Sudden Transitions in Attitudes

HAN L. J. van der MAAS

ROGIER KOLSTEIN
JOOP van der PLIGT
University of Amsterdam

Both the dynamic approach and catastrophe modeling have been warmly welcomed in
research on attitudes and opinions. In this article, the authors discuss a general method-
ology for testing catastrophe models and apply it to the dynamics of attitude formation
and change. First, by making use of the so-called catastrophe flags, converging support
for the catastrophe model can be attained. Each flag relates to a specific hypothesis
about attitudinal change. Second, fitting stochastic catastrophe models to data enables
one to carry out a direct test of catastrophe models. Results of analyzing large data sets
on political attitudes support the validity of the general catastrophe model of attitude
change in which transitions in attitudes are a function of involvement and information.
Present results suggest that in the case of political attitudes, involvement might well be
correlated with attitude. A more refined approach to the measurement of information
and involvement is suggested.

Keywords: attitude change; polarization; ambivalence; stochastic catastrophe theory;

catastrophe flags; maximum likelihood estimation

INTRODUCTION

The dynamical system approach is the subject of increasing inter-

est in psychology in general. More recently, the approach has been
applied to attitudes, and several authors have stressed the potential of
the dynamical system approach. Petty, Wegener, and Fabrigar (1997)
consider the dynamical system approach to the study of attitudes an
important new research area. McGarty and Haslam (1997) refer to
the dynamical system approach as one of the most promising new
developments in social psychology. Catastrophe theory (Thom 1972;

SOCIOLOGICAL METHODS & RESEARCH, Vol. 32, No. 2, November 2003 125-152
DOI: 10.1177/0049124103253773
© 2003 Sage Publications
125
126 SOCIOLOGICAL METHODS & RESEARCH

Zeeman 1976; Gilmore 1981), part of the dynamical system approach,

has been applied to attitudes to explain and describe transitions in
attitudes (see, e.g., Zeeman 1976; Flay 1978; Latané and Nowak 1994;
Tesser and Achee 1994).
According to these authors, attitudes are unstable when involve-
ment is high and arguments, pro and contra an attitude, balance each
other out. In such a transitional state, small changes in pro/contra
information may lead to large, sudden jumps in attitudes. Although the
application of catastrophe theory holds promise, several researchers
stress the need for sound statistical methods that can be applied to
catastrophe hypotheses and dynamical system hypotheses in gen-
eral (Beek, Verschoor, and Kelso 1997; Burlingame and Hope 1997;
Kruglanski, Clement, and Jost 1997).
Several early investigators (Zeeman 1976; Flay 1978; Tesser 1980)
proposed appealing catastrophe models but never tested them directly.
This was mainly because standard statistical tests cannot simply be
applied to catastrophe models. These early investigators were heavily
criticized (Sussman and Zahler 1978), mainly for the lack of testing.
In response to this criticism, several researchers developed methods
to test catastrophe models.
A first line of research involves the testing of specific predictions
derived from catastrophe models (the so-called catastrophe flags)
(Gilmore 1981). The more that these catastrophe flags are detected,
the more convincing is the claim of a catastrophe. As will be shown,
this line of research mainly consists of the integration and reinterpreta-
tion of empirical results and ideas put forward in the attitude literature.
A second line of research focuses on techniques that apply explicit
statistical formulations of the catastrophe models. These techniques
involve the direct statistical fit of the catastrophe model to empiri-
cal data (Cobb 1978, 1980, 1981; Guastello 1988; Oliva, Desarbo,
Day, and Jedidi 1987). As explained below, there are good reasons to
prefer the method of Cobb. Computational improvements make his
method very useful (Hartelman 1997; Hartelman, van der Maas, and
Molenaar 1998).
We discuss and apply both lines of testing within the catastrophe
model for attitude change to cross-sectional data. With these applica-
tions, we intend to test a model of attitude change and demonstrate
that testing catastrophe models in psychology is feasible.
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 127

