Measurement Error Masks Bipolarity in Affect Ratings

Abstract

For years, affect researchers have debated about the true dimensionality of mood. Some have argued that positive and negative moods are largely independent and can be experienced simultaneously. Others claim that mood is bipolar, that joy and sorrow represent opposite ends of a single dimension. The 3 studies presented in this article suggest that the evidence that purportedly shows the independence of seemingly opposite mood states, that is, low correlations between positive and negative moods, may be the result of failures to consider biases due to random and nonrandom response error. When these sources of error are taken into account using multiple methods of mood assessment, a largely bipolar structure for affect emerges. The data herein speak to the importance of a multi-method approach to the measurement of mood.

Full text

Journal of Personality and Social Psychology
1993,
Vol. 64, No.
6,1029-1041
Co,
ipyrighl 1993 by Ihe American Psychological Association, Inc.
Measurement Error Masks Bipolarity in Affect Ratings
Donald Philip Green, Susan Lee Goldman, and Peter Salovey
For years, affect researchers have debated about the true dimensionality of mood. Some have
argued that positive and negative moods are largely independent and can be experienced simulta-
neously. Others claim that mood
is
bipolar, that joy and sorrow represent opposite ends of a single
dimension. The
3
studies presented in this article suggest that the evidence that purportedly shows
the independence of seemingly opposite mood
states,
that
is,
low
correlations between positive and
negative moods, may be the result of failures to consider biases due to random and nonrandom
response
error.
When these sources of error
are
taken into account using multiple methods of mood
assessment, a largely bipolar structure for affect
emerges.
The data herein speak to the importance
of a multimethod approach to the measurement of mood.
For some time now, it has been popular for investigators of
mood and emotion to assert that positive and negative affect are
independent. The idea is that someone can experience joy
tinged with sorrow, hatred tempered by love, and anger coin-
cident with kindness. A foray into the recent mood literature
reveals conclusions such as
the
following:
"Periodic factor analy-
ses.
.
.
have produced a strongly similar pattern of results over
the years: two large factors, one for positive and the other for
negative affect" (Moore & Isen, 1990, pp. 4-5). The view that
positive and negative affect are not opposite ends of a single
dimension (i.e., are not strongly negatively correlated) but, in-
stead, represent nearly orthogonal dimensions of mood can be
traced to several highly influential studies of well-being by
Bradburn and his colleagues (Bradburn,
1969;
Bradburn &
Cap-
lovitz, 1965).
We
suspect that the empirical assumptions on which the "in-
dependent factors" view of positive and negative mood rest are
unsound. In our view, the independence of positive and nega-
tive affect is a statistical artifact. The conclusion that positive
and negative affect are largely uncorrelated fails to take account
of the errors of measurement that arise in mood assessment—
Donald Philip Green, Department of Political
Science,
Yale
Univer-
sity; Susan Lee Goldman and Peter Salovey, Department of Psychol-
ogy, Yale University.
This article benefited considerably from comments by Edward
Diener, Jay Hull, John
D.
Mayer, Alexander Rothman, David Watson,
and four anonymous reviewers regarding a draft. Thanks to Jon Cow-
den, who assisted in the preparation of the manuscript.
This research was supported
by
a grant to Donald Philip Green from
Yale's Institution for Social and Policy Studies, by a National Science
Foundation Minority Predoctoral Fellowship to Susan Lee Goldman,
and by the following grants to Peter Salovey: National Institutes of
Health Grants BRSG S07 RRO7O15 and CA 42101 and National
Science Foundation Grant BNS-9058020.
The survey instruments used to collect the mood data used in this
article as well as the data themselves are available from us on request.
Correspondence concerning this article should be addressed to
Donald Philip Green, Department of Political Science, Yale Univer-
sity,
P.O.
Box
3532,
Yale
Station,
New Haven, Connecticut 06520-3532,
or to Peter Salovey, Department of Psychology, Yale University, P.O.
Box 11A, Yale Station, New Haven, Connecticut 06520-7447.
errors that occur because people have difficulty translating
their moods into survey responses. We argue that widely used
methods of mood assessment are statistically unreliable and
that, moreover, the measurement error associated with the as-
sessment of positive and negative feelings is not random. In-
stead, the errors of measurement in both scales tend to be
correlated because they emerge from the same sources. We
shall demonstrate that when random and nonrandom measure-
ment error is taken into account, the independence of positive
and negative affect, however defined, proves ephemeral. When
one adjusts for random and systematic error in positive and
negative affect, correlations between the two that at first seem
close to 0 are revealed to be closer to -1.00 and support a
largely bipolar structure.
Measurement of Mood
Historically, early affect researchers assumed the intuitively
appealing idea that positive and negative moods represent op-
posite poles of one underlying dimension. To these investiga-
tors,
it seemed likely that a happy person was one who was not
sad and that a sad individual could not simultaneously be
happy. Such notions led Guilford (1954), echoing Wundt
(1897),
to declare that "an affective scale is a bipolar one" (p.
264).
Affect was thought to be essentially the same as the first
dimension of the semantic differential—evaluation—with simi-
lar pleasant and unpleasant poles (Osgood, Suci, & Tannen-
baum, 1957).
However, early factor analytic work rarely confirmed this
bipolar formulation. When investigators asked subjects to rate
mood adjectives, pleasant and unpleasant items tended not to
load on opposite ends of one dimension but rather formed two
separate dimensions, each with a single pole. It seemed that
people either felt pleasant or not or felt unpleasant or not, and
their ratings of positive and negative affect seemed largely inde-
pendent
(e.g.,
Borgatta,
1961;
Clyde,
1963;
McNair
&
Lorr,
1964;
Nowlis &
Nowlis,
1956).
These investigators assumed that bipo-
larity could only be found when it was forced onto the data by
deliberately labeling the ends of rating scales with adjectives
that were antonyms, but that when affect was measured with
only a single adjective for each scale item, separate positive and
1029

