Simplifying the Implication-Realization Model of Melodic Expectancy

Author(s): E. Glenn Schellenberg

Source: Music Perception: An Interdisciplinary Journal, Vol. 14, No. 3 (Spring, 1997), pp.
295-318
Published by: University of California Press
Stable URL: http://www.jstor.org/stable/40285723
Accessed: 16-05-2016 12:00 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at
http://about.jstor.org/terms

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted
digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about
JSTOR, please contact [email protected].

University of California Press is collaborating with JSTOR to digitize, preserve and extend access to
Music Perception: An Interdisciplinary Journal

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
Music Perception 1 997 by the regents of the
Spring 1997, Vol. 14, No. 3, 295-318 university of California

Simplifying the Implication-Realization Model of

Melodic Expectancy

E. GLENN SCHELLENBERG

University of Windsor

Results from previous investigations indicate that the implication-real-

ization (I-R) model (Narmour, 1990) of expectancy in melody may be
overspecified and more complex than necessary. Indeed, Schellenberg's
(1996) revised model, with two fewer predictor variables, improved pre-
dictive accuracy compared with the original model. A reanalysis of data
reported by Cuddy and Lunney (1995) provided similar results. When
the principles of the I-R model were submitted to a principal-compo-
nents analysis, a solution containing three orthogonal (uncorrelated) fac-
tors retained the accuracy of the model but was inferior to the revised
model. A separate principal-components analysis of the predictors of the
revised model yielded a two-factor solution that did not compromise the
revised model's predictive power. Consequently, an even simpler model
of melodic expectancy was derived. These results provide further evi-
dence that redundancy in the I-R model can be eliminated without loss
of predictive accuracy.

Narmour known
cognition (1990, 1992)
as the proposes a theory of melodic
implication-realization perception
(I-R) model. and
Narmour's
model is based on the idea that when listeners hear a melody they typically
form expectancies about its continuation. These expectancies are presumed
to result from a combination of innate and learned factors (see also Jones,
1990). The inclusion and specification of innate factors in the model is,
arguably, one of its most important contributions to the psychology of
music. The influence of these factors on melodic expectancies has been
shown to generalize across listeners from different musical cultures tested
in a wide variety of musical and nonmusical contexts (Cuddy & Lunney,
1995; Schellenberg, 1996), consistent with their proposed universality.
Nonetheless, the I-R model appears to be overspecified and more complex
than necessary (Schellenberg, 1996). The present report examines the de-
gree to which the model can be simplified while retaining its predictive
accuracy. Simplicity is one of the main criteria by which psychological models

Address correspondence to E. Glenn Schellenberg, Department of Psychology, Univer-

sity of Windsor, Windsor, Ontario, Canada N9B 3P4. (e-mail: [email protected])

295

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
296 E. Glenn Schellenberg

are evaluated (e.g., Cutting, Bruno, Brady, & Moore, 1992). Indeed, the
principle of parsimony (Occam's razor) is a scientific canon stating that a
simpler explanation of any phenomenon is preferable to a more complex
explanation. Simpler models, with fewer parameters, also "stand a better
chance of being scientifically replicable" (Bentler &c Mooijaart, 1989, p.
315). Hence, a simplified model of melodic expectancy is more likely to
describe a universal phenomenon (i.e., one that could be replicated across
all listeners in all musical contexts).
Learning is obviously involved in the expectancies that arise while listen-
ing to melodies. For example, a listener familiar with Western music (i.e.,
music from the common practice period, or from traditional or popular
idioms) would generally expect a Western melody to stay "in key." The
claim of innate components in melodic expectancy is more contentious.
According to the I-R model, listeners, regardless of cultural background or
exposure, form common expectancies that stem from innate psychological
principles of perception and cognition. Any melodic interval that is not
perceived as complete (or closed) is considered an implicative interval, such
intervals imply that some tones are more likely than others to follow. (For
a detailed account of factors causing closure, see Narmour, 1990.) The
interval between the second tone of an implicative interval and the follow-
ing tone is considered a realized interval. At its most basic (i.e., tone-to-
tone) level, the I-R model describes how implicative intervals imply real-
ized intervals and how some realized intervals are more implied than others.
Narmour (1990, 1992) claims that these implications result from five per-
ceptual predispositions (acting in combination with learned factors): registral
direction, intervallic difference, registral return, proximity, and closure. A
cursory description follows (for a more detailed description, see Schellenberg,
1996).
The first two predispositions (registral direction and intervallic differ-
ence) form the core of the I-R model; from these, the basic melodic struc-
tures are derived. Both principles depend on whether the implicative inter-
val is small (five semitones or fewer) or large (seven semitones or more),
because small intervals are considered to have different implications than
large intervals. (The tritone [six semitones] is considered a threshold inter-
val that is neither small nor large.) The principle of registral direction states
that small intervals imply a continuation of pitch direction (e.g., a small
upward implicative interval implies an upward realized interval), whereas
large intervals imply a change of direction (e.g., a large upward implicative
interval implies a downward realized interval or a unison). The principle of
intervallic difference states that small implicative intervals imply similarly-
sized realized intervals (defined as the same size two semitones if the
realized interval changes registral direction, the same size three semitones

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 Simplifying the Implication-Realization Model of Melodic Expectancy 297

otherwise), whereas large implicative intervals imply relatively smaller re-

alized intervals.
The remaining three predispositions are not "principles" of the I-R model
in the same way that the first two are. Nonetheless, all five are considered
part of a "genetic code" (Narmour, 1989) for the perception of melodies.
In keeping with earlier reports (Cuddy & Lunney, 1995; Schellenberg, 1996),
all five will be considered principles. The principle of registral return de-
scribes a melodic archetype in which the second tone of the realized inter-
val is proximate in pitch (by two semitones or fewer) to the first tone of the
implicative interval, creating a symmetric (aba) or approximately symmet-
ric (aba') pattern of pitches about a point in time. For example, the se-
quence C4-G4-C4 is symmetric, whereas C4-G4-Ctt4 is approximately sym-
metric. Patterns become less archetypal (and less implied) as they deviate
from exact mirror-image symmetry. The proximity principle states that small
realized intervals (defined by Narmour as five semitones or fewer) are more
implied than large intervals and that implications are stronger for rela-
tively smaller realized intervals. For example, intervals of five semitones
are more implied than intervals of six semitones, intervals of four semitones
are more implied than five semitones, and so on. Finally, the principle of
closure describes how listeners perceptually segment melodies as a func-
tion of pitch direction and interval size. Closure occurs when a melody
changes direction (i.e., the implicative and realized intervals are in different
directions) or when a relatively smaller realized interval follows a larger
implicative interval; these two contributing factors are not mutually exclu-
sive. All five principles apply to both upward and downward implicative
intervals.
Schellenberg (1996) quantified each of the principles for 263 different
combinations of implicative and realized intervals, represented by the grid
in Figure 1. Only intervals one octave or smaller are considered in the grid;
melodic intervals larger than octaves are relatively rare across musical cul-
tures (Dowling & Harwood, 1986). Implicative intervals of 6 semitones
(tritones) and 12 semitones (octaves) are also excluded because Narmour
(1990, 1992) considers tritones to be a threshold interval and octaves a
unique interval by virtue of octave equivalence. Each principle applies to
both upward and downward implicative intervals, the grid making no dis-
tinction between pitch directions. Hence, each square in the grid (with one
exception) represents two different combinations of implicative and real-
ized intervals. For example, the square in the bottom left corner represents
an upward implicative interval of 1 1 semitones followed by a downward
realized interval of 12 semitones (e.g., C4-B4-B3) or a downward implica-
tive interval of 11 semitones followed by an upward realized interval of 12
semitones (C4-C)t3-Ctt4). Only the middle square in the top row, representing