THE CATASTROPHE THEORY OF ATTITUDES

Zeeman (1976) introduced catastrophe theory in psychology and

also formulated the specific hypothesis that (attitude) change can be
described by catastrophe theory (Thom 1972; Poston and Stewart
1978; Gilmore 1981; Castrigiano and Hayes 1993). This hypothesis
implies that smooth changes in independent or control variables may
lead to abrupt, discontinuous changes in attitudes. To understand the
details of Zeeman’s hypothesis, some understanding of catastrophe
theory is required.
There are seven different families of catastrophe models, based
on the number of control and dependent variables. In the social
sciences, the so-called cusp catastrophe model is the most frequently
used model because it is the simplest catastrophe model that gives
rise to sudden discontinuities. It consists of two control variables and
one behavioral variable. Control variables change at least an order
of magnitude slower than the behavioral variable. The cusp model is
based on the following nonlinear deterministic dynamical system:

dX/dt = −dV (X; α, β)/dX.

That is, behavior X (attitude in our case) changes over time t

according to the derivative of the cusp potential function:
1 1
V (X; α, β) = X4 − βX2 − αX,
4 2
which has as equilibria (first derivative to zero)

X3 − βX − α = 0,

where α and β are the control variables. When the values of the
control variables remain fixed, the system seeks an equilibrium state
(called a point attractor). This means that some minimum or max-
imum of a certain quantity (e.g., energy, profit, or cognitive disso-
nance) is obtained. Catastrophe theory explains how these equilibria
change as a function of the control variables. This change of equilib-
ria may lead to discontinuous, abrupt changes in behavior. The two
control variables of the cusp are called the normal (α) and the splitting
factor (β).
128 SOCIOLOGICAL METHODS & RESEARCH

X = attitude
Equilibria of cusp
neutral point
(0,0,0)

α = information

Sudden jump

Inaccessible region
β = involvement

α
B
Bifurcation lines

Figure 1: The Cusp Model of Attitude Change

Figure 1 gives a graphical representation of the cusp model. The

folded surface represents the equilibria of the system. The plane below
is the control plane. The surface between the bifurcation lines is called
the bifurcation set. Outside this set, only one stable state exists. Within
the set exist three states, but the middle state is repelling (i.e., a state
that is inaccessible). Sudden catastrophic jumps occur when the nor-
mal variable is smoothly changed at high values of the splitting vari-
able. A typical example is the freezing of water. The condition of
water is the dependent variable, and the two independent variables are
temperature (≈ normal variable) and pressure (≈ splitting variable).
Smooth changes in temperature lead to sudden transitions between
the two possible states (Path A). If pressure is low (Path B), a contin-
uous change between water and ice is possible.
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 129

The catastrophe theory of attitudes (Zeeman 1976; Flay 1978;

Latané and Nowak 1994) states that information and involvement can
serve as the control factors (α and β) in the cusp model. The behav-
ioral aspect of an attitude is the equilibrium state based on informa-
tion and involvement. Latané and Nowak (1994) state that this holds
for stereotypes as well (see also Eiser 1994). The information vari-
able can be perceived as a position on a dimension, which combines
factors such as previous experience, self-interest, genetic disposition,
and environmental effects (Zeeman 1976). Involvement is an orga-
nizing property that may be voluntary or involuntary (Zeeman 1976).
Latané and Nowak state that, because attitudes represent some kind
of organized body of knowledge (such as a schema or category),
increasing organization should lead to attitudes becoming more cate-
gorical. Involvement features as the splitting factor because it breaks
the unimodal structure of the attitude distribution into two (opposing)
attitudinal positions. Involvement can also be related to the literature
on attitude strength. For instance, Krosnick and Petty (1995) argue
that involvement is one of the determinants of attitude importance
and attitude strength.
Latané and Nowak (1994) attempted to test the cusp catastrophe for
attitude change. In a first study, they reanalyzed data (from Stouffer
et al. 1950) consisting of the opinions of 75,000 soldiers about 7,000
attitude statements. Latané and Nowak found that as involvement
increased, bimodality in the frequency became more extreme. To
test this statistically, they focused on attitude variance, assuming that
increased variance would provide some evidence of multimodality
and divergence. However, whereas bimodality almost always implies
increased variance, increased variance does not necessarily imply
bimodality. As a consequence, their statistical test is not completely
convincing. In their second study, Latané and Nowak asked 100 under-
graduate psychology students to judge a list of political statements
in terms of favorableness and importance. They found that as judged
importance increases, so does favorableness. Their results showed
a significant overall (positive) correlation between importance and
extremity of opinion. A replication of this study showed similar results
(Latané and Nowak 1994). This correlational test is also circuitous.
Although encouraging, these studies thus provide only modest sup-
port for the catastrophe model. By reanalyzing these data with the
130 SOCIOLOGICAL METHODS & RESEARCH