1030
D.
GREEN, S. GOLDMAN, AND P. SALOVEY
negative dimensions would be recovered (R.
F.
Green
&
Gold-
fried,
1965).
Consequently, various mood
scales were
developed
on the basis of
the
view that affect was best characterized by at
least two monopolar factors (e.g., Izard's, 1972, Differential
Emotions Scale; McNair, Lorr, & Droppleman's, 1971, Profile
of Mood
States,
and
Nowlis's,
1965,
Mood Adjective Checklist).
It
was
not long, however, before the independent factors view
was
challenged. Bentler
(1969)
constructed an adjective version
of the semantic differential containing 141 evaluative terms.
Two hundred subjects were asked to describe the emotional
meaning of 200 different concepts using these terms
(as
well as
other terms loading on the other two traditional dimensions of
the semantic differential, activity and potency; there were 352
total adjectives). This massive matrix was subjected to non-
metric multidimensional scaling (Bentler, 1966) to identify the
latent ordinal dimensions associated with observed adjective
ratings. Once the biasing effects of response style were offset by
controlling for the number of adjectives checked by a subject, a
pattern of correlations emerged that supported a bipolar struc-
ture.
Positively valenced adjectives were inversely correlated
with negatively valenced ones.
In the years that followed, investigators identified other sys-
tematic sources of error variance that, when accounted for,
transformed data that appeared
to
contain independent, mono-
polar positive and negative affect factors into dimensions that
were in fact better characterized by bipolarity. For instance,
Meddis (1972) identified a set of systematic sources of error
variance that tends to operate in adjective rating
scales,
namely
that (a) many contained a "don't know" category that was not
really in the middle of the scale
(e.g.,
Thayer, 1967), which sub-
jects often used to indicate noncomprehension of the adjective
or
its
irrelevance to the rating task and
(b)
many adjective rating
scales were not symmetrical (e.g., McNair
&
Lorr, 1964), often
containing two or three levels of acceptance (e.g., extremely,
quite a
bit,
and
a little)
and only one level of rejection
(e.g.,
not at
all).
Meddis
(1972)
claimed that such asymmetries in
scale
con-
struction were "suppressing negative correlations and prejudic-
ing the factor analysis against the discovery of bipolar factors"
(p.
180). After correcting for such sources of systematic error,
Meddis (1972) obtained strong evidence for
a
small set of
bipo-
lar mood factors.
The piece de resistance for the view that affect is bipolar was
a now classic study whose title declared, simply, "Affect Space is
Bipolar" (Russell, 1979; see also related work, Russell, 1978;
Russell
&
Mehrabian,
1977).
Russell
(1979)
argued that bipolar-
ity was suppressed in most studies of affect because of a series
of measurement issues that created systematic error, including
those described by Meddis (1972), and (a) the sample of emo-
tion words included on scales often underrepresented one end
of a bipolar continuum, (b) instructions often asked subjects to
rate how they feel over extended periods of
time,
(c) response
formats often resulted in bimodal rather than normal distribu-
tions for each item (the modal response was often not at all),
and
(d)
items in close proximity on the scale often showed spuri-
ously inflated intercorrelations. After correcting these short-
comings, Russell (1979) found strong evidence that these de-
fects in measurement had previously obscured the fact that af-
fect space is bipolar and defined, in part, by a strong, primary
bipolar factor, Pleasure-Displeasure, and a secondary bipolar
factor, Aroused-Sleepy.
After 20 years, during which the original independent view
was largely replaced by this bipolar perspective, it was some-
what unexpected that the 1980s would be characterized by a
resurgence of interest in a model of affect claiming relative
independence of positive and negative factors. Three indepen-
dent laboratories published prominent articles in fairly quick
succession, all providing evidence that, at least under certain
conditions, mood
was
best characterized
by two
broad indepen-
dent dimensions, positive affect and negative affect (Diener &
Emmons, 1984; Warr, Barter & Brownbridge, 1983; Zevon &
Tellegen, 1982).
Zevon and Tellegen (1982), taking an ideographic approach
to the study of mood, asked a small group of subjects to com-
plete a daily 60-item mood adjective checklist for 90 consecu-
tive
days.
P-factor analytic techniques were used to analyze the
data, and each subject's first two factors were derived and com-
pared (using within-individual or P-factor analysis). In 21 of
their 23 cases, these factors were largely independent positive
and negative affect dimensions. (These factors were not in a
strict sense monopolar; their opposite ends reflected the adjec-
tives
sleepy
and
calm)
Zevon and Tellegen (1982) characterized
these two factors as "descriptively bipolar but affectively unipo-
lar dimensions" (p. 112). These investigators did not see their
results as necessarily contradicting those offered by the "bipo-
lar"
camp.
Rather, they viewed the issue
as,
in part, a rotational
one.
If one chose to rotate the Zevon and Tellegen (1982) posi-
tive and negative dimensions
45°,
a bipolar Pleasant-Unpleas-
ant dimension (and
a
unipolar arousal dimension) would result.
(See Watson, 1988, pp. 128-129, for
a
discussion of how dimen-
sional labeling varies with factor rotation.) And as Watson
(1988) argued in several studies building on the Zevon and
Tellegen (1982) approach, the selection of adjectives deter-
mines the factor analytic solution and hence the emergence of
independent or bipolar mood factors
(see
also Larsen
&
Diener,
1992;
Watson & Tellegen, 1985).
Another perspective promoting the "independence" view
was offered by Diener and Emmons (1984). They noted that
positive and negative affect could be strongly and inversely re-
lated at any given moment in time and yet still be independent
in terms of how people reflect on their moods over a longer
period of
time.
Using daily mood reports that extended from 30
to 70 days, they discovered that positive and negative affect
were inversely correlated only during intense emotional experi-
ences.
However, as a longer and longer time frame was consid-
ered (from moments, to days, to
3-week
intervals), the correla-
tion between positive and negative affect decreased dramati-
cally. Diener and Emmons (1984) concluded that under
conditions of low intensity and over longer time frames, posi-
tive and negative affect might not be polar opposites (see also
Diener & Iran-Nejad, 1986, for evidence supporting this view
and Watson, 1988, for evidence against it), and mood might be
best characterized by dimensions of intensity of affect and fre-
quency of
positive
and negative affect (Diener, Larsen, Levine,
& Emmons, 1985).
A
third perspective on the independence of
positive
and nega-
tive affect was provided by Warr et
al.
(1983). They argued that
because good and bad events do not tend to be associated

BIPOLARITY OF MOOD
1031
across persons (experiencing positive outcomes does not neces-
sarily mean that one might not also experience negative out-
comes) the dominant reactions to these events, positive and
negative affect, will be independent
as
well.
On the other hand,
if a response format is provided that asks subjects to indicate
the
proportion
of
time
they had experienced positive or negative
affect, the two would more likely be inversely correlated (Warr
etal., 1983).
Following
the
publication of these three
lines
of work,
as
well
as a
related
series
of studies suggesting that separate personality
dimensions, extraversion versus neuroticism, undergirded and
perhaps caused the independence of positive and negative af-
fect (Costa
&
McCrae,
1980;
Emmons
&
Diener, 1986; Larsen,
1989;
McCrae
&
Costa, 1983), a virtual cottage industry devel-
oped with the goal of demonstrating that positive and negative
affect were indeed independent across a variety of contexts.
Separate pleasant and unpleasant factors seemed to character-
ize the emotional experiences of bereaved individuals who had
recently lost a spouse (Porritt & Bartrop, 1985). Independent
positive and negative dimensions could be recovered in cross-
cultural mood data, for example, in a large sample of Japanese
subjects (Watson, Clark,
&
Tellegen, 1984). Relatively indepen-
dent positive and negative affect factors were discovered in tra-
ditional, multidimensional mood scales like the Multiple
Mood Adjective Checklist, although it was acknowledged that
response sets may have contributed to the observed indepen-
dence (Gotlib & Meyer, 1986). Data from adolescents who
carried electronic beepers and who were asked to report on
their moods at random times revealed frequency rates of
posi-
tive and negative affect that were not correlated with each other
(Larson, 1987). And, in perhaps the largest sample studied,
Mayer and Gaschke (1988) confirmed that a two-dimensional
structure of mood characterized the responses of nearly 1,600
undergraduates who completed three different mood scales (al-
though, like Zevon and Tellegen [1982] before them, Mayer
and Gaschke [1988] noted that bipolar Pleasant-Unpleasant
and Arousal-Calm factors were simply rotated variants of the
separate Positive Affect-Tired and Negative Affect-Relaxed
factors). In recent years, the popularity of separate, orthogonal,
positive or negative dimensions of mood has not waned, al-
though the degree of independence
is
thought to vary
as a
func-
tion of
the
particular adjectives chosen (Watson, 1988), the in-
tensity of the affects considered, and the time frame during
which they were measured (Diener
&
Iran-Nejad,
1986),
as
well
as
the response format used to measure them (Warr et
al.,
1983).
Independence of Positive-Negative
Affect: Systematic
and
Random Measurement Error
Although we acknowledge the potential theoretical value of
considering positive and negative affect from an independent
dimensions perspective (especially the importance for under-
standing the underpinnings of subjective well-being, see
Diener, 1984, in press), it is the purpose of this article to call
attention to methodological issues that may have caused the
field prematurely to undervalue the traditional, bipolar view of
affect.
In the study described earlier, Bentler (1969) noted that a
systematic source of error variance, the acquiescent response
style or the tendency to check adjectives of all kinds, masked
the bipolar nature of semantic space, dramatically reducing
correlations between the opposite ends of bipolar constructs.
Writing more than 20 years ago, Bentler (1969) made the pre-
scient observation that "rating scales
in
general are
quite
suscep-
tible to an extremity response style. If an extremity response
style
existed...
its
effects would be to attenuate the potentially
high
negative
correlation between polar oppositional
terms.
. .
polar oppositional semantic tendencies might
be
negated
by
the
existence of nonoppositional, or one-sided, irrelevant response
tendencies"
(pp.
34-35).
Bentler
was
in fact identifying
a
partic-
ular exemplar of exactly the problem that we address here.
The purpose of the present set of studies is to demonstrate
that questionnaire items about mood that are worded similarly
and placed close together evoke similar response biases, giving
rise to error
covariation.
When this source of nonrandom mea-
surement error is combined with random error of measure-
ment, correlations between positive and negative affect scales
that bear
no
resemblance to true correlations
may be
generated.
Constructs that are truly bipolar (large negative correlations
between them) may appear to be independent (correlations
near zero). Our hypothesis is that raw correlations and factor
analyses that do not take the special properties of measurement
error into account tend to suggest that positive and negative
affect are largely independent. When systematic and random
sources of error variance are accounted for, a bipolar model of
affect emerges.
A
Brief Look at How Measurement Error Distorts
Measures
of
Association
When constructs are mismeasured, even strong underlying
associations may turn up weak or incorrectly signed. Consider,
by way of illustration, the product-moment correlation be-
tween happiness, denoted £,, and sadness, £
2
:
If these latent
constructs were observed without error, the correlation would
be (/ = individual)
(1)
Suppose, however, that our information about these two vari-
ables
is
imperfect. Instead of observing
£,,
and
£
2(
>
we
observe
;q,
and x
2i
, respectively:
(2)
(3)
where
8
U
and
5
2
,
represent the error in measuring happiness and
sadness, respectively.
Assume for the moment that the latent mood factors are
independent of the measurement error terms and that the
errors associated with the measure of happiness are indepen-
dent of the errors associated with the measure of
sadness.
Thus,
in the absence of sampling error,

