Documentos de Académico
Documentos de Profesional
Documentos de Cultura
William H. Batchelder
Xiangen Hu
University of Memphis
The fuzzy logic model of perception (FLMP) is analyzed from a measurement-theoretic perspective.
FLMP has an impressive history of fitting factorial data, suggesting that its probabilistic form is
valid. The authors raise questions about the underlying processing assumptions of FLMP. Although
FLMP parameters are interpreted as fuzzy logic truth values, the authors demonstrate that for several factorial designs widely used in choice experiments, most desirable fuzzy truth value properties
fail to hold under permissible rescalings, suggesting that the fuzzy logic interpretation may be unwarranted. The authors show that FLMP's choice rule is equivalent to a version of G. Rasch's (1960)
item response theory model, and the nature of FLMP measurement scales is transparent when stated
in this form. Statistical inference theory exists for the Rasch model and its equivalent forms. In fact,
FLMP can be reparameterized as a simple 2-category logit model, thereby facilitating interpretation
of its measurement scales and allowing access to commercially available software for performing
statistical inference.
ture by feature, with prototypes representing each relevant response category. The results of these comparisons are said to be
"fuzzy logic truth values," indicating the degree of match of
each stimulus feature to a corresponding prototype feature. The
fuzzy truth values then are represented by parameters in a probabilistic model that predicts the classification probabilities for
each stimulus.
In this article we examine the probabilistic classification process of FLMP from a measurement-theoretic perspective. Our
analysis has substantial consequences for the model, some positive and some negative. In particular, we show that the fuzzy
logic parameter values cannot be recovered uniquely from the
classification probabilities. Although it is possible to set up
scales of measurement on the basis of the classification probabilities, these scales fail to satisfy the properties needed to justify
their interpretation as fuzzy logic truth value scales. We also
show that the basic probability formula of FLMP is identical
with that of the well-known model of item response theory developed by Rasch (1960) and studied extensively by psychometricians, and in various equivalent forms, in the foundations of
measurement literature. Neither the Rasch formulation nor the
others that we cover have been analyzed previously in terms of
FLMP.
Although our analysis directly challenges the interpretation
of FLMP parameters as fuzzy truth values, it helps to explain
why the model frequently does a good job of fitting data in factorial pattern classification experiments. Indeed, some of the
equivalent formulations have been quite successful in analogous
applications, and they have the added benefit of having been
studied extensively in the psychometrics literature from a statistical standpoint. Thus, rather than questioning the ability of
FLMP to fit data, our article calls attention to the need for more
Court S. Crowther, Department of Linguistics, University of California, Los Angeles; William H. Batchelder, Department of Cognitive Sciences, University of California, Irvine; Xiangen Hu, Department of Psychology, University of Memphis.
Portions of this article were presented at the 25th annual meeting of
the Society for Mathematical Psychology, Stanford University, Palo
Alto, California, August 22, 1992 (Crowther & Hu, 1992). We gratefully acknowledge comments from Jean-Claude Falmagne, Christolf
Klauer, Ece Kumbasar, R. Duncan Luce, and David M. Riefer on earlier drafts. The research presented in this article was supported by a
National Science Foundation (NSF) training grant to the Institute for
Mathematical Behavioral Sciences at University of California, Irvine;
by a National Institutes of Health training grant to the Phonetics Laboratory at University of California, Los Angeles; and by NSF Grant SBR9309667.
Correspondence concerning this article should be addressed to William H. Batchelder, Department of Cognitive Sciences, Social Sciences
Tower, University of California, Irvine, California 92717. E-mail may
be sent via Internet to whbatche@aris.ss.uci.edu.
396
397
tor levels Q and Oj, respectively, are supportive of a TI response. The "additive complements" of q and o j; (1 - q) and
(1 - Oj), respectively, represent the degree to which Q and Oj
are supportive of a T2 response. After feature evaluation, the
fuzzy truth values, Cj and Oj, are passed on to the feature integration stage.
During feature integration, the feature values are combined,
and the degree to which the resultant combination matches
each relevant prototype is determined. In most applications
multiplication is used as the operation for combining feature
values (see Massaro, 1987, chapter 7). In this case, the products, CjOj and (1 - c;)( 1 - DJ), are taken to represent the degree
that the resultant feature combination matches the prototypes
corresponding to categories T! and T2, respectively, and they
are both interpreted as fuzzy truth values in the [0,1] interval.
Finally, in the pattern classification stage, the relative goodness of match between each feature combination and each relevant prototype is determined using the following "relative goodness rule" (Massaro & Friedman, 1990, p. 230), hereafter
RGR, which is motivated as a variant of Luce's (1959) choice
rule:
Pij(Ci, Oj) =
CjOj
(D
where PJJ(CJ, QJ) is the probability that stimulus (Q, Oj) is categorized as T,, 1 < i <; /, 1 < j < ./. The RGR, then, is the ratio
of the support (or evidence) for one alternative (numerator) to
the total support for all relevant alternatives (denominator).
A concrete example from the literature illustrates the operation of the model and facilitates subsequent discussion. In an
audiovisual speech perception experiment reported in Massaro
and Cohen (1983), one visual factor (or cue) and one acoustic
factor were combined factorially to create stimuli that ranged
perceptually from /ba/ to /da/. There were two levels of the
visual factor, C. At one level, the visual component of the stimulus was a speaker saying /ba/, and the corresponding visual
feature was (closed lips). The other level was a speaker saying
/da/, and the corresponding visual feature was (open lips), or,
in the notation of FLMP, (1 - closed lips), which is the additive
complement of (closed lips). A seven-level continuum comprised the acoustic factor, O. One end of the continuum contained rising second and third formant transitions that were
more appropriate for /ba/, and the corresponding acoustic feature was (rising F2-F3). The other end of the continuum contained stimuli with falling second and third formant transitions
that were more appropriate for /da/, and the corresponding
acoustic feature was (falling F2-F3), or (1 - rising F2-F3).
Thus, there were 14 stimuli (two levels of the visual factor coupled with seven levels of the acoustic factor), and the participants' task was to classify each stimulus as either "ba" or "da."
The prototypes can be defined as:
/ba/: (closed lips )A( rising F2-F3);
/da/: (1 - closed lips)A( 1 - rising F2-F3).
Let Q be the fuzzy truth value that represents the extent to
which the visual component of the stimulus, Q, is perceived as
containing "closed lips", and Oj the extent to which the acoustic
component of the stimulus, O j; is perceived as containing "ris-
398
p("ba"|Q,Oj) = -
Theorem 1
CiOj
RMSD =
IX.J
cf =
(3)
and
(4)
(5)
(2)
Ci
(6a)
>D
(6b)
and
(6c)
Equations 3 and 4 for cf and o* can then be derived explicitly
from Equations 6b and 6c. The derivation involves transforming (1 - Ci)/C; and (1 - OJ)/DJ in Equation 6a in the manner of
Equations 6b and 6c, and letting D = 1 + B. Solving for cf and
of, Equations 3 and 4 are obtained through simple algebraic
manipulation. The proof of the second part of Theorem 1 is
tedious, and because it adds little to the argument we omit the
details.1
From a technical standpoint, Theorem 1 shows that FLMP is
1
To prove the second part of Theorem 1, one equates expressions as
in Equation 5 and rewrites the parameters as in Equation 6a, and then
it can be shown that a unique D in Equations 6b and 6c can be obtained.
MEASUREMENT-THEORETIC
399
ANALYSIS OF FLMP
Corollary 1. Suppose that Equation 1 holds for some parameters <Ci> and <0j>. Let x e (0, 1) be arbitrary and arbitrarily pick any particular ck (or Oi). Then there is exactly one
set of parameters <c* >, <o*> that satisfies Equations 1 and 5
andck = x(orc>i = x).
Proof. Without loss of generality, pick Ck e C. First, note
that for all 0 < x < 1,
B(x) =
x(l-ck)
(7)
400
ba
-O- B = Orig.
0.9
-A- B = -0.96
/n
0.8-3
-B- B = 1.57
0.7
-A- B = 11
0.6
0.5
//
0.4
0.3
0.2
0.1
da
ba
Visual Factor Level
Figure 1. Relative evidence for the syllable "da" for the acoustic (top panel) and visual (bottom panel)
factors. The original scale values from Massaro and Cohen (1983) are plotted with open circles (B = orig.),
and the other markers represent scale values computed with three different choices for B using Equations 4
and 3, respectively. Open triangle, B = -0.96; open square, B = 1.57; filled triangle, B = 11. Orig. = original.
The definitions of meaningfulness given in Roberts (1979) and Suppes and Zinnes (1963) are not accepted by all measurement theorists
as covering all applications of the concept of meaningfulness. In fact,
considerable foundational work has since occurred to provide a more
adequate sense of meaningfulness (e.g., Luce, Krantz, Suppes, & Tversky, 1990). However, the situation in the current article is sufficiently
simple that the Suppes and Zinnes (1963) and Roberts (1979) definition can be regarded as adequate.
401
5B
(8)
and
(9)
Clc-Ol
i( 1 - Ck)
= 0,
(10)
which always has a solution BO > -1, for 0 < ck, o, < 1. From
Equations 8 and 9, it is clear that c* > o* for B < B0 and o* >
c k t forB>B 0 .
Corollary 2 shows that it is meaningless to compare even ordinally the magnitudes of fuzzy truth values between respective
levels of the two factors. In fact, it follows by a minor extension
of the preceding argument that for any set of proportions <? >,
there are two sets of parameter values <Cu, DIJ> and <c2i, O2j>,
both of which minimize RMSD in Equation 2 and satisfy cu <
QI j and c2i > o2 j, for all 1 < i < / and 1 < j < /. In other words,
rescalings can completely reverse all interfactor ordering on the
parameters. This is illustrated in Figure 1: For B = 11 (filled
triangles), all levels of the acoustic factor are above .5, whereas
all levels of the visual factor are below .5; however, for B = 0.96
(open triangles), this ordering is completely reversed.
1 -x'
(11)
for 0 < x < 1. Then L(c f ) and L(OJ) can be interpreted as the
log-odds ratios (e.g., Luce, 1959) for factor levels Q and Oj, respectively. It is easy to see from Equations 3 and 4 that
cf) = L(Ci)-log(l+B)
(12)
and
So far, we have discovered that an FLMP parameter corresponding to any particular level of either factor can be rescaled to any particular value in (0, 1); however, this step
determines the value of B and therefore rescales all the other
values. We have also shown that ordinal interfactor compari-
4
In more recent applications, Massaro has suggested using variation
in the response proportions themselves rather than the parameters. This
operational basis for the definition of factor impact is independent of
FLMP and is not subject to our concerns.
402
L(of) = L( 0 j ) + log(l
(13)
forB>-l.
From Equations 12 and 13, it is clear that the ranges of the
log-odds ratios for both factors can be ordinally compared. Specifically, define
= L(c max )-L(c min )
and
= L(o m a x )-L(o m i n ),
where cmax, c min , 0,*, and omin are defined as before. It is meaningful to assert that, say, RC > RO, because from Equations 12
and 13, the values of RC and RO do not depend on the value of
B, and therefore are invariant under rescaling.
(14)
= cr
(15)
iff C]<( 1 + B) < C|( 1 + B). Because B > -1, we can conclude
that Equation 15 holds iff ck < Ci. A similar argument shows
that Op < oJ iff Op < oq.
Corollary 3 shows that the separate scales for each factor established by Equation 1 satisfy Goguen's (1969) ordinal properties; however, because both scales are yoked and set by a single
constant B, they are much stronger than ordinal scales. More
importantly, because of Corollary 2, they cannot be merged
onto a single scale and retain interscale ordinal properties. Obviously, interscale ordinality is a desirable property if the estimates are to be considered on a single dimension or in a common currency, such as fuzzy truth values intended to indicate
degree of support for a certain prototype. In the next section we
discuss some of the steps that one might consider taking to avoid
the model nonidentifiability problem.
Pi(Q) =
Ci + ( 1 -
(16)
and this condition coupled with PJ(OJ) = QJ uniquely determines the parameter values.
However, there are three considerations involved with including single-factor conditions that may limit their applicability.
First, the logic of the approach requires the assumption that the
scale values in the single-factor experiment given by Equation
(17)
AT
403
Cj Oj Uk
Cj Oj Uk + ( 1 - Ci)( 1 - Oj)( 1 - U k )
, (18a)
-Uk)
uk
where pyk(Cj, Oj, u k ) is the probability of identifying stimulus
(Q, Oj, U k ) as an instance of T,. Let DI, D2, D3 > 0 be such
DI D2 D3 = 1. If the following transformations,
(l-cD_(l-Ci)
D,,
cf
Ci
U8b)
'D 3 ,
(18d)
and
uk
404
CfOj
CjOj + Cj ( 1 - Oj) + ( 1 - Ci)0j + ( 1 - Ci)( 1 - Oj)
= CjOj,
and similarly
Figure 1 are monotonic and defined by a single parameter, although one might think their shapes differ. In any case, many
aspects of the relationship between any such curves can be understood by analyzing Equations 3 and 4. We think it is difficult
to maintain that the "shape" of functions of the Q and Oj do not
change with the scale parameter, B.
Fixing a parameter value to resolve the nonidentifiability
problem may entail negative consequences depending on just
how one wants to use the parameter values. In particular, we see
no way to set a particular scale value a priori in a manner that
guarantees the properties of the fuzzy logic interpretation of the
parameters given in the quote cited from Massaro and Friedman ( 1 990, pp. 23 1 -232 ) . Criteria such as equating parameter
value ranges used by Oden ( 1979, Footnote 4) could always be
used; however, they have only arbitrary and unsystematic implications for the fuzzy logic interpretation of the parameters.
On the other hand, it may be reasonable to fix a parameter at
a certain value and maintain a fuzzy logic interpretation in certain circumstances. For example, if one had sufficient reason
for believing that the psychological value of a particular level of
one of the factors should be "completely ambiguous," then it
seems reasonable to set the corresponding fuzzy truth value to
0.5 before starting the estimation procedure. Our survey of the
applications of FLMP did not yield many situations with a factor level that was obviously "completely ambiguous," and even
in cases where there was such a level, one worries that response
bias for one of the two categories in the presence of ambiguity
might enter and thus thwart this approach to the nonidentifiability problem. The next section provides some insight into the
type of indirect measurement that is entailed by Equation 1 .
(19)
where -oo < 0;, ft < oo , 0i is the "ability" of the ith participant,
and ft is the "difficulty" of item j.
5
Actually, as noted earlier, Oden fixed the value of one parameter in
a different form of the model, and he did not discuss the effect of this
step on interpreting the parameters as fuzzy truth values. Our Equations
3 and 4 provide this interpretation.
(20)
and
(21)
yield 0 < 3j, bj < 1 and, from Equation 19,
(22)
which is equivalent to Equation 1. In this reformulation, bj
should be thought of as an "item easiness" rather than as an
item difficulty parameter because bj decreases with increasing
0J.
The connection between Equation 1 and Equation 19 is interesting because the indirect measurement scale denned by
Equation 19 is obvious, namely, it is called a difference scale
(Suppes & Zinnes, 1963), where an arbitrary constant can be
added to both participant ability and item difficulty parameters
without changing the p(Xjj = 1). It is not surprising that the
equivalent model in Equations 1 and 22 also defines a scale up
to a single constant, albeit a nonstandard one as seen in Theorem 1. In fact, most of the points made in Theorem 1 and Corollaries 1, 2, and 3 are transparent for the equivalent Rasch
(1960) model in Equation 19.
Equation 1 for the two-factor experiment not only arises from
the Rasch (1960) model, but it is well studied in the foundations
of measurement literature. For example, Falmagne (1985) describes a probabilistic conjoint measurement model for two responds satisfying the condition
(23)
where f and g are nonnegative functions defined on the factors 7
and J, respectively, anti Equation 23 represents the probability
of a fixed response (see Equation 1 ) given the factorial stimulus
Falmagne ( 1985, p. 148) shows that under certain technical
conditions, Equation 23 is equivalent to a quadruple condition,
namely, for all levels i, i' e 7 and j, j ' e /,
PC Pi-j-
Pij'Pi-j
(24)
(25)
405
these authors have drawn formal parallels between the two-category, two-factor FLMP and several other mathematical models
of information integration and classification currently in use
in psychology, such as Bayesian and connectionist approaches.
They have shown that, among the models compared to FLMP,
those that could be written in "likelihood product" form make
predictions identical to those of FLMP. Thus, Equation 1 of
FLMP admits to many different interpretations, including ones
discussed by Cohen and Massaro (1992). Indeed, the basic
equation of FLMP for two factors is one of the most studied
equations in psychology, not only in the area of classification,
but in test theory and foundations of measurement.
Thus, for the models examined by Cohen and Massaro
(1992) and Massaro and Friedman (1990), those that can be
put into likelihood product form do not have identifiable parameters for two-category, two-factor experiments. With choice
models or information integration models, the nonidentifiability usually can be captured by scale constants. Consequently,
although the likelihood product models may share the property
of having nonidentifiable parameters, the arguments given
above regarding nonidentifiability are damaging in principle
only to models like FLMP, where parameters are intended to be
interpreted in very specific ways that are not invariant under
rescaling. If a model's parameters are not to be interpreted as
such, they are not subject to the same criticism. The consequences of parameter nonidentifiability depend on how one
wishes to interpret the parameters in question.
Perhaps the most important point to be made in this section
is that a formula such as Equation 1 admits to numerous interpretations. To justify a particular psychological interpretation
such as that offered by FLMP, it is necessary to provide strong
evidence that the parameters are parameterizing the psychological process claimed by the model. In the case of FLMP, in light
of its many equivalent forms, we argue that, even with the problem of parameter nonuniqueness covered in Theorem 1 and its
corollaries aside, the interpretation of the parameters of Equation 1 as functioning as fuzzy logic truth values in individuals'
mental processing is not convincing.
Statistical Inference
That the model represented in Equation 1 is equivalent to
the two-parameter Rasch (1960) model in Equation 19 (with
parameters compressed to the unit interval) has numerous
practical advantages. For more than two decades, Psychometrica and other quantitative journals have published many articles concerning estimation of the Rasch model family, including
generalizations that allow more than one participant parameter
and more than one test-item parameter (e.g., Holland, 1990;
Lewis, 1986). In fact, in the psychometric literature, the Rasch
model in Equation 19 is extremely well understood statistically
from just about every inferential perspective. It is therefore desirable that users of FLMP and other equivalent models are able
to use this extensive statistical theory.
Although FLMP and the Rasch (1960) model are formally
equivalent, there are some important differences between the
test theory situation and the information integration situation.
First, in test theory, participants generally are drawn from a
population at random, and sometimes it facilitates the statistical
406
inference of the Rasch model to make this population assumption explicit. On the other hand, unlike the subject ability factor
in test theory, in an information integration experiment, the
levels of both factors are selected quite intentionally by the
experimenter.
A second difference is that, in many situations the levels
within a factor can be ordered on some physical scale (e.g.,
length of a crossbar, pitch of a vowel sound, and so on). From a
psychophysical perspective, these physical orderings may suggest intrascale orderings that can be imposed on the estimation
problem. In general, estimation of categorical models under ordinal constraints is a standard topic in inference (e.g., Agresti,
1984; Robertson, Wright, & Dykstra, 1988). In fact, there have
been several efforts (e.g., Fischer & Formann, 1982; Fischer &
Tanzer, 1994) to incorporate ordinal constraints into the estimation of the Rasch model.
Finally, a third difference between the test theory situation
and FLMP is that much of the estimation theory for the Rasch
model concerns the case where only one observation (correct or
error) is obtained for each combination of participant and item.
In FLMP applications, one usually obtains repeated observations at each combination of factor levels. Thus, the data in
FLMP are usually response proportions rather than single occurrences. However, it is of great value, we think, to note that it
is perfectly feasible to work with FLMP when only one observation is made at each factor-level combination, and in fact some
of the most sophisticated applied statisticians in psychology
have devoted considerable energy to estimation theory in just
this case.
Fortunately, there is a way to reformulate the Rasch model,
and hence FLMP, as a simple logit model. In this way, data from
a factorial experiment with repeated observations can be analyzed with FLMP and other equivalent models using standard,
readily available software. Agresti (1990, Section 4.3.2) describes a logit model for an / X J X 2 data structure. The data
structure exactly matches the two-factor, two-response situation
for FLMP with experimentally preset numbers of observations
falling into each cell of the design. In terms of the logit function
in Equation 11, the model in Agresti (1990) can be written for
a particular response as
L(Pij) = a + |8p + ftf,
(26)
Theorem 2
Consider the FLMP for a two-factor, two-response paradigm
for some set of parameters (Q), <Oj>, leading to Equation 1.
Then it is equivalent to the logit model in Equation 26.
Proof
Deriving Equation 26 from Equation 1: Letpij(Ci,Oj) = pyin
Equation 1, and use the logit function in Equation 11 to write
= \og-
1 /
Next, define the parameter means L(C) = - Z L(Cj)andL(0)
/ i-l
1 J
= - Z L(OJ). It is then easy to define a as the grand mean of
(27)
Next, pick any real number r arbitrarily and define aj = /3f +
(a + r) and bj = r - 0. Rewriting Equation 27, we get
(28)
Equation 28 is in the form of the Rasch (1960) model in Equation 19, and this was shown to be equivalent to Equation 1 of
FLMP.
Theorem 2 is the key to relating FLMP in Equation 1 to standard statistical theory. Converting to logits lays bare the essence
of FLMP as an additive model that assumes no interaction.
Logit models are highly studied, including various models for
ordered factor levels, and standard methods exist to transform
logit models into useful log-linear forms (e.g., Agresti, 1984,
1990). A variety of standard, readily available software exists
to analyze such models (e.g., Agresti, 1990, Appendix A). This
software can perform all aspects of statistical inference, including parameter estimation, goodness of fit, and hypothesis
testing.
Conclusion
Our analysis of FLMP has focused on the indirect scales of
measurement implied by Equation 1 for a two-factor, twocategory experiment as well as related experimental designs.
In the two-factor case, we have shown that the equation
defines two indirect scales of measurement in the sense of
Suppes and Zinnes (1963) that are interrelated and unique
up to a single quantity B > -1. The model represents a natural version of probabilistic conjoint measurement that has
proven very useful in fitting data from a variety of two-factor
MEASUREMENT-THEORETIC
ANALYSIS OF FLMP
407
408