cusp-fitting technique, we will provide a more direct test of this

model.
Gilmore (1981) derived eight behavioral properties of the cusp
model, the so-called catastrophe flags. These catastrophe flags are
(multi) modality, sudden jump, inaccessibility, hysteresis, divergence,
anomalous variance, divergence of linear response, and critical slow-
ing down. In the next section, we review each of the flags with respect
to attitudinal change.

HYPOTHESIS TESTING BY USING THE CATASTROPHE FLAGS

Catastrophe theory provides a parsimonious description of many phe-

nomena. By making use of catastrophe flags, many theories can be
accommodated in a single model (van der Maas and Molenaar 1992).
Below, these flags are related to specific hypotheses within the theory
of attitude formation and change.
Multimodality of behavior means that for specific values of the con-
trol variables, a multimodal distribution of the behavioral variable is
likely to occur. Usually, multimodality is studied in cross-sectional
samples, but it can also be evaluated with individual repeated measure-
ments (sudden jump flag). Multimodality of attitudes refers to a mul-
timodal distribution of attitude-related behavior within a given group.
Flay (1978) gives several examples of bimodality in which involve-
ment is the splitting factor. One example concerns student evaluations
of courses. According to Flay, overall ratings will be extreme (posi-
tive or negative) when workload (involvement) is rated as high. One
way to assess the modality of attitudes in a statistically sound way is
to use finite mixture models (Titterington, Smith, and Makov 1985).
It is, for instance, possible to typify the behavior by using latent class
analysis (McCutcheon 1987; van der Maas 1998).
Inaccessibility is strongly related to modality. It implies that inter-
mediate (behavioral) states between two opposite attitudes are rare.
People will behave according to one or the other attitude but will
not readily assume intermediate positions. In Flay’s (1978) example,
students show less neutral attitudes when workload is high. The rela-
tion between involvement and attitude extremity is also addressed in
the classic work of Sherif, Sherif, and Nebergall (1965).
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 131

A sudden jump (Path A, Figure 1) takes place when the amount

of information in favor of the opposite attitude is increased continu-
ously. Beyond some threshold, people suddenly switch from one atti-
tude to the other. This relates to conversion in the conversion model
of stereotype change (Rothbart 1981). However, instead of an atti-
tude change after a dramatic counterattitudinal piece of information
(as proposed by the conversion model), a sudden jump can occur
given a very minor piece of information or following a continuous
flow of counterattitudinal information. There is some anecdotal sup-
port for the hypothesis concerning the sudden transition of attitudes
(e.g., the case of Patty Hearst) (see Zimbardo, Ebbesen, and Maslach
1977). When involvement is relatively low, attitude change is contin-
uous. This resembles the bookkeeping model of stereotype change
(Rothbart 1981). This model states that any gradual incoming infor-
mation is weighted and added, which results in continuous change
of the stereotype. Tourangeau (1987) and Saris (1997) showed that
public opinion, on issues of low interest, can easily be swayed by
specific contextual information.
Hysteresis means that a sudden jump will occur at different val-
ues of the normal variable, depending on the direction of change
in this variable (Path A). For instance, in disturbance-free condi-
tions, water freezes at –4 ◦ C and thaws at 0 ◦ C. Hysteresis in the
attitude model means that the informational value at which people
change their attitude (the threshold) depends on the person’s initial
position and the direction of change in information. For instance,
people with a positive attitude toward abortion change to a neg-
ative attitude only when information is strongly against abortion.
People with a negative attitude change to a positive attitude only
when information is strongly in favor of abortion. Reardon and
Tesser (1982) provide some experimental support of hysteresis.
Respondents with a history of conformity (i.e., “weak” feelings
about prior information) jump to nonconformity only if informa-
tion elicits strong feelings of disagreement. On the other hand,
respondents with a history of nonconformity (i.e., strong feelings
about prior information) jump to conformity only if they are con-
fronted with counterattitudinal information with which they moder-
ately agree. The catastrophe model is unique in that it provides an
explanation for the phenomenon of hysteresis. Hence, we consider
132 SOCIOLOGICAL METHODS & RESEARCH

