plane moves inward
to landmarks on the skull. This instrument will measure distances from the axis to the landmark.
One of the fundamental problems that osteologists face in metric studies lies in the
measurements they use. A researcher should not generate new measurements that others cannot
duplicate, and it has been suggested that osteologists should only include established
measurements (Pearson and Davin, 1924). Few researchers have specifically discussed the
benefits of using non-standard instrumentation and the information that can be collected with
them (Pearson, 1934; Woo and Morant, 1934; Howells, 1960, 1973; Rightmire, 1970, 1976; Gill
22
FIGURE 3.2. Non-Standard Instrumentation. PaleoTech™ PaleoCal-1 coordinate caliper /
simometer (a) and PaleoTech™ radiometer (b).
(a)
(b)
23
et al., 1988; Brues, 1990; Cunha and van Vark, 1991). The consensus among researchers who
have used non-standard measurements is that “unconventional measurements . . . are capable of
higher discriminatory power than the more traditional ones” (Jantz and Owsley, 1994:197).
However, as seen in the literature, most researchers remain loyal to standard instrumentation and
measurements.
Selection of Measurements
When selecting measurements for a predictive discriminant analysis, one must consider
three important factors (DiBennardo, 1986): (1) does a combination of measurements better
discriminate between two groups compared to a single measurement?; (2) what is the construct
of the measurements (the theoretical foundation by which the measurements separate the
groups)?; and (3) how well will the LDF created from the measurements perform?
Unfortunately, these factors are not always addressed and the justification of selection criteria for
particular measurements is frequently left unstated.
Studies in craniometrics fall into one of three categories with regards to measurement
selection. The first category is that no explanation is provided. Cunha and van Vark (1991)
examined a series of crania from turn of the 20th
century Portuguese crania. They took 61
measurements as defined by Howells (1973) and were able to correctly identify the sex of 80%
of the individuals. In their study of Chinese crania, Song and colleagues (1992) chose a mixture
of 38 standard and non-standard measurements and produced a hit-rate of 97%. Steyn and İşcan
(1998) selected 12 “standard” measurements in their study of identifying the sex of South
African individuals and were able to do so with a success rate of 86%. Although the hit-rates of
the aforementioned studies are relatively high, the authors give no exact reason why the variables
they used were selected.
24
The second category is that measurements are chosen because they are easy to record.
Giles and Elliot (1962, 1963) chose the eight standard variables to measure in developing their
methodology because of “ease of recording.” In similar fashion, Wright (1992) reduced Howells
set to 29 variables that are the easiest to measure. He used these measurements to develop a
computer program called CRANID, which assigns an unknown skull to an ancestral group, not
unlike FORDISC.
The third category of craniometric studies are those in which measurements are selected
to describe a certain region of the cranium. When Rightmire (1976) examined the skulls of
Bantu-speaking African groups, he selected 37 measurements, some standard and others
specifically designed to measure certain features of the midface, vault, and brow. In an
assessment of craniometric relationships between Plains Indian groups within a cultural-
evolutionary framework, Key (1983) recorded 65 measurements. Most of these came from
Howells’ (1973) set, but 9 were created by the author “to more fully measure particular
morphological complexes” (Key, 1983:40). In his studies of cranial variation, Gill (1984)
noticed that the greatest difference between Northwest Plains Indians and other groups existed in
the nasal bridge. He took six unconventional measurements of the nasal region to develop a
method that is used by many today (Gill, 1995; Gill et al., 1988; Gill and Gilbert, 1990; Curran,
1990). Ross and associates (2004) selected landmarks “that would reveal the overall cranial
morphology of the crania” to help identify Cuban American skulls in forensic contexts.
Number of Measurements. In the early years of PDA, researchers were limited in the
number of response variables they could use because all of the calculations had to be done by
hand, taking a very long time to complete. Rao (1949) stated that, with regards to the number of
variables used, “It does not seem to be, always, the more the better. . .” He noticed that the
discriminatory power leveled off with the addition of more variables and then began to drop
25
when even more were added. Bronowski and Long (1952) agreed with Rao and noted that
additional information from more measurements becomes insignificant after a certain point
because the measurements become more correlated with one another. Using a computer
simulation, Dunn and Varady (1966) found that for a fixed number of variables, as the sample
size used to create the linear discriminant functions increases, so does the probability of correct
classification of the objects. For a fixed sample size, the probability of correct classification
decreases as the number of variables increase from two to ten. The phenomenon of decreased
classificatory power with an increase in the number of response variables has become known as
“Rao’s paradox” (Kowalski, 1972) as noted by other researchers (van Vark, 1976; Johnson et al.,
1989). Others describe an “optimum measurement of complexity,” a phrase used to explain how
the combination of the nature and the number of variables affects discriminant functions (van
Vark and van der Sman, 1982; van Vark and Schaafsma, 1992). The nature of the variables
refers to the shape and size of the cranium that the measurements represent. Although they
discuss how the number of variables and the sample size balance out to reach an optimum level
of discrimination, the authors do not give any exact numbers.
Summary
It is essential that one be aware of the types of measurements of the cranium that can be
taken in order to focus and improve a study. One is only limited by the equipment available and
by the time available to complete the study. Despite inconsistencies between researchers
regarding how standard and non-standard measurements and instruments are selected, there does
appear to be some general agreement on the use of specific types of measurements. Certain
measurements can only be used for certain populations (Howells, 1957) and the variables that are
eventually selected are determined by the questions that the researcher is asking (i.e., ancestry or
26
sex identification) (Pietrusewsky, 2000). Research has shown that the use of non-standard
measurements allows the biological anthropologist to better answer specific questions with
regards to craniometric ancestry determination.
27
CHAPTER 4: MATERIALS AND METHODS
The Study Sample
This study examined three different ancestral groups from different time periods: recent
African Americans (AA), recent European Americans (EA), and prehistoric Coyotero Apache
(CA). The Coyotero Apache crania were drawn from the Indiana University Anthropology
Department Osteological Collection in Bloomington, IN. The collection includes ~6,000 human
skeletons from numerous prehistoric and historic archaeological sites in the United States. The
series of crania used in this study are from the Edward Palmer Arkansaw Mounds located just
outside of Little Rock, AR, comprised of ~100 crania. It is a Late Woodland to Mississippian
site dating from A.D. 700-950 (Jeter, 1990). Nineteen males and thirty six females were selected
from the collection. These crania were the most complete specimens available.
The African American and European American samples were drawn from the Hamman-
Todd Osteological Collection located at the Cleveland Museum of Natural History (CMNH) in
Cleveland, OH. This collection has one of the largest primate skeletal samples in the world.
Specifically, the collection contains 3,100 modern human skeletons from unclaimed bodies at the
Cuyahoga County Morgue and city hospitals. They date from the late 1800s and early 1900’s
and many have known dates of birth, dates of death, age, weight, height, cause of death, sex, and
ancestry. From the AA subsample, 25 males and 24 females were selected, with ages ranging
from 21 to 46 years old and a mean age of 28.9 years. Twenty six males and twenty five females
were selected from the EA subsample, with ages ranging from 19 to 48 years old and a mean age
of 33.7 years. A total of 100 crania were measured from the CMNH. Table 4.1 shows the
distribution of the sample by collection, age, and sex.
28
TABLE 4.1. Distribution of Individuals by Collection, Ancestral Group, Sex, and Age.
Collection IU CMNH
Group Coyotero Apache African American European American
Sex M F M F M F
Count 19 36 25 24 26 25
Min. Age NA 21 19
Max. Age NA 46 48
Mean Age NA 28.9 33.7
Only complete, adult crania with no major deformities were included in this study. Sex,
ancestry, date of birth, date of death, and age at death was recorded from the CMNH records. If
age at death was not known, adult status was determined by the presence of an erupted third
molar. If the sex of a cranium was not known, it was determined by the recorder at the time of
analysis based on gross morphological features, including size of mastoid process, supraorbital
margin form, nuchal crests, and overall cranial robusticity. It should be noted that none of the
Apache specimens had known ages at death or sex, although sex was determined by the author.
Taking the Measurements and Instrumentation
A battery of 65 measurements was taken on each cranium. These measurements are
listed on the “Craniometrics Recording Form” in Appendix A, and their definitions are given in
Appendix B. All measurements were recorded to the nearest millimeter and on the left side of
the skull in the case of bilateral measurements. If the left side was damaged to the degree that a
measurement could not be taken, then the right side was used and a note was made on the form
with the letter “R”. If an area was damaged or resorbed by more than 2 mm, then the
measurement was taken and recorded on the form in parentheses. If more than 2 mm of bone
was missing, the measurement was not taken at all.
29
The protocol for data collection was developed at the University of Indianapolis
Archeology and Forensics Laboratory (AFL) using modern human crania and at the Glenn A.
Black Laboratory of Archaeology at Indiana University using prehistoric crania as test subjects.
These crania were used to refine measurement techniques, including locating landmarks, caliper
point placement, and the order in which the measurements were taken. Dr. Stephen Nawrocki
assisted the author in clarifying measurement definitions.
The measurements were derived from three sources: (1) FORDISC 2.0 (FD2) (Ousley
and Jantz, 1996), (2) Howells (1973), and (3) Gill (1984). The 24 FD2 measurements are those
used in a standard FD2 analysis of the cranium. The Howells subset includes 57 measurements,
20 of which are also FD2 measurements. Even though 20 measurements are identical in the two
subsets, they were examined separately as part of each measurement group during the statistical
analysis. Howells also used trigonometry to calculate 13 angles from his 57 measurements. In
this study, however, the angles were omitted. Although angles help to describe more complex
features of the cranium, a computer-aided multivariate analysis will elucidate which variables
describe the complex features without having to perform a calculation prior to data entry
(Corruccini, 1975; Campbell, 1978). The third subset of measurements was used by Gill and
describes the shape of the nasal area. Gill used a simometer for the measurements in his study.
Two of the six measurements are the same as the FD2 and Howells sets.
A Paleo-Tech™ spreading caliper, Mitutoyo™ sliding dial caliper, Paleo-Tech™
PaleoCal-1 coordinate caliper, and a Paleo-Tech™ radiometer were used to take all
measurements of the crania from IU. These instruments were borrowed from the AFL. At the
CMNH, a Mitutoyo™ direct input tool and foot pedal were used in conjunction with a
Mitutoyo™ sliding digital caliper. This caliper was connected directly to a laptop computer so
the measurements could be input directly into a Microsoft™ Office Excel 2003 spreadsheet.
30
Appendix B identifies which instrument was used for each of the measurements.
Statistical Analysis
Three assumptions were made about the data prior to analysis. First, all of the crania
were assumed to be correctly classified as to ancestry as recorded in the collections’ records.
Second, the data is assumed to follow a multivariate normal distribution. Third, the distributions
are uncontaminated. Contamination can take be expressed in two ways: scale contamination
and location contamination. Scale contamination occurs when the instruments vary more than
usual and give erroneous values. Location contamination occurs when instruments slip from the
landmark (Lachenbruch and Goldstein, 1979). To test the data for multivariate normality, a
Box’s M test was performed.
Four sets of discriminant function formulae were constructed with a forward stepwise
method using the statistics computer software package SPSS 13.0. Predictive discriminant
analysis requires that all specimens have all measurements or else they will not be included in
the calculation of the functions. Measurements that could not be taken were substituted with the
‘linear trend at point’ function of SPSS. This function replaces missing values with the linear
regression estimate for that point. The existing series is regressed on an index variable scaled 1
to n. Missing values are replaced with their predicted values. Only 12 (0.001%) of the 10,075
measurements taken were missing and subsequently replaced.
The stepwise selection method followed İşcan and Steyn’s (1999) data entry
methodology, with p values of p = 0.05 to enter and p = 0.15 to remove. This seemingly liberal
cutoff value does not increase the overall error rate (MacLachlan, 1980). Forward stepwise PDA
involves adding a variable that meets the required p-value for the function. The function is
recalculated using the remaining variables and if a new variable meets the cutoff, it is added to
31
the function. This process is repeated until only the statistically significant variables are
included in the final formulae. The functions are built step-by-step in a forward, additive
manner, starting with zero measurements. Stepwise analysis is a useful tool to explore which
variables are helpful in classifying objects (Snapinn and Knoke, 1989).
The crania were divided into three different ancestry groups: (1) African American, (2)
European American, and (3) Coyotero Apache. The data were entered four times to produce
four sets of discriminant function formulae for a total of eight functions. The first set of
functions used the 24 standard measurements of the cranium included in a FD2 analysis. The
second set used the 57 Howells measurements. The third set of formulae used Gill’s 6
measurements of the nasal region. The fourth and final set combined all 65 measurements.
The weights of the variables for each function were calculated. Eigenvalues and Wilk’s λ
values were calculated to test the significance of the discriminant functions. Structure matrices
were constructed to determine how heavily the variables loaded on the functions.
The data were entered into the program with unequal prior probabilities. The prior
probability is the chance that a cranium is selected and correctly randomly assigned to its group.
As there are three groups in this study, the probability of correctly assigning a cranium to its
group by chance alone is 0.333. However, because the sizes of the groups in this study are not
equal, the prior probability must be weighted by the sample size. Therefore, the prior probability
for the African American group is 0.316, the European American group is 0.329, and the
Coyotero Apache group is 0.355. If the power of the discriminant functions is greater than the
prior probability, then it is successful at discriminating between groups.
The four sets of discriminant function formulae were compared side by side to determine
which set was better able to correctly separate groups and sort the individuals. The ability of the
discriminant functions to correctly classify individuals was assessed by testing the functions on
32
the original sample via simple resubstitution as well as a leave-one-out method (LOO).
Resubstituting the entire original sample directly into the discriminant functions will generally
overestimate the ability of the formulae to separate these groups in the population as a whole
(Lachenbruch, 1967; Lachenbruch and Goldstein, 1979). Resubstitution rates will therefore be
used only as a general baseline measure of the formulae’s performance. To avoid the problem of
“statistical incest” and properly test the functions, one must use cases that were not included in
the original study sample (Smith, 1947; Lachenbruch and Mickey, 1968). Unfortunately, this
requires an entirely new dataset. LOO more accurately tests the functions without having to use
a new dataset (van Vark, 1976; Feldesman, 1997). The computer removes one case from the
sample and creates a discriminant function using the remaining cases. The case that was
removed is then entered back into the function to see if will be properly classified into its
original group. The computer then repeats this process for each of the cases in the sample. The
LOO method ensures that the cases used to produce the functions are not used when testing them
for classificatory ability. The success of the discriminant functions is expressed as a percentage
of cases in the original sample that are correctly classified into their groups.
33
CHAPTER 5: RESULTS
Stepwise Statistics
The resulting forward stepwise analyses included 12 of the 24 FD2 variables, 12 of the 57
Howells variables, 4 of the 6 Gill variables, and 15 of the 65 All Measurements variables (Table
5.1). The Howells, Gill, and All Measurements sets each included a mixture of standard and
non-standard measurements. All of the selected FD2 variables were standard measurements,
while 8 of the 12 Howells variables, 3 of the 4 Gill variables, and 8 of the 15 All Measurements
variables were non-standard measurements.