1032
D.
GREEN,
S.
GOLDMAN,
AND P.
SALOVEY
cov (4,, 8
ki
) = cov
(5,,-, 5
2l
)
= 0. (4)
The correlation between ^, and x
2i
would then be
(£„.,
var<5
2
(5)
It follows directly that
\r
XltX2
\
<
|r
J]>fa
|
unless
van?,,
=
var<5
2
,
= 0. In
other words, random error produces correlation coefficients
that are biased toward zero.
This,
of course,
is
hardly
news,
especially to researchers work-
ing in
the factor analytic
tradition.
The
statistical problem, how-
ever, grows more complex when the data contain nonrandom
response biases, whether due to acquiescence (Bentler, 1969),
extreme response style (Diener et
al.,
1985), or idiosyncratic use
of response options (D. P. Green, 1988). In determining how
systematic response biases contribute to the distortion of corre-
lation coefficients, consider the following case. As before, we
assume that the measurement errors are independent of the
two moods, cov(^,, d
ki
) = 0, but this time we allow response
biases to affect both measures, such that
cov(5,,,
8
2i
)
+ 0.
The correlation between x,, and x
2
, becomes
cov
(<5,,,
5
2j
)
var<5
2
(6)
Depending
on the
sign
and
magnitude
of
cov(<5,,,
8
2i
) and the
reliability
of the two
proxies,
r
XiX2
may be
greater than, less
than,
or
equal
to
r
f|{2
.
In
sum, random measurement error atten-
uates correlation coefficients; nonrandom error, however,
may
produce correlations that have
the
incorrect sign. Exploratory
or confirmatory factor analyses
of
data collected using
a
single
measurement approach often produce misleading results,
un-
less special allowances
are
made
for
nonrandom error (Bank,
Dishion, Skinner,
&
Patterson,
1990;
Dillon, Kumar,
& Mu-
lani,
1987).
The
data presented below demonstrate that
mea-
surement error
in
single-method assessments
of
happy
and sad
mood states
can
lead
one to
overestimate their degree
of
inde-
pendence. When multiple methods of data collection
are
used,
confirmatory factor analysis indicates that happy
and sad
mood states
are
largely bipolar.
Study
1:
The Bipolarity
of
Mood
at
Two Time Points
In
the
current study,
we
used
a
two-phase longitudinal design
to examine
the
influence
of
response bias, both within
and
across time,
on the
dimensionality
of
mood. Data were
col-
lected
at two
time points
1
week apart.
At
each time, subjects
completed four short affect measures. Each
of the
four affect
measures differed
in
terms
of
response format
and
assessed
both positive
and
negative mood. This multimethod approach
to mood assessment allowed
us to
estimate
the
possible contri-
bution
of
systematic method-specific influences
on (a) the
reli-
ability
of
affect ratings
and (b) the
dimensional structure
of
mood.
To test
the
dimensional nature
of
mood,
we
performed
con-
firmatory factor analysis (CFA) using both LISREL
VI and VII
(Joreskog
&
Sorbom, 1986,1988). This procedure finds the com-
bination
of
parameters that maximizes
the
likelihood
of ob-
taining
the
variance-covariance matrix of the observed sample
data. CFA allows
us to
specify,
a
priori,
the
theoretical relations
among latent mood factors, observed measures,
and
unique
factors (also known
as
random errors). When constructs
are
measured with multiple indicators,
it is
possible
to
estimate
both
the
interfactor correlations while simultaneously estimat-
ing intramethod correlations among
the
errors of measurement
(Bollen, 1989). Thus,
CFA
allows
us to
consider both random
and systematic variation when estimating
the
relations among
the latent factors. Finally,
CFA
makes
it
possible
to
assess
the
statistical
fit of
one model
in
relation
to
other models that
in-
voke different
a
priori assumptions.
Method
Subjects. Subjects were recruited from an introductory psychology
course
at a
large northeastern university during
the
spring
of
1991.
Participation in the study
was
not
a
course requirement,
no
extra credit
incentive
was
offered for experimental participation, and subjects who
chose
to
participate did not receive financial compensation.
Of the 232 students
who
attended class during the
first
phase of data
collection, 209 consented
to
participate
in a
study
on
mood.
At the
second data collection,
one
week later,
147 of
the
209
participants
returned questionnaires. However, because some items
in
either
the
first or second set
of
measures were left unanswered, complete infor-
mation was
not
available
for
8 subjects. As
a
result, the analyses pre-
sented here are based on 139 subjects
(71
women and 68 men).
Procedure.
The
procedure
for
administering
the
questionnaires
was similar
at
both data collection times. Before
the
start
of
class,
subjects were asked to complete four mood measures to describe how
they were feeling "this morning."' The mood measures were identical
at each time, and the procedure took no more than 10 min.
Mood
measures.
There are three common response formats used
to measure current emotional experiences: (a) the adjective checklist,
(b) the response options format, and
(c)
the H-point Likert
scale.
These
formats were used in designing four measures of mood for the current
study:
a
mood adjective checklist,
a
response-options format that pre-
sented a list of statements to which the subjects indicated their degree
of agreement
by
choosing
a
number ranging from
1 (strong
disagree-
ment)
to
5
(strongagreement),
a response-options format that presented
statements to be rated from
1 (very well)
to
4 (not at all)
according
to
the
degree
to
which they described
the
subject's mood,
and a
semantic
differential Likert scale. Words
and
phrases used
to
create measures
were drawn from lists
of
empirically derived synonyms (e.g., Izard,
1977;
McNair et al., 1971)
to
represent either happy or sad mood. The
actual terms used,
as
well
as the
question formats,
are
presented
in
Appendix
A.
Results
Measurement
model.
The dimensionality of mood states is
conventionally assessed by means of the correlation coeffi-
cient. As the relevant correlation is between positive and nega-
tive
affect factors, rather than survey measures, statistical analy-
sis requires a measurement model linking unobserved moods
to observed indicators. The model we posit, which is presented
in formal algebraic terms in Appendix B, supposes the ob-
served score for each individual to be a linear combination of
mood factor, systematic response bias, and random error. Note
that our model does not assume that the data contain random
1
Note that
our
measurement approach asks subjects
to
assess
the
general character of their
moods,
rather than the frequency with which
they have experienced moods of different sorts (Larson, 1987).

BIPOLARITY OF MOOD
1033
and nonrandom error; if
these
measurement problems are not
in evidence, the results will indicate this. The important as-
sumptions of our model are that systematic response errors are
uncorrelated with mood factors and that measurement errors
are uncorrelated across items with different question formats.
These
assumptions,
like those of most
any
quantitative analysis,
are not unassailable; but as we demonstrate below, our results
are highly robust across a spectrum of different statistical as-
sumptions.
An
illustrative example. Let us
begin,
however,
by
disregard-
ing the redundancy with which mood
is
measured in our study;
for purposes of illustration, let us pretend we have only the 10
adjective checklist measures for happiness and sadness at one
point in time. To meet the statistical identification require-
ments
of
CFA,
we
would create
two
subscales for happiness and
for sadness. Arbitrarily arraying the mood adjectives in alpha-
betical order (the order in which they were presented), we
would create one happiness scale by summing responses to the
words
cheerful
and
elated;
another scale by adding responses to
glad,
happy, and
joyful.
Two corresponding sadness scales
would be created in similar fashion. The correlation matrix of
these four scales is listed in Table 1. Notice that on casual in-
spection, the matrix seems to suggest two mildly negatively
correlated factors, one for the happy items and another for the
sad. Indeed, when we posit a two-factor model that assumes
only random error (cf. Long,
1983),
that
is
what
we
find.
The x
2
for this model is 0.01 (df=
l,N=139,p=
.92), and the esti-
mated latent correlation between happiness and sadness
is
just
-.34.
Moreover, the statistical
fit
for the two-factor model com-
pletely dominates the fit associated with a nested one-factor
alternative that assumes only random error
is
present. Comput-
ing the
difference between the
x
2
s
associated with the
two
mod-
els,
a statistic that is itself distributed x
2
, we obtain
65.26,
(df=
1,
N
=
139, p
<
.001), which argues for a two-factor model.
The abysmal performance of the model that
assumes a
single,
bipolar mood state seems to suggest that the mood states of
happiness and sadness vary more or less independently of one
another. But the one-factor model can be resuscitated by relax-
ing the assumption that the adjective checklist items contain
only random
error.
The x
2
associated with the one-factor model
can be driven from 65.27 to 0.14 if we assume that a consistent
pattern of response bias (and hence covariance among the
errors of measurement) runs throughout the adjective check-
Table
1
Descriptive Statistics
and
Intercorrelations
for
Illustrative
Example
Described in
Study
1
Ratings
1.
Happy
1
2.
Happy 2
3.
Sadl
4.
Sad2
1
.57**
-.15*
-.21*
2
-.18*
-.26**
3
.56**
4 ?M
0.39
0.48
0.40
0.39
SD
0.60
0.72
0.66
0.71
Note. Happyl = summed ratings for cheerful and elated; Happy2 =
summed ratings for
glad,
happy,
and joyful; Sadl
=
summed ratings for
blue and discouraged; Sad2 = summed ratings for feeling
low, low,
and
sad.
*p<.05.
**p<.001.
list. Note that even with this highly restrictive assumption—
namely, each of the items absorbs the same amount of system-
atic measurement error variance—we can obtain
a
statistical fit
that is on par with the two-factor model. In other words, fit
statistics for these two nonnested alternative models may pro-
vide little guidance
as
to whether mood
states are
in fact bipolar
(cf. Burke,
Brief,
George, Roberson, & Webster, 1989, p. 1095).
The most sensible solution to this problem of indeterminacy
would
be
to posit
a
pair of nested
models
that
allows
for varying
degrees of nonrandom error while permitting
two
possibly dis-
tinct mood factors to
emerge.
The difficulty
is
that the parame-
ters of
these
models will not be identified unless
we
have addi-
tional measures of mood—measures that are subject to differ-
ent sorts of response bias. The essential feature of the present
study is that it uses a variety of mood measures so as to over-
come the identification problem. In short, using multiple mea-
sures,
we develop a series of nested models that sustains the
potential distinction between opposite mood states while also
permitting nonrandom measurement error.
Correcting
the
observed correlation
between happiness and
sadness. The first and most parsimonious measurement
model assumes that measurement error associated with each of
the four types of survey questions is random. Each latent vari-
able
is
assumed to take on the metric of the adjective checklist,
leaving three unstandardized loadings to be estimated for each
latent mood factor. In addition, we estimate one measurement
error variance for each of the
scales,
for
a
total of
16
parameters.
The four latent moods (happiness and sadness at both points in
time) are assumed to be intercorrelated, adding another
10
pa-
rameters to be estimated. In all, the basic measurement model
involves 38 parameters (12 factor loadings + 16 measurement
error variances + 10 mood factor covariances).
Tables
2
and
3
offer
a
striking contrast between the raw inter-
item correlations and the interfactor correlations we obtain
after correcting for random measurement error. Consider, for
example, the correlation between the adjective checklist scale
for happiness and that for sadness at the first point in time.
Although this correlation is observed to be -.25 (see Table 2),
the underlying correlation
is
estimated to be -.84
(see
Table
3).
This distortion is directly attributable to the poor reliability of
the mood adjective checklist scales. The adjective checklist,
however, is not the only scale containing measurement error.
The strongest negative correlation observed in the first wave,
—.69
between an
agree-disagree scale
and a Likert
scale,
never-
theless understates the actual degree of bipolarity of mood.
Had we summed all our various mood measures into a single
index, the observed correlation between happy and sad mood
would have been -.72. Summing all measures other than the
adjective checklist yields a correlation of-.74.
Mismeasurement need not be solely a matter of random
error. Accordingly, the second model permits nonrandom error
between survey items of similar response format. For example,
the model allows for the possibility that a propensity to check
boxes influenced individuals' scores on each of
the
four adjec-
tive checklist scales. Similar allowance for response tendency
was made for each of the other three measures, for a total of 24
additional parameters to be estimated.
Relaxing the assumption of random error produces a dra-
matic and statistically significant improvement in fit over the