this flag to be sufficient to conclude that a catastrophe model

holds.
When the splitting factor increases, the multimodality of the behav-
ior becomes more extreme. This is called divergence because two
initially proximate behaviors become increasingly opposed to each
other. Divergence in the attitude model means that two (groups of)
persons with initially almost the same neutral attitude (an attitude
based on an intermediate value of information) can diverge into two
opposites when involvement increases. This corresponds to the notion
of polarization (Latané and Nowak 1994; Latané, Nowak, and Liu
1994; Tesser 1976). Attitudes that differ slightly may diverge into
strong opposite attitudes. An example of divergence can be found
in an experiment by Teahan (1975), who found a polarization effect
among Black and White police officers. At their entrance into the
police academy, attitudes of these officers were similar. After an
18-month period, both groups showed more interest in their own group
and more hostility toward the other group.
Anomalous variance is an increase in the variance of behavior that
occurs in the neighborhood of the bifurcation set (e.g., in the area
where sudden transitions are possible). In this area, we can expect
large fluctuations in behavior. For instance, people who have just
arrived in some new group (e.g., immigrants who have settled in a
new country) have to come to understand the attitudes of the new
group and to give meaning to the behaviors that surround them. They
will fluctuate between “old” and “new” behaviors until they settle
down in the new stable behavioral state. This relates to findings con-
cerning attitudinal ambivalence. Ambivalence occurs when people
consider attitudinal information both pro and contra a given subject.
Ambivalent attitudes are attitudes that simultaneously contain posi-
tive and negative cognitions (see Eagly and Chaiken 1993; Kaplan
1972; Thompson, Zanna, and Griffin 1995).
The following flags differ from those discussed above because
they are associated with perturbations of the system. Divergence of
linear response means that, after a strong counterattitudinal message,
people in the neighborhood of the bifurcation area show large fluctu-
ations in behavior before returning to their stable former attitudinal
state or settling in the new one. This corresponds to the conversion
model (Rothbart 1981) in which a single counterattitudinal message
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 133

may result in attitude change. After some oscillations, the respondent

may resort to his or her original viewpoint but may also jump to a
new attitude. Depending on the complexity of the issue, subtyping
(i.e., creation of new subgroups) is also possible (Weber and Crocker
1983). The counterattitudinal message is evaluated as a subcategory,
which leaves the overall stereotype intact but perhaps a little less
stable. Catastrophe theory predicts that when a sufficient number of
counterattitudinal messages are given, even subtyping will eventually
cease, and a jump to the other attitude will occur. This corresponds
to the view of Pettigrew (1981), although other researchers (e.g.,
Taylor 1981) believe that subtyping will lead only to new, more
detailed stereotypes.
Critical slowing down is closely related to the former catastrophe
flag. It means that, after a perturbation, it takes a certain amount of
time before stable behavior returns. Reaction times to the attitudinal
items can be used to test for this. In the neighborhood of the bifurca-
tion area, reaction times are much slower than outside the bifurcation
area (van der Maas, Raijmakers, Hartelman, and Molenaar 1999). In
the case of ambivalence, reaction times when judging the attitudinal
object or issue are also found to be much slower (van der Pligt, De
Vries, Manstead, and van Harreveld 2000). Vallacher, Nowak, and
Kaufman (1994) explored mixed valence representations both con-
ceptually and operationally. Their “mouse approach,” in which they
register the exact computer mouse movements in response to attitu-
dinal information, enables one to track the time course of switching
between different attitudinal positions.

FITTING CATASTROPHE MODELS

Several statistical methods to fit catastrophe models have been devel-

oped. First, we briefly discuss the main techniques.