TABLE 5.1. Variables Included by Stepwise Analysis. Variables are listed in the order they
were entered into the functions; the * indicates a non-standard measurement. See Appendix A
for measurement abbreviations.
Measurement Subset
Step FD2 Howells Gill All
1 PAC FRS* ZOB* FRS*
2 NLB ZMB* DKB ZOB*
3 BBH PRR* ALC* PAC
4 MAL ZOR* NZS* STB*
5 OBH PAC ZMB*
6 XCB STB* FMB*
7 GOL FMB* FOB
8 AUB BNL OBH
9 WFB OBH AVR*
10 BPL ASB* NZS*
11 FOB NLB NLB
12 ZYB SSR* ASB*
13 ZYB
14 OCF
15 BBH
34
Linear Discriminant Functions
The discriminant functions produced by the stepwise analysis for each of the
measurement sets are given in Equations 5.1, 5.2, 5.3, 5.4, 5.5, 5.6, 5.7, and 5.8. Because three
groups were included in this study, only two discriminant functions were produced for each
measurement set. The coefficients are multiplied by the value of a measurement, all products are
summed, and then a constant is added to give a discriminant function score. The scores from
Functions 1 and 2 from each measurement set were then used to create scatterplots of the crania
in discriminant space (Figures 5.1, 5.2, 5.3, and 5.4). The closer the specimens of a group are
clustered together on the scatterplot, the smaller the amount of variation there is in that group. If
there is no overlap between the group clusters then the discriminant functions are able to
completely separate the groups. The more overlap that exists between groups and the greater the
spread of the clusters, the less successful the functions are at classifying the specimens.
In order to use the equations and scatterplots to determine the ancestry of an unknown
cranium, one must first select which measurement set to use. If the measurements included in an
FD2 analysis are available to be recorded, one could use Equations 5.1 and 5.2. The values of
the measurements are plugged in to the equations and solved for Function 1 and Function 2. The
two resulting scores are then plotted on the FD2 scatterplot (Figure 5.1). The Function 1 score
corresponds to the x axis value and the Function 2 score corresponds to the y axis value. The
closer the cranium is to a group centroid, the more likely it belongs to that group.
EQUATION 5.1. FD2 Function 1 = -0.089GOL – 0.073XCB + 0.067ZYB + 0.081BBH +
0.042BPL – 0.088MAL + 0.084AUB – 0.094WFB + 0.190NLB + 0.130OBH –
0.044PAC + 0.018FOB + 1.335
35
EQUATION 5.2. FD2 Function 2 = 0.004GOL – 0.078XCB + 0.031ZYB – 0.033BBH +
0.069BPL + 0.070MAL – 0.049AUB – 0.025WFB + 0.182NLB + 0.253OBH +
0.040PAC – 0.140FOB – 5.223
EQUATION 5.3. Howells Function 1 = -0.100BNL – 0.140NLB – 0.121OBH + 0.034PAC –
0.028STB + 0.051ASB – 0.188ZMB + 0.165FMB + 0.156FRS – 0.054SSR + 0.012PRR
+ 0.142ZOR + 1.514
EQUATION 5.4. Howells Function 2 = -0.072BNL + 0.090NLB + 0.186OBH + 0.057PAC –
0.069STB – 0.077ASB – 0.025ZMB + 0.101FMB + 0.030FRS – 0.129SSR + 0.264PRR
– 0.065ZOR – 9.265
EQUATION 5.5. Gill Function 1 = -0.217DKB + 0.186ZOB – 0.162NZS + 0.146ALC – 6.500
EQUATION 5.6. Gill Function 2 = 0.375DKB + 0.070ZOB + 0.023NZS – 0.159ALC – 7.498
EQUATION 5.7. All Measurements Function 1 = -0.148ZOB + 0.032NZS – 0.066ZYB –
0.044BBH – 0.131NLB – 0.148OBH + 0.044PAC – 0.142ZMB + 0.189FMB +
0.183FRS + 0.016OCF – 0.006FOB - 0.013STB + 0.051ASB + 0.050AVR + 1.232
EQUATION 5.8. All Measurements Function 2 = 0.110ZOB – 0.104NZS + 0.021ZYB –
0.027BBH + 0.016NLB + 0.225OBH + 0.067PAC – 0.043ZMB + 0.040FMB +
0.036FRS + 0.03OCF9 – 0.173FOB – 0.114STB – 0.060ASB + 0.097AVR – 4.212
36
6420-2-4-6
Function 1
6
4
2
0
-2
-4
Fu
ncti
on
2
3
2
1
Group Centroid
Coyotero Apache
European American
African American
Group
FIGURE 5.1. Canonical Discriminant Functions: FD2 Measurements.
37
420-2-4-6
Function 1
4
2
0
-2
-4
Fu
ncti
on
2
3
2
1
Group Centroid
Coyotero Apache
European American
African American
Group
FIGURE 5.2. Canonical Discriminant Functions: Howells Measurements.
38
43210-1-2-3
Function 1
3
2
1
0
-1
-2
-3
Fu
ncti
on
2
32
1
Group Centroid
Coyotero Apache
European American
African American
Group
FIGURE 5.3. Canonical Discriminant Functions: Gill Measurements.
39
6420-2-4-6
Function 1
4
2
0
-2
-4
Fu
ncti
on
2
3
2
1
Group Centroid
Coyotero Apache
European American
African American
Group
FIGURE 5.4. Canonical Discriminant Functions: All Measurements.
40
Discriminant Function Coefficients. The standardized discriminant function
coefficients for each of the measurement sets were calculated. They indicate the contribution of
a variable, also known as the variable weight, to the discriminant function. The higher the
absolute value of the coefficient, the greater the contribution of the variable to the function. The
coefficients do not take into account any shared contribution amongst variables, rather, only the
unique contribution of each variable is considered.
The four most heavily weighted variables for the FD2 set are maximum cranial length
(GOL), minimum frontal breadth (WFB), biauricular breadth (AUB), and basion-bregma height
(BBH) for Function 1, and orbital height (OBH), maximum cranial breadth (XCB), basion-
prosthion length (BPL), and nasal breadth (NLB) for Function 2. All standardized discriminant
function coefficients for the FD2 set are given in Table 5.2. The four most heavily weighted
variables for the Howells set are bimaxillary breadth (ZMB), bifrontal breadth (FMB),
zygoorbitale radius (ZOR), and frontal subtense (FRS) for Function 1, and prosthion radius
(PRR), subspinale radius (SRR), bifrontal breadth (FMB), and bistephanic breadth (STB) for
Function 2 (Table 5.3). The four most heavily weighted variables for the Gill set are mid-orbital
breadth (ZOB), maxillofrontal breadth (DKB), alpha chord (ALC), and naso-zygoorbital
subtense (NZS) for Function 1, and interorbital breadth (DKB), alpha chord (ALC), mid-orbital
breadth (ZOB), and naso-zygoorbital subtense (NZS) for Function 2. All standardized
discriminant function coefficients for the Gill set are given in Table 5.4. The four most heavily
weighted variables for the All Measurements set are bifrontal breadth (FMB), bimaxillary
breadth (ZMB), frontal subtense (FRS), and bizygomatic breadth (ZYB) for Function 1, and
bistephanic breadth (STB), mid-orbital breadth (ZOB), parietal chord (PAC), and orbital height
(OBH) for Function 2 (Table 5.4).
41
TABLE 5.2. Standardized Discriminant Function Coefficients: FD2 Set. The bold values
are the four most heavily weighted variables.
Variable Function 1 Function 2
GOL -0.685 0.030
XCB -0.427 -0.456
ZYB 0.432 0.202
BBH 0.456 -0.184
BPL 0.266 0.437
MAL -0.287 0.230
AUB 0.467 -0.272
WFB -0.471 -0.124
NLB 0.378 0.363
OBH 0.268 0.523
PAC -0.322 0.291
FOB 0.047 -0.357
TABLE 5.3. Standardized Discriminant Function Coefficients: Howells Set. The bold
values are the four most heavily weighted variables. The * indicates a non-standard
measurement.
Variable Function 1 Function 2
BNL -0.456 -0.327
NLB -0.278 0.180
OBH -0.251 0.384
PAC 0.250 0.417
STB* -0.177 -0.440
ASB* 0.258 -0.391
ZMB* -0.906 -0.120
FMB* 0.740 0.452
FRS* 0.508 0.097
SSR* -0.264 -0.628
PRR* 0.059 1.319
ZOR* 0.608 -0.280
42
TABLE 5.4. Standardized Discriminant Function Coefficients: Gill Set. The * indicates a
non-standard measurement.
Variable Function 1 Function 2
DKB -0.551 0.953
ZOB* 0.901 0.340
NZS* -0.447 0.064
ALC* 0.526 -0.573
TABLE 5.5. Standardized Discriminant Function Coefficients: All Measurements Set. The bold values are the four most heavily weighted variables. The * indicates a non-standard
measurement.
Variable Function 1 Function 2
ZOB* -0.231 0.534
NZS* 0.087 -0.286
ZYB -0.424 0.136
BBH -0.248 -0.154
NLB -0.261 0.033
OBH -0.307 0.465
PAC 0.320 0.489
ZMB* -0.681 -0.207
FMB* 0.844 0.180
FRS* 0.597 0.117
OCF 0.131 0.313
FOB -0.016 -0.443
STB* -0.081 -0.727
ASB* 0.257 -0.302
AVR* 0.221 0.427
Test for Multivariate Normality. A Box’s test of equality of covariance was run to test
the homogeneity of the covariance matrices, which are used to construct the discriminant
functions. Homogeneous matrices verify the assumption of multivariate normality. The FD2
and All Measurements functions are significant at the p = 0.05 level or better indicating that the
43
covariance matrices do differ, hence, the sample is not multivariate normal. However, a
predictive discriminant analysis can still be robust when the assumption of homogeneous
covariance is not met (Huberty, 1994). The significance values for the Howells and Gill
functions do not reach the p = 0.05 level, indicating that the covariance matrices do not differ
and therefore, the samples are multivariate normal. Table 5.6 presents the Box’s M values and
levels of significance.
TABLE 5.6. Box’s Test of Equality of Covariance.
Measurement Set Box's M F df1 df2 p
FD2 299.391 1.3705 156 60544.955 0.006
Howells 206.898 1.178 156 6054.955 0.064
Gill 30.768 1.48 20 81443.852 0.077
All 361.841 1.297 240 59776.961 0.001
Significance of Functions. Eigenvalues for each function were also calculated. The
eigenvalue reflects the relative importance of the functions that are used to predict the ancestry
of a given cranium. The eigenvalue is the root of the vector of the covariance matrix. If there is
more than one discriminant function, the first eigenvalue will be the largest and most important,
and the second will be the next most important in explanatory power, and so on. The
eigenvalues assess relative importance because they reflect the percentage of variance explained
in the dependent variables, cumulating to 100% across all functions. A higher eigenvalue
indicates a more important function. The percentage of variance explained by Function 1 across
all measurement sets ranged from 66.4 to 92.5. The percentage of variance explained by
Function 2 across all measurement sets ranged from 7.5 to 33.6.
Wilk’s λ values were calculated to test the significance of the discriminant function as a
44
whole. The smaller the Wilk’s λ value, the more significant the function is. All discriminant
functions are significant at the 0.05 level. The eigenvalues of the discriminant functions along
with the percent of variance explained, Wilk’s λ values, and levels of significance are given in
Table 5.7.
TABLE 5.7. Variance Explained by the Functions and Tests for Significance.
Measurement
Set Function Eigenvalue
% of
variance Wilk's λ
Chi-
Squared df p
FD2 1 2.550 66.5 0.123 306.570 24 0.000
2 1.283 33.5 0.438 120.954 11 0.000
Howells 1 3.909 71.5 0.080 370.774 24 0.000
2 1.559 28.5 0.391 137.680 11 0.000
Gill 1 0.688 92.5 0.561 86.953 8 0.000
2 0.056 7.5 0.947 8.1995 3 0.042
All 1 4.027 66.4 0.066 395.195 30 0.000
2 2.036 33.6 0.329 161.047 14 0.000
Structure Matrices. Pooled within-groups correlations between discriminating variables
and standardized discriminant functions were calculated. The variables are ordered by absolute
size of correlation within a function. The correlations are known as structure coefficients or
discriminant loadings. One can describe a function by considering the set of variables that loads
most heavily within it, such as vault shape or prognathism.
In the FD2 set the variables most heavily loaded for Function 1 are parietal chord (PAC),
cranial length (GOL), bizygomatic breadth (ZYB), basion-bregma height (BBH), minimum
frontal breadth (WFB), and biauricular breadth (AUB). These variables generally measure facial
width and vault length. The variables most heavily loaded for Function 2 are maximum alveolar
length (MAL), basion-prosthion length (BPL), nasal breadth (NLB), orbital height (OBH),
45
maximum cranial breadth (XCB), and foramen magnum breadth (FOB). These variables
measure maxillary prognathism and facial size. Structure coefficients for the FD2 set are
presented in Table 5.8.
In the Howells set the variables most heavily loaded for Function 1 are frontal subtense
(FRS), bimaxillary breadth (ZMB), parietal chord (PAC), zygoorbitale radius (ZOR), and
basion-nasion length (BNL). These variables generally measure cranial length. The variables
most heavily loaded for Function 2 are prosthion radius (PRR), bistephanic breadth (STB), nasal
breadth (NLB), orbital height (OBH), subspinale radius (SSR), bifrontal breadth (FMB), and
biasterionic breadth (ASB). These variables generally measure prognathism and cranial width.
Table 5.9 gives the structure coefficients for the Howells set.
In the Gill set the variables most heavily loaded for Function 1 are mid-orbital breadth
(ZOB), alpha chord (ALC), and naso-zygoorbital subtense (NZS). These three variables are
measures of nasal region width and projection. The variable most heavily loaded for Function 2
is maxillofrontal breadth (DKB). Structure coefficients for the Gill set are presented in Table
5.10.
In the All Measurements set the variables most heavily loaded for Function 1 are frontal
subtense (FRS), bimaxillary breadth (ZMB), parietal chord (PAC), bizygomatic breadth (ZYB),
basion-bregma height (BBH), and naso-zygoorbital subtense (NZS). These variables generally
measure lower facial width and vault shape. The variables most heavily loaded for Function 2
are molar alveolar radius (AVR), mid-orbital breadth (ZOB), bistephanic breadth (STB), nasal
breadth (NLB), orbital breadth (OBH), foramen magnum breadth (FOB), bifrontal breadth
(FMB), biasterionic breadth (ASB), and occipital fraction (OCF). These variables generally
measure upper facial width. Table 5.11 gives the structure coefficients for the All Measurements
set.
46
TABLE 5.8. Structure Matrix: FD2 Set. The bold values indicate the largest absolute
correlation between each variable and any discriminant function.
Variable Function 1 Function 2
PAC -0.441 0.273
GOL -0.377 0.258
ZYB 0.238 0.074
BBH 0.232 -0.087
WFB -0.216 -0.033
AUB 0.189 -0.125
MAL -0.071 0.492
BPL 0.098 0.409
NLB 0.268 0.378
OBH 0.175 0.355
XCB 0.002 -0.286
FOB 0.156 -0.277
TABLE 5.9. Structure Matrix: Howells Set. The bold values indicate the largest absolute
correlation between each variable and any discriminant function. The * indicates a non-standard
measurement.
Variable Function 1 Function 2
FRS* 0.527 0.020
ZMB* -0.388 0.198
PAC 0.331 0.324
ZOR* 0.204 0.178
BNL -0.156 0.037
PRR* -0.005 0.484
STB* 0.213 -0.325
NLB -0.244 0.292
OBH -0.168 0.288
SSR* -0.017 0.274
FMB* 0.005 0.244
ASB* 0.085 -0.169
47
TABLE 5.10. Structure Matrix: Gill Set. The bold values indicate the largest absolute
correlation between each variable and any discriminant function. The * indicates a non-standard
measurement.
Variable Function 1 Function 2
ZOB* 0.747 0.573
ALC* 0.504 0.014
NZS* -0.200 0.131
DKB 0.050 -0.253
TABLE 5.11. Structure Matrix: All Measurements Set. The bold values indicate the largest
absolute correlation between each variable and any discriminant function. The * indicates a non-
standard measurement.
Variable Function 1 Function 2
FRS* 0.519 0.004
ZMB* -0.380 0.184
PAC 0.330 0.275
ZYB -0.193 0.026
BBH -0.177 -0.100
NZS* 0.075 -0.053
AVR* 0.121 0.338
ZOB* -0.206 0.337
STB* 0.206 -0.290
NLB -0.237 0.261
OBH -0.162 0.256
FOB -0.104 -0.239
FMB* 0.008 0.213
ASB* 0.082 -0.150
OCF* -0.039 0.052
Classification Accuracy
The classification accuracy of each set of discriminant functions was calculated using two
internal analyses: resubstitution and the Leave-One-Out (LOO) method. The FD2 discriminant
48
functions correctly classified 85.7% of the African American (AA) specimens, 88.2% of the
European American (EA) specimens, and 96.4% of the Coyotero Apache (CA) specimens via
resubstitution. AA crania were sometimes misclassified as EA, EA were misclassified as AA,
and CA were misclassified equally among the AA and EA. The LOO method for the FD2
functions correctly classified 81.6% of the AA crania, 86.3% of the EA crania, and 89.1% of the
CA crania. AA crania were sometimes misclassified as EA, EA were misclassified as AA, and
CA were misclassified as EA. All counts and percentages for the FD2 functions are presented in
Table 5.12.
The Howells discriminant functions correctly classified 91.8% of the AA specimens,
90.2% of the EA specimens, and 96.4% of the CA specimens via resubstitution. The LOO
method for the Howells functions correctly classified 87.8% of the AA crania, 90.2% of the EA
crania, and 92.7% of the CA crania. AA crania were sometimes misclassified as EA, EA were
misclassified as AA, and CA were misclassified equally among the AA and EA for both
resubstitution and LOO. All counts and percentages for the Howells functions are presented in
Table 5.13.
The Gill discriminant functions correctly classified 42.9% of the AA specimens, 76.5%
of the EA specimens, and 69.1% of the CA specimens via resubstitution. The LOO method for
the Gill functions correctly classified 42.9% of the AA crania, 76.5% of the EA crania, and
65.5% of the CA crania. AA crania were sometimes misclassified as CA, EA were misclassified
as AA, and the CA crania were misclassified as AA for both resubstitution and LOO. All counts
and percentages for the Gill functions are presented in Table 5.14.
The All Measurements discriminant functions correctly classified 91.8% of the AA
specimens, 96.1% of the EA specimens, and 98.2% of the CA specimens via resubstitution. AA
crania were sometimes misclassified as EA, EA were misclassified as AA, and CA were
49
misclassified as EA. The LOO method for the All Measurements functions correctly classified
89.8% of the AA crania, 94.1% of the EA crania, and 96.4% of the CA crania. AA crania were
sometimes misclassified as EA, EA were misclassified as AA, and CA crania were misclassified
equally among the AA and EA. All counts and percentages for the All Measurements functions
are presented in Table 5.15.
The average correct classification for each of the measurement sets was also calculated
(Table 5.16). The FD2 functions correctly classified 90.3% of the crania via resubstitution and
85.8% via LOO. The Howells functions correctly classified 92.9% of the crania via
resubstitution and 90.3% via LOO. The Gill functions correctly classified 63.2% of the crania
via resubstitution and 61.9% via LOO. The All Measurements functions correctly classified
95.5% of the crania via resubstitution and 93.5% via LOO.
TABLE 5.12. Classification Accuracy for FD2 Discriminant Functions. The bold values
indicate cases correctly classified.
Predicted Group Membership
Total Group AA EA CA
Resubstitution Count AA 42 4 3 49
EA 5 45 1 51
CA 1 1 53 55
% AA 85.7 8.2 6.1 100.0
EA 9.8 88.2 2.0 100.0
CA 1.8 1.8 96.4 100.0
LOO Count AA 40 6 3 49
EA 5 44 2 51
CA 1 5 49 55
% AA 81.6 12.2 6.1 100.0
EA 9.8 86.3 3.9 100.0
CA 1.8 9.1 89.1 100.0
50
TABLE 5.13. Classification Accuracy for Howells Discriminant Functions. The bold values
indicate cases correctly classified.
Predicted Group Membership
Total Group AA EA CA
Resubstitution Count AA 45 3 1 49
EA 4 46 1 51
CA 1 1 53 55
% AA 91.8 6.1 2.0 100.0
EA 7.8 90.2 2.0 100.0
CA 1.8 1.8 96.4 100.0
LOO Count AA 43 5 1 49
EA 4 46 1 51
CA 2 2 51 55
% AA 87.8 10.2 2.0 100.0
EA 7.8 90.2 2.0 100.0
CA 3.6 3.6 92.7 100.0
TABLE 5.14. Classification Accuracy for Gill Discriminant Functions. The bold values
indicate cases correctly classified.
Predicted Group Membership
Total Group AA EA CA
Resubstitution Count AA 21 13 15 49
EA 8 39 4 51
CA 13 4 38 55
% AA 42.9 26.5 30.6 100.0
EA 15.7 76.5 7.8 100.0
CA 23.6 7.3 69.1 100.0
LOO Count AA 21 13 15 49
EA 8 39 4 51
CA 13 6 36 55
% AA 42.9 26.5 30.6 100.0
EA 15.7 76.5 7.8 100.0
CA 23.6 10.9 65.5 100.0
51
TABLE 5.15. Classification Accuracy for All Measurements Discriminant Functions. The
bold values indicate cases correctly classified.
Predicted Group Membership
Total Group AA EA CA
Resubstitution Count AA 45 4 0 49
EA 2 49 0 51
CA 0 1 54 55
% AA 91.8 8.2 0.0 100.0
EA 3.9 96.1 0.0 100.0
CA 0.0 1.8 98.2 100.0
LOO Count AA 44 4 1 49
EA 2 48 1 51
CA 1 1 53 55
% AA 89.8 8.2 2.0 100.0
EA 3.9 94.1 2.0 100.0
CA 1.8 1.8 96.4 100.0
TABLE 5.16. Average Classification Accuracy of Measurement Sets.
Measurement Set Resubstitution LOO
FD2 90.3% 85.8%
Howells 92.9% 90.3%
Gill 63.2% 61.9%
All Measurements 95.5% 93.5%
Casewise Statistics
Appendix C provides the casewise statistics for each cranium for each measurement set
following the LOO method of classification. This data includes actual group membership,
predicted group membership, posterior probabilities, typicality probabilities, D2 distances from
the group centroid, and discriminant function scores. Posterior probabilities indicate how likely
52
it is that a cranium belongs to the group that it has been assigned to. Typicality probabilities
indicate how similar the cranium is to other crania in its group. Smaller D2 values indicate that
the cranium is very similar to the average cranium of that group. Larger D2 values indicate that
the cranium is unlike the average group member.
53
CHAPTER 6: DISCUSSION AND CONCLUSIONS
The purpose of this study was to determine if taking non-standard measurements of the
cranium would be helpful in determining ancestry when utilizing predictive multiple
discriminant function analysis. It was hypothesized that a combination of standard and non-
standard measurements would increase the classificatory power of the linear discriminant
function created from the measurements. The inclusion of non-standard measurements would
perhaps more completely describe the shape of the cranium by measuring features missed by
standard measurements.
Stepwise Analysis
The use of a forward stepwise variable selection method shows that a large number of
variables is not necessary when constructing discriminant functions. Twelve of the FD2
variables, 45 of the Howells variables, 2 of the Gill variables, and 50 of the All Measurements
variables were not included in the final analysis (Table 5.1). The variables that were eliminated
did not make a significant contribution to the discriminant function. It is proposed here that all
discriminant analyses employ a stepwise process for variable selection so as to only include
statistically significant variables and eliminate any ‘statistical noise’ resulting from superfluous
variables. Stepwise analysis is a feature that is not included in FORDISC 2.0. The inability to
cull out variables that are not important discriminators might produce discriminant functions that
are not as powerful as those produced with a stepwise analysis.
For the Howells, Gill, and All Measurements sets, the stepwise method selected
combinations of standard and non-standard measurements to classify the crania. Also, in the All
Measurements set, measurements from the FD2, Howells, and Gill sets were included. The three
54
same standard variables were included in the FD2, Howells, and All Measurements sets: parietal
chord (PAC), nasal breadth (NLB), and orbital height (OBH). Parietal chord measures the length
of the parietal bones along the sagittal suture, which would be a measure of partial cranial length.
African Americans have a larger value (i.e., a longer vault) and Native Americans have a smaller
value (i.e., a shorter vault). Nasal breadth is regarded as an indicator of ancestry, with narrow
nasal apertures being characteristic of European Americans and wide nasal apertures being
characteristic of African Americans. Orbital height does not indicate much by itself except for
the size of the orbital cavity. Although all three of these variables appeared in three sets of
functions, they were not the most heavily weighted as seen by the standardized discriminant
function coefficients (Tables 5.2, 5.3, and 5.5). In the FD2 set, NLB and OBH were heavily
weighted for Function 2. In the Howells set, only PAC was heavily weighted for Function 1.
OBH and PAC were heavily weighted for Function 1 of the All Measurements set. The
inclusion of parietal chord, orbital height, and nasal breadth in the analysis, coupled with their
high discriminant weights, demonstrates that they are important discriminators between the three
ancestral groups in this study.
Data Constructs
Placing a label on the discriminant functions allows one to replace the abstract nature of
the functions with something more comprehensible. A construct is defined as a concept that
comprises the underlying structure of variables (Huberty, 1994). In discriminant analysis the
constructs can be found by examining the structure matrices and the largest absolute correlations
between the variables and the discriminant functions. A construct may help to identify regions
of the cranium that are most useful in separating the groups. Labels were given to the functions
based on how the measurements describe skull shape and size (Table 6.1). Function 2 of both
55
the FD2 and Howells sets used prognathism and measurements of cranial width to separate the
groups. African groups tend to be more prognathic, whereas European groups tend to be more
orthognathic. The functions derived from the All Measurements set split the cranium into a
lower facial width component and an upper facial width component. Lower facial width was
comprised of measurements of the zygomatics, which would be useful in separating Native
Americans from other groups, since they tend to have more laterally projecting zygoma. The
Gill functions were restricted to the nasal region, with European Americans having a narrower,
more projecting nasal profile and African Americans having a wider, more rounded nasal profile.
These data constructs are concordant with non-metric analyses of the cranium in that they
describe recognized differences in size and shape between ancestral groups. The only difference
between studies of discrete traits and this study is that constructs are being attached to
continuous values.
TABLE 6.1. Discriminant Function Constructs.
Measurement Set Function Data Construct
FD2 1 facial width / vault length
2 maxillary prognathism / facial size
Howells 1 cranial length
2 prognathism / width
Gill 1 nasal width & projection
2 maxillofrontal breadth
All Measurements 1 lower facial width / vault shape
2 upper facial width
Classificatory Power of Discriminant Functions
The success of the discriminant functions is assessed by their classification accuracy. As
56
seen in Table 5.16, all measurement sets performed better than the ~ 33.3% prior probability of
correct random assignment to one of the three groups. The set with the lowest accuracy is the
Gill set, followed by the standard measurements of the FD2 set. The second best set was the
mixed measurements of the Howells set, and the set with the highest classification accuracy is
the mixed measurements of the All Measurements set. The combination of standard and non-
standard measurements increased the classificatory power of the linear discriminant function
formulae; therefore, the original hypothesis is accepted.
Taking a closer look at the performance of each measurement set in predicting group
membership of the crania via the leave-one-out method, patterns within the discriminant
functions emerge. The FD2, Howells, and All Measurements functions performed best with the
Coyotero Apache (CA) crania, with 89.1%, 92.7%, and 96.4% correct classification,
respectively. The separation of the CA from the European (EA) and African Americans (AA)
can be seen in the scatterplots of the discriminant function scores (Figures 5.1, 5.2, and 5.4).
This result is expected given the nature of the sample. The EA and AA groups come from
individuals that lived in the same area of the country within the last 100 years, whereas the CA
group lived ~1,300 years ago. The CA crania were noticeably more robust than the other groups,
with pronounced muscle attachments. Also, most of the CA crania had undergone some minor
plastic deformation from being compressed in the soil for so long, as well as a minor degree of
cultural deformation. Both of these phenomena caused the crania to be slightly compressed in
the sagittal plane, giving them slightly flatter frontal and occipital bones and resulting in a higher
vault.
Surprisingly, the Gill set performed best on the EA crania, with 76.5% correct
classification. Gill originally used the six measurements of the nasal region to differentiate
Native American crania from other groups, yet only 65.5% of the CA crania were correctly
57
classified. This result is most likely due to the fact that Gill specifically used Northwest Plains
Indian crania to develop the technique, while the current study examined Native American crania
from the Midwest Plains. All four sets of functions had a difficult time classifying the African
American crania, with consistently lower hit rates compared to the other groups. Perhaps the
African American subsample is more variable in cranial form than either the European American
or Coyotero Apache samples. This greater variability can be seen by examining the spread of the
African American specimens on the discriminant function score scatterplots in Figures 5.1, 5.2,
5.3, and 5.4. The African American group is the least tightly clustered group of the three, with
overlap into the European American and Coyotero Apache clusters.
General Conclusions
Taking non-standard measurements would be most useful in the case of incomplete
crania. Limiting data collection to standard measurements would greatly reduce the number of
measurements that could be taken of a particular part of the cranium. For example, the facial
skeleton is the most often damaged portion of the cranium, leaving just the neurocranium for
analysis. There are 12 FORDISC measurements that focus on the neurocranium. An additional
11 neurocranial Howells’ measurements could be used in conjunction with the FORDISC
measurements to aid in identification. Even though more measurements is not necessarily better,
as seen in the results of this study, the combination of standard and non-standard measurements
could aid in the identification of an incomplete cranium.
Twelve non-standard measurements were selected in this study as being important
discriminating variables. These include frontal subtense (FRS), bimaxillary breadth (ZMB),
prosthion radius (PRR), zygoorbitale radius (ZOR), bistephanic breadth (STB), bifrontal breadth
(FMB), biasterionic breadth (ASB), subspinale radius (SSR), mid-orbital breadth (ZOB), alpha
58
chord (ALC), naso-zygoorbital subtense (NZS), and molar alveolus radius (AVR). FRS, NZS,
ZMB, FMB, ZOB, and ALC were taken with the coordinate calipers. PRR, ZOR, SSR, and
AVR are all radial measurements that required the use of the radiometer. STB and ASB were
taken using a sliding caliper. Although ZMB, FMB, ZOB, and ALC were taken using the
coordinate calipers, one can take these four measurements using a sliding caliper. The
coordinate caliper was used in this study because a subtense was measured immediately after and
directly from the breadth/chord measurement. This fact demonstrates the diversity of the
coordinate caliper. It is possible to take up to three measurements at once: chords/breadths,
subtenses, and fractions. With regards to time constraints and data collection, it would take
longer to switch instruments for each measurement instead of using one instrument to take them
all.
The twelve non-standard measurements provide a detailed description of the cranium that
is not achieved with standard measurements. ZMB, ZOR, FMB, ZOB, ALC, and NZS are all
measures of the face. It is not surprising that these measurements were selected because the
facial skeleton tells researchers the most about the population affiliation of an individual. PRR,
SSR, and AVR all measure the projection of the maxilla. These three more fully describe which
portions of the maxilla project the farthest. FRS, STB, and ASB measure aspects of the vault
and add more detail to the overall shape of the vault. Although standard measurements are able
to discriminate between groups with a relatively high level (85.8%) of accuracy by themselves,
the inclusion of non-standard measurements increases the accuracy level (93.5%) of the
functions by ferreting out more subtle shape and size differences.
It is suggested here that biological anthropology laboratories should purchase
instrumentation such as coordinate calipers and radiometers so that non-standard measurements
can be taken to record data that would be missed by just using spreading and sliding calipers.
59
Another instrument that could be used as a replacement for the four aforementioned instruments
is a MicroScribe® digitizer. This instrument works in combination with a personal computer to
record cranial landmarks and their positions in three-dimensional space. The values of the
measurements are subsequently calculated by the computer using the coordinate data. This
instrument allows for much quicker data collection, however, the device is much more expensive
than all of the other calipers put together. Still, the MicroScribe® digitizer could be used with
the latest version of FORDISC: FORDISC 3.0 (FD3). FD3 will expand its battery of
measurements to include those used by Howells and this study (Ousley and Jantz, 2005). As
FORDISC is the ‘industry standard’ for computerized analysis of unknown human crania, it
seems reasonable to postulate that biological anthropology laboratories that adopt FD3 will begin
recording Howells non-standard measurements and entering them into the program more
frequently. As shown in this study, the inclusion of non-standard measurements should increase
the predictive power of FORDISC.
Suggestions for Future Study
Potential research in predictive discriminant analysis for ancestry determination of human
crania could include studying groups other than those used in this study to determine if the same
measurements would be utilized in a stepwise analysis. More population-specific functions need
to be created to more accurately predict ancestral group membership. Perhaps there are certain
measurements that are useful to discriminate between groups no matter what their temporal
distance or geographical location. Arcs could be incorporated in future studies, which only
require flexible measuring tapes and minimal time to record. Although existing measurements
do an excellent job of describing the variation among and between groups, new measurements
could be developed that might increase the power of the discriminant functions. Also, it is
60
important to discover the effects that the sex of a specimen has on the resulting discriminant
functions. Will including sex as a grouping criterion along with ancestry increase or decrease the
classificatory power of the functions?
61
LITERATURE CITED
Afifi AA, Clark V, May S (2004) Computer-Aided Multivariate Analysis. 4th
Ed. Boca Raton:
Chapman & Hall/CRC.
Ashton EH, Healy MJR, and Lipton S (1957) The descriptive use of discriminant functions in
physical anthropology. Proc. Roy. Soc. B. 146:552-572.
Ayers HG, Jantz RL, Moore-Jansen PH (1990) Giles & Elliot race discriminant functions
revisited: A test using recent forensic cases. In GW Gill and S Rhine (eds). Skeletal
Attribution of Race. Maxwell Museum Anthropology, Anthropological Papers, No. 4:
Albequerque, NM. pp. 65-71.
Barnard MM (1935) The secular variations of skull characters in four series of Egyptian skulls.
Annals of Eugenics. 6:352-371.
Bass WM (1995) Human Osteology: A Laboratory and Field Manual, 4th
Edition. Missouri
Archaeological Society, Inc.
Benington RC, Pearson K (1911) Cranial type-contours. Biometrika. 8:123-201.
Benington RC, Pearson K (1912) A study of the Negro skull with special reference to the Congo
and Gaboon crania. Biometrika. 8:292-339.
Birkby WH (1966) An evaluation of race and sex identification from cranial measurements. Am.
J. Phys. Anthrop. 24:21-28.
Brace CL (1995) Region does not mean “race” – reality versus convention in forensic
anthropology. J. For. Sci. 40:171-175.
Bronowski J, Long WM (1952) Statistics of discrimination in anthropology. Am. J. Phys.
Anthrop. 10:385-394.
Broom R (1923) A contribution to the craniology of the yellow-skinned races of South Africa. J.
Roy. Anthrop. Inst. Great Britain and Ireland. 53:132-149.
Brues AM (1990) The once and future diagnosis of race. In GW Gill and S Rhine (eds). Skeletal
Attribution of Race. Maxwell Museum Anthropology, Anthropological Papers, No. 4:
Albequerque, NM. pp. 1-7
Buikstra JE, Ubelaker DH (eds) (1994) Standards for Data Collection from Human Skeletal
Remains. Arkansas Archeological Survey no. 44. Fayetteville, Arkansas.
Campbell NA (1978) Multivariate analysis in biological anthropology: Some further
considerations. J. Hum. Evol. 7:197-203.
62
Carpenter JC (1976) A comparative study of metric and non-metric traits in a series of modern
crania. Am J. Phys. Anthrop. 45:337-344.
Church MS (1995) Determination of race from the skeleton through forensic anthropological
methods. For. Sci. Rev. 7:2-39.
Corruccini RS (1975) Multivariate analysis in biological anthropology: Some considerations. J.
Hum. Evol. 4:1-19.
Cunha E, van Vark GN (1991) The construction of sex discriminant functions from a large
collection of skulls of known sex. Int. J. Anthrop. 6:53-66.
Curran BK (1990) The application of measure of midfacial projection for racial classification. In
GW Gill and S Rhine (eds). Skeletal Attribution of Race. Maxwell Museum
Anthropology, Anthropological Papers, No. 4: Albequerque, NM. pp. 55-57.
de Mérejkowsky C (1882) Sur un nouveau caractére anthropologique. Bulletins de la Société
d’Anthropologie de Paris. 293-304.
DiBennardo R (1986) The use and interpretation of common computer implementations of
discriminant function analysis. In KJ Reichs (ed). Forensic Osteology: Advances in the
Identification of Human Remains. Springfield: Charles C. Thomas. 171-195.
DiBennardo R, Taylor JV (1983) Multiple discriminant function analysis of sex and race in the
postcranial skeleton. Am. J. Phys. Anthrop. 61:305-314.
Dunn OJ, Varady PD (1966) Probabilities of correct classification in discriminant analysis.
Biometrics. 22:908-924.
Feldesman MR (1997) Bridging the chasm: Demystifying some statistical methods used in
biological anthropology. In NT Boaz and LD Wolfe (eds). Biological Anthropology: The
State of the Science. Bend OR: International Institute for Human Evolutionary Research,
pp. 73-99.
Fisher RA (1936) The use of multiple measurements in taxonomic problems. Annals of
Eugenics. 3:179-188.
Fisher RA (1938) The statistical utilization of multiple measurements. Annals of Eugenics.
8:376-386.
Fisher TD, Gill GW (1990) Application of the Giles & Elliot discriminant function formulae to a
cranial sample of Northwestern Plains Indians. In GW Gill and S Rhine (eds). Skeletal
Attribution of Race. Maxwell Museum Anthropology, Anthropological Papers, No. 4:
Albequerque, NM. pp. 59-63.
Giles E (1966) Statistical techniques for race and sex determination: Some comments in
defense. Am. J. Phys. Anthrop. 25:85-86.
63
Giles E, Elliot O (1962) Race identification from cranial measurements. J. For. Sci. 7:147-157.
Giles E, Elliot O (1963) Sex determination by discriminant function. Am. J. Phys. Anthrop.
21:53-68.
Gill GW (1984) A forensic test case for a new method of geographical race determination. In TA
Rathbun and JE Buikstra (eds). Human Identification: Case Studies in Forensic
Anthropology. Springfield: Charles C. Thomas, pp. 229-339.
Gill GW (1986) Craniofacial criteria in forensic race identification. In KJ Reichs (ed). Forensic
Osteology: Advances in the Identification of Human Remains. Springfield: Charles C.
Thomas, pp. 143-159.
Gill GW (1995) Challenge on the frontier: Discerning American Indians from whites
osteologically. J. For. Sci. 40:783-788.
Gill GW (1998) Craniofacial criteria in forensic race identification. In KJ Reichs (ed). Forensic
Osteology: Advances in the Identification of Human Remains. 2nd
Ed. Springfield:
Charles C. Thomas, pp. 143-159.
Gill GW, Gilbert BM (1990) Race identification from the midfacial skeleton: American blacks
and Whites. In GW Gill and S Rhine (eds). Skeletal Attribution of Race. Maxwell
Museum Anthropology, Anthropological Papers, No. 4: Albequerque, NM. pp. 47-53.
Gill GW, Hughes SS, Bennett SM, Gilbert BM (1988) Racial identification from the midfacial
skeleton with special reference to American Indians and whites. J. For. Sci. 33:92-99.
Gill GW, Rhine S (1990) Skeletal Attribution of Race. Maxwell Museum Anthropology,
Anthropological Papers, No. 4: Albequerque.
Howells WW (1937) The designation of the principle anthrometric landmarks of the head and
skull. Am. J. Phys. Anthrop. 22:477-494.
Howells WW (1957) The cranial vault: Factors of size and shape. Am. J. Phys. Anthrop. 15:19-
48.
Howells WW (1960) Criteria for selection of osteometric measurements. Am. J. Phys. Anthrop.
30:451-458.
Howells WW (1966) The Jomon population of Japan. A study by discriminant analysis of
Japanese and Ainu crania. In Craniometry and Multivariate Analysis. Papers of the
Peabody Museum, Harvard University, vol. 57, no. 1, pp. 1-43.
Howells WW (1969) The use of multivariate techniques in the study of skeletal populations. Am.
J. Phys. Anthrop. 31:311-314.
64
Howells WW (1970) Multivariate analysis for the identification of race from crania. In TD
Stewart (ed). Personal Identification in Mass Disasters. Washington: Smithsonian
Institution, pp. 111-118.
Howells WW (1972) Analysis of patterns of variation in crania of recent man. In R Tuttle (ed).
The Functional and Evolutionary Biology of Primates. Chicago: Aldine-Atherton, pp.
123-151.
Howells WW (1973) Cranial Variation in Man: A Study by Multivariate Analysis of Patterns of
Difference Among Recent Human Populations. Papers of the Peabody Museum of
Archaeology and Ethnology, Harvard University, vol. 67.
Hrdlička A (1920) Anthropometry. Philadelphia: The Wistar Institute of Anatomy and Biology.
Huberty CJ (1994) Applied Discriminant Analysis. New York: John Wiley & Sons, Inc.
Hursh TM (1976) The study of cranial form: Measurement techniques and analytical methods.
In E Giles and JS Friedlander (eds). The Measures of Man: Methodologies in Biological
Anthropology. Peabody Museum Press, Cambridge, pp. 465-491.
İşcan MY, Steyn M (1999) Craniometric determination of population affinity in South Africans.
Int. J. Legal Med. 112:91-97.
Jantz RL, Ousley SD (1993) FORDISC 1.0: Computerized Forensic Discriminant Functions.
Knoxville, University of Tennessee.
Jantz RL, Owsley DW (1994) White traders in the upper Missouri: Evidence from the Swan
Creek site. In DW Owsley and RL Jantz (eds). Skeletal Biology in the Great Plains:
Migration, Warfare, Health, and Subsistence. Washington: Smithsonian Institute Press,
pp. 189-201.
Jeter MD (1990) Edward Palmer’s Arkansaw Mounds. The University of Arkansas Press:
Fayetteville.
Johnson DR, O'Higgins P, Moore WJ, McAndrew TJ (1989) Determination of race and sex of
the human skull by discriminant function analysis of linear and angular dimensions – an
appendix. For. Sci. Int. 41:41-53.
Kachigan SK (1991) Multivariate Statistical Analysis: A Conceptual Introduction 2nd
Ed. New
York: Radius.
Kennedy KAR (1995) But professor, why teach race identification if races don't exist? J. For.
Sci. 40:797-800.
Klepinger L, Giles E (1998) Classification or confusion: Statistical interpretation in forensic
anthropology. In KJ Reichs (ed). Forensic Osteology: Advances in Identification of
Human Remains 2nd
Ed. Springfield: Charles C. Thomas, pp. 427-440.
65
Key PJ (1983) Craniometric Relationships Among Plains Indians. Report of Investigations No.
34, The University of Tennessee Department of Anthropology, Knoxville.
Key PJ (1994) Relationships of the Woodland period on the northern and central plains: the
craniometric evidence. In DW Owsley and RL Jantz (eds). Skeletal Biology in the Great
Plains: Migration, Warfare, Health, and Subsistence. Washington: Smithsonian Institute
Press, pp. 179-187.
Kowalski CJ (1972) A commentary on the use of multivariate statistical analysis in
anthropometric research. Am. J. Phys. Anthrop. 30:19-131.
Kitson E (1931) A study of the Negro skull with special reference to the crania from Kenya
Colony. Biometrika. 23:271-314.
Krogman WM, İşcan MY (1986) The Human Skeleton in Forensic Medicine. 2nd
Ed.
Springfield: Charles C. Thomas.
Lachenbruch PA (1967) An almost unbiased method of obtaining confidence intervals for the
probability of misclassification in discriminant analysis. Biometrics. 23:639-645.
Lachenbruch PA, Mickey MR (1968) Estimation of error rates in discriminant analysis.
Technometrics. 10:1-11.
Lachenbruch PA, Goldstein M (1979) Discriminant analysis. Biometrics. 35:69-85.
Mahalanobis PC, Majumdar DN, Rao CR (1949) Anthropometric survey of the United
Provinces, 1941: a statistical study. Sankhyá: The Indian J. Stats. 9:89-342.
Martin R (1928) Lehrbuch der Anthropologie. Band I. Jena, Germany: Verlag Gustave Fischer.
McLachlan GJ (1980) On the relationship between the F test and the overall error rate for
variable selection in two-group discriminant analysis. Biometrics. 36:501-510.
McLachlan GJ (1992) Discriminant Analysis and Statistical Pattern Recognition. New York:
John Wiley and Sons.
Moore-Jansen PM, Ousley SD, Jantz RL (1994) Data Collection Procedures for Forensic
Skeletal Material, 3rd
Edition. Report of Investigations no. 48. The University of
Tennessee, Knoxville. Department of Anthropology.
Nawrocki SP (1993). The concept of race in contemporary physical anthropology. In The
Natural History of Paradigms: Science and the Process of Intellectual Evolution, ed. by J.
Langdon & M. McGann. University of Indianapolis Press, pp. 222-234.
Novotný V, İşcan MY, Loth SR (1993) Morphologic and osteometric assessment of age, sex, and
race from the skull. In MY İşcan, RP Helmer (eds). Forensic Analysis of the Skull. New
York: Wiley-Liss, pp. 71-88.
66
Ousley SD, Jantz RL (1996) FORDISC 2.0: Personal Computer Forensic Discriminant
Functions. Knoxville: University of Tennessee.
Ousley SD, Jantz RL (2005) The Next FORDISC: FORDISC 3. Abstract in the Proceedings of
the American Academy of Forensic Sciences. Vol. XI.
Patriquin ML, Steyn M, Loth SR (2002) Metric assessment of race from the pelvis of South
Africans. For. Sci. Int. 127:104-113.
Pearson K (1934) On simometers and their handling. Biometrika. 26:265-268.
Pearson K, Davin AG (1924) On the biometric constants of the human skull. Biometrika.
16:328-363.
Pietrusewsky M (1990) Craniofacial variation in Australian and Pacific populations. Am. J Phys.
Anthrop. 82:319-340.
Pietrusewsky M (2000) Metric analysis of skeletal remains: Methods and applications. In MA
Katzenberg and SR Saunders (eds). Biological Anthropology of the Human Skeleton.
New York: Wiley-Liss, pp. 375-415.
Rao CR (1946) Tests with discriminant functions in multivariate analysis. Sankhyá: The Indian
J. of Stat. 7:407-414.
Rao CR (1948) The utilization of multiple measurements in problems of biological classification.
J. Roy. Stat. Soc. B. 10:159-203.
Rao CR (1949) On some problems arising out of discrimination with multiple characters.
Sankhyá: The Indian J. of Stats. 9:343-366.
Rao CR (1952) Advanced Statistical Methods in Biometric Research. New York: Wiley.
Reichs KJ (1986) Forensic Osteology: Advances in the Identification of Human Remains.
Springfield: Charles C. Thomas.
Rightmire GP (1970) Bushmen, Hottentot, and South African Negro crania studied by distance
and discrimination. Am. J. Phys. Anthrop. 33:169-195.
Rightmire GP (1976) Metric versus discrete traits in African skulls. In E Giles and JS
Friedlander (eds). The Measures of Man: Methodologies in Biological Anthropology.
Cambridge: Peabody Museum Press, pp. 383-407.
Ross AH, Slice DE, Ubelaker DH, Falsetti AB (2004) Population affinities of 19th
century Cuban
crania: Implications for identification criteria in south Florida Cuban Americans. J. For.
Sci. 49:11-16
67
Sauer NJ (1992) Forensic anthropology and the concept of race: if races don’t exist, why are
forensic anthropologists so good at identifying them? Soc. Sci. Med. 34:107-111.
Smith CA (1947) Some examples of discrimination. Ann. Eugen. 18:272-282.
Snapinn SM, Knoke JD (1989) Estimation of error rates in discriminant analysis with selection
of variables. Biometrics. 45:289-299.
Snow CC, Hartman S, Giles E, Young FA (1979) Sex and race determination of crania by
calipers and computer: A test of the Giles and Elliot discriminant functions in 52 forensic
science cases. J. For. Sci. 24:448-460.
Song H, Lin ZQ, Jia JT (1992) Sex diagnosis of Chinese skulls using multiple stepwise
discriminant function analysis. For. Sci. Int. 54:135-140.
Stewart TD (1948) Medico-legal aspects of the skeleton I: Age, sex, race and stature. Am. J.
Phys. Anthrop. 6:315-322.
Steyn M, İşcan MY (1998) Sexual dimorphism in the crania and mandibles of South African
whites. For. Sci. Int. 98:9-16.
Taylor JV, Roh L, Goldman AD (1984) Metropolitan Forensic Anthropology Team (MFAT)
case studies in identification: 2. Identification of a Vietnamese trophy skull. J. For. Sci.
29:1253-1259.
Thieme FP, Schull WJ (1957) Sex determination from the skeleton. Hum. Bio. 29:242-273.
Townsend GC, Richards LC, Carroll A (1982) Sex determination of Australian Aboriginal skulls
by discriminant function analysis. Australian Dental J. 27:320-326.
van Vark GN (1976) A critical evaluation of the application of multivariate statistical methods to
the study of human populations from their skeletal remains. Homo. 27:94-113.
van Vark GN (1994) Multivariate analysis: Is it useful for hominid studies? Courier
Forschungsinstitut Senckenberg. 171:289-294.
van Vark GN, van der Sman PGM (1982) New discrimination and classification techniques in
anthropological practice. Z. Morph. Anthrop. 73:21-36.
van Vark GN, Schaafsma W (1992) Advances in the quantitative analysis of skeletal
morphology. In SR Saunders and MA Katzenberg (eds). Skeletal Biology of Past
Peoples: Research Methods. New York: Wiley-Liss, pp. 225-257.
von Bonin G (1931) A contribution to the craniology of the Easter Islanders. Biometrika. 23:249-
270.
68
von Bonin G (1936) On the craniology of Oceania. Crania from New Britain. Biometrika.
28:123-148.
von Bonin G, Morant GM (1938) Indian races in the United States. A survey of previously
published cranial measurements. Biometrika. 30:94-129.
Wright RVS (1992) Correlation between cranial form and geography in Homo sapiens: CRANID
– a computer program for forensic and other applications. Archaeology in Oceania.
27:128-134.
Woo TL, Morant GM (1932) A preliminary classification of Asiatic races based on cranial
measurements. Biometrika. 24:108-134.
Woo TL, Morant GM (1934) A biometric study of the “flatness” of the facial skeleton in man.
Biometrika. 26:196-250.
Zambrano CJ, Kolatorowicz A, Nawrocki SP (2005) Performance of FORDISC 2.0 using
inaccurate measurements. Poster presented at the 57th
annual meeting of the American
Academy of Forensic Sciences. New Orleans LA.
Zugibe FT, Taylor J, Weg N, DiBennardo R, Costello JT, Deforest P (1985) Metropolitan
Forensic Anthropology Team (MFAT) case studies in identification: 3. Identification of
John J. Sullivan, the missing journalist. J. For. Sci. 30:221-231.
69
APPENDIX A: CRANIOMETRICS RECORDING FORM
Date: _______________ Ancestry: AA EA NA
Collection: ___________________ Sex: M F
Collection Accession #: ______________ DOB: ______________
Study Specimen #: ______________ DOD: ______________
Age: ______
(All measurements taken on the left side, to the nearest mm.)
1 max. cranial length GOL 35 cheek height WMH
2 max. cranial breadth XCB 36 bimaxillary breadth ZMB
3 bizygomatic breadth ZYB 37 bimaxillary subtense SSS
4 basion-bregma height BBH 38 bifrontal breadth FMB
5 cranial base length BNL 39 nasio-frontal subtense NAS
6 basion-prosthion length BPL 40 dacryon subtense DKS
7 max. alveolar breadth MAB 41 naso-dacryl subtense NDS
8 max. alveolar length MAL 42 supraorbital projection SOS
9 biauricular breadth AUB 43 glabella projection GLS
10 upper facial height NPH 44 simotic chord WNB
11 min frontal breadth WFB 45 simotic subtense SIS
12 upper facial breadth UFBR 46 frontal subtense FRS
13 nasal height NLH 47 frontal fraction FRF
14 nasal breadth NLB 48 parietal subtense PAS
15 orbital breadth OBB 49 parietal fraction PAF
16 orbital height OBH 50 occipital subtense OCS
17 biorbital breadth EKB 51 occipital fraction OCF
18 interorbital breadth DKB 52 vertex radius VRR
19 frontal chord FRC 53 nasion radius NAR
20 parietal chord PAC 54 subspinale radius SSR
21 occipital chord OCC 55 prosthion radius PRR
22 for. mag. length FOL 56 dacryon radius DKR
23 for. mag. breadth FOB 57 zygoorbitale radius ZOR
24 mastoid height MDH 58 frontomalare radius FMR
25 mastoid width MDB 59 ectoconchion radius EKR
26 nasio-occipital length NOL 60 zygomaxillare radius ZMR
27 max frontal breadth XFB 61 molar alveolus radius AVR
28 bistephanic breadth STB 62 maxillofrontal breadth DKB
29 min cranial breadth WCB 63 naso-maxillofrontal subtense NDS
30 biasterionic breadth ASB 64 mid-orbital breadth ZOB
31 bijugal breadth JUB 65 naso-zygoorbital subtense NZS
32 malar length, inf. IML 66 alpha chord ALC
33 malar length, max. XML 67 naso-alpha subtense NLS
34 malar subtense MLS
70
APPENDIX B: CRANIAL MEASUREMENT DEFINITIONS
Definitions for #’s 1-61 are taken from Howells (1973).
Definitions for #’s 62-67 are taken from Gill (1984).
Abbreviations for #’s 64-67 were created by the author.
Instrument Codes:
Sliding caliper: A Spreading caliper: B Coordinate Caliper: C Radiometer: D
1. Maximum cranial length (GOL): the distance of glabella from opisthocranion. B
2. Maximum cranial breadth (XCB): the maximum width of the skull perpendicular to the mid-
sagittal plane wherever it is located. B
3. Bizygomatic breadth (ZYB): the direct distance between both zygia located at their most
lateral points of the zygomatic arches. B
4. Basion-bregma height (BBH): the direct distance from the lowest point on the anterior
margin of the foramen magnum, basion to bregma. B
5. Cranial base length (BNL): the direct distance from nasion to basion. B
6. Basion-prosthion length (BPL): the direct distance from basion to prosthion. B
7. Maximum alveolar breadth (MAB): the maximum breadth across the alveolar borders of the
maxilla measured on the lateral surfaces at the location of the first maxillary molar. B
8. Maximum alveolar length (MAL): the direct distance from prosthion to alveolon. A
9. Biauricular breadth (AUB): the least exterior breadth across the roots of the zygomatic
processes. B
10. Upper facial height (NPH): the direct distance from nasion to prosthion. A
11. Minimum frontal breadth (WFB): the direct distance between the two frontotemporale. B
12. Upper facial breadth (UFBR): the direct distance between the two frontomalare temporalia.
13. Nasal height (NLH): the direct distance from nasion to nasospinale. A
14. Nasal breadth (NLB): the maximum breadth of the nasal aperture. A
15. Orbital breadth (OBB): the laterally sloping distance from dacryon to ectoconchion. A
16. Orbital height (OBH): the direct distance between the superior and inferior orbital margins.
A
17. Biorbital breadth (EKB): the direct distance from one ectoconchion to the other. C
18. Interorbital breadth (DKB): the direct distance between right and left dacryon. C
19. Frontal chord (FRC): the direct distance from nasion to bregma. C
20. Parietal chord (PAC): the direct distance from bregma to lambda. C
21. Occipital chord (OCC): the direct distance from lambda to opisthion. C
22. Foramen magnum length (FOL): the direct distance of basion from opisthion. A
23. Foramen magnum breadth (FOB): the distance between the lateral margins of the foramen
magnum at the point of greatest lateral curvature. A
24. Mastoid height (MDH): the projection of the mastoid process below, and perpendicular to,
the eye-ear plane in the vertical plane. A
25. Mastoid width (MDB): width of the mastoid process at its base, through its transverse axis.
A
26. Nasio-occipital length (NOL): greatest cranial length in the median sagittal plane, measured
from nasion. B
71
27. Maximum frontal breadth (XFB): the maximum breadth at the coronal suture, perpendicular
to the median plane. B
28. Bistephanic breadth (STB): breadth between the intersections, on either side, of the coronal
suture and the inferior temporal line marking the origin of the temporal muscle. A
29. Minimum cranial breadth (WCB): the breadth across the sphenoid at the base of the
temporal fossa, at the infratemporal crests. A
30. Biasterionic breadth (ASB): direct measurement from one asterion to the other. A
31. Bijugal breadth (JUB): the external breadth across the malars at the jugalia, i.e., at the
deepest points in the curvature between the frontal and temporal processes of the malars.
A
32. Malar length, inferior (IML): the direct distance from zygomaxillare anterior to the lowest
point of the zygo-temporal suture on the external surface, on the left side. A
33. Malar length, maximum (XML): total direct length of the malar in a diagonal direction,
from the lower end of the zygo-temporal suture on the lateral face of the bone, to
zygoorbitale, the junction of the zygo-maxillary suture with the lowest border of the
orbit, on the left side. C
34. Malar subtense (MLS): the maximum subtense from the convexity of the malar angle to the
maximum length of the bone, at the level of the zygomaticofacial foramen, on the left
side. C
35. Cheek height (WMH): the minimum distance, in any direction, from the lower border of the
orbit to the lower margin of the maxilla, mesial to the masseter attachment, on the left
side. A
36. Bimaxillary breadth (ZMB): the breadth across the maxillae, from one zygomaxillare
anterior to the other. C
37. Bimaxillary subtense (SSS): the projection or subtense from subspinale to the bimaxillary
breadth. C
38. Bifrontal breadth (FMB): the breadth across the frontal bone between frontomalare anterior
on each side, i.e., the most anterior point on the fronto-malar suture. C
39. Nasio-frontal subtense (NAS): the subtense from nasion to the bifrontal breadth. C
40. Dacryon subtense (DKS): the mean subtense from dacryon (average of two sides) to the
biorbital breadth. C
41. Naso-dacryl subtense (NDS): the subtense from the deepest point in the profile of the nasal
bones to the interorbital breadth. C
42. Supraorbital projection (SOS): the maximum projection of the left supraorbital arch
between the midline, in the region of glabella or above, and the frontal bone just anterior
to the temporal line in its forward part, measured as a subtense to the line defined. C
43. Glabella projection (GLS): the maximum projection of the midline profile between nasion
and supraglabellare (or the point at which the curve of the profile of the frontal bone
changes to join the prominence of the glabellar region), measured as a subtense. C
44. Simotic chord (WNB): the minimum transverse breadth across the two nasal bones, or
chord between the naso-maxillary sutures at their closest approach. C
45. Simotic subtense (SIS): the subtense from the nasal bridge to the simotic chord, i.e., from
the highest point in the transverse section which is at the deepest point in the nasal
profile. C
46. Frontal subtense (FRS): the maximum subtense, at the highest point on the convexity of the
frontal bone in the midplane, to the nasion-bregma chord. C
72
47. Frontal fraction (FRF): the distance along the nasion-bregma chord, recorded from nasion,
at which the frontal subtense falls. C
48. Parietal subtense (PAS): the maximum subtense, at the highest point on the convexity of the
parietal bones in the midplane, to the bregma-lambda chord. C
49. Parietal fraction (PAF): the distance along the bregma-lambda chord, recorded from
bregma, at which the parietal subtense falls. C
50. Occipital subtense (OCS): the maximum subtense, at the most prominent point on the basic
contour of the occipital bone in the midplane. C
51. Occipital fraction (OCF): the distance along the lambda-opisthion chord, recorded from
lambda, at which the occipital subtense falls. C
52. Vertex radius (VRR): the perpendicular to the transmeatal axis from the most distant point
on the parietals (including bregma or lambda), wherever found. D
53. Nasion radius (NAR): the perpendicular to the transmeatal axis from nasion. D
54. Subspinale radius (SSR): the perpendicular to the transmeatal axis from subspinale. D
55. Prosthion radius (PRR): the perpendicular to the transmeatal axis from prosthion. D
56. Dacryon radius (DKR): the perpendicular to the transmeatal axis from the left dacryon. D
57. Zygoorbitale radius (ZOR): the perpendicular to the transmeatal axis from the left
zygoorbitale. D
58. Frontomalare radius (FMR): the perpendicular to the transmeatal axis from the left
fromtomalare anterior. D
59. Ectoconchion radius (EKR): the perpendicular to the transmeatal axis from the left
ectoconchion. D
60. Zygomaxillare radius (ZMR): the perpendicular to the transmeatal axis from the left
zygomaxillare anterior. D
61. Molar alveolus radius (AVR): the perpendicular to the transmeatal axis from the most
anterior point on the alveolus of the left first molar. D
62. Maxillofrontal breadth (DKB): distance between maxillofrontale left and right. C
63. Naso-maxillofrontal subtense (NDS): subtense from the maxillofrontal points to the deepest
point on the nasal bridge. C
64. Mid-orbital breadth (ZOB): the distance between zygoorbitale left and right. C
65. Naso-zygoorbital subtense (NZS): subtense from the zygoorbital points to the deepest point
along the nasal bridge. C
66. Alpha cord (ALC): the point alpha is the deepest point on the maxilla, left and right, on a
tangent run between the naso-maxillary suture where it meets the nasal aperture, and
zygoorbitale. C
67. Naso-alpha subtense (NLS): subtense from the alpha points to the deepest point on the nasal
bridge. C
73
APPENDIX C: CASEWISE STATISTICS
Group 1 = African American, Group 2 = European American, Group 3 = Coyotero Apache
TABLE C.1. FD2 Set
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
1 1 3 0.215 0.993 15.496 1.604 -0.148
2 1 1 0.250 0.999 14.841 -0.876 3.348
3 1 1 0.945 0.990 5.364 -1.664 1.598
4 1 1 0.522 1.000 11.084 -1.681 3.030
5 1 1 0.246 1.000 14.917 -1.033 3.527
6 1 1 0.005 0.948 28.187 -5.457 0.721
7 2 2 0.846 0.992 7.171 -1.316 -1.833
8 1 1 0.902 0.923 6.272 -0.709 1.084
9 1 2 0.887 0.914 6.527 -0.673 -0.712
10 1 1 0.228 0.843 15.258 -0.016 1.685
11 1 3 0.565 0.717 10.579 0.994 2.297
12 1 1 0.085 0.944 19.132 -2.501 0.991
13 2 2 0.608 0.824 10.096 -1.733 -0.773
14 1 1 0.574 0.986 10.476 -2.910 1.349
15 1 1 0.762 0.931 8.296 -0.356 1.457
16 1 1 0.537 0.994 10.904 -1.950 1.794
17 2 2 0.418 0.729 12.348 -2.003 -0.643
18 1 1 0.785 0.984 8.000 -1.128 1.582
19 1 1 0.378 0.607 12.874 -1.131 0.312
20 1 1 0.631 0.996 9.824 -2.233 1.852
21 1 1 0.970 0.987 4.595 -1.197 1.601
22 2 2 0.915 0.547 6.021 -2.286 -0.312
23 2 2 0.050 0.817 21.059 -0.083 -1.100
24 1 1 0.414 0.994 12.403 -2.222 1.736
25 2 2 0.000 0.998 57.335 -4.416 -2.713
26 2 1 0.254 0.815 14.780 -1.329 0.202
27 2 2 0.348 0.999 13.288 -1.217 -2.551
28 2 1 0.681 0.964 9.258 -2.430 0.711
29 1 1 0.189 0.990 16.041 -0.760 2.073
30 1 2 0.444 0.542 12.020 -0.346 0.093
31 1 2 0.028 0.784 22.962 -3.220 -0.344
32 1 2 0.962 0.548 4.877 -1.224 -0.045
33 1 1 0.252 0.616 14.817 -1.419 0.309
34 1 1 0.672 0.858 9.354 -1.511 0.685
35 1 1 0.008 0.977 27.047 -1.282 1.519
36 1 1 0.987 0.997 3.802 -1.126 2.183
37 2 2 0.032 0.751 22.554 -1.953 -0.789
74
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
38 2 2 0.294 0.903 14.098 -1.240 -0.997
39 1 1 0.189 0.994 16.058 -2.251 1.746
40 1 1 0.834 0.978 7.338 -2.418 1.244
41 1 1 0.840 0.998 7.265 -2.070 2.180
42 1 1 0.009 1.000 26.651 -1.519 4.033
43 2 2 0.881 0.943 6.638 -0.034 -1.573
44 1 1 0.950 0.935 5.219 -1.287 0.984
45 2 2 0.472 0.914 11.672 0.080 -1.603
46 1 1 0.950 0.810 5.225 0.209 1.724
47 1 2 0.686 0.561 9.198 -1.142 0.016
48 2 1 0.265 0.684 14.591 -0.765 0.083
49 1 1 0.374 0.995 12.933 -1.656 1.902
50 1 1 0.530 0.942 10.983 -1.645 1.022
51 1 3 0.124 0.647 17.739 0.311 0.305
52 2 2 0.113 0.984 18.086 -0.469 -1.872
53 1 2 0.719 0.788 8.811 -2.339 -0.469
54 1 1 0.548 0.965 10.779 -2.376 1.110
55 1 2 0.347 0.941 13.305 -1.058 1.147
56 2 2 0.758 0.968 8.334 -1.537 -1.376
57 2 2 0.912 0.950 6.084 -0.770 -1.162
58 2 2 0.335 0.986 13.488 -0.105 -2.370
59 2 2 0.052 0.656 20.860 0.394 -1.400
60 2 2 0.431 0.956 12.180 -2.055 -1.356
61 2 1 0.471 0.990 11.686 -0.922 -1.811
62 1 1 0.986 0.956 3.868 -1.886 1.035
63 1 1 0.253 1.000 14.793 -1.709 3.689
64 1 2 0.850 0.881 7.110 0.079 1.890
65 2 2 0.906 0.978 6.197 -0.414 -1.648
66 2 1 0.566 0.975 10.567 -1.866 -1.516
67 2 1 0.656 0.763 9.537 -1.185 0.215
68 1 2 0.930 0.955 5.700 -0.266 1.820
69 2 2 0.037 0.993 22.049 -0.849 -2.014
70 2 1 0.836 0.571 7.323 -1.727 -0.298
71 1 1 0.840 0.595 7.264 -1.921 0.112
72 1 3 0.518 0.967 11.130 -2.920 1.065
73 2 1 0.092 0.735 18.875 0.810 -1.213
74 1 2 0.767 0.944 8.225 -0.417 1.498
75 2 2 0.416 0.986 12.379 -0.804 -1.700
76 2 1 0.069 0.977 19.878 0.114 -2.606
77 1 2 0.654 0.982 9.571 -1.743 1.410
78 2 1 0.747 0.998 8.475 -2.140 -2.399
79 1 2 0.894 0.565 6.404 0.237 0.967
80 2 2 0.710 0.676 8.913 -1.470 -0.448
81 2 2 0.390 0.978 12.719 -0.337 -1.794
82 2 1 0.616 0.914 10.003 -1.842 -1.068
83 2 2 0.086 0.900 19.089 -1.255 0.359
84 2 2 0.930 0.986 5.705 -0.797 -1.672
75
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
85 2 2 0.419 0.993 12.339 -0.903 -1.919
86 2 2 0.666 0.970 9.428 -1.956 -1.458
87 2 2 0.375 0.971 12.918 -0.107 -1.965
88 2 2 0.747 1.000 8.471 -1.626 -2.862
89 2 2 0.273 0.999 14.443 -0.164 -3.985
90 2 3 0.139 0.921 17.304 0.653 -2.862
91 2 2 0.855 0.749 7.034 1.242 -1.613
92 2 2 0.127 0.948 17.648 -2.552 -1.398
93 2 2 0.300 0.946 14.015 -1.680 -1.256
94 2 2 0.487 0.996 11.495 -1.163 -2.096
95 2 2 0.239 0.685 15.037 -0.246 -0.559
96 2 2 0.987 0.948 3.785 -0.880 -1.125
97 2 2 0.772 0.867 8.171 -0.488 -0.821
98 2 2 0.116 0.951 17.980 0.685 -3.246
99 2 2 0.350 0.992 13.273 -0.775 -1.929
100 2 3 0.978 0.945 4.253 -1.113 -1.110
101 3 3 0.159 0.998 16.750 2.381 0.418
102 3 3 0.064 0.958 20.173 1.490 0.463
103 3 3 0.452 0.998 11.923 2.320 0.980
104 3 3 0.859 0.808 6.978 0.921 0.785
105 3 3 0.393 0.968 12.680 1.517 0.582
106 3 3 0.354 0.987 13.213 1.886 1.083
107 3 3 0.206 0.994 15.688 2.052 0.909
108 3 3 0.138 0.983 17.321 1.811 -0.134
109 3 1 0.375 0.688 12.924 -0.367 0.354
110 3 3 0.087 0.992 19.050 1.943 0.636
111 3 2 0.002 0.537 30.409 0.557 -0.153
112 3 2 0.530 0.517 10.986 0.677 -0.699
113 3 3 0.429 0.994 12.213 2.015 0.127
114 3 2 0.000 0.946 158.399 1.806 -1.136
115 3 3 0.879 0.969 6.668 1.921 2.168
116 3 3 0.608 1.000 10.091 3.228 1.362
117 3 3 0.231 1.000 15.190 2.808 0.460
118 3 3 0.801 1.000 7.798 3.129 -0.488
119 3 3 0.451 0.993 11.931 1.956 0.604
120 3 3 0.047 0.997 21.244 2.651 -0.957
121 3 3 0.642 0.996 9.707 2.277 1.450
122 3 3 0.980 0.997 4.194 2.180 0.643
123 3 3 0.932 1.000 5.658 2.727 0.647
124 3 3 0.283 0.976 14.275 1.646 0.793
125 3 3 0.682 0.829 9.245 0.983 -0.081
126 3 3 0.996 1.000 3.004 3.403 -0.040
127 3 2 0.012 0.832 25.592 1.131 -1.878
128 3 3 0.918 0.997 5.964 2.156 0.774
129 3 3 0.834 0.998 7.351 2.344 0.467
130 3 3 0.930 0.992 5.712 2.207 -0.779
131 3 3 0.627 0.902 9.869 1.951 -1.809
76
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
132 3 3 0.697 1.000 9.068 4.154 1.554
133 3 3 0.059 1.000 20.458 3.298 -0.664
134 3 3 0.846 0.999 7.178 2.611 0.859
135 3 3 0.811 0.966 7.664 1.593 -0.358
136 3 3 0.696 0.992 9.086 2.483 -1.263
137 3 3 0.861 0.962 6.943 1.404 0.534
138 3 3 0.883 0.999 6.592 2.381 0.452
139 3 3 0.782 0.914 8.034 1.837 -1.536
140 3 3 0.231 0.958 15.202 1.481 0.731
141 3 3 0.000 1.000 52.434 3.189 1.562
142 3 2 0.172 0.528 16.436 1.403 -1.944
143 3 3 0.954 0.998 5.129 2.256 0.238
144 3 3 0.020 0.984 24.080 1.989 1.329
145 3 3 0.308 0.943 13.890 1.854 -1.159
146 3 3 0.531 0.999 10.973 2.631 0.177
147 3 3 0.173 0.995 16.408 2.194 1.169
148 3 3 0.003 0.979 29.774 2.536 -1.638
149 3 3 0.974 1.000 4.449 3.161 0.541
150 3 3 0.556 0.905 10.689 1.868 -1.605
151 3 3 0.951 0.993 5.209 2.082 -0.367
152 3 3 0.493 1.000 11.418 3.121 0.907
153 3 3 0.844 1.000 7.197 2.764 0.921
154 3 3 0.004 0.894 28.900 1.722 -0.967
155 3 3 0.955 1.000 5.101 3.118 0.215
TABLE C.2. Howells Set
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
1 1 1 0.584 0.999 10.365 0.808 2.310
2 1 1 0.117 1.000 17.970 1.351 2.726
3 1 1 0.929 0.998 5.722 0.820 2.126
4 1 1 0.229 0.965 15.235 1.271 1.295
5 1 1 0.004 0.998 29.009 0.146 3.336
6 1 1 0.074 0.998 19.648 3.907 2.399
7 2 2 0.317 0.968 13.751 1.515 -1.139
8 1 1 0.363 0.538 13.080 1.338 0.312
9 1 2 0.857 0.999 7.014 1.636 -1.648
10 1 1 0.187 0.961 16.083 1.433 1.281
11 1 3 0.792 0.675 7.909 -0.856 1.095
12 1 1 0.166 1.000 16.592 2.055 3.123
13 2 2 0.854 0.988 7.047 2.582 -1.312
14 1 1 0.691 0.776 9.135 2.715 0.714
15 1 1 0.349 0.996 13.274 1.158 1.939
77
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
16 1 1 0.575 0.989 10.468 0.772 1.611
17 2 2 0.150 0.989 16.997 2.909 -1.366
18 1 1 0.605 0.974 10.120 0.794 1.329
19 1 1 0.668 0.960 9.407 1.792 1.249
20 1 1 0.966 0.955 4.738 1.429 1.139
21 1 1 0.853 0.997 7.075 1.874 2.008
22 2 2 0.559 0.556 10.655 0.704 -0.193
23 2 1 0.159 0.694 16.769 0.690 0.034
24 1 1 0.054 1.000 20.777 1.275 3.568
25 2 2 0.002 0.998 30.498 3.020 -1.931
26 2 2 0.733 0.805 8.641 1.001 -0.513
27 2 2 0.593 0.999 10.262 0.992 -2.389
28 2 2 0.314 0.630 13.795 2.622 -0.163
29 1 1 0.000 1.000 36.289 0.401 3.561
30 1 2 0.438 0.425 12.093 -0.259 -0.144
31 1 1 0.543 0.912 10.838 1.963 1.016
32 1 2 0.761 0.506 8.301 2.463 0.278
33 1 2 0.324 0.547 13.646 0.921 0.180
34 1 1 0.604 0.561 10.133 2.362 0.384
35 1 1 0.010 0.989 26.101 0.701 1.718
36 1 1 0.141 0.963 17.224 1.403 1.307
37 2 2 0.011 0.799 25.941 1.244 -0.673
38 2 2 0.343 0.975 13.366 2.117 -1.157
39 1 1 0.043 1.000 21.503 0.639 3.213
40 1 1 0.970 0.973 4.588 1.833 1.336
41 1 1 0.918 1.000 5.969 1.287 2.931
42 1 1 0.003 1.000 29.732 1.898 3.609
43 2 2 0.436 0.970 12.120 0.389 -1.456
44 1 1 0.436 0.786 12.124 -0.037 0.823
45 2 3 0.134 0.945 17.430 -1.347 -2.434
46 1 1 0.646 0.889 9.654 0.373 0.859
47 1 1 0.414 0.508 12.407 2.162 0.333
48 2 2 0.251 0.927 14.824 1.595 -0.875
49 1 1 0.454 0.957 11.894 1.585 1.228
50 1 1 0.602 1.000 10.156 1.682 2.803
51 1 1 0.014 0.795 25.236 -0.106 1.253
52 2 2 0.268 0.740 14.540 -0.062 -0.959
53 1 2 0.638 0.920 9.745 3.160 -0.319
54 1 1 0.206 0.984 15.681 2.406 1.639
55 1 1 0.095 0.999 18.736 2.411 2.490
56 2 2 0.310 0.997 13.857 2.641 -1.773
57 2 2 0.446 0.998 11.998 0.665 -2.373
58 2 2 0.541 0.999 10.861 1.560 -2.118
59 2 2 0.344 0.998 13.350 1.684 -1.914
60 2 2 0.342 0.967 13.378 3.075 -0.986
61 2 2 0.178 1.000 16.298 1.474 -3.034
62 1 1 0.293 1.000 14.114 1.082 2.838
78
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
63 1 1 0.007 1.000 27.399 1.264 2.602
64 1 1 0.714 0.967 8.868 0.732 1.239
65 2 2 0.491 0.999 11.442 2.657 -2.186
66 2 2 0.561 0.967 10.631 2.460 -1.018
67 2 1 0.496 0.756 11.392 1.154 0.231
68 1 1 0.180 0.999 16.251 0.605 2.816
69 2 2 0.163 0.997 16.666 1.305 -1.898
70 2 1 0.115 0.908 18.028 2.330 0.532
71 1 1 0.447 0.850 11.984 1.170 0.776
72 1 1 0.483 0.996 11.548 2.273 1.985
73 2 2 0.501 0.973 11.334 0.813 -1.257
74 1 1 0.284 0.936 14.269 0.216 1.208
75 2 2 0.816 0.999 7.588 2.611 -2.067
76 2 2 0.336 0.998 13.465 0.892 -2.134
77 1 1 0.515 0.995 11.163 2.000 1.949
78 2 2 0.480 0.994 11.577 2.893 -1.526
79 1 1 0.820 0.974 7.538 0.328 1.402
80 2 2 0.677 0.807 9.296 1.805 -0.455
81 2 2 0.761 0.988 8.305 1.036 -1.463
82 2 2 0.867 0.994 6.849 2.001 -1.581
83 2 1 0.296 0.734 14.070 1.419 0.178
84 2 2 0.906 0.995 6.187 0.598 -1.956
85 2 2 0.148 1.000 17.048 1.515 -2.492
86 2 2 0.462 0.999 11.799 1.670 -2.288
87 2 2 0.892 0.992 6.445 1.045 -1.577
88 2 2 0.742 1.000 8.533 1.225 -2.536
89 2 2 0.930 1.000 5.719 2.208 -2.928
90 2 2 0.013 1.000 25.501 1.630 -3.466
91 2 2 0.945 0.986 5.349 0.264 -1.858
92 2 2 0.618 0.832 9.977 2.266 -0.476
93 2 2 0.210 0.883 15.608 3.250 -0.586
94 2 2 0.296 0.984 14.070 2.254 -1.289
95 2 2 0.409 0.996 12.470 2.536 -1.686
96 2 2 0.655 0.985 9.557 0.966 -1.416
97 2 2 0.392 0.954 12.688 0.247 -1.430
98 2 2 0.585 0.998 10.351 0.541 -2.497
99 2 2 0.040 0.993 21.807 1.968 -1.592
100 2 2 0.977 0.959 4.316 1.735 -0.953
101 3 1 0.009 0.631 26.704 -0.597 0.299
102 3 3 0.225 1.000 15.316 -2.811 -0.248
103 3 3 0.426 1.000 12.249 -3.553 1.576
104 3 3 0.911 0.985 6.093 -1.543 -0.014
105 3 3 0.598 0.930 10.205 -1.205 0.009
106 3 3 0.067 0.986 20.005 -1.797 0.367
107 3 3 0.167 1.000 16.553 -3.150 -0.907
108 3 3 0.277 0.987 14.387 -1.888 -1.345
109 3 3 0.220 0.926 15.398 -1.322 0.322
79
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
110 3 2 0.259 0.872 14.679 -0.358 -1.028
111 3 1 0.163 0.778 16.661 -0.375 0.544
112 3 3 0.423 1.000 12.291 -3.061 -1.559
113 3 3 0.398 1.000 12.605 -2.456 -0.170
114 3 3 0.961 1.000 4.922 -2.834 -1.014
115 3 3 0.796 1.000 7.864 -3.354 1.975
116 3 3 0.413 1.000 12.412 -3.363 0.432
117 3 3 0.175 1.000 16.356 -4.618 -0.440
118 3 3 0.514 0.999 11.177 -2.587 -1.430
119 3 3 0.641 1.000 9.716 -3.271 -1.301
120 3 3 0.506 1.000 11.272 -3.464 -0.253
121 3 3 0.414 0.995 12.407 -2.034 0.600
122 3 3 0.331 1.000 13.543 -2.723 -0.188
123 3 3 0.238 0.967 15.066 -2.125 1.842
124 3 3 0.487 1.000 11.499 -3.166 0.545
125 3 3 0.514 0.974 11.169 -1.525 0.309
126 3 3 0.796 1.000 7.862 -4.172 -0.084
127 3 3 0.393 1.000 12.673 -3.332 -0.696
128 3 3 0.944 1.000 5.393 -2.914 0.617
129 3 3 0.768 1.000 8.218 -2.786 0.532
130 3 3 0.980 1.000 4.166 -2.501 -0.876
131 3 3 0.362 1.000 13.103 -3.439 -1.860
132 3 3 0.462 1.000 11.793 -4.873 0.915
133 3 3 0.293 1.000 14.124 -2.740 0.346
134 3 3 0.740 0.997 8.555 -2.003 0.257
135 3 3 0.625 1.000 9.898 -2.685 0.197
136 3 3 0.545 1.000 10.815 -3.919 -1.919
137 3 3 0.241 0.941 15.003 -1.469 -1.108
138 3 3 0.926 1.000 5.802 -3.845 -0.530
139 3 3 0.565 0.526 10.578 -0.662 -0.667
140 3 3 0.106 0.997 18.327 -2.465 1.138
141 3 3 0.829 0.999 7.414 -2.291 0.543
142 3 2 0.343 0.541 13.372 -0.682 -0.907
143 3 3 0.952 1.000 5.162 -2.357 -0.275
144 3 3 0.226 1.000 15.290 -2.492 -0.499
145 3 3 0.667 1.000 9.418 -3.243 -1.797
146 3 3 0.728 0.999 8.705 -2.677 1.444
147 3 3 0.288 0.966 14.205 -2.081 1.781
148 3 3 0.000 0.696 37.997 -1.330 -1.063
149 3 3 0.444 1.000 12.026 -3.679 0.792
150 3 3 0.748 1.000 8.465 -2.881 -0.606
151 3 3 0.992 0.990 3.399 -1.604 -0.399
152 3 3 0.049 1.000 21.096 -4.092 1.854
153 3 3 0.953 1.000 5.152 -3.432 -0.134
154 3 3 0.001 1.000 32.183 -4.985 -0.527
155 3 3 0.824 1.000 7.478 -4.193 -0.219
80
TABLE C.3. Gill Set
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
1 1 3 0.290 0.850 4.978 2.865 0.435
2 1 1 0.013 0.473 12.600 -0.114 1.027
3 1 3 0.167 0.852 6.470 1.674 -1.202
4 1 3 0.260 0.568 5.277 1.241 0.762
5 1 1 0.436 0.514 3.782 0.668 1.078
6 1 1 0.755 0.510 1.893 -0.532 1.361
7 2 2 0.805 0.867 1.623 -2.015 -0.014
8 1 2 0.015 0.791 12.374 -1.954 2.391
9 1 2 0.613 0.896 2.679 -1.840 -0.837
10 1 3 0.894 0.592 1.099 0.936 -0.144
11 1 3 0.464 0.498 3.590 0.955 0.763
12 1 1 0.786 0.518 1.724 -0.248 1.085
13 2 2 0.767 0.879 1.832 -1.823 -1.199
14 1 1 0.892 0.487 1.116 -0.339 0.892
15 1 2 0.008 0.884 13.853 -1.644 -0.222
16 1 1 0.979 0.531 0.445 0.091 0.954
17 2 1 0.320 0.602 4.695 -0.002 1.352
18 1 2 0.849 0.408 1.373 -0.325 -0.903
19 1 2 0.698 0.519 2.203 -0.858 0.691
20 1 1 0.929 0.439 0.870 0.089 0.238
21 1 2 0.575 0.385 2.898 -0.220 -1.290
22 2 1 0.134 0.549 7.044 -0.398 1.118
23 2 2 0.770 0.785 1.812 -1.441 -0.826
24 1 3 0.889 0.528 1.132 0.807 0.151
25 2 3 0.720 0.509 2.085 0.250 -0.999
26 2 2 0.885 0.591 1.160 -1.147 0.786
27 2 2 0.616 0.517 2.662 -0.724 -0.417
28 2 2 0.057 0.828 9.158 -1.435 -2.579
29 1 2 0.893 0.432 1.110 -0.486 -0.163
30 1 2 0.012 0.630 12.800 -0.601 -2.970
31 1 2 0.232 0.541 5.593 -1.012 1.333
32 1 1 0.448 0.581 3.704 -0.458 1.916
33 1 1 0.851 0.438 1.360 0.163 0.241
34 1 2 0.537 0.454 3.127 -0.426 -0.831
35 1 1 0.545 0.450 3.077 -0.102 0.484
36 1 3 0.555 0.806 3.017 1.340 -1.388
37 2 2 0.565 0.681 2.960 -1.392 0.574
38 2 2 0.506 0.516 3.316 -0.611 -1.343
39 1 3 0.104 0.529 7.685 1.126 0.973
40 1 1 0.073 0.575 8.578 1.680 2.658
41 1 3 0.480 0.504 3.487 0.923 0.657
42 1 1 0.914 0.513 0.970 -0.233 1.001
43 2 2 0.974 0.730 0.493 -1.351 -0.191
44 1 1 0.582 0.610 2.859 0.708 1.804
45 2 3 0.678 0.361 2.317 -0.106 -0.562
81
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
46 1 2 0.895 0.575 1.098 -0.786 -0.520
47 1 3 0.443 0.679 3.737 0.727 -1.301
48 2 2 0.861 0.838 1.302 -1.965 0.448
49 1 1 0.845 0.431 1.398 0.127 0.192
50 1 3 0.944 0.525 0.754 0.995 0.495
51 1 1 0.867 0.440 1.264 -0.050 0.300
52 2 2 0.305 0.791 4.836 -1.824 0.605
53 1 2 0.984 0.631 0.379 -0.941 -0.538
54 1 2 0.980 0.712 0.432 -1.206 -0.350
55 1 3 0.679 0.585 2.309 1.195 0.426
56 2 2 0.452 0.785 3.673 -1.621 -0.055
57 2 2 0.934 0.624 0.835 -1.094 0.159
58 2 2 0.180 0.685 6.274 -1.619 1.214
59 2 2 0.201 0.965 5.975 -2.753 -1.108
60 2 2 0.136 0.911 7.005 -2.026 -1.460
61 2 2 0.953 0.376 0.690 -0.333 -0.364
62 1 1 0.086 0.431 8.163 0.002 0.483
63 1 1 0.893 0.538 1.107 0.389 1.037
64 1 3 0.981 0.609 0.416 0.841 -0.518
65 2 2 0.623 0.607 2.620 -1.130 0.424
66 2 2 0.768 0.657 1.827 -1.325 0.674
67 2 1 0.060 0.630 9.059 -1.005 2.654
68 1 3 0.471 0.475 3.544 0.659 0.403
69 2 3 0.973 0.562 0.506 1.059 0.320
70 2 2 0.292 0.761 4.953 -1.825 1.047
71 1 3 0.378 0.684 4.213 1.315 -0.101
72 1 1 0.354 0.532 4.405 0.987 1.476
73 2 3 0.678 0.489 2.313 0.681 0.277
74 1 1 0.769 0.523 1.816 0.048 0.948
75 2 2 0.627 0.927 2.599 -2.475 -0.073
76 2 1 0.713 0.639 2.123 0.174 1.687
77 1 1 0.555 0.464 3.016 0.034 0.535
78 2 2 0.110 0.841 7.539 -1.642 -1.310
79 1 1 0.952 0.482 0.693 0.754 0.777
80 2 2 0.533 0.888 3.153 -1.978 -0.776
81 2 2 0.663 0.665 2.398 -1.000 -1.177
82 2 2 0.534 0.569 3.144 -1.307 1.526
83 2 1 0.700 0.472 2.192 0.799 0.690
84 2 1 0.876 0.488 1.213 -0.178 0.649
85 2 2 0.733 0.766 2.017 -1.313 -1.176
86 2 2 0.718 0.475 2.098 -0.682 0.031
87 2 2 0.133 0.820 7.052 -1.407 -2.360
88 2 2 0.843 0.866 1.409 -1.909 -0.402
89 2 2 0.873 0.832 1.231 -1.814 -0.028
90 2 1 0.899 0.409 1.070 -0.428 0.257
91 2 2 0.375 0.780 4.235 -1.305 -1.715
92 2 2 0.017 0.954 12.098 -2.899 0.100
82
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
93 2 2 0.811 0.890 1.586 -1.945 -0.947
94 2 1 0.969 0.401 0.542 -0.056 -0.100
95 2 2 0.510 0.764 3.290 -1.284 -1.399
96 2 2 0.361 0.814 4.351 -1.824 0.237
97 2 2 0.965 0.451 0.583 -0.567 -0.124
98 2 2 0.784 0.489 1.738 -0.739 0.164
99 2 2 0.751 0.429 1.920 -0.591 0.225
100 2 2 0.593 0.593 2.796 -0.951 -0.274
101 3 3 0.006 0.421 14.491 0.162 -1.359
102 3 3 0.943 0.604 0.762 0.757 -0.753
103 3 3 0.946 0.503 0.739 0.769 0.196
104 3 3 0.631 0.791 2.577 1.452 -1.376
105 3 1 0.696 0.375 2.216 -0.066 -0.333
106 3 2 0.448 0.976 3.702 -2.866 -0.623
107 3 3 0.940 0.589 0.787 0.798 -0.524
108 3 1 0.727 0.519 2.050 0.689 0.897
109 3 1 0.867 0.434 1.268 -0.147 0.204
110 3 1 0.642 0.555 2.517 0.153 1.002
111 3 3 0.266 0.481 5.216 1.195 0.923
112 3 3 0.902 0.582 1.048 1.069 0.056
113 3 3 0.915 0.610 0.968 1.157 -0.025
114 3 2 0.992 0.741 0.269 -1.399 0.119
115 3 3 0.788 0.575 1.715 1.441 0.693
116 3 3 0.962 0.508 0.607 0.623 -0.142
117 3 3 0.256 0.876 5.323 2.463 -1.064
118 3 3 0.392 0.507 4.107 0.995 0.440
119 3 1 0.851 0.579 1.360 0.553 1.288
120 3 1 0.434 0.522 3.801 0.394 0.720
121 3 3 0.926 0.743 0.888 1.754 -0.302
122 3 3 0.115 0.885 7.430 2.942 -0.595
123 3 3 0.952 0.587 0.696 1.319 0.449
124 3 3 0.278 0.883 5.092 2.446 -1.200
125 3 3 0.415 0.577 3.933 1.937 1.298
126 3 3 0.167 0.913 6.458 2.309 -2.022
127 3 3 0.313 0.749 4.763 1.902 -0.242
128 3 3 0.672 0.373 2.349 -0.082 -0.910
129 3 3 0.149 0.504 6.765 0.158 -1.908
130 3 1 0.760 0.539 1.870 1.021 1.305
131 3 1 0.844 0.440 1.399 -0.073 0.206
132 3 3 0.704 0.652 2.173 1.389 -0.043
133 3 3 0.017 0.936 12.066 3.112 -1.544
134 3 3 0.904 0.754 1.040 1.640 -0.589
135 3 3 0.621 0.426 2.634 0.151 -0.708
136 3 3 0.408 0.521 3.985 1.040 0.408
137 3 3 0.871 0.652 1.245 1.170 -0.384
138 3 2 0.857 0.516 1.323 -0.839 0.604
139 3 3 0.899 0.508 1.067 0.800 0.192
83
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
140 3 3 0.757 0.795 1.886 1.988 -0.542
141 3 3 0.912 0.769 0.984 1.599 -0.818
142 3 3 0.957 0.622 0.657 0.953 -0.505
143 3 1 0.083 0.548 8.233 -0.721 1.881
144 3 1 0.981 0.418 0.419 0.079 0.005
145 3 1 0.229 0.406 5.625 -0.288 0.046
146 3 2 0.476 0.356 3.514 -0.170 -0.884
147 3 3 0.747 0.647 1.940 1.456 0.112
148 3 3 0.006 0.924 14.282 2.783 -1.678
149 3 3 0.957 0.579 0.651 0.814 -0.388
150 3 1 0.988 0.485 0.331 -0.047 0.574
151 3 1 0.359 0.644 4.363 0.425 1.637
152 3 2 0.030 0.497 10.685 -0.303 -1.967
153 3 2 0.087 0.372 8.127 -0.144 -0.958
154 3 3 0.003 0.779 15.762 2.367 -0.131
155 3 3 0.326 0.893 4.640 2.254 -1.670
TABLE C.4. All Measurements Set
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
1 1 1 0.068 0.970 23.834 0.167 1.749
2 1 1 0.007 1.000 31.717 1.639 2.586
3 1 1 0.956 1.000 7.069 0.736 2.741
4 1 1 0.204 1.000 19.226 1.343 3.366
5 1 1 0.018 0.998 28.665 -0.315 4.149
6 1 1 0.026 1.000 27.308 4.164 2.630
7 2 2 0.946 0.881 7.397 1.402 -0.574
8 1 1 0.578 0.910 13.309 1.832 0.910
9 1 2 0.721 1.000 11.435 1.804 -2.099
10 1 1 0.471 0.983 14.730 1.725 1.380
11 1 3 0.381 0.517 16.017 -1.076 2.336
12 1 1 0.292 1.000 17.456 1.659 2.614
13 2 2 0.729 0.997 11.331 2.643 -1.581
14 1 1 0.804 0.969 10.247 2.735 1.225
15 1 1 0.001 0.824 38.999 1.186 0.948
16 1 1 0.905 0.974 8.435 1.432 1.194
17 2 2 0.553 0.872 13.644 2.326 -0.583
18 1 1 0.393 0.983 15.841 1.137 1.370
19 1 1 0.474 0.949 14.696 1.726 1.079
20 1 1 0.960 0.988 6.900 1.935 1.430
21 1 1 0.494 1.000 14.418 2.051 2.423
22 2 1 0.627 0.694 12.681 1.300 0.103
23 2 2 0.201 0.999 19.293 1.119 -2.012
84
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
24 1 1 0.056 0.999 24.552 1.564 2.287
25 2 2 0.003 1.000 34.673 3.529 -2.183
26 2 2 0.333 0.877 16.767 0.992 -0.701
27 2 2 0.789 0.999 10.475 1.297 -1.985
28 2 2 0.156 0.913 20.427 2.837 -0.744
29 1 1 0.000 1.000 41.184 0.908 2.940
30 1 2 0.684 0.941 11.934 0.227 -0.796
31 1 2 0.250 0.949 18.240 2.522 -0.291
32 1 1 0.809 0.773 10.168 2.077 0.585
33 1 1 0.297 0.937 17.376 0.895 1.022
34 1 1 0.707 0.528 11.634 2.234 0.305
35 1 1 0.115 1.000 21.720 1.267 2.432
36 1 1 0.665 0.995 12.184 2.186 1.714
37 2 2 0.168 0.999 20.115 1.802 -1.978
38 2 2 0.519 0.990 14.082 2.280 -1.292
39 1 1 0.066 1.000 23.940 1.227 3.153
40 1 1 0.423 0.999 15.398 2.070 2.275
41 1 1 0.858 1.000 9.354 1.605 2.959
42 1 1 0.034 1.000 26.425 2.407 3.267
43 2 2 0.507 0.992 14.250 0.415 -1.871
44 1 1 0.855 0.962 9.416 -0.243 2.042
45 2 3 0.053 0.763 24.798 -1.275 -3.193
46 1 1 0.000 0.999 44.481 0.920 2.207
47 1 1 0.485 0.613 14.535 2.127 0.437
48 2 2 0.404 0.995 15.673 1.693 -1.542
49 1 1 0.302 0.999 17.284 1.427 2.296
50 1 1 0.913 1.000 8.258 1.292 3.008
51 1 1 0.023 0.637 27.734 -0.579 2.002
52 2 2 0.241 0.997 18.419 0.267 -2.840
53 1 2 0.866 0.866 9.211 2.892 -0.191
54 1 1 0.036 0.994 26.213 2.284 1.761
55 1 1 0.077 1.000 23.325 2.398 2.475
56 2 2 0.795 0.990 10.383 2.258 -1.275
57 2 2 0.976 0.994 6.224 0.752 -1.543
58 2 2 0.318 1.000 17.014 1.261 -2.385
59 2 2 0.257 1.000 18.100 1.658 -2.757
60 2 2 0.162 0.993 20.278 3.365 -1.392
61 2 2 0.265 1.000 17.950 1.414 -2.422
62 1 1 0.326 0.999 16.873 0.975 2.207
63 1 1 0.802 1.000 10.283 1.229 3.515
64 1 1 0.542 0.997 13.784 1.389 1.804
65 2 2 0.127 1.000 21.303 1.906 -2.638
66 2 2 0.893 0.999 8.690 2.109 -1.957
67 2 1 0.676 0.918 12.035 1.258 0.492
68 1 1 0.040 0.985 25.781 -0.522 3.543
69 2 2 0.121 0.998 21.537 1.010 -1.905
70 2 2 0.266 0.599 17.941 2.187 -0.237
85
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
71 1 1 0.776 0.958 10.663 1.011 1.072
72 1 1 0.538 0.998 13.831 2.101 2.022
73 2 2 0.102 0.602 22.212 0.375 -0.447
74 1 1 0.405 0.977 15.661 0.334 1.484
75 2 2 0.582 0.999 13.268 2.444 -1.975
76 2 2 0.016 0.909 29.056 -0.119 -2.182
77 1 1 0.868 0.999 9.178 1.844 2.296
78 2 2 0.557 1.000 13.590 2.944 -2.517
79 1 1 0.977 0.980 6.158 0.023 1.781
80 2 2 0.818 0.989 10.030 1.238 -1.301
81 2 2 0.842 0.999 9.639 1.326 -2.039
82 2 2 0.775 0.783 10.682 2.320 -0.371
83 2 2 0.248 0.684 18.294 1.218 -0.388
84 2 2 0.508 0.991 14.238 0.821 -1.476
85 2 2 0.423 0.988 15.400 1.229 -1.316
86 2 2 0.556 0.983 13.595 1.997 -1.173
87 2 2 0.625 1.000 12.709 0.909 -3.336
88 2 2 0.881 1.000 8.936 2.151 -2.333
89 2 2 0.567 1.000 13.461 1.425 -3.801
90 2 2 0.074 0.999 23.514 0.850 -2.348
91 2 2 0.784 0.988 10.553 0.030 -2.394
92 2 2 0.033 0.998 26.506 2.714 -1.810
93 2 2 0.177 0.999 19.857 3.108 -2.032
94 2 2 0.765 0.995 10.823 2.030 -1.458
95 2 2 0.816 0.998 10.060 2.127 -1.729
96 2 2 0.980 0.986 5.985 1.340 -1.196
97 2 2 0.786 0.966 10.523 0.047 -1.767
98 2 2 0.211 0.997 19.067 0.330 -2.638
99 2 2 0.035 0.996 26.250 2.804 -1.592
100 2 2 0.986 0.995 5.589 1.859 -1.471
101 3 2 0.002 0.964 35.268 0.008 -0.764
102 3 3 0.065 1.000 24.033 -2.678 -0.908
103 3 3 0.490 1.000 14.469 -2.963 1.580
104 3 3 0.916 0.992 8.178 -1.683 0.279
105 3 3 0.220 0.938 18.860 -1.314 0.299
106 3 3 0.013 0.959 29.661 -1.485 -0.258
107 3 3 0.245 1.000 18.341 -2.752 -0.521
108 3 3 0.508 0.999 14.235 -2.196 -0.610
109 3 3 0.243 0.991 18.377 -1.994 0.893
110 3 3 0.046 0.560 25.317 -0.933 0.561
111 3 1 0.041 0.495 25.704 -0.869 0.596
112 3 3 0.201 0.996 19.292 -2.590 -2.259
113 3 3 0.237 1.000 18.503 -2.620 0.201
114 3 3 0.112 1.000 21.860 -3.443 -0.356
115 3 3 0.944 1.000 7.458 -3.143 1.639
116 3 3 0.862 1.000 9.288 -2.977 -0.453
117 3 3 0.325 1.000 16.904 -4.686 -0.437
86
Case Actual Predicted Typicality Posterior D2 Discriminant Score
Number Group Group Prob. Prob. Function 1 Function 2
118 3 3 0.650 1.000 12.384 -3.383 -1.193
119 3 3 0.189 1.000 19.570 -3.701 -1.673
120 3 3 0.342 1.000 16.617 -3.491 -0.924
121 3 3 0.280 0.997 17.682 -2.435 1.305
122 3 3 0.141 1.000 20.868 -3.438 1.218
123 3 3 0.761 0.999 10.888 -2.904 1.623
124 3 3 0.584 1.000 13.235 -3.296 0.221
125 3 3 0.546 0.998 13.729 -2.084 0.308
126 3 3 0.942 1.000 7.513 -4.057 -0.283
127 3 3 0.103 0.999 22.194 -2.407 -0.869
128 3 3 0.962 1.000 6.842 -2.707 0.452
129 3 3 0.464 1.000 14.831 -2.694 0.248
130 3 3 0.870 1.000 9.148 -2.428 0.478
131 3 3 0.576 1.000 13.339 -3.315 -1.375
132 3 3 0.649 1.000 12.394 -5.099 1.192
133 3 3 0.002 1.000 35.780 -3.030 -0.419
134 3 3 0.875 0.996 9.055 -2.059 0.846
135 3 3 0.783 0.999 10.557 -2.186 -0.095
136 3 3 0.150 1.000 20.590 -4.103 -0.866
137 3 3 0.396 0.943 15.799 -1.269 0.192
138 3 3 0.480 1.000 14.607 -3.600 -1.222
139 3 3 0.752 0.925 11.004 -1.162 -0.492
140 3 3 0.117 1.000 21.659 -2.976 1.482
141 3 3 0.823 0.995 9.953 -1.998 0.823
142 3 3 0.682 0.830 11.966 -0.987 -0.600
143 3 3 0.596 1.000 13.086 -3.032 -1.202
144 3 3 0.332 1.000 16.777 -3.069 -0.039
145 3 3 0.048 1.000 25.145 -3.563 -1.488
146 3 3 0.699 0.999 11.736 -2.349 0.588
147 3 3 0.089 0.996 22.756 -2.105 0.606
148 3 3 0.000 0.681 49.885 -1.534 -1.212
149 3 3 0.933 1.000 7.754 -3.713 0.844
150 3 3 0.855 1.000 9.407 -3.258 -1.307
151 3 3 0.924 0.988 7.987 -1.540 0.151
152 3 3 0.030 1.000 26.870 -3.428 -0.028
153 3 3 0.885 1.000 8.853 -2.516 -0.142
154 3 3 0.002 1.000 35.838 -3.494 -0.165
155 3 3 0.674 1.000 12.062 -4.561 -0.024