1034
D.
GREEN, S. GOLDMAN, AND P. SALOVEY
Table 2
Raw Interitem Correlations Among
All
Indicators in Study
I
Indicator
10
11
12
13 14
15 16
1.
2.
3.
4.
5.
8.
9.
10.
11.
12.
13.
14.
15.
16.
Tl H-ACL
Tl H-A/D
Tl H-Desc
Tl H-Likert
Tl
S-ACL
6. Tl S-A/D
7.
Tl
S-Desc
TIS-Likert
T2 H-ACL
T2 H-A/D
T2 H-Desc
T2 H-Likert
T2
S-ACL
T2 S-A/D
T2
S-Desc
T2
S-Likert
.60
.49
.55
-.25
-.39
-.35
-.40
.68
.71
-.49
-.64
-.53
-.56
.67
-.51
-.59
-.54
-.60
-.50
-.69
-.59
-.66
.62
.64
.56
.69
.67
.74 —
.13
.00
-.01
.02
.16
.01
.03
.11
.10
.08
.08
.16
.01
.09
.02
.03
.08
.13
.21
.20
-.13
-.14
-.10
-.13
.05
.06
.06
.16
.00
-.04
.01
-.03
.04
-.15
-.17
-.19
.21
.18
.11
.14
-.02
-.14
-.13
-.21
.17
.18
.09
.10
-.01
-.07
-.10
-.16
.18
.19
.19
.16
-.00
-.11
-.09
-.14
.08
.15
.12
.15
.60
.55
.63
-.25
-.37
-.41
-.42
—
.78
.82
-.53
-.67
-.59
-.59
—
.79
-.63
-.62
-.60
-.67
—
-.57
-.63
-.64
-.68
—
.61
.62
.61
—
.70
.64
—
.69
Note. The letter before each format descriptor refers to the latent factor on which each indicator loads. Across-time correlations are enclosed in
box. Tl = first data collection. T2 = second data collection. H
=
happy; S = sad. ACL
=
Adjective Check List; A/D
=
agree-disagree format; Desc =
descriptive statements format; Likert = Likert scale.
first model (difference in x
2
= 76.5, df= 24, p
<
.001,
see Table
4).
Other
results,
not reported in Table
4,
are
also
of
note.
As
we
anticipated, the checklist scores are susceptible to nonrandom
error. All six of the potential error correlations between check-
list scales proved to be positive, as expected, and five were
statistically significant using a one-tailed / test
(a =
.05). There
is evidence of nonrandom response effects among the other
items as well. Three of the remaining 18 covariances between
the errors of measurement proved to be significant at the .05
level (two-tailed), testifying to the existence of systematic re-
sponse effects that occur within a survey and persist over time.
2
Interestingly enough, the estimated interfactor correlations
do not change much when our CFA model takes nonrandom
error into account. This finding took us by surprise and led us
to an interesting methodological insight: By measuring moods
in a redundant fashion, one typically insulates a CFA analysis
from the biasing effects of model misspecification. (The statis-
tical basis for this conjecture can be found in a technical report
by
D.
P Green, Goldman,
&
Salovey, 1992). In short, although
assuming that the data contain only random error leads to the
wrong CFA
model, redundant measurement
increases the
num-
ber of elements in the covariance matrix that
CFA
analyzes that
are free from the effects of nonrandom error. As a result, the
CFA estimates from the "wrong" model approximate the true
parameters. In the present case, our use of four different mea-
sures of each mood produces a covariance matrix of
120
off-
diagonal elements, only 24 of which are contaminated by
nonrandom error. Naturally, allowing for nonrandom error
produces a better
fitting
statistical model, but the random error
model does an adequate job of estimating the underlying pa-
rameters of interest.
Although relaxing the assumptions of our CFA model to in-
corporate both random and nonrandom error
does
not alter the
estimated intermood correlations, it does change our explana-
tion of why the observed correlations between adjective check-
lists are not stronger. For example, the observed correlation
between adjective checklist measures of happiness and sadness
in the second wave of our study is -.25. The corresponding
latent correlation
is
estimated to be
—.85
(random error model)
or -.84 (nonrandom error model). Our estimates based on a
CFA model that allows for both random and nonrandom error
suggest that random error drives the latent correlation down to
-.42;
nonrandom error reduces it further to
-.29.
3
Lest one think that CFA transforms all weak correlations
into strong
ones,
we
find
that the over-time stability of the four
mood states
is
quite weak. The latent degree of sadness at Time
1
displays
a
correlation of just .20 with sadness
1
week later, and
the other over-time correlations are even weaker. The
finding
is
of substantive interest because it suggests that moods are not
enduring characteristics, even over short periods of
time.
This
finding has important implications for interpreting responses
to questions that ask subjects to report their moods over the
past month or year. If moods are highly transitory, such mea-
sures may be better viewed as tapping mood tendencies or
mood predispositions.
2
Had we only allowed for nonrandom error between common for-
mats within a given wave, thereby ignoring the problem of response
biases that persist over time, the result would have been a significant
improvement in fit (p < .001) over the random error model but no
change in the estimated interfactor correlations.
3
The slight difference between this correlation and the actual ob-
served correlation is an unexplained residual attributed to sampling
variability.