GUASTELLO’S POLYNOMIAL REGRESSION TECHNIQUE

Guastello’s (1988) polynomial regression technique takes

dX/dt = −dV /dx = −X3 + βX + α = 0 as a starting point. The
idea is that given some measurement of X, β, and α at t = 1, X at
134 SOCIOLOGICAL METHODS & RESEARCH

t = 2 can be derived from the formula for dX/dt. Setting dt equal to

1 gives the following regression equation:
y1 − y2 = c0 + c1 y13 + c2 y12 + c3 βy1 + c4 α,
where y1 = (X1 − λ)/σ and y2 = (X2 − λ)/σ.
Guastello (1988) uses reverse hierarchical entry (entering higher
order terms first) in computing the regression equation and compares
the model with two alternative linear regression models. Guastello’s
method has been criticized by Alexander, Herbert, Deshon, and
Hanges (1992) on a number of points. Their most important criticism
relates to the fact that the analysis of completely random data (e.g.,
random numbers for y1 and y2 ) yields a proportion of explained vari-
ance of about .49 (this can be easily be demonstrated by means of
simulation). This is a consequence of using the difference between y1
and y2 as the dependent variable and a function of y1 as the predic-
tor. Given some weak assumptions, if the correlation between y1 and
y2 is 0, then the correlation between y1 and the difference between
y1 and y2 is about .72 (for details, see Alexander et al. 1992). In
combination with the reverse hierarchical entry method, in which the
cubic term is entered before the linear term, this statistical artifact
leads to the incorrect conclusion that the cubic term is a significant
predictor in the regression when the data consist entirely of noise.
In his reply, Guastello (1992) argues that differences between con-
strained random numbers are indeed catastrophically distributed. He
used a fit of random numbers to the cusp model to support his view
and found the above-mentioned proportion of explained variance of
.49 (see fourth row of Table 2 in Guastello 1992), as predicted by
Alexander et al. (1992). However, that some types of chaotic time
series, which look like completely random time series, can be gener-
ated with cusp-like equations (which is what Guastello argues) does
not imply that computer-generated random numbers (uniformly dis-
tributed) are catastrophically distributed.
Consider also the alternative linear models used by Guastello
(1988). In both linear models, the difference between y1 and y2 is not
a function of y1 . We think that the correct alternative linear model
is Guastello’s nonlinear regression equation (see above) without the
quadratic and the cubic term (e.g., c1 = c2 = 0). For random data,
this linear model also yields an explained variance of about .5, which
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 135

illustrates two things: (a) this type of difference regression equation

can be misleading,1 and (b) the good fit of the random numbers is not
related to the cubic term, and hence with the cusp model, but is an
artifact of the difference method.

GEMCAT

Oliva et al. (1987) proposed a multivariate methodology for

estimating catastrophe models. Lange, Oliva, and McDade (2000)
improved this methodology and developed a program called
Gemcat II. It is claimed that the program can fit any of the catastrophe
models to data. It needs to be added that experience is mostly limited
to the cusp. In Gemcat II, the cusp variables X, β, and α are all lin-
ear functions of one or more measured variables. The coefficients of
these linear functions are estimated by minimizing

et2 = t [Xt3 − βt Xt − αt ]2 ,

where et is the error, ideally zero, for each case or data point.
This technique does not rely on difference scores and, consequently,
does not suffer from the main problem of Guastello’s (1988) approach.
However, despite its advantages above Guastello’s technique, there
are two major problems with Gemcat. The first problem is that et is
not the quantity that should be minimized in case of implicit regres-
sion equations. What should be minimized is the squared difference
between the dependent variable Xt and the prediction of Xt . As a
consequence, the least squares statistical theory cannot be applied to
the et term in the Gemcat approach. Hypothesis testing is therefore
difficult.
The second problem is more fundamental and also concerns
Guastello’s (1988) technique. The derivative of the cusp potential
function is used as a basis for the computation of fit. The problem
of the derivative is that it does not distinguish between maxima and
minima of the potential function. The maxima, however, represent the
inaccessible states (see Figure 1), and data points in this area should
contribute greatly to the error in fit of the cusp model. Unfortunately,
in Gemcat and Guastello’s technique, this is not the case since both
techniques do not distinguish between minima and maxima of the
cusp potential function. This point is illustrated in Figure 2. As a
136 SOCIOLOGICAL METHODS & RESEARCH

