Professional Documents
Culture Documents
net/publication/336170377
Facet Theory
CITATIONS READS
2 801
1 author:
Samuel Shye
Hebrew University of Jerusalem
71 PUBLICATIONS 1,604 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Samuel Shye on 25 February 2020.
Facet Theory
Facet Theory (FT) is a general approach to empirical sciences that advocates the integration of
formal analysis of research contents with multivariate statistical data analysis, for the purpose of
formulating and testing laws in those sciences. FT was initially proposed by Louis Guttman as a
conceptualization with no clearly testable hypotheses, on the one hand; and those involving
statistical assumptions or intricacies not warranted by the nature of the observations, on the other
hand [22,23,49]. FT is especially geared to research problems with a large number of interacting
variables, hence it has been found useful mostly in the social, psychological and life sciences.
More generally, FT is concerned with systems that are sufficiently complex to be considered
"behavioral". For the study of such systems, FT offers techniques of formal research design
and measurement (Multiple Scaling by Partial Order Scalogram Analysis by base Coordinates--
POSAC). This article reviews the development, principles and techniques of FT and illustrates
1. Content FT: Concepts as Spaces. The empirical observation of N subjects with respect to
their responses to each of n questions results in an Nxn data matrix wherein element aij records
subject i's response to the j-th question. The system of observations is a mapping PxQ → R from
2
the cartesian product of population P of subjects and the set Q of questions, into the set R of
E.g., in the above, P is the population facet and Q is the question facet.
Example 1: Suppose members of a population P are presented with the following question:
q1: What personal contribution would you be willing to make in order to help preserve the
forests? (possible responses: 1.pay, 2.do organizational work, 3.plant trees, 4.none of the
above).
If more than one choice out of the first three is to be permitted, q1 may be broken down into
Given the overall structure of the questions, the mapping is, in effect, PxQ1 → R where Q1 is the
set {pay, do organizational work, plant trees} which differentiates among the three questions.
Additional environmental concerns such as the preservation of whales, or of water quality, may
be included to expand the observational design. The new mapping would be:
PxQ1xQ2 → R, where Q2={forests, whales, water quality}. (Following the inclusion of Q2, Q1
is modified by replacing its 3rd element, 'plant trees', by the more general 'act directly'.) Each of
the facets Q1 and Q2 classifies the expanded set of nine questions by a content criterion; hence
Q1 and Q2 are called content facets. The mapping sentence [20, 34, 49] for the observational
Q1: Contribution
pay
Is citizen pi(piP) willing to do _ org ' l _ work in order to
act − directly
forests
help preserve whales → {y1,n1,y2,n2,...,y9,n9}
water − quality
where the range contains the possible responses y (yes), and n (no), for each of the nine
questions. E.g., that citizen p23 is not willing to act directly to preserve whales would be recorded
by mapping the point [p23, act directly, whales] of the domain into the appropriate point (say, n8)
Mapping sentences (and the investigated contents) may be expanded systematically in two ways:
by extension--adding a new element to an existing facet (e.g. 'air quality' may be added to Q2); or
by intension--adding a new facet having at least two elements (e.g. in the above mapping
Early FT was cocerned with reproducing the observed individual responses from derived
proposed by Guttman to be the criterion for the completeness of the facet design [19]. However,
specific scores are too volatile an aspect of data. A more promising direction seemed to be the
estimation of pairwise correlations from content considerations. Work by Foa [14, 15] and by
Guttman [19] led to the formulation of the contiguity hypothesis, which guided facet theorists for
many years: The correlation between two variables increases with the similarity between the
facet elements defining them. Here, 'contiguity' refers to proximity (semantic and correlational)
Indeed, the origins of FT may be traced back to a correlational theory, the simplex [18], which
was proposed as a revision (in lieu of the alternative revision offered by multiple-factor analysis)
of Spearman's general factor theory of intelligence. The simplex theory for mental ability tests
hypothesizes as follows: Let the sequence of n tests t1,t2,...,tn be simply ordered with respect to
their complexity (a psychological-substantive content facet). Then, for every three tests ti, tj, tk,
(i<j<k) the partial correlation rik.j=0. This implies that in the formula for partial correlation the
-log rik = (-log rij) + (-log rjk). Letting the distance between any two tests ti, tj be dij = - log rij, we
and the tests can be represented as points on a one dimensional euclidean metric space so that
Example 2. Let rij = 1/4, rjk = 1/8. The simplex hypothesis (1) specifies that rik equals their
product:
5
Taking the log (for simplicity, log2 rather than loge) of (3) and multiplying by -1 we get: 2 + 3 =
...*----ti----*----tj----*----*----tk----*----...
<----2----><-------3------->
<------------5------------->
FT takes the line to represent a content facet, i.e., an underlying conceptual differentiation (e.g.,
test complexity facet). Eq.(2), with base 2 replacing e, reproduces the observed correlations from
the distances. Once a simplex theory has been established, correlations can be predicted from
Ordinal simplex and other topological configurations. Since the metric version of the
simplex, illustrated above, was too stringent for psychological data, ways to "soften" the model
were proposed [18, 36, 37, 38]. The ordinal simplex, e.g., requires that inter-test distances along
the line would reproduce only the rank order of the n(n-1)/2 correlations between the n variables.
But in cases where facet analysis can predict order among facet elements only, and not among
their pairwise distances, a weaker condition suffices for mapping tests on the line--and a weaker
Additional correlational structures were proposed and discovered in empirical data. The
6
circumplex was defined as the pattern of circularly ordered tests and was identified, e.g., in the
study of color perception [36, 38]. The radex was defined as a 2-dimensional configuration of
tests, combining concentric circumplexes with simplexes emanating from a common origin [18,
1, 53]. These parametrized formulations, too, gave way to the less stringent ordinal or nonmetric
Multidimensional Scaling (MDS). For data analysis the problem was this: Given a set of n
objects (typically variables) with a matrix of pairwise similarity coefficients, rij (typically
correlations) between them, find a mapping of the objects into a suitable geometric space, where
dij is the distance in that space between objects i and j, such that:
Further specifications on the mapping (e.g., the kind of distance function) as well as variations of
Following work by coombs [8, 9, 10] on interstimulus similarity and pair comparisons,
algorithms for solving this problem were proposed by Kruskal [26, 27], Guttman [21] and others.
Guttman's Smallest Space Analysis (SSA) was programmed [30] to attain, for a given
dimensionality, the best fit of the inter-object distances to the given coefficient matrix. These
computer programs have been successfully applied to research data in psychology, sociology and
Subuniverses in SSA Concept space. Experience with mapping sentences has led to
interpreting points in the 'question space' Q as defining classes of questions rather than single
questions.
Example 3. In Example 1, the point [pay, water quality] may represent the subset of all specific
7
questions that are classifiable into 'pay' by Q1 and into 'water quality' by Q2; e.g.: "Are you
willing to pay higher taxes in order to help preserve the quality of river water?" In principle,
then, there could be infinitely many questions that belong to [pay, water quality]. This led to
redefining Q as the set of all questions (observed or not) that conform to the mapping sentence
semantic structure (including its range facet). When the questions all have a common-meaning
range (CMR), Q is referred to as the "content universe". Hence, the set of all questions with a
given CMR define a content universe such as intelligence [35, 24], commitment to work [2] or
variables, was replaced by the general regional hypothesis [42, 45] which concerns contiguity, in
(i) The SSA concept representation space can be partitioned into connected regions each
(ii) If two classes of variables have the same facet designation in all but one, say, the h-th
facet, and if the two distinguishing facet elements, qhl and qhl+1 are conceptually specified
The challenge for researchers in substantive domains, then, is to identify the empirically relevant
The re-interpretation of Q as an abstract content space has led to interpreting it, as well as each
of its content facets, as a content continuum. Thus in the discovery of the radex of attitudes
The order specification for each of the facets actually holds for a continuum rather than just
for a list of discrete entities. The elements exhibited ... are only discrete points or perhaps
intervals on those continua which correspond to words existing in our language (p.7).
Evidence for lawfulness in complex behavioral data has been accumulated [2, 5, 7, 16, 24, 31,
35, 43, 50]. Simple, replicable partition patterns were defined in n-spaces [42]. Thus, in 2-space
(i) Partition into ordered stripes, by parallel straight lines (a 2-d generalization of the simplex);
(ii) Partition into circularly ordered sectors, by radii emanating from a common center (a 2-d
Content facets that are found to conform to these partitions are called respectively, axial, angular
and radial facets. Regional hypotheses are tested with respect to a given partition pattern. In
practice, specific content facets (e.g. 'life areas') have been identified, that play a similar role
(e.g., angular) across changing experimental contexts and even across different research domains
[28]. The computerized procedure for testing regional hypotheses, Faceted SSA [48, 49, 4],
depicts the concept space with observed variables marked as points, and then finds, for a pre-
specified facet, the best partition of the concept space. For that partition it computes the Facet
A particular content universe may be decomposed into a number of independent facets. Thus, the
radex is a possible combination of a radial facet and an angular facet. The duplex is a
combination of two axial facets; the cylindrex--a combination in 3-space of the radex and an
axial facet, etc etc. (For a systematic description of partitions patterns see [42].)
Example 4. For the environmental attitude space of Example 1, the following is a possible
9
hypothesis: Facet Q1 is radial and Facet Q2 is angular, and together Q1xQ2 form a radex.
A rationale for this yet untested hypothesis is that the environmental concerns (Q2) are likely to
be circularly ordered, as has been found in other investigated domains, while kinds of
contribution (Q1) would be linearly ordered with pay in the inner disk (where the environmental
concerns are least differentiated); organizational work, in an intermediate ring; and direct action
(being the most specific to the environmental concern), in the outermost ring, the ring that allows
FORESTS
act directly
organizational work
w
pay
WHALES
WATER
QUALITY
10
Modern FT: A New Scientific Imagery. Given an investigated behavioral concept (such as
intelligence, positive attitude towards an object, adjustability etc) that may be attributed to
investigated subjects, FT conceives of the concept as the semantic space of all the variables that
assess it. In this imagery, each of the variables that comprise the concept is represented by a
point in an analog geometric space. The finite set of observed variables is but a sample from the
entire set compising the concept. If that sample is taken with the aid of a mapping sentence, it is
"stratified", relative to the a priori content facets. SSA enables inferences about the structure of
the concept from the sample of observed variables to the "population" of variables that comprise
the concept. This portrayal is encapsuled in the basic assumption of modern FT:
The Continuity Principle (Shye). The mapping of the universe of variables to a connected
I.e., every point in the concept space represents a possible variable of that concept. The spatial
imagery of concepts implies that spatial orientation between regions--rather than correlations
between variables--is the way FT assesses affinity between conceptual components. Moreover,
partitionability of the concept-space may now be regarded as a new kind of statistic, whose
"values" are the particular partition patterns. Being a more general ("softer") aspect of the data
than correlations, or even correlation ranking, partitionability leads to more stable lawfulness and
theories (e.g., in intelligence research [35, 50, 51, 46], achievement motive research [43, 13]; and
FT and Complex Systems. FT has been regarded as a meta-theory or as scaffolding for the
construction of general scientific theories. Its application to the study of general action system
theory has proved useful for identifying the recursive "functioning mode facet" {expressive,
numerous studies this facet has been found to play an angular role, in accordance with theoretical
predictions. Faceted action system theory may well tie in with certain aspects of deterministic
chaos theory. In both theories, regularities are sought not in the phenomena immediately
observed but in their transformed abstract space (concept space and phase space resp.); and in
range (CMR) represent well the investigated content universe, they may be used for measuring
subjects with respect to that concept. Let A1,...,An be the ranges of the n observed variables, each
Aj ordered from high to low with respect to the common meaning; and let A = A1xA2x...xAn be
the cartesian product of all the range facets, Aj (j=1...n). A system of observations is a mapping
from the observed subjects P to A: P → A i.e., each subject pi gets a score from each Aj (j=1...n):
pi [ai1,ai2,...,ain] a(pi). The point a(pi) in A is also called the profile of pi, and the set A’
(A’A) of observed profiles is called scalogram. FT's solution to the measurement problem is
based on the observation that the essential measurement information is embodied in the relations
of comparability and incomparability, defined between observed profiles thus: Two different
profiles ai=[ai1,ai2,...,ain] and aj=[aj1,aj2,...,ajn], are comparable, aiSaj, (with ai greater than aj, ai>aj)
if and only if aik ajk for all k=1...n, and aik' ajk' for some k'. Two different profiles are
incomparable, ai$aj, if neither ai>aj nor aj>ai. A, and therefore its subset A', form a partly ordered
set [3].
12
Measurement requires a minimal coordinate space X, that would provide yardsticks into which
Definition (Shye, [40, 44]). The p.o. dimensionality of scalogram A' is the smallest m (m
n) for which there exist m facets X1...Xm (each Xi is ordered) and there exists a 1 - 1
mapping Q:X'→A' from X' (X' X=X1x...xXm) to A' such that a>a' if and only if x>x'
The coordinate scales, Xi (i=1...m) represent underlying fundamental variables whose meanings
must be inferred in any specific application. The well known Guttman scale [17,52] (example:
1111, 1121, 1131, 2131, 2231, 2232) is simply a 1-d scalogram, i.e. one all of whose profiles are
comparable. In FT, only a Guttman scale, if confirmed, justifies measurement by a single score.
The Guttman scale (which orders subjects in a unidimensional measurement space) must not be
confused with the simplex (which orders variables in a unidimensional content space).
The procedure of identifying and interpreting the coordinate scales X1...Xm is called multiple
scaling. Multiple scaling is facilitated by Partial Order Scalogram Analysis by base Coordinates
(POSAC) for which algorithms and computer programs have been devised [44, 4]. In practice, a
particular dimensionality is attempted and a solution that best accommodates condition (12) is
sought. The POSAC/LSA program finds an optimal solution in 2-d coordinate space, then goes
on to analyze by Lattice Space Analysis (LSA) the role played by each of the variables in
structuring the POSAC 2-space, thereby facilitating interpretation of the derived coordinate
scales, X1, X2 [40, 44]. Recent developments include the algorithms for computerized
partitioning of the POSAC space by the range facet of each variable, which induces natural
intervals on the coordinate scales, X1, X2, [49, 4]. Applications of POSAC/LSA can be found,
3. Concept/Measurement Space Duality. Based on the same data matrix, Faceted SSA concept
space and POSAC measurement space are mathematically related. Proved relationships rely on
the introduction of a new kind of coefficient, E*, the coefficient of structural similarity [40,44].
While E* assesses pairwise similarity between variables, it does depend on variations in the
remaining n-2 variables observed. That is, in the spirit of FT, E* depends on the sampled
special version of SSA with E* as the similarity coefficient, and with lattice ("city block")
distance function. This suggests a plausible solution to a Guttman’s challenge [22]: What
correlation coefficient should one use in mapping the content space? E*, which links the two
4. FT in Science and Statistics Today. Aspects of FT have made their way to the current
mainstream of social science and psychology. Examples are the Guttman Scale, the cartesian
design of questionnaires and the notion of facet (e.g., in Guilford's cube of intelligence); and the
widespread use of SSA (MDS), recently included in commercial statistical software packages.
Increasingly, statisticians and research scientists are becoming familiar with modern FT as a
References
[1] Anderson, T. W. (1959). Some stochastic process models for intelligence test scores. In K. J.
Arrow, S. Karlin, and P. Suppes (eds.), Mathematical Methods in the Social Sciences. Stanford,
[2] Aranya N. Jacobson, D. & Shye, S. (1976). Smallest Space Analysis of potential work
[3] Birkhoff, G. (1948) Lattice Theory (revised ed.). New York: American Mathematical
[4] Borg I. & Shye S. (1995). Facet Theory: Form and Content. Thousand Oaks CA: Sage.
[5] Borg I. & Leutner D. (1983). Dimensional models for the perception of rectangles.
[6] Canter D. (Ed.). (1985). Facet Theory: Approaches to Social Research. New York: Springer.
[8] Coombs, C. H. (1954). A method for the study of interstimulus similarity. Psychometrika, 19,
183-194.
[11] Coombs, C. H. & Kao, R. C. (1955). Nonmetric factor analysis. Engng. Res. Bull. Ann
[13] Elizur, D., (1986). Achievement motive and sport performance. International Review of
[14] Foa, U. (1958). The contiguity principle in the structure of interpersonal relations. Human
[15] Foa, U. (1965). New developments in facet design and analysis. Psychological Review, 72,
262-274.
[16] Galinat, W. & Borg, I. (1987). On symbolic temporal information: beliefs about experience
15
[17] Guttman, L. (1944). A basis for scaling qualitative data. American Sociological Review,
9:139-150.
[18] Guttman, L. (1954). A new approach to factor analysis: the radex. In P.F. Lazarsfeld (ed.)
Mathemetical Thinking in the Social Sciences. New York: Free Press, 216-257.
[19] Guttman, L. (1959a). Introduction to facet design and analysis. Proceedings of the Fifteenth
[20] Guttman, L. (1959b). A structural theory for intergroup beliefs and action. American
[21] Guttman, L. (1968). A general nonmetric technique for finding the smallest coordinate
[22] Guttman, L. (1977). What is not what in statistics. The Statistician, 26, 81-107.
[23] Guttman, L. (1982). What is not what is theory construction. In R Hauser D Mechanic & A
Haller (eds.) Social Structure and Behavior. New York: Academic Press 331-348. Reprinted in
Levy, S. (ed) Louis Guttman on Theory and Methodology: Selected Writings. Aldershot:
Dartmouth.
[24] Guttman, L. and Levy, S. (1991). Two structural laws for intelligence tests. Intelligence, 15,
79-103).
[25] Hox, J.J., Mellenbergh,G.M. and Swanborn, P.G. (Eds.) (1995). Facet Theory: Analysis and
[28] Levy, S. (1985). Lawful roles of facets in social theories. In D. Canter (ed.) Facet Theory:
[29] Levy, S. (Ed.) (1994). Louis Guttman on Theory and Methodology: Selected Writings.
Aldershot: Dartmouth.
[30] Lingoes, J.C. (1973). The Guttman-Lingoes Nonmetric Program Series. Ann Arbor, MI:
Mathesis Press.
[31] Marsden, P.V. & Laumann, E.O. (1978). The social structure of religious groups: a
replication and methodological critique. In S. Shye (ed.) Theory Construction and Data Analysis
[32] Mevorach-Levy & Shye, S. (1993). Faceted systemic theory of distributive justice: quality
and utility, too. Proceedings of the fourth International Facet Theory Conference, Prague.
[33] Russett, B. & Shye, S. (1993). Aggressiveness, involvement and commitment in foreign
policy attitudes. In D. Caldwell & T.J. McKeown (eds.), Diplomacy, Force and Leadership:
[34] Schlesinger, I.M. (1978). On some properties of mapping sentences. In S. Shye (ed.) Theory
Construction and Data Analysis in the Behavioral Sciences. San Francisco: Jossey-Bass.
[35] Schlesinger, I. M. & Guttman, L. (1969). Smallest Space Analysis of intelligence and
[36] Shepard, R.N. (1962). The analysis of proximities: multidimensional scaling with an
unknown distance functio. Part I: Psychometrika, 27, 125-140. Part II: Psychometrika, 27, 219-
245.
[37] Shepard, R.N. (1966). Metric structures in ordinal data. Journal of Mathematical
17
Psychology, 1966,3,287-315.
[38] Shepard, R.N. (1978). The circumplex and related topological manifolds in the study of
perception. In S. Shye (ed.) Theory Construction and Data Analysis in the Behavioral Sciences.
[39] Shye, S. (1971). A Note on Mapping Sentences and their Derivation through the Use of
SSA. Jerusalem: The Israel Institute of Applied Social Research (A technical Report). (Also in
Abu Gosh, S. & Shye, S. (1971) The Impact of the Elective Local Council System on Arab
Traditional Society in Israel. Jerusalem: The Israel Institute of Applied Social Research.
[40] Shye S. (1976). Smallest Space Analysis of Profiles and their Relationship to Partial Order
Scalogram Analysis of the Variables. Jerusalem: The Hebrew University. (Technical Report).
[41] Shye, S. (Ed.) (1978a). Theory Construction and Data Analysis in the Behavioral Sciences.
[42] Shye, S. (1978b). Facet anlysis and regional hypotheses. In S. Shye (ed.) Theory
Construction and Data Analysis in the Behavioral Sciences. San Francisco: Jossey-Bass.
[43] Shye S. (1978c) Achievement motive: a faceted definition and structural analysis.
[44] Shye, S. (1985a). Multiple Scaling: The Theory and Application of Partial Order Scalogram
[45] Shye, S. (1985b). Nonmetric multivariate models for behavioral action systems. In D.
Canter (ed.) Facet Theory: Approaches to Social Research. New York: Springer.
[46] Shye, S. (1988). Inductive and deductive reasoning: astructural reanalysis of ability tests.
[47] Shye, S. (1989). The systemic quality of life model: a basis for urban renewal evaluation.
18
[48] Shye, S. (1991). Faceted SSA: A Computer Program for the PC. Jerusalem: The Israel
[49] Shye, S. & Elizur, D. (1994). Introduction to Facet Theory. Thousand Oaks, CA: Sage.
[50] Shye, S., Yanai, J., & Pick, C.G. (1994). Directional consistency: determinant of learned
[51] Snow, R.E., Kyllonen, P.C. & Marshalek, B. (1984). The topography of ability and learning
[52] Stouffer S. A. et al. (1950). Measurement and Prediction: Studies in social psychology in
World War II, Vol. IV. Princeton, NJ: Princeton University Press.
[53] van den Wollenberg, A.L. (1978). Nonmetric representation of the radex in its factor pattern
parametrization. In S. Shye (ed.) Theory Construction and Data Analysis in the Behavioral
Samuel Shye