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
298 E. Glenn Schellenberg

-.s2g
"I fail*

rt > ^ m

'C f^ JS o x ti

lii-ii
s -s I H
s s i e^ &
j> I '6 ?<
> J3 v^^S e
u -2- - ^

Ill Id 3
s S -8 sU<

uiiii
iiiiii

s 2 ^ > e^
g o 5 S * -.

<H1J,

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 Simplifying the Implication-Realization Model of Melodic Expectancy 299

two consecutive unisons (an implicative interval of 0 semitones followed

by a realized interval of 0 semitones), is a single combination.
Schellenberg (1996) quantified values for each of the five principles to
predict listeners' responses in three experiments.1 As shown in Figure 2,
each pair of implicative and realized intervals was given a numerical value
for each of the five principles. The first two principles (registral direction
and intervallic difference) were coded as dummy variables. Instances where
a realized interval conformed to the implication generated by the implica-
tive interval were coded as 1; other instances (representing violations of the
implications) were coded as 0. The other three principles (registral return,
proximity, and closure) were coded such that instances of relatively greater
conformity to the implications received relatively higher values. For ex-
ample, the proximity principle was coded 6 for instances of maximum prox-
imity (i.e., when the realized interval was a unison), 5 when the realized
interval was one semitone, 4 when the realized interval was two semitones,
and so on. Although these numerical values appear to be rather arbitrary,
they represent accurately the principles of the I-R model (E. Narmour, per-
sonal communications, February 1990, April 1990, May 1991, June 1991).
In each experiment, Schellenberg's (1996) listeners heard fragments of
melodies (piano timbre), each of which ended with an implicative interval,
and rated (on a 7-point scale) how well individual test tones added to the
end continued the fragments. The test tones were always within an octave
of the last tone of each fragment. The fragments were from British folk
songs (Experiment 1), Webern lieder (Experiment 2), and Chinese pentatonic
folk songs (Experiment 3). Listeners in Experiments 1 and 2 varied in mu-
sical training, those in Experiment 3 varied in cultural background (Ameri-
can or Chinese). Standard multiple regression was used to model the out-
come variable (listeners' ratings) as a function of five predictor variables
(the principles of the I-R model). Potential experiential influences of tonal-
ity were minimized by including a tonality covariate in the analyses where
appropriate (in Experiments 1 and 3). In general, the I-R principles suc-
cessfully predicted response patterns, which did not differ as a function of
musical style, formal musical training, or cultural background. In short,
the results were consistent with Narmour's (1990, 1992) claim of innate
influences on melodic expectancies. A summary of Schellenberg's (1996)
results with the I-R model is provided in Table 1 (Model 1).
In an earlier study of melodic expectancy, Carlsen (1981) tested music
students from Germany, Hungary, and the United States. The students heard
two-tone stimulus intervals and were required to sing continuations of these

1 . These experiments were initially conducted in collaboration with Carol L. Krumhansl.

Krumhansl (1995) provided a summary of the data for a music audience. Schellenberg
(1996) provided a complete account for a psychology audience.

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 300 E. Glenn Schellenberg

Fig. 2* Schellenberg's (1996) quantification of the five principles of the implication-realiza-

tion model. Reprinted from Cognition, 58(1), E. G. Schellenberg, "Expectancy in melody:
Tests of the implication-realization model," 75-125, 1996, with kind permission of Elsevier
Science-NL, Sara Burgerhartstraat 25, 1055 KV Amsterdam, The Netherlands.

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 Simplifying the Implication-Realization Model of Melodic Expectancy 301

Table 1
Multiple Regression Results from Data of Schellenberg
(1996, Appendixes A, B, and C)
Degrees of Freedom

R R2 F Model Residual

Experiment 1 (N=120)*
Model 1: Implication-realization (I-R) model .826 .683 40.49 6 113
Model 2: Revised model .871 .759 90.65 4 115
Model 3: Principal-components model (I-R) .818 .669 58.03 4 115
Model 4: Principal-components model (revised) .871 .759 121.62 3 116
Model 5: Two-factor model .871 .759 121.88 3 116
Experiment 2 (N = 200)
Model 1: I-R model .679 .461 33.18 5 194
Model 2: Revised model .726 .527 72.73 3 196
Model 3: Principal-components model (I-R) .669 .447 52.86 3 196
Model 4: Principal-components model (revised) .724 .524 108.38 2 197
Model 5: Two-factor model .721 .520 106.80 2 197
Experiment 3 (N= 132)*
Model 1: I-R model .830 .690 46.27 6 125
Model 2: Revised model .869- .755 97.66 4 127
Model 3: Principal-components model (I-R) .829 .688 70.00 4 127
Model 4: Principal-components model (revised) .868 .754 130.62 3 128
Model 5: Two-factor model .866 .750 127.89 3 128

Note. All ps < .0001. For each regression model, the number of predictor variables equals
the degrees of freedom for the model.
* These models included a covariate to control for effects of tonality.

intervals. Although Carlsen was not attempting to test the I-R model, his
stimuli can be considered implicative intervals. Intervals between the sec-
ond tone of his stimuli and the first tone of responses can be considered
realized intervals. Reanalyses of Carlsen's data revealed a good fit to the I-
R model (Cuddy & Lunney, 1995; Schellenberg, 1996) and no differences
as a function of cultural background (Schellenberg, 1996). A reanalysis of
Unyk and Carlsen's (1987) replication of the study by Carlsen (1981) with
American music students also revealed that response patterns were consis-
tent with predictions of the I-R model (Schellenberg, 1996).
Nonetheless, strong correlations among some of the quantified principles
of the I-R model raise the possibility of redundant predictors. Pairwise cor-
relations between predictors are provided in Table 2, calculated for the 263
combinations of implicative and realized intervals shown in Figure 1. The
correlations reveal that intervallic difference, proximity, and closure form
a highly collinear set of predictors. Thus, inclusion of all three principles in
the model may be unnecessary.
Schellenberg (1996)'s revised model (Table 1, Model 2), which contained
two fewer principles than the original I-R model (Table 1, Model 1), in fact
predicted his listeners' response patterns with greater accuracy. Quantified

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 302 E. Glenn Schellenberg

Table 2
Pairwise Correlations between Principles of the Implication-Realization
Model and the Revised Model (N= 263)
Intervallic Registral
Implication-Realization Model Difference Return Proximity Closure

Registral direction .144* .032 .014 .010

Intervallic difference .035 .628** .320**
Registral return -.028 .125*
Proximity .380**
Registral Return-
Revised Model Revised Proximity Revised
Registral direction-revised .336** -.047
Registral return-revised .004
__

**p<.0001.

values for the three revised principles are provided in Figure 3. Because of
intercorrelations among the original principles of intervallic difference, prox-
imity, and closure, Schellenberg omitted intervallic difference and closure
and revised proximity, forming a new predictor variable that he designated
proximity-revised. Proximity-revised simply represents the size of the real-
ized interval (in semitones); intervals are less implied as they become larger.
The remaining two principles were also modified. Registral direction was
revised so that it applies only to large intervals, with large implicative inter-
vals implying a realized interval in a different direction. Finally, Schellenberg
revised registral return as an all-or-none (dummy) variable, eliminating the
distinction between symmetric and approximately symmetric pitch patterns.
Although Schellenberg derived his revised principles on the basis of the
data from the first of his three experiments, the revised model proved supe-
rior to the original at predicting listeners' responses across all three experi-
ments.

Reanalysis of Data from Cuddy and Lunney (1995)

Cuddy and Lunney (1995) used a task similar to that of Schellenberg

(1996), the principal difference being that their listeners were required to
rate test tones as continuations of two-tone implicative intervals rather
than fragments of real melodies. Their intervals were major seconds, minor
thirds, major sixths, and minor sevenths, each of which was presented in
ascending and descending forms (a total of eight different intervals) with a
piano timbre. For each interval, listeners rated all 25 chromatic tones within

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 Simplifying the Implication-Realization Model of Melodic Expectancy 303

Fig. 3. Schellenberg's (1996) quantification of the three principles of his revised model.
Reprinted from Cognition, 55(1), E. G. Schellenberg, "Expectancy in melody: Tests of the
implication-realization model," 75-125, 1996, with kind permission of Elsevier Science-
NL, Sara Burgerhartstraat 25, 1055 KV Amsterdam, The Netherlands.

an octave from the second tone of the implicative interval (total of 200
ratings). The first tone of each interval was a dotted quarter note, the sec-
ond tone was an eighth note, and the test tone was a quarter note, such
that durations stood in a ratio of 3:1:2. Again, standard multiple regres-
sion analyses revealed that the I-R model successfully predicted response
patterns well above chance levels even though the stimulus contexts were
musically impoverished.
On the basis of their findings, Cuddy and Lunney (1995) revised registral
direction so that it applied only to large intervals, although their method of
requantification was different from that of Schellenberg (1996). They also
used the revised version of registral return as coded in Figure 3. Their most
complete explanation of response patterns was provided by a multiple re-

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 304 E. Glenn Schellenberg

gression model that contained a revised version of the I-R model (closure
was excluded, registral direction and registral return were revised) and three
co varites (one controlling for variation in ratings due to pitch height, the
other two controlling for influences of tonality).
Because Cuddy and Lunney (1995) did not use the revised model as
coded here, an initial analysis provided a direct comparison of the revised
and original I-R models. The results are summarized in Table 3 (Models 1
and 2). As with Schellenberg's (1996) data, the revised model did not result
in a loss of predictive accuracy compared with the original I-R model. Rather,
72.5% of the variance was explained by the revised model (plus the three
covariates) compared with 64.0% for the I-R model (plus covariates).
To summarize, the I-R model can successfully predict ratings in melodic
expectancy tasks well above chance levels regardless of differences in musi-
cal styles and groups of listeners. Nonetheless, considered in combination
with Schellenberg's (1996) findings, the reanalysis of Cuddy and Lunney's
(1995) data confirms that the I-R model can be substantially simplified
without loss of predictive power.

Methods and Objectives

The objective of the present report was to derive the simplest possible
model of expectancy in melody while retaining the predictive power of
more complex models. Specifically, principal-components analysis was used
to measure the number of nonredundant (i.e., orthogonal, or uncorrelated)
dimensions embodied in the I-R model. A "principal-components" model
consisting solely of these unique dimensions was evaluated for relative effi-

Table 3
Multiple Regression Results from Data of Cuddy and Lunney
(1995, Appendix)
Degrees of Freedom

R R2 F Model Residual

Model 1: Implication-realization (I-R) model .800 .640 42.38 8 191

Model 2: Revised model .851 .725 84.67 6 193
Model 3: Principal-components model (I-R) .794 .630 54.74 6 193
Model 4: Principal-components model (revised) .849 .721 100.04 5 194
Model 5: Two-factor model .851 .724 101.53 5 194

All ps < .0001 (N = 200). For each regression model, the number of predictor variables
equals the degrees of freedom for the model. Each model included three covariates. Two of
these held constant differences in responding due to the tonal implications of the stimulus pat-
terns; the third controlled for influences of the pitch height of the test tones.

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 Simplifying the Implication-Realization Model of Melodic Expectancy 305

cacy in predicting the data of Cuddy and Lunney (1995) and those of
Schellenberg (1996); the principal-components model was also compared
with the original I-R model and Schellenberg's revised model. Finally, a
separate principal-components analysis was conducted on the principles of
the revised model; from this analysis, a further simplified model of melodic
expectancy was derived.

PRINCIPAL-COMPONENT ANALYSIS

A brief description of principal components analysis follows for the ben-

efit of readers unfamiliar with multivariate statistical techniques. (For a
detailed account, see Tabachnick and Fidell [1996, chap. 13]; see Diekhoff
[1992, chap. 16] for a concise overview. The current summary is derived
from both sources.) Principal-components analysis is one of two statistical
methods generally referred to as factor analysis (the other being the com-
mon factor model). In factor analysis, the goal is to reduce a set of vari-
ables (e.g., the principles of the I-R model) to a smaller set of orthogonal
(uncorrelated) variables. These uncorrelated variables are called factors ,
each of which is a different linear combination of the original variables,
derived such that each factor is uncorrelated (r = 0) with every other factor.
Hence, factors are latent variables, not actually present in the set of origi-
nal variables. Rather, they represent the underlying "structure" of the origi-
nal set (i.e., the number of orthogonal dimensions contained in the set).
The first factor in a principal-components solution is the linear combi-
nation of original variables that accounts for the largest proportion of the
variance in the entire set; this proportion is measured with an eigenvalue. A
factor's eigenvalue is the sum of the squared correlations between the fac-
tor and each original variable. Thus, the first factor is the linear combina-
tion of original variables that has the largest possible eigenvalue. The sec-
ond factor is a different linear combination of the original variables that is
orthogonal to the first factor and explains the next largest proportion of
variance in the original variables (having the second largest eigenvalue).
The third factor is the linear combination orthogonal to the first and sec-
ond factors that explains the next largest proportion of variance (having
the third largest eigenvalue), and so on.
To explain all of the variance in a set of variables, one needs as many
factors as there are variables unless some of the variables are perfectly cor-
related. Typically, however, the first few factors explain a large proportion
of variance in the original set and are considered to be an adequate account
of the set's structure (i.e., an adequate solution). There are no hard and fast
rules by which to determine the "correct" number of factors to include in a
solution. Nonetheless, a factor with an eigenvalue of less than 1 explains
less variance than the variance found in a single original variable. Because

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 306 E. Glenn Schellenberg

the goal is to reduce the number of variables, it is common to constrain a

solution such that only factors with eigenvalues greater than or equal to 1
are included.
Factors retained in a solution are rotated to increase their interpretabil-
ity. Varimax rotation is the most common method. Rotation does not af-
fect the orthogonality of the factors, but makes the solution more inter-
prtable by increasing high pairwise correlations between factors and
original variables and by decreasing low correlations. After rotation, it is
easier to identify which variables are most strongly associated with each
factor.

Principal-Components Analysis of I-R Model

The five principles of the I-R model were submitted to a principal-com-

ponents analysis with varimax rotation to determine (1) whether a simpli-
fied version of the I-R model (i.e., no redundancy among predictors) re-
tained its predictive accuracy and (2) whether the solution was similar to
the revised model proposed by Schellenberg (1996). The data for this analysis
consisted solely of the quantified values of the predictor variables provided
in Figure 2 (i.e., the analysis did not involve the outcome variable - listen-
ers' ratings). The analysis included all 263 combinations of implicative and
realized intervals represented by the grid in Figure 1. Factors with eigenval-
ues less than 1 were excluded from the solution, which contained three
factors. The results of the analysis are summarized in Table 4. The first
factor, which accounted for the largest proportion of the variance (eigen-
value of 1.90), explained 37.9% of the variance in the original five prin-
ciples. The second and third factors (eigenvalues of 1.04 and 1.02, respec-
tively) accounted for 20.9% and 20.5% of the variance, respectively. Thus,
79.3% of the variance in the five principles of the I-R model can be ex-
plained by three orthogonal factors. The first factor was highly correlated
with the principles of proximity, intervallic difference, and closure. The
second factor was highly correlated with registral return and moderately
correlated with closure. The third factor was highly correlated with registral
direction and exhibited small but significant correlations with intervallic
difference and closure.
The next set of analyses examined how well the three orthogonal factors
from the principal-components analysis predicted melodic expectancy data
from previous investigations (Cuddy &c Lunney, 1995; Schellenberg, 1996).
Hence, listeners' ratings were modeled as a function of three latent vari-
ables (i.e., the factors identified in the principal-components analysis) rather
than the predictor variables specified by the I-R model. These analyses de-
termined whether the proportion of variance in the I-R model left unex-

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 Simplifying the Implication-Realization Model of Melodic Expectancy 307

Table 4
Results from Principal-Components Analysis of the Implication-
Realization Model

Standardized Correlation
Principle Coefficient Coefficient with Factor
Factor 1
Registral direction -.068 -.034 .037
Intervallic difference .876 .436 .835***
Registral return -.055 -.043 .001
Proximity .242 .472 .876***
Closure .495 .342 55***
Constant -1.079
Factor 2
Registral direction .024 .012 .032
Intervallic difference -.157 -.078 -.041
Registral return 1.200 .925 .962***
Proximity -.072 -.140 -.107
Closure .412 .284 .322***
Constant -.554
Factor 3
Registral direction 1.922 .962 .983***
Intervallic difference .308 .153 .201 * *
Registral return .032 .024 .042
Proximity -.031 -.061 -.016
Closure -.244 -.168 -.130*
Constant -.828

*p < .05
**p<.005
***p<.0001

plained by the principal components solution (i.e., about one-fifth of the

total variance) was necessary to explain listeners' response patterns. As
shown in Tables 1 and 3, the principal-components (I-R) model (Model 3)
was comparable in goodness-of-fit to the I-R model (Model 1) across data
sets. In each case, substituting the three orthogonal factors from the princi-
pal-components (I-R) model for the five predictors of the I-R model re-
sulted in a loss of predictive accuracy of less than 2%. Thus, although the
three orthogonal factors leave approximately one fifth of the variance in
the I-R model unexplained, this unexplained variance is relatively unim-
portant in explaining ratings of melodic expectancy.
Similarities between the principal-components (I-R) model and the re-
vised model were examined by measuring pairwise correlations between
the orthogonal factors and the revised principles. Each factor from the prin-
cipal-components (I-R) model was most highly correlated with a different
revised principle: the first factor with the revised proximity principle (r =
~.86, p < .0001), the second factor with the revised registral return prin-
ciple (r = .89, p < .0001), and the third factor with the revised registral

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
308 E. Glenn Schellenberg

direction principle (r = .59, p < .0001). These correlations reveal that the
principles of the revised model, which was derived from listeners' expect-
ancy ratings, are associated with those of the mathematically derived prin-
cipal components (I-R) model. Nonetheless, the revised registral direction
principle was also weakly correlated with the first and second factors (r =
.23, p < .0005, and r = .44, p < .0001, respectively). Moreover, the revised
model (Model 2) consistently outperformed the principal-components (I-
R) model (Model 3) across the analyses reported in Tables 1 and 3, ex-
plaining between 5.3% and 9.5% more variance in each case. Hence, the
revised model is not simply a reflection of the latent structure of the origi-
nal I-R model. Rather, the changes incorporated into the revised model
actually improve its predictive power relative to the original I-R model.
Subsequent analyses focused on potential simplification and improvement
of the revised model.

Principal-Components Analysis of the Revised Model

Although Schellenberg (1996) found that his revised model of melodic

expectancy explained response patterns better than the I-R model, the re-
vised model also contains some redundancy. As shown in Table 2, the re-
vised principles of registral direction and registral return are correlated,
implying that inclusion of both principles may be unnecessary. Accord-
ingly, a second principal-components analysis was used to derive an even
simpler model (the principal-components [revised] model) with no redun-
dancy and no loss of predictive power.
As with the I-R model, the three principles of the revised model were
submitted to a principal-components analysis with varimax rotation (same
263 combinations of implicative and realized intervals, same eigenvalue
criterion); the results are summarized in Table 5. A solution containing two
orthogonal factors was obtained. The first factor explained 44.5% of the
variance in the three principles of the revised model (eigenvalue of 1.34),
the second factor explained 33.5% (eigenvalue of 1.00). Hence, the two-
factor solution explained 78.0% of the total variance in the revised model.
High correlations with the revised principles of registral direction and
registral return revealed that the first factor was essentially a linear combi-
nation of these two principles. By contrast, the second factor was almost
perfectly correlated with the revised proximity principle.
The two factors uncovered in the principal-components solution were
used to predict the data collected by Cuddy and Lunney (1995) and by
Schellenberg (1996). As indicated in Tables 1 and 3, substitution of the two
orthogonal factors from the principal-components (revised) model (Model
4) for the three predictors of the revised model (Model 3) resulted in a loss

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 Simplifying the Implication-Realization Model of Melodic Expectancy 309

Table 5
Results from Principal-Components Analysis of the Revised Model
Standardized Correlation
Principle Coefficient Coefficient with Factor
Factor 1
Registral direction-revised .880 .608 .814*
Registral return-revised 1.618 .616 .820*
Proximity-revised .004 .013 -.013
Constant -.323
Factor 2
Registral direction-revised -.099 -.068 -.087
Registral return-revised .219 .083 .065
Proximity-revised .274 .993 .996*
Constant -1.740

*p < .0001.

of predictive accuracy of 0.3% or less. Thus, the proportion of variance in

the revised model that is unexplained by the principal-components analysis
(about one fifth of the total variance) is not needed to explain the available
data. Moreover, these results imply that a model limited to two factors or
principles might provide the best account of processes underlying the for-
mation of tone-to-tone expectancies when listening to melodies.

A Simplified Model of Expectancy in Melody

The two-factor model of melodic expectancy is the outcome of the re-

sults reported in the preceding sections. Quantified values of the model's
two principles are provided in Figure 4. The principles of the two-factor
model were obtained by modifying the factors from the principal-compo-
nents (revised) model such that they would be both theoretically relevant
and easy to use. Thus, the two-factor model is essentially a validation of
the principal-components (revised) model.
The first principle of the two-factor model is pitch proximity. This prin-
ciple states that when listeners hear an implicative interval in a melody,
they expect the next tone to be proximate in pitch to the second tone of the
implicative interval (i.e., they expect a small realized interval). The pitch-
proximity principle is identical to Schellenberg's (1996) proximity-revised
principle (Figure 3), which was found to be almost identical (r = 1) to the
second factor in the principal-components analysis of his revised model.
The principle, which measures the size of the realized interval in semitones
(unison = 0, minor second = 1, major second = 2, etc.), is simple to quan-
tify. Because larger intervals are assigned higher values, negative associa-

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 310 E. Glenn Schellenberg

Fig. 4. Quantification of the principles of the two-factor model.

tions between the principle and expectancy data are expected. Although
the principle is specified here using semitones, any logarithmic pitch scale
could be substituted. Thus, the pitch-proximity principle can be used with
scales whose intervals are not multiples of semitones. With the Thai scale,
for example, where consecutive tones of the seven-tone scale are equidis-
tant or approximately equidistant (Ellingson, 1992; Morton, 1980), one
would code a unison as 0, an interval of one scale step as 1, an interval of
two scale steps as 2, and so on. Moreover, the pitch-proximity principle
does not suffer from the arbitrary assumption of Narmour's (1990, 1992)
original concept of proximity, in which all realized intervals larger than
five semitones are considered to be equally nonproximate (see Figure 2).
Expectancies are typically considered to represent learned schmas,
whereas perceptual grouping based on proximity is considered to reflect a
"primitive" (unlearned) process (Bregman, 1990). Nonetheless, because
proximity tends to predict grouping in vision (Koffka, 1935; Kohler, 1947)
and audition (Bregman, 1990), the pitch-proximity principle could stem
from a hard-wired perceptual predisposition that also influences expectan-
cies when listening to melodies, as Narmour claims. For example, listeners
perceive tones that are proximate in pitch to be similar (Kallman, 1982).
Moreover, research on "auditory scene analysis" (Bregman, 1990) indi-
cates that proximity is a critical factor in the perception of auditory pat-
terns that appears to be independent of learning. Tones proximate in pitch
tend to be heard as originating from the same source or object; nonproximate

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 Simplifying the Implication-Realization Model of Melodic Expectancy 311

tones are typically interpreted as originating from different sources. Thus,

a melody could be relatively incoherent (i.e., difficult to process and repre-
sent; see Bharucha & Pryor, 1986; Schellenberg & Trehub, 1996) if its
component tones are heard as emanating from multiple sources. Indeed,
Deutsch (1978) reported more efficient processing of sequences with small
intervals than of sequences with larger intervals.
The pitch-proximity principle can also be interpreted as a principle of
conjunct motion (i.e., motion by scale step; see Aldwell & Schachter, 1989).
The present results demonstrate that listeners typically expect melodies to
continue in scale steps (the smallest possible intervals in a given mode)
rather than in leaps (larger intervals). When a melody departs from con-
junct motion, smaller leaps in pitch are less unexpected than larger leaps. It
remains unclear, however, whether listeners' expectancies for proximate
tones are innate rather than learned. Because small intervals are common
in melodies (Dowling &c Harwood, 1986), listeners could learn to expect
typically sized intervals. Nonetheless, the cross-cultural predominance of
small intervals in music (Dowling &c Harwood, 1986) is consistent with
the idea that the pitch-proximity principle describes a musical universal.
Further, proximity is a robust predictor of responses from listeners from
different musical cultures (Carlsen, 1981; Schellenberg, 1996), tested with
various musical styles (Schellenberg, 1996) or in relatively nonmusical con-
texts (Carlsen, 1981; Cuddy & Lunney, 1995).
The second principle of the two-factor model, pitch reversal, is based on
the first factor obtained in the principal-components analysis of the revised
model (Table 5). This factor was essentially a linear combination of the
revised principles of registral direction and registral return (i.e., Factor 1 =
0.880 [registral direction-revised] + 1.618 [registral return-revised] + 0.004
[proximity-revised] - 0.323, see Table 5); both revised principles were highly
correlated with this factor. For ease of use, the coefficients for the revised
principles of registral direction and registral return were rounded (from
0.880 to 1 and from 1.618 to 1.5, respectively), and the relatively small
constant and proximity-revised coefficient were ignored, giving the follow-
ing formula:

pitch-reversal = registral direction-revised + (1.5) registral return-revised.