18

e 16

14

12

10

-4 -2 -2 0 2 4
X

Figure 2: Error, e, as Function of X, According to the Gemcat Optimization Function

(for α = 0 and β = 2)
NOTE: At X = 0, the potential function has a maximum and data points are not permitted in
this so-called inaccessible state. However, the graph shows a minimum at X = 0, implying that
data points around X = 0 do not contribute to the total error. Both in Gemcat and Guastello’s
(1988) technique, this occurs because they do not distinguish between minima and maxima of
the cusp potential function.

consequence, the fit of the cusp model to any data set will be overes-
timated. It is difficult to solve this problem as long as the derivative of
the potential function is used to fit data. The next technique applies the
potential function of the cusp and does not suffer from this problem.

COBB’S METHOD

Cobb (1978, 1980, 1981) and Cobb and Zacks (1985) developed
a catastrophe fitting technique that is based on a stochastic interpre-
tation of catastrophe theory. The deterministic equation of motion,
dX/dt, is extended with a stochastic noise term (i.e., a Wiener
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 137

innovation). Assuming a constant variance of the innovation, the

probability density function (PDF) for the cusp can be written as
follows (for derivation, see Cobb and Zacks 1985; Hartelman 1997):


1 4 1 2
F (X|α, β) = D exp − X + βX + αX , (1)
4 2

X = (Y − λ)/σ, (2)

α = a0 + a1 Z1 + a2 Z2 + . . . an Zn , (3)

β = b0 + b1 Z1 + b2 Z2 + . . . bn Zn , (4)
where D is an integration constant depending on α and β. Param-
eters λ and σ scale the observed behavioral variable Y to X, and
α and β are linear functions of the observed control variables Z1 to
Zn . The behavioral variable is univariate, but the dependent variables
may be multivariate. The parameters λ, σ, a0 . . . an and b0 . . . bn are
estimated using a maximum likelihood procedure (see Cobb and
Watson 1980).
The inaccessible mode of the deterministic cusp model is reflected
in the PDF by an area of low probability between the two high-
probability modes (e.g., a bimodal distribution). Fitting the PDF
of equation (1) to data does, in principle, not differ from fitting
other probability densities such as the normal distribution. Concep-
tually, this procedure is much better than the other techniques since
it is a relevant stochastic generalization of deterministic catastrophe
theory. Yet, it has not been used often. A first reason might be that the
derivation of the PDF is not easy to understand, but the most impor-
tant reason is that Cobb’s computer program often breaks down for
nonapparent reasons.
This last problem was solved by Hartelman (1997) and Hartelman
et al. (1998). Hartelman replaced the expectation-maximization rou-
tine that Cobb used by a quasi-Newton optimization routine, which
is more reliable in this context. The gradient of the objective function
is determined by a finite difference approximation. Hartelman also
replaced the numerical integrator of Cobb by the numerical integra-
tor D01AMF of the NAG library. These computational improvements
138 SOCIOLOGICAL METHODS & RESEARCH

lead to a much better performance of Cobb’s technique, although

sometimes sensitivity to starting values remains a problem (as with
maximum likelihood estimation of many other distributions).
The starting value problem may be solved by trying several sets
of starting values. All solutions reported in this article are maxima of
the likelihood function obtained by varying starting values. Hartel-
man’s (1997) program gives unstandardized and standardized (given
standardization of observed variables) estimates of the parameters.
These parameters or weights of the observed independent variables
on the normal and splitting factors indicate how strongly the control
variables are associated with the normal and splitting factor (as well
as the independence between the control variables). Furthermore, the
data points can be placed in the control plane, spanned by α and β.
This shows in which part of the cusp data points are located.
The fit of the cusp model is compared with the fit of a multivariate
logistic and a linear regression model. The logistic model allows for
rapid change of the dependent variable but does not incorