JOURNAL OF PERSONALITY ASSESSMENT, 1995,64(1), 21-50
Copyright
Q
1995, Lawrence Erlbaum Associates, Inc.
Domains and Facets: Hierarchical
Personality Assessment Using the
Revised
NEO
Personality Inventory
Paul
T.
Costa,
Jr.
and
Robert
R.
McCrae
Gerontology Research Center
National Institute on Aging, NZH
Baltimore,
MD
Personality traits are organized hierarchically, with narrow, specific traits com-
bining to define broad, global factors. The Revised NEO Personality Inventory
(NEO-PI-R; Costa
&
McCrae, 1992c) assesses personality at bo1.h levels, with
six specific facet scales in each of five broad domains. This article describes
conceptual issues in specifying facets of a domain and reports evidence on the
validity of NEO-PI-R facet scales. Facet analysis-the interpretation of a scale
in terms of the specific facets with which it correlates-is illustrated using
alternative measures of the five-factor model and occupational scales. Finally,
the hierarchical interpretation of personality profiles is discussed. Interpreta-
tion on the domain level yields a rapid understanding of the individual; inter-
pretation of specific facet scales gives a more detailed assessment.
The Revised NEO Personality Inventory (NEO-PI-R; Costa
&
McCrae,
1992c) is a 240-item questionnaire designed to operationalize the five-factor
model of personality (FFM; Digman, 1990; McCrae
&
John, 1992). Over the
past decade, the FFM has become a dominant paradigm in personality psy-
chology, yet most attention has been focused on the
EIig Five factors them-
selves, to the neglect of the specific traits that define these factors. In this
article we emphasize the facet scales of the NEO-PI-R, discussing the logic
behind their development, the evidence of their discriminant validity, and
their utility in interpreting the nature of other personality scales. We also
address the complexities of interpreting profiles from an
instrument that
provides both a global and a detailed assessment of an individual's personal-
ity. The first part of the article may appeal chiefly to the personality theorist,
22
COSTA AND McCRAE
the middle part to the researcher, and the last to the clinician interested in the
assessment of individuals.
A HIERARCHICAL MODEL OF PERSONALITY
STRUCTURE
In an article on the cross-cultural invariance of personality structure,
Paunonen, Jackson, Trzebinski, and Forsterling (1992) concluded that "If
one desires a broad overview of personality dimensions, we regard the
five-factor model as most promising, but if one's theoretical or pragmatic
requirements are for a more differentiated, detailed perspective, perhaps
other measurement models should be considered" (p. 455). The same senti-
ment has been expressed by many others (Briggs, 1989; Buss, 1989;
Mershon
&
Gorsuch, 1988), all of whom noted the greater precision of
measurement, if narrower focus, of more specific traits.
What these critics of the FFM have failed to do, however, is to agree upon
which specific traits should be measured. Many alternative sets of primary
traits have been proposed, from the 16 factors of
Cattell to the 20 Murray
needs measured by Jackson's (1984) Personality Research Form. Although
most of these scales can be interpreted in terms of the FFM, they were
constructed without reference to it and do not represent a systematic carving
up of the five-factor space. In this article we describe an approach to the
assessment of traits at both general and specific levels explicitly guided by
the FFM: The domain-and-facet approach of the
NEO-PI-R.
The Logic of Domains and Facets
As Goldberg (1993) noted, there is a long tradition of identifying different
levels of specificity in personality trait assessment. Conceptually, this is
usually illustrated by the combination of discrete behaviors to form specific
traits, and the combination of groups of covarying traits to form broad
dimensions of personality. Factor analysts such as Guilford, Cattell, and
Eysenck all adopted such a hierarchical model, although Guilford and Cattell
emphasized the lower level traits and Eysenck the higher. In the usual factor
analytic approach, test items were factored, usually using oblique rotations,
and the obtained factor scores were then factored themselves to yield second
order factors. Third order factors were occasionally reported.
In practice, this bottom-up scheme presented several difficulties. Most
important was the specification of the initial pool of items. What should be
included? Even large item pools may omit important aspects of personality.
For example, McCrae, Costa, and Piedmont (1993) reported that there are
relatively few items in the California Psychological Inventory that measure
Agreeableness, and
J.
H.
Johnson, Butcher, Null, and
K.
N. Johnson's (1984)
DOMAINS AND FACETS
23
item factor analysis of the Minnesota Multiphasic Personality Inventory
(Hathaway
&
McKinley, 1983) found no factors related lo Conscientiousness.
The lexical approach, in which the body of trait names in the natural
language has been adopted as an exhaustive enumeration of traits, has
proven to be the most fruitful guide to a comprehensive model of personal-
ity; it was in analyses based on trait terms that the FFM was first discovered.
But the lexical approach has distinct limitations as the basis of a hierarchical
model of personality, first because some specific traits are not well repre-
sented in the natural language (McCrae,
1990), and second because trait
terms are found at every level of breadth (John, Ha~mpson,
&
Goldberg,
1991), from extremely narrow (e.g., sanctimonious, sedentary, sirupy) to
extremely broad
(e.g., kind, weak, able). Broad terms naturally covary with
many narrower terms, whereas narrower terms may inot covary with each
other. The result is that when representative lists of trait adjectives are
factored, the broader terms account for the lion's share of the covariance,
and only five broad factors typically emerge (Goldberg, 1990).
These problems are minimized by a top-down approach to hierarchical
assessment. In the program of research that lead to the development of the
NEO-PI-R, we began by looking for the broadest and most pervasive themes
that recurred in personality measures. Eysenck's Extraversion (E) and Nm-
roticism
(N)
had already been identified as the Big Two by Wiggins (1968),
and we proposed that Openness to Experience (0) also qualified as a major
dimension of personality (Costa
&
McCrae, 1978).
A
few years later we
recognized the need for Agreeableness (A) and Conscientiou~sness (C).
Rather than use the term factors, which might apply to any level in Ihe
hierarchy, we chose to call
N,
E, 0, A, and C domains, a term defined as "a
sphere of concern or function" (Morris, 1976, p. 389). Intellectual curiosity,
need for variety, and aesthetic sensitivity all concerned some aspect of
experiencing the world, and thus belonged in the domain of
0. Although tlhis
terminology is somewhat unusual, it is not unparalleled: A.bout the same
time, and quite independently, Digman (1979) presented a paper entitled
"The Five Major Domains of Personality Variables: Analyses of Personality
Questionnaire Data in the Light of the Five Robust Factors ]Emerging from
Studies of Rated Characteristics."
We regarded domains as multifaceted collections of specific cognitive,
affective, and behavioral tendencies that might be grouped in many different
ways, and we used the term facet to designate
the lower level traits corre-
sponding to these groupings.' Our working metaphor was the mathematical
set, which could be divided into subsets by selecting different combinations
of elements.
his
usage should be distinguished from that of Guttman (1954), who used the term
facet
to
refer to one of several conceptual factors that, when crossed, yielded a set of variables. A
well-known example is Guilford's (1967) structure of intellect model, which uses Operation,
Content, and Product as facets in Guttman's sense.
Guidelines for Faceting Domains
Consider the set of attributes that together define the domain of
N,
such as
chronic tendencies to feel tense, worried, irritable. There are many possible
ways to group these attributes into what we might consider specific traits
(see Figure 1). We could treat them singly, recognizing for example the
difference between tension and apprehension, as Spielberger (1972) did; or
we might combine these two with other traits like
shy
and
guilt-prone
to form
a broader anxiety cluster that might be contrasted with depression and
hostility clusters, as Zuckerman and Lubin (1965) suggested.
Hofstee, De Raad, and
Goldberg (1992), noting that many traits in the
lexicon have appreciable loadings on two of the five basic factors, have
suggested that facets can be identified by their location on the ten cir-
cumplexes formed by pairs of the five factors. Traits such as
irritable
and
touchy,
which are primarily located in the domain of
N,
might be grouped
together because they share a secondary loading on low A.
Each of these ways of identifying specific traits within the domain of
N
is
reasonable, but the differences among them explain why there is so little
consensus on lower level traits (Briggs, 1989). In fact, with only twelve
elements in a set, there are 4,094 different proper,
non-null subsets. The
ways in which a domain as broad as
N
could be subdivided is virtually
limitless.
This is not to say that the identification of specific facets is not useful.
Even if there is an element of arbitrariness in the way in which a domain is
subdivided, there are still good reasons to make distinctions. Any meaning-
LESS
MEANINGFUL
MORE
MEANINGFUL
Hclplcss
FIGURE
1
An illustration of how traits in the domain of Neuroticism might be
grouped into facets. Overlapping groupings are less meaningful than mutually exclusive
groupings.
DOMAINS AND FACETS
25
ful specification of facets should provide more information than the undif-
ferentiated global domain scale, and some specifications are more meaning-
ful than others.
Perhaps most obviously, facets should represent the more closely coviury-
ing elements within the domain, not arbitrary combinalions of elements; and
they should be mutually exclusive, with each element in the domain assigned
to only a single facet, as shown in Figure 1. Both these goals are facilitated
by factor analyses of items within the domain, because factor analysis
identifies discrete clusters of covarying items. This kind of item factor
analysis was one of the steps in the
development of NEO-PI-R facets
(Costa, McCrae,
&
Dye, 1991).
Ideally, all facets should be of comparable scope and breadth in content.
It would make little sense to insist upon fine distinctions in some aspects of
personality while allowing only coarse distinctions elsewhere. Narrower
facets are desirable insofar as they measure specific traits
with great fidelity,
but if the full domain is to be covered, the number of narrow facets needed
might become unmanageable.
Exhausting the domain would seem to be desirable: Just as the FFM is
intended to be a comprehensive taxonomy of all personality traits, so each
set of facets might aspire to be
a
comprehensive specification of the contents
of a domain. But this requirement can be problematic. For example, hypo-
chondriasis has been regarded as
a
facet of M (Eysenck
&
Wilson, 1976), but
we deliberately omitted somatic complaints from the NEO-PI-R because we
wanted an uncontaminated measure of N to predict health complaints. It is
sometimes difficult to know the boundary between a domain of personality
and its external correlates.
Finally, the facets of each domain should be as consistent as possible with
existing psychological constructs. It is in combing the literature that
we
identify traits relevant to each domain, and, where empirically supportable,
it makes sense to retain these initial constructs. They are familiar to person-
ality psychologists, and their previous use suggests that they will have some
utility. The MEO-PI-R N fwets of Anxiety, Angry Hostility, Depression,
Self-Consn;iousness, Impulsiveness, and Vulnerability all hawe clear roots in
the psychological literature (Costa
&
McCrae, 1980).
A
Complication: Overlapping Domains
The elegance of a purely hierarchical model of personality structure is
marred by the fact that the domains themselves are not mutually exclusive.
That is to say, there are traits that appear to lie within two or more domains.
In the language of factor analysis, personality cannot be adequately
cle-
scribed by simple structure; some traits load on more tlhan one factor.
This phenomenon is most clearly illustrated by the Interpersonal Cir-
cumplex (Leary, 1957; Wiggins, 19791, a circular arrangement of traits