When quantified in this manner, the pitch-reversal principle remains al-

most perfectly correlated with the first factor from the principal compo-
nents analysis of the revised model (r = .997, p < .0001). Moreover, it is
uncorrelated with the pitch-proximity principle (r = .029).
As illustrated in Figure 4, pitch reversal can have one of five values: -1,
0, 1, 1.5, and 2.5. The greater range of values for large implicative intervals
(-1, 1, and 2.5) than for small implicative intervals (0 and 1.5) indicates
that the principle is a relatively stronger determinant of expectancies for

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 312 E. Glenn Schellenberg

large intervals. Although the pitch-reversal principle is simply an additive

combination of two principles from the revised model (registral direction-
revised and registral return-revised), it is justified for two reasons: (1) it
collapses two overlapping (correlated) principles - both of which specify a
reversal of pitch direction - into a single principle, and (2) it results in a
very parsimonious model of melodic expectancy with no redundancy among
predictors.
The pitch-reversal principle extends the pitch-proximity principle to re-
lations between noncontiguous tones. In addition to expecting the next
tone in a melody to be proximate in pitch to the tone heard most recently,
the principle claims that listeners often expect the next tone to be proxi-
mate in pitch to the tone that preceded the most recently heard tone. In
other words, listeners often expect the second tone of a realized interval to
be proximate to the first tone of the implicative interval. Hence, the pitch-
reversal principle describes expectancies for proximate pitch relations slightly
more global than those described by the pitch-proximity principle.
The pitch-reversal principle also describes expectancies that arise when
a melody violates the pitch-proximity principle (i.e., when a large implica-
tive interval is heard). Once the coherence of a melody has been "threat-
ened" by disjunct motion (a melodic leap), the principle asserts that listen-
ers expect a reversal of pitch direction. If this expectancy is considered
jointly with the expectancy for small intervals (as described by the pitch-
proximity principle), the overall expectancy is for the resulting gap in pitch
to be filled (i.e., listeners expect a change of direction and a relatively small
interval). Filling the gap helps to restore the integrity and coherence of the
melodic line. An expectancy for "gap-fill" melodic patterns was articu-
lated explicitly by Meyer (1973) and is consistent with rules of music theory
and voice leading (e.g., Aldwell & Schachter, 1989). Moreover, previous
investigations (Rosner 6c Meyer, 1982, 1986; Schmuckler, 1989) have pro-
vided evidence supporting the psychological validity of gap-fill expectan-
cies.
Human vocal limitations may be another factor contributing to both
principles of the two-factor model. Large intervals are more difficult to
sing accurately than are small intervals (Bregman, 1990), so the pitch-prox-
imity principle could stem, in part, from an expectancy for intervals that
are easily sung. Moreover, if a tone following a large interval in a melody
does not change direction, it is more likely to exceed a singer's vocal range
than a tone that changes direction. Hence, the expectancy for pitch rever-
sals could reflect knowledge of vocal range limitations.
Analyses involving predictions of data collected by Cuddy and Lunney
(1995) and Schellenberg (1996) by means of the two-factor model are sum-
marized in Tables 1 and 3 (Model 5). Because the model was derived from
the principal-components analysis of the revised model, its fit to the vari-

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 Simplifying the Implication-Realization Model of Melodic Expectancy 313

ous data sets was virtually identical to that of the principal-components

(revised) model (Tables 1 and 3, Model 4), which, in turn, was almost
identical in predictive power to the revised model (Tables 1 and 3, Model
2). Collapsing two principles (registral direction-revised and registral re-
turn-revised) into one (pitch-reversal) meant that the coefficients for the
principles could not vary freely for each data set, making it impossible for
the two-factor model to exceed the revised model in predictive power. None-
theless, the loss of a degree of freedom had little consequence for predictive
accuracy across data sets. This finding provides empirical validation of the
pitch-reversal principle in particular and the two-factor model in general.
Although the two-factor model is substantially different from the origi-
nal I-R model, it can still be considered a revision of the original. Both
factors are based on ideas articulated by Narmour (1990, 1992). More-
over, some aspects of the two-factor model reflect central tenets of the I-R
model. For example, the pitch-reversal principle retains Narmour's distinc-
tion between small and large intervals, his proposal that large intervals
imply a reversal of pitch direction, and his suggestion that pitch patterns
symmetric about a point in time are melodic archetypes. Nonetheless, the
two-factor model can be considered superior to the original on the basis of
its greater simplicity in accounting for melodic expectancies. Two of the
principles from the original I-R model (i.e., intervallic difference and clo-
sure) appear to have little psychological validity, and none of the remaining
principles as originally formulated is retained in the two-factor model. In-
deed, the present report makes it clear that the original I-R model provides
neither the most complete nor the most elegant description of the available
data.
It is important to note, however, that the stimulus contexts examined in
the present study were severely constrained. Specifically, the contexts and
analyses focused on pitch relations at very local (tone-to-tone) levels, with
rhythmic factors held constant. Rhythm plays a major role in determining
expectancies (Jones, 1981, 1982, 1987, 1990; Jones & Boltz, 1989), and
several studies have reported that rhythmic factors interact with those based
on pitch (e.g., Boltz, 1991, 1993; Jones, Boltz, & Kidd, 1982; Jones,
Summerell, &c Marshburn, 1987). Because influences of pitch proximity
were found to extend to noncontiguous tones (i.e., tones separated by one
tone), more global properties of melodies are also likely to influence listen-
ers' expectancies (Schmuckler, 1990).
In sum, the two-factor model proposes that tone-to-tone expectancies
are determined primarily by proximity; upcoming tones in a melody are
expected to be proximate to tones heard previously. When listeners hear
successive tones that are nonproximate (relatively distant in pitch), they
expect the next tone to fill in the gap. Despite the success of this very simple
nonredundant model across data sets, it is conceivable that a more com-

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 314 E. Glenn Schellenberg

plex model with built-in redundancies might be required to explain me-

lodic expectancies in other contexts. For example, Narmour (1992) claims
that "the redundancy built into our melodic cognitive systems" (p. 239)
helps to "eliminate perceptual processing 'mistakes'" (p. 244). Hence, the
present focus on parsimony may have resulted in a model that is applicable
only to contexts similar to the ones examined here, where "processing mis-
takes" may be relatively unlikely. Moreover, a melody's rhythmic and glo-
bal pitch characteristics are likely to interact with the local pitch principles
of the two-factor model. Nonetheless, the two-factor model provides a very
simple and elegant explanation of the results from the studies examined in
the present report. Future research could examine the predictive power of
the two-factor model in other contexts. Interested researchers should bear
in mind, however, that for stimulus contexts with tonal implications,
covariates for influences of tonality may be required. Developmental and
cross-cultural approaches would be particularly useful in assessing whether
the two-factor model is universally applicable and whether learning and
exposure to music are necessary for the principles to become operative.2

References

Aldwell, E., c Schachter, C. (1989). Harmony and voice leading (2nd d.). San Diego:
Harcourt Brace Jovanovich.
Bentler, P. M., & Mooijaart, A. (1989). Choice of structural model via parsimony: A ratio-
nale based on precision. Psychological Bulletin, 106, 315-317.
Bharucha, J. J., & Pryor, J. H. (1986). Disrupting the isochrony underlying rhythm: An
asymmetry in discrimination. Perception & Psychophysics, 40, 137-141.
boltz, M. (lyyl). borne structural determinants or melody recall. Memory cr Cognition,
19,239-251.
Boltz, M. (1993). The generation of temporal and melodic expectancies during musical
listening. Perception & Psychophysics, 53, 585-600.
Bregman, A. S. (1990). Auditory scene analysis. Cambridge, MA: MIT Press.
Carlsen, J. C. (1981). Some factors which influence melodic expectancy. Psychomusicology,
1, 12-29.
Cuddy, L. L., 6c Lunney, C. A. (1995). Expectancies generated by melodic intervals: Percep-
tual judgments of melodic continuity. Perception & Psychophysics, 57, 451-462.
Cutting, J. E., Bruno, N., Brady, N. P., & Moore, C. (1992). Selectivity, scope, and simplic-
ity of models: A lesson from fitting judgments of perceived depth. Journal of Experimen-
tal Psychology: General, 121, 364-381.
Diekhoff, G. (1992). Statistics for the social and behavioral sciences: Unwartate, bwanate,
multivariate. Dubuque, IA: Wm. C. Brown.
Deutsch, D. (1978). Delayed pitch comparisons and the principle of proximity. Perception
& Psychophysics, 23, 227-230.
Dowling, W. J., 6c Harwood, D. L. (1986). Music cognition. Orlando, FL: Academic Press.
Ellingson, T. (1992). Transcription. In H. Myers (Ed.), Ethnomusicology: An introduction
(pp. 110-152). New York: Norton.

2. Funding for this research was provided by the Natural Sciences and Engineering Re-
search Council of Canada. Emmanuel Bigand, Katy Dunham, Mari Jones, Bill Thompson,
and Sandra Trehub provided helpful comments on earlier versions of the manuscript.

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 Simplifying the Implication-Realization Model of Melodic Expectancy 315

Jones, M. R. (1981). Music as a stimulus for psychological motion: Part I. Some determi-
nants of expectancies. Vsychomusicolozy, 1 (2), 34-51.
Jones, M. R. (982). Music as a stimulus for psychological motion: Part II. An expectancy
model. Vsychomusicology, 2 (1), 1-13.
Jones, M. R. (1987). Dynamic pattern structure in music: Recent theory and research. Per-
ception & Psychophysics, 41, 621-634.
Jones, M. R. (1990). Learning and the development of expectancies: An interactionist ap-
proach. Psychomusicology, 9, 193-228.
Jones, M. R., c Boltz, M. (1989). Dynamic attending and responses to time. Psychological
Review, 96, 459-491.
Jones, M. R., Boltz, M., & Kidd, G. (1982). Controlled attending as a function of melodic
and temporal context. Perception & Psychophysics, 32, 211-218.
Jones, M. R., Summerell, L., & Marshburn, E. (1987). Recognizing melodies: A dynamic
interpretation. The Quarterly Journal of Experimental Psychology, 39 A, 89-121.
Kallman, H. J. (1982). Octave equivalence as measured by similarity ratings. Perception &
Psychophysics, 32, 37-49.
Koffka, K. (1935). Principles of Gestalt psychology. London: Routledge & Kegan Paul.
Kohler, W. (1947). Gestalt psychology: An introduction to new concepts of modern psy-
chology. New York: Liveright.
Krumhansl, C. L. (1995). Music psychology: Influences from music theory. Music Theory
Spectrum, 17, 53-80.
Meyer, L. B. (1973). Explaining music: Essays and explorations. Chicago: University of
Chicago Press.
Morton, D. (1980). Thailand. In S. Sadie (Ed.), The new Grove dictionary of music and
musicians (Vol. 18, pp. 712-722). London: Macmillan.
Narmour, E. (1989). The "genetic code" of melody: Cognitive structures generated by the
implication-realization model. Contemporary Music Review, 1989,4, 45-64.
Narmour, E. (1990). The analysis and cognition of basic melodic structures: The implica-
tion-realization model. Chicago: University of Chicago Press.
Narmour, E. (1992). The analysis and cognition of melodic complexity: The implication-
realization model. Chicago: University of Chicago Press.
Rosner, B. S., & Meyer, L. B. (1982). Melodic processes and the perception of music. In D.
Deutsch (Ed.), The psychology of music (pp. 317-341). New York: Academic Press.
Rosner, B. S., & Meyer, L. B. (1986). The perceptual roles of melodic process, contour and
form. Music Perception, 4, 1-40.
Schellenberg, E. G. (1996). Expectancy in melody: Tests of the implication-realization model.
Cognition, 58, 75-125.
Schellenberg, E. G., & Trehub, S. E. (1996). Children's discrimination of melodic intervals.
Developmental Psychology, 32, 1039-1050.
Schmuckler, M. A. (1989). Expectation in music: Investigation of melodic and harmonic
processes. Music Perception, 7, 109-150.
Schmuckler, M. A. (lyyu). lhe performance or global expectancies, fsychomusicology, y,
122-147.
Tabachnick, B. G., 6c Fidell, L. S. (1996). Using multivariate statistics (3rd d.). New York:
Harper Collins.
Unyk, A. M., & Carlsen, J. C. (1987). The influence of expectancy on melodic perception.
Psychomusicology, 7, 3-23.

Appendix
The tables in this appendix provide detailed statistics from the multiple regression analy-
ses. For each predictor variable (or set of predictors) in each model, the squared semipartial
correlation (sr2 or SR2), F ratio, and p value are reported. The squared semipartial correla-
tion represents the proportion of variance in the outcome variable that is uniquely ex-
plained by each predictor (or set of predictors).

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 316 E. Glenn Schellenberg

Table Al
Analyses of Data from Schellenberg (1996)
sr> F df p

Experiment 1 (AT =120)

Model 1: Implication-Realization (I-R) Model
model R2 = .683
Registral direction .029 10.12 1,113 <.005
Intervallic difference .024 8.22 1,113 <.005
Registral return .067 23.63 1,113 <.0001
Proximity .089 31.56 1,113 <.0001
Closure .036 12.56 1,113 <.001
Tonality covariate .015 5.13 1,113 <.O5
Model 2: Revised Model
model R2 = .759
Registral direction-revised .084 40.29 1,115 <.0001
Registral return-revised .052 24.91 1,115 <.0001
Proximity-revised .472 225.56 1,115 <.0001
Tonality covariate .034 16.30 1,115 <.0001
Model 3: Principal-Components Model (I-R)
model R2 = .669
Factor 1 .516 179.12 1,115 <.0001
Factor 2 .077 26.57 1,115 <.0001
Factor 3 .036 12.22 1,115 <.001
Tonality covariate .016 5.51 1,115 <.O5
Model 4: Principal-Components Model (Revised)
model R2 = .759
Factor 1 .214 102.89 1,116 <.0001
Factor 2 .490 235.45 1,116 <.0001
Tonality covariate .035 16.64 1,116 <.0001
Model 5: Two-Factor Model
model R2 = .759
Pitch proximity .476 229.15 1,116 <.0001
Pitch reversal .207 99.80 1,116 <.0001
Tonality covariate .034 16.41 1,116 <.0001

Experiment 2 (N = 200)
Model 1: Implication-Realization (I-R) Model
model R2 = .461
Registral direction .093 33.63 1,194 <.0001
Intervallic difference .017 5.93 1,194 <.O5
Registral return .035 12.42 1,194 <.001
Proximity .071 25.69 1,194 <.0001
Closure .016 5.60 1,194 <.O5
Model 2: Revised Model
model R2 = .527
Registral direction-revised .094 38.90 1,196 <.0001
Registral return-revised .016 6.50 1,196 <.O5
Proximity-revised .358 148.16 1,196 <.0001

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
 Simplifying the Implication-Realization Model of Melodic Expectancy 317

Table Al (Continued)
st2 F df p
Model 3: Principal-Components Model (I-R)
model JR2 = .447
Factor 1 .311 110.24 1,196 <.0001
Factor 2 .036 12.94 1,196 <.0005
Factor 3 .103 36.61 1,196 <.0001
Model 4: Principal-Components Model (Revised)
model R2 = .524
Factor 1 .144 59.44 1,197 <.0001
Factor 2 .381 157.53 1,197 <.0001
Model 5: Two-Factor Model
model R2 = .520
Pitch proximity .364 149.38 1,197 <.0001
Pitch reversal .144 59.26 1,197 <.0001

Experiment 3 (N=132)
Model 1: Implication-Realization (I-R) Model
model R2 = .690
Registral direction .010 3.97 1,125 <.O5
Intervallic difference .065 25.85 1,125 <.0001
Registral return .031 12.42 1,125 <.001
Proximity .067 26.76 1,125 <.0001
Closure .048 19.31 1,125 <.0001
Tonality covariate .009 3.26 1,125 <.l
Model 2: Revised Model
model R2 = .755
Registral direction-revised .077 39.73 1, 127 <.0001
Registral return-revised .015 7.75 1, 127 <.01
Proximity-revised .570 295.10 1,127 <.0001
Tonality covariate .010 4.96 1,127 <.O5
Model 3: Principal-Components Model (I-R)
model R2 = .688
Factor 1 .615 250.42 1,127 <.0001
Factor 2 .043 17.52 1,127 <.0001
Factor 3 .017 6.86 1,127 <.01
Tonality covariate .008 3.41 1,127 <.l
Model 4: Principal-Components Model (Revised)
model R2 = .754
Factor 1 .122 63.10 1,128 <.0001
Factor 2 .599 311.29 1,128 <.0001
Tonality covariate .010 5.09 1,128 <.O5
Model 5: Two-Factor Model
model R2 = .750
Pitch proximity .579 296.22 1,128 <.0001
Pitch reversal .120 61.41 1,128 <.0001
Tonality covariate .010 5.15 1,128 <.O5

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms
318 E. Glenn Schellenberg

Table A2
Analyses of Data from Cuddy and Lunney (1995, Appendix) N= 200
st2 F df p

Model 1: Implication-Realization (I-R) Model

model R2 = .640
Registral direction .008 4.25 1,191 <.O5
Intervallic difference .022 11.46 1,191 <.001
Registral return .026 13.34 1,191 <.0005
Proximity .034 17.85 1,191 <.0001
Closure .008 3.89 1,191 <.l
Tonality and height covariates .362 63.93 3,191 <.0001
Model 2: Revised Model
model R2 = .725
Registral direction-revised .018 12.56 1,193 <.0005
Registral return-revised .025 17.38 1,193 <.0001
Proximity-revised .304 213.02 1,193 <.0001
Tonality and height covariates .379 88.56 3,193 <.0001
Model 3: Principal-Components Model (I-R)
model R2 = .630
Factor 1 .246 127.96 1,193 <.0001
Factor 2 .023 11.89 1,193 <.001
Factor 3 .014 7.03 1,193 <.01
Tonality and height covariates .359 62.43 3,193 <.0001
Model 4: Principal-Components Model (Revised)
model R2 = .721
Factor 1 .070 48.12 1,194 <.0001
Factor 2 .304 210.83 1,194 <.0001
Tonality and height covariates .376 86.85 3,194 <.0001
Model 5: Two-Factor Model
model R2 = .724
Pitch proximity .303 212.41 1,194 <.0001
Pitch reversal .066 55.81 1,194 <.0001
Tonality and height covariates .378 88.37 3,194 <.0001

This content downloaded from 130.132.173.56 on Mon, 16 May 2016 12:00:17 UTC
All use subject to http://about.jstor.org/terms