BIPOLARITY OF MOOD
1035
Table 3
Factor Loadings
and
Interfactor Correlations
for Random
Error
Model
at Time
1
and Time 2
in
Study
1
Table 4
Factor Loadings and Interfactor Correlations for Correlated
Error Model at Time
1
and Time 2 in Study 1
Indicator
Standardized factor
loadings"
H-ACL
H-A/D
H-Desc
H-Likert
S-ACL
S-A/D
S-Desc
S-Likert
Interfactor correlations
Tl happy
Tl sad
T2 happy
T2sad
Tl
Happy
.63
.78
.86
.85
-.84**
.16
-.02
Sad
.72
.85
.83
.83
-.18
.23*
T2
Happy
.66
.87
.92
.89
—
-.85**
Sad
.75
.83
.81
.83
•—
Indicator
Standardized factor
loadings"
H-ACL
H-A/D
H-Desc
H-Likert
S-ACL
S-A/D
S-Desc
S-Likert
Interfactor correlations
Tl happy
Tl sad
T2 happy
T2sad
Tl
Happy
.65
.80
.84
.86
—
-.84**
.14
-.04
Sad
.71
.86
.83
.83
—
-.17
.20*
T2
Happy
.69
.85
.92
.89
—
-.84**
Sad
.72
.85
.82
.80
Note. The letter before each format descriptor refers to the latent
factor on which each indicator loads. H = happy; S = sad; T = time;
ACL = Adjective Check List; A/D = agree-disagree format; Desc =
descriptive format; Likert = Likert scale.
a
X
2
(98,
N
=
139) = 162.39, p < .001. Goodness of Fit Index = .87.
Adjusted Goodness of Fit Index =
.83.
Root-mean-squared residual =
.05.
*p<.Q5.
**/?<.001.
Note. The letter before each format descriptor refers to the latent
factor on which each indicator loads. H = happy; S = sad; T = time;
ACL = Adjective Check List; A/D = agree-disagree format; Desc =
descriptive format; Likert = Likert scale.
'
x
2
(74,
N=
139) =
85.90,
p
=
ns.
Goodness of Fit Index
=
.93.
Adjusted
Goodness of Fit Index = .87. root-mean-squared residual = .05.
*p<.05. **p<.00\.
Study 2: Replication Using Varied Question Order
During
the
fall of
1991,
we replicated our initial study
using a
sample of students enrolled in an undergraduate psychology
course at the same university. The administration of
the
study
was the same, except that the sequence of the different formats
for mood assessment was varied according to a Latin-square
design. Of
the
320 students in the class, 285 completed mood-
assessment surveys. Eliminating respondents who offered in-
valid or missing answers to any of the items left 250 valid cases.
The second study confirmed the main conclusions of our
initial CFA analysis: Happiness and sadness appear to be bipo-
lar in structure, and appearances to the contrary may be attrib-
uted to random and nonrandom error (Tables 5 and 6). The
observed correlation between the happy and sad mood adjec-
tive scales is -.40. When we take random error into account,
this correlation jumps to -.92. Allowing for nonrandom re-
sponse errors between items of similar question wording and
response format cut the x
2
from 75.20 to 29.75 with an expen-
diture of just four degrees of freedom
(TV
= 250, p < .001).
Again, amid strong evidence of nonrandom error, there was
little change in the estimated correlation between happy and
sad mood states
as we
relaxed the constraints of the
CFA
model
Study 3: Replication Using Different Time Horizon and
Adjectives
Five weeks after the administration of Study
2,
we replicated
that study using the same pool of subjects. This time, however,
the time frame of our mood assessment questions
was
changed
from "how you have been feeling since this morning" to "how
you have been feeling
over the
past month"
so as
to test whether
a longer time horizon diminishes the bipolarity of mood ob-
served above (as suggested by Diener & Iran-Nejad, 1986).
Again, the sequence of the different types of mood assessment
was varied according to a Latin-square design. Because the
questionnaire was administered close to the end of
the
semes-
ter, attendance was up, and 304 students completed mood as-
sessment surveys.
The data provided little support for the notion that the bipo-
larity of mood is a function of the time frame of the mood
assessment (Tables 7 and 8). A CFA model assuming random
measurement error produced a disattenuated correlation of
Table 5
Raw Interitem Correlations for Happy and Sad
Words:
Study 2
Indicator
1.
H-ACL
2.
H-A/D
3.
H-Desc
4.
H-Likert
5.
S-ACL
6. S-A/D
7.
S-Desc
8.
S-Likert
1
_
.65
.63
.67
-.40
-.50
-.55
-.61
2
.71
.76
-.62
-.66
-.59
-.68
3
.73
-.58
-.58
-.69
-.68
4
—
-.57
-.66
-.60
-.75
5
.67
.54
.65
6
—
.55
.76
7
.60
8
—
Note. The letter before each format descriptor refers to the latent
factor on which each indicator loads. H = happy; S = sad; ACL =
Adjective Check List; A/D = agree-disagree format; Desc = descrip-
tive format; Likert = Likert scale.

1036
D.
GREEN, S. GOLDMAN, AND P. SALOVEY
Table 6
Standardized Factor
Loadings and
Interfactor Correlations
for
Random and
Correlated Error Models
for
Happy
and Sad
Words:
Study 2
Table 8
Standardized
Factor
Loadings and
Interfactor Correlations
for
Random and
Correlated Error Models
for
Happy
and Sad
Words:
Study 3
Model and indicator
Random error
3
H-ACL
H-A/D
H-Desc
H-Likert
S-ACL
S-A/D
S/Desc
S-Likert
Correlated error
b
H-ACL
H-A/D
H-Desc
H-Likert
S-ACL
S-A/D
S-Desc
S-Likert
Factor loadings
Happy
.74
.86
.83
.89
.75
.87
.82
.87
Sad
.75
.83
.71
.89
.76
.84
.70
.88
Model and indicator
Random error
a
1.
H-ACL
2.
H-A/D
3.
H-Desc
4.
H-Likert
5.
S-ACL
6. S-A/D
7.
S-Desc
8. S-Likert
Correlated error
b
1.
H-ACL
2.
H-A/D
3.
H-Desc
4.
H-Likert
5.
S-ACL
6. S-A/D
7.
S-Desc
8. S-Likert
Factor loadings
Happy
.67
.79
.87
.86
.67
.80
.86
.85
Sad
.72
.85
.73
.84
.72
.85
.72
.82
Note. The letter before each format descriptor refers to the latent
factor on which each indicator loads. H = happy; S = sad; ACL =
Adjective Check List; A/D = agree-disagree; Desc = descriptive for-
mat; Likert = Likert scale.
a
X
2
(19,
N
=
250) = 75.20, p < .001; Goodness of Fit Index = .93;
Adjusted Goodness of Fit Index =
.86;
RMSR
=
.04.
Happy-sad inter-
factor correlation = -.92 (p < .001).
V(15,
N = 250) = 29.75, p = .025; Goodness of Fit Index = .97;
Adjusted Goodness of Fit Index =
.93;
RMSR
=
.03.
Happy-sad inter-
factor correlation = -.91 (p
<
.001).
Note. The letter before each format descriptor refers to the latent
factor on which each indicator loads. H = happy; S = sad; ACL =
Adjective Check List; A/D = agree-disagree format; desc = descrip-
tive format; Likert = Likert scale.
a
x
2
(19, N = 304) = 135.50, p < .001; Goodness of Fit Index = .91;
Adjusted Goodness of
Fit
Index =
.83;
root-mean-squared residual =
.06.
Happy-sad interfactor correlation = -.87 (p
<
.001).
b
X
2
(15,
N
=
304) = 25.54, p = .04; Goodness of Fit Index = .98; Ad-
justed Goodness of Fit Index =
.95;
root-mean-squared residual = .03.
Happy-sad interfactor correlation = -.87 (p
<
.001).
-.86 between happiness and sadness. Allowance for both ran-
dom and nonrandom error increased this estimate to -.87.
In keeping with previous findings, CFA assuming random
and nonrandom error produced similar substantive conclu-
sions.
Again, however, there can be no doubt about the preva-
lence
of nonrandom error in these mood
assessments.
The esti-
mated covariance between response errors associated with the
two mood checklist scales is more than 8 times its standard
Table 7
Raw
Interitem Correlations
for
Happy
and Sad
Words:
Study 3
Indicator
1
5
1.
H-ACL
2.
H-A/D
3.
H-Desc
4.
H-Likert
5.
S-ACL
6. S-A/D
7.
S-Desc
8. S-Likert
—
.60
.59
.59
-.10
-.44
-.44
-.47
—
.69
.67
-.46
-.53
-.48
-.59
.74
-.53
-.63
-.59
-.65
-.50
-.64
-.59
-.66
.68
.52
.60
.60
.69
.61 —
Note. The letter before each format descriptor refers to the latent
factor on which each indicator loads. H = happy; S = sad; ACL =
Adjective Check List; A/D = agree-disagree format; Desc = descrip-
tive format; Likert = Likert scale.
error, suggesting that respondents have markedly different pro-
pensities to check off adjectives whether due to individual dif-
ferences in acquiescence, expressiveness, or emotional intensity
(Bentler, 1969; Diener, Larsen, Levine, & Emmons, 1985).
Viewing the three studies together, we found that all four of the
within-wave correlations between the checklist measures were
significantly inflated by nonrandom error (p < .05). Like
Bentler (1969) before us, we advise caution when researchers
analyze data obtained with a checklist format.
4
Results for the other formats suggest that they vary in their
susceptibility to nonrandom error. The Likert scales produced
negatively correlated measurement errors in all four
cases,
two
significantly so. The same pattern of results obtained for the
agree-disagree format. The self-description items, however,
tended to produce positive error correlations
(3
of
4),
although
only one proved significantly positive. One feature that distin-
guished the self-description items from the Likert scales and
agree-disagree format is that the latter offered subjects a non-
4
A
PsycLIT scan for the years 1987-1991 listed more than 25 ab-
stracts that mentioned the mood adjective checklist as the primary
method by which mood was assessed. In addition, a recently devel-
oped mood measure uses an adjective checklist format (Matthews,
Jones,
&
Chamberlain, 1990).

BIPOLARITY OF MOOD
1037
committal middle response option. As D. P. Green (1988) has
argued, negative error covariance
is
typical of
items with
attrac-
tive middle alternatives.
The strongest case for
the
independence of positive and nega-
tive mood has been advanced by Watson, Clark, and Tellegen
(1988) using adjectives drawn from the mood dimensions
formed by the 45° rotation described in the introduction and
labeled
Positive
Affect and
Negative
Affect.
Positive affect de-
fined in this fashion comprises moods such
as
excited or enthu-
siastic rather than happy; negative affect refers to moods such
as distressed or nervous rather than sad. Because both mood
states operate in a state of arousal, we would not expect the
latent dimensions to bear a perfect negative correlation with
one another. In essence, arousal represents a third, oblique fac-
tor related positively to both positive and negative affect. The
question of interest is whether positive and negative affect are
orthogonal, as a two-dimensional exploratory factor analytic
solution of the Positive and Negative Affect Schedule (the
PANAS) would suggest (Watson, Clark, & Tellegen, 1988).
To address this question, we included in the questionnaire
administered in Study
3 a series
of mood
items
derived from the
PANAS.
Again, four formats were used to measure mood. The
time frame specified was "the past month," so as to give the
independence hypothesis its best opportunity to acquit
itself.
The words used to assess positive and negative affect are listed
in Appendix
C.
Our statistical analysis
is
based on
305
subjects
with valid responses.
Our analysis offers no support for the view that positive and
negative affect, operationalized using the PANAS adjectives,
represent orthogonal dimensions (see Tables 9 and 10). The
observed correlations between the two mood states range from
-.11 to -.47, and correction for measurement error places the
estimated interfactor correlation at -.57 (random error) or-.58
(nonrandom error). A formal statistical test of the hypothesis
that the two factors are orthogonal yields an unequivocal an-
swer: A x
2
difference test with
1
degree of freedom comparing
the nonrandom error model with a similar model in which the
interfactor correlation is constrained to be zero produces a
value of
81.23,
p
<
.001.
Positive and negative affect, although
perhaps not as bipolar as happy and sad moods, cannot be
regarded as orthogonal factors.
Table 9
Raw
Interitem Correlations
for
Positive and
Negative Affect
Words:
Study 3
Table 10
Standardized Factor
Loadings and
Interfactor Correlations
for
Random and
Correlated Error Models
for
Positive
and
Negative
Words:
Study 3
Indicator
1.
2.
3.
4.
5.
6.
7.
8.
P-ACL
P-A/D
P-Desc
P-Likert
N-ACL
N-A/D
N-Desc
P-Likert
1
.44
.62
.65
-.11
-.25
-.35
-.31
2
—
.55
.54
-.20
-.27
-.31
-.31
3
.64
-.23
-.30
-.45
-.40
4
—
-.14
-.37
-.44
-.47
5
—
.50
.46
.52
6
—
.61
.61
7
.61
8
—
Model and indicator
Random error"
P-ACL
P-A/D
P-Desc
P-Likert
N-ACL
N-A/D
N-Desc
N-Likert
Correlated error
b
P-ACL
P-A/D
P-Desc
P-Likert
N-ACL
N-A/D
N-Desc
N-Likert
Factor loading
Positive
.75
.65
.80
.84
.71
.65
.81
.80
Negative
.60
.77
.79
.80
.60
.77
.78
.79
Note. The letter before each format descriptor refers to the latent
factor on which each indicator
loads.
P =
positive;
N =
negative;
ACL =
Adjective Check List; A/D = agree-disagree format; desc = descrip-
tive format; Likert = Likert scale.
a
X
2
(19,
N = 305) = 124.09, p < .001. Goodness of Fit Index = .91.
Adjusted Goodness of Fit Index = .84. root-mean-squared residual =
.07.
Positive-negative interfactor correlation = -.57 (p
<
.001).
b
X
2
(15,
N = 305) = 41.59, p < .001. Goodness of Fit Index = .97.
Adjusted Goodness of Fit Index = .92. root-mean-squared residual =
.04.
Positive-negative interfactor correlation = -.58.
General Discussion
The findings presented in these studies underscore the im-
portance of multimethod research designs (Campbell & Fiske,
1959).
Although this methodological principle guides research
in some areas of social science (cf. Bank, Dishion, Skinner, &
Patterson,
1990),
it
is often honored in the
breach.
This
is
partic-
ularly true of mood research, where multimethod designs are
seldom used to assess the structure of affective states. As a
result, the empirical analysis of mood has been prone to statis-
tical artifacts, such as the finding that positive and negative
moods are weakly correlated. When random and nonrandom
sources of error are taken into account using multiple methods
of mood assessment, a largely bipolar structure for affect
emerges.
5
Previous researchers in this area have not been oblivious to
the problem of measurement error. Most, in fact, have taken
pains to assess the reliability of the mood adjective checklist
Note. The letter before each format descriptor refers to the latent
factor on which each indicator
loads.
P = positive affect; N = negative
affect; ACL = Adjective Check List; A/D = agree-disagree format;
Desc = descriptive format; Likert = Likert scale.
5
Although the negative interfactor correlations we obtain between
pleasant and unpleasant moods are extremely high, they are not -1.0.
In none of our studies does a nested x
2
difference test enable us to
accept the null hypothesis that one rather than two factors generated
the
data.
Following Bentler
(1969),
we
dub these mood states "approxi-
mately" bipolar.

1038
D.
GREEN, S. GOLDMAN, AND P. SALOVEY
scales.
The problem
is
that without redundancy in method, one
cannot readily assess reliability. Reliability assessment as it is
usually performed assumes random measurement error, but
additive scales, particularly those constructed from adjective
checklists, may well contain systematic response bias as well.
Because nonrandom error can lead to inflated reliability esti-
mates (D. P Green & Citrin, in press), researchers are led to
believe that observed correlations are
less
attenuated than they
really are, which only reinforces the tendency to conclude that
opposite affective states are relatively independent.
6
Another drawback of the single measurement approach is
that it can throw off assessments of discriminant
validity.
How-
ever tempting it might be to regress some theoretically telling
dependent variable on pleasant and unpleasant mood scales,
random and nonrandom error
are
likely to produce misleading
results. As Achen (1985) has demonstrated, the coefficients
from this regression will tend to be biased, sometimes severely.
Thus,
even if happiness and sadness were perfectly bipolar,
ordinary least squares regression might suggest that both have
independent predictive effects of different magnitudes. A re-
view
of studies demonstrating that positive and negative moods
are differentially related to various criteria—traits, life satisfac-
tion, and clinical diagnoses—lies beyond the scope of this arti-
cle.
We
should note, however, that this sort of validity research
brings its own set of statistical complications to the study of
mood structure.
Perhaps the strongest argument in favor of a multimethod
approach is that it greatly enhances the robustness of statistical
tests of bipolarity. Despite strong evidence that our data are
contaminated with nonrandom error, the precise way in which
measurement error is modeled has little bearing on the sub-
stantive conclusions generated by CFA. Across a variety of
model specifications, the underlying structure of pleasant and
unpleasant mood states turns out to be approximately bipolar,
at least for the short to medium time frames measured
here.
As
we noted in the introduction, this is by no means a new conclu-
sion. A vast number of investigators, however, operate under
the opposite assumption. A cursory scan of recent work using
PsycLIT netted no fewer than 193 empirical studies between
1987 and 1991 that measured positive and negative mood as
distinct dimensions. In addition, many studies and clinical
practices have operated under the assumption that positive af-
fect and negative affect, as defined by Watson (1988) and mea-
sured
by the
PANAS adjectives, are not only distinct, but orthog-
onal dimensions. Although we do not contend that these con-
structs are perfectly bipolar, the correlation between them, net
of measurement error, is substantial.
Beyond these substantive conclusions, however, lie some
practical insights about how one might develop more effective
measures of mood states. An alternative to Bentler's (1969)
method of administering an adjective checklist with a large
number of items (with which to "partial out" the effects of
nonrandom error) is a series of short question batteries of dif-
ferent format. Granted, the adjective checklist scale will con-
tain more random noise when the number of adjectives is re-
duced, but the abundance of across-method correlations makes
reliability correction possible and, as we have seen, robust.
Although the multimethod design we are proposing is rela-
tively modest in comparison with the ambitious multisource
research design proposed by Bank et al. (1990), it is clearly at
odds with current practice. It is conventional to ask subjects to
respond to many items of the same type so as to streamline the
survey instrument and minimize the amount of instructions
that must be given to the respondent. In our
view,
the benefits
of using a one-format questionnaire are small in comparison
with the costs of potential biases due to systematic response
effects. Three methodologically distinct 2-item assessments of
mood states, for example, are likely to be much more statisti-
cally informative than a single 20-item assessment.
7
The latter,
after all, provides limited opportunity to gauge and counteract
the effects of systematic response bias.
The methodological argument we have advanced extends
beyond the scope of mood research. Consider, for example, a
topic that has puzzled political scientists since the early 1980s:
the structure of public sentiment toward partisan and ideologi-
cal groups. Intuition suggests that people who like liberals
would tend to dislike conservatives; those who like Democrats
dislike Republicans. But at first glance, the data suggest other-
wise.
The weak zero-order correlations between these evalua-
tive dimensions seem to indicate that people who feel warmly
toward conservatives do not tend to dislike liberals (Conover
&
Feldman, 1981). Similarly, survey data suggest that feelings
about Democrats are relatively independent of feelings about
Republicans (Weisberg, 1980). Yet, when these
findings
are re-
evaluated using a multimethod CFA, random and nonrandom
measurement error conceal underlying correlations of -
.80
and
up(D.
P.
Green, 1988).
Perhaps the best illustration of
how
measurement error can
distort the strength of negative associations comes from the
National Election Survey of
1964,
in which respondents were
asked to rate the "warmth" they felt toward
a
variety of political
6
A good illustration of how standard approaches to reliability en-
hancement fail
to
overcome problems of nonrandom error
is
presented
in the Bentler (1969) study: His scales comprised dozens of
items
and
therefore appeared highly reliable. Yet, because his mood scales were
fraught with nonrandom error, he obtained weak bivariate correlations
between opposite mood
states.
Bentler
was able
to correct this problem
by devising another highly reliable measure of the response bias itself
and using this measure to "partial out" the positive correlation among
the measurement errors. The techniques described here achieve the
same result but do not require the investigator to administer a pro-
digious mood adjective checklist inventory.
7
Consider, for
example,
two alternative
ways
of studying the dimen-
sionality of positive and negative affect, as denned by Watson (1988).
One approach (see Study 3) would be to use three sets of two-item
indices (semantic differential, self-description, and agree-disagree).
This design takes only minutes to administer, facilitates a two-factor
model that allows for nonrandom error with
5
degrees of freedom, and
suggests an interfactor correlation of-.58. Another approach would
be to administer only the original 20-item PANAS scale. This battery
takes about the same amount of time to complete but does not provide
the data analyst any latitude in the way measurement error may be
modeled (cf. Green
&
Citrin, in
press).
Although with appropriate cor-
rection for mismeasurement, both approaches yield the same answer,
the former presents substantially fewer modeling
problems.
Moreover,
data collected in a single-method fashion give little diagnostic infor-
mation that might alert the researcher to the problems created by mis-
measurement.

BIPOLARITY
OF
MOOD
1039
and social groups on
a
"feeling thermometer" ranging from 0 to
100.
Two well-known groups on this list were the Ku Klux
Klan (KKK) and the National Association for the Advance-
ment of Colored People (NAACP). It is hard to imagine that
people who felt warmly toward the NAACP in 1964 could feel
anything but contempt for the KKK and vice versa, yet the
correlation between these two items turned out to be only -.31
(N=
1,454).
Far from indicating the multidimensional charac-
ter of public sentiment toward these two groups, the data attest
to the low reliability and systematic response biases that plague
this type of survey measure
(D. P.
Green, 1988).
Anomalous results of this kind should direct our attention
back to the
way
in which the data were generated.
Were
the data
collected using one or many methods? How reliable are the
measures? The point is not that two-factor solutions should be
rejected out of hand, but rather that this type of research find-
ing requires special methodological attention. Although the
present analysis does not resolve the debate about the structure
of mood, we do provide evidence to suggest that claims of
inde-
pendence deserve a second look. Until the issue is resolved,
investigators would best be advised to measure mood using
several different methods within the same study (i.e., varied
question and response formats, and if feasible, data obtained
through psychophysical assessment or clinical observation),
even if this means sacrificing the extensiveness of any single
battery of measures.
References
Achen, C. H.
(1985).
Proxy variables and incorrect signs on regression
coefficients.
Political
Methodology,
11,
299-316.
Bank, L., Dishion, T., Skinner, M.,
&
Patterson, G. R. (1990). Method
variance
in
structural equation modeling: Living with "glop."
In
G. R. Patterson
(Ed.),
Depression
and
aggression in
family interaction
(pp.
247-279). Hillsdale, NJ: Erlbaum.
Bentler,
P.
M.
(1966).
Multidimensional homogeneity
scaling.
Paper
pre-
sented
at the
Special Meeting
of the
Psychometric Society,
New
York.
Bentler,
P.
M. (1969). Semantic space is (approximately) bipolar.
Jour-
nal
of
Psychology,
71,
33-40.
Bollen,
K. A.
(1989).
Structural equations
with latent
variables.
New
York: Wiley.
Borgatta,
E.
F. (1961). Mood, personality
and
interaction.
Journal
of
General
Psychology,
64,105-137.
Bradburn, N. M.
(1969).
The structure
of psychological
well-being.
Chi-
cago:
Aldine.
Bradburn,
N.
M., & Caplovitz, D. (1965).
Reports
on
happiness.
Chi-
cago:
Aldine.
Burke,
M. J.,
Brief,
A.
P., George,
J.
M., Roberson, L., & Webster,
J.
(1989).
Measuring affect
at
work: Confirmatory factor analyses
of
competing mood structures with conceptual linkage to cortical regu-
latory systems. Journal
of
Personality
and
Social
Psychology,
57,
1091-1102.
Campbell, D. T., & Fiske, D. W (1959). Convergent and discriminant
validation
by the
multitrait-multimethod matrix.
Psychological
Bulletin,
56,81-105.
Clyde,
D.
J.
(1963).
Manual for the Clyde Mood
Scale.
Coral
Gables,
FL:
Biometric Laboratory, University
of
Miami.
Conover,
P.,
&
Feldman,
S.
(1981).
The origins and meaning of liberal-
conservative self-identifications. American Journal
of
Political
Science,
25,617-645.
Costa, P.
T, &
McCrae,
R. R.
(1980). Influence
of
extraversion
and
neuroticism
on
subjective well-being: Happy and unhappy people.
Journal
of
Personality
and
Social
Psychology,
38,
668-678.
Diener,
E.
(1984). Subjective well-being.
Psychological
Bulletin,
95,
542-575.
Diener,
E. (in
press). Challenges
in
measuring subjective well-being
and ill-being.
Psychological
Bulletin.
Diener,
E.,
&
Emmons, R.
A.
(1984).
The independence of positive and
negative affect. Journal
of
Personality
and
Social
Psychology,
47,
1105-1117.
Diener,
E., &
Iran-Nejad,
A.
(1986).
The
relationship
in
experience
between various types
of
affect.
Journal
of
Personality
and Social
Psychology,
50,1031-1038.
Diener, E., Larsen, R. J., Levine,
S.,
&
Emmons R. A. (1985). Intensity
and frequency: Dimensions underlying positive and negative affect.
Journal
of
Personality
and
Social
Psychology,
48,1253-1265.
Dillon, W R., Kumar, A.,
&
Mulani,
N.
(1987).
Offending estimates in
covariance structure
analysis:
Comments on the causes of and solu-
tions to Heywood cases.
Psychological
Bulletin,
101,126-135.
Emmons,
R. A., &
Diener,
E.
(1986). Influence
of
impulsivity
and
sociability on subjective
well-being.
Journal of Personality and Social
Psychology,
50,1211-1215.
Gotlib, I. H.,
&
Meyer, J.
P.
(1986).
Factor analysis of the multiple affect
adjective check list:
A
separation
of
positive
and
negative affect.
Journal
of
Personality
and
Social
Psychology,
50, 1161-1165.
Green,
D.
P.
(1988).
On the dimensionality of public sentiment toward
partisan
and
ideological groups. American Journal
of
Political
Science,
32, 758-780.
Green, D. P., & Citrin, J.
(in
press). Measurement error and the struc-
ture
of
attitudes:
Are
positive
and
negative judgments opposites?
American Journal
of
Political Science
Green,
D.
P.,
Goldman,
S.
L.,
&
Salovey,
P.
(1992).
The
effects
ofnonran-
dom
error
on
confirmatory
factor
analysis:
How
multiple-methods
improve
robustness.
Unpublished manuscript, Yale University, New
Haven, CT.
Green, R. F,
&
Goldfried, M. R.
(1965).
On the bipolarity of semantic
space.
Psychological
Monographs,
79
(Whole No. 599).
Guilford,
J. P.
(1954).
Psychometric
methods
(2nd ed). New
York:
McGraw-Hill.
Izard,
C. E.
(1972).
Patterns
of
emotions.
San
Diego, CA: Academic
Press.
Izard, C. E. (1977). Human
emotions.
New York: Plenum Press.
Joreskog,
K. G,
& Sdrbom, D. (1986). LISREL
VI.
Mooresville,
IN:
Scientific Software.
Joreskog,
K. G,
& Sbrbom, D. (1988). LISREL
VII.
Mooresville,
IN:
Scientific Software.
Larsen,
R.
J. (1989, August).
Personality as
an
affect dispositional
sys-
tem.
Paper presented
at
the 97th Annual Convention of the Ameri-
can Psychological Association, New Orleans, LA.
Larsen,
R. J., &
Diener,
E.
(1992). Promises
and
problems with
the
circumplex model of emotion.
Review
of
Personality
and
Social
Psy-
chology,
13,
25-59.
Larson,
R.
W (1987).
On the
independence
of
positive
and
negative
affect within hour-to-hour experience.
Motivation
and
Emotion,
11,
145-156.
Long, J.
S.
(1983).
Confirmatory factor
analysis.
Sage University Paper
series on
Quantitative Applications
in
the Social Sciences
(Series
No.
07-033). Beverly Hills, CA: Sage.
Matthews,
G,
Jones,
D.
M.,
&
Chamberlain, A. G.
(1990).
Refining the
measurement of mood: The UWIST mood adjective checklist.
Brit-
ish Journal
of
Psychology,
81, 17-42.
Mayer,
J.
D,
&
Gaschke,
Y.
N.
(1988).
The experience and meta-exper-
ience
of
mood. Journal
of
Personality
and Social
Psychology,
55,
102-111.
McCrae,
R.
R.,
&
Costa,
P.
T. (1983). Psychological maturity and sub-

1040
D.
GREEN,
S.
GOLDMAN,
AND P.
SALOVEY
jective well-being: Toward
a
new
synthesis.
Developmental
Psychol-
ogy,
19,
243-248.
McNair, D. M., & Lorr, N. (1964).
An
analysis
of
mood
in
neurotics.
Journal
of
Abnormal
and
Social
Psychology,
69, 620-627.
McNair, D. M., Lorr, M., & Droppleman, L.
F.
(1971).
Manual:
Profile
of mood
states.
San
Diego, CA: Educational and Industrial Testing
Service.
Meddis, R.
(1972).
Bipolar factors in mood adjective
checklists.
British
Journal
of
Social
and
Clinical
Psychology,
11,178-184.
Moore,
B.
S.,
&
Isen, A. M. (1990). Affect and social behavior.
In
B.
S.
Moore &
A. M.
Isen (Eds.), Affect and
social behavior
(pp.
1-21).
Cambridge, England: Cambridge University Press.
Nowlis,
V
(1965).
Research with the Mood Adjective Check List. In S.
S. Tomkins & C.
E.
Izard (Eds.),
Affect,
cognition,
and
personality
(pp.
352-389). New York: Springer.
Nowlis,
V, &
Nowlis,
H. H.
(1956).
The
description
and
analysis
of
mood.
Annals
of the
New York Academy
of
Sciences,
65, 345-355.
Osgood,
C,
Suci, G., & Tannenbaum, P. (1957). The
measurement
of
meaning.
Urbana: University
of
Illinois Press.
Porritt, D,
&
Bartrop, R.
W
(1985).
The independence of pleasant and
unpleasant affect: The effects of bereavement.
Australian Journal of
Psychology,
37, 205-213.
Russell,
J. A.
(1978). Evidence
of
convergent validity
on the
dimen-
sions
of
affect. Journal
of
Personality
and
Social
Psychology,
36,
1152-1168.
Russell,
J.
A. (1979). Affective space
is
bipolar.
Journal
of Personality
and
Social
Psychology,
37, 1161-1178.
Russell,
J. A., &
Mehrabian,
A.
(1977). Evidence
for a
three-factor
theory of emotions.
Journal
of
Research
in
Personality,
11,
273-294.
Thayer,
R. E.
(1967). Measurement
of
activation through self-report.
Psychological
Reports,
20, 663-678.
Warr,
P.,
Barter, J.,
&
Brownbridge, G. (1983). On the independence of
positive
and
negative affect.
Journal
of
Personality
and Social
Psy-
chology,
44,
644-651.
Watson, D. (1988). The vicissitudes
of
mood measurement: Effects of
varying
descriptors,
time frames, and response formats
on
measures
of positive and negative affect. Journal of Personality and Social Psy-
chology,
55,
128-141.
Watson, D, Clark, L. A.,
&
Tellegen, A. (1984). Cross-cultural conver-
gence
in
the structure of mood: A Japanese replication and a com-
parison with
U.S.
findings.
Journal of Personality and Social
Psychol-
ogy,
47,
127-144.
Watson,
D.,
Clark, L. A.,
&
Tellegen,
A.
(1988).
Development and vali-
dation of a brief measure of positive and negative affect: The PANAS
scales.
Journal
of
Personality
and
Social
Psychology,
54,1063-1070.
Watson,
D,
& Tellegen,
A.
(1985). Toward
a
consensual structure
of
mood.
Psychological
Bulletin,
98,
219-235.
Weisberg, H.
F.
(1980).
A
multidimensional conceptualization of party
identification.
Political
Behavior,
2, 33-60.
Wundt,
W
(1897).
Outlines
of psychology
(C.
H. Judd,
Trans.).
Leipzig,
Germany, Wilhelm Englemann.
Zevon, M.
A.,
&
Tellegen,
A.
(1982).
The structure of mood
change:
An
idiographic/nomothetic analysis.
Journal
of
Personality
and
Social
Psychology,
43, 111-122.
Appendix A
Items Used to Measure Happy and Sad Moods in Studies
1
-3
Study
1
Adjective Checklist
Cheerful
Elated
Glad
Happy
Joyful
Blue
Discouraged
Feeling low
Low
Sad
Agree-Disagree Response Option (5 points)
I have been
in a
cheerful mood.
All
in
all, I've been feeling despondent.
"Describes
Me"
Response Option (4 points)
I have been
in a
good mood.
I have felt sad and dispirited.
zctive Checklist
Cheerful
Contented
Happy
Pleased
Satisfied
Warmhearted
Studies
2 and 3
Blue
Depressed
Downhearted
Gloomy
Sad
Unhappy
Agree-Disagree Response Option (5 points)
I've been
in
good spirits.
All
in
all, I've been feeling kind
of
depressed.
"Describes
Me"
Response Option (4 points)
I have been
in a
good mood.
I've felt sad and dispirited.
Unipolar
Liken Scale
(7 points)
Happy-not happy
Discouraged-not discouraged
Unipolar
Likert Scale
(7 points)
Happy-not happy
Discouraged-not discouraged

BIPOLARITY
OF
MOOD
Appendix B
Table Bl
Measurement Equations for
Confirmatory Factor
Analyses
Factor
£,
=
positive affect factor, Time
1
=
f
2
=
negative affect factor, Time
1
=
Random error
x,
=
X
H
£,
+
5,
Nonrandom error
x,
=
X,,|,
X
2
=
X
2
,{,
x
l
~
X3ifi
x
4
=
X
41
^,
X
i
=
X52I2
1041
= positive affect factor, Time 2
=
8
8
2
X,
=
X
93
£
3
+
5,
x
l0
=
^10,3^3
+ 5,
0
•*1I
=
X[ 13I3 + 5
H
5
12
X
\2
~
£
4
=
negative affect factor, Time 2
=
Note. The measures
x^,
%,
x?,
and jq
3
represent mood adjective checklist
scales.
Similarly, the x
2
, x
6
,
x,
0
,
and
x,
t
represent agree-disagree
items;
x
3
,
x
7
, x,,, and ^5 are self-description
items;
and
x
4
,
Xg,
x,
2
, and x,
6
are semantic differential Likert
scales.
The
Xs are
factor
loadings,
and the
5s are
individual elements in the
error matrix. All measures depicted within braces are assumed
to
load on the mood factor named to the
left. The
X*.
associated with
the
mood adjective checklist measures
are
assumed
to be
unity,
to set the
metric for the latent factors. The four factor variances and
six
interfactor covariances
(0,
y
)
are free parame-
ters.
The
16
measurement error variances
(0£
t
)
are free parameters. Analyses in which allowance is made
for nonrandom error assume that all 5^ ending
in
the same subscript letter are potentially correlated. This
adds an additional 24 free parameters. The measurement models used in Studies 2-3 are identical, except
that they involve only two factors and Measures x,-Xg.
Appendix C
Items Used to Measure Positive Affect and Negative
Affect in Study 3
Adjective Checklist
Active
Alert
Determined
Excited
Interested
Proud
Afraid
Hostile
Irritated
Jittery
Nervous
Upset
Agree-Disagree Response Option (5 points)
I have been feeling very focused and
"on
task."
I've had trouble paying attention.
For some reason, I've been feeling sort
of
nervous.
I feel "calm, cool, and collected."
"Describes Me" Response Option (4 points)
I have felt very inspired.
I have had very little interest
in
things around me.
I have felt rather distressed.
I have been feeling calm and relaxed.
Unipolar Likert Scale (7 points)
Alert-not alert
Enthusiastic-not enthusiastic
Distressed-not distressed
Scared-not scared
Received March
12, 1992
Revision received October 30,
1992
Accepted December
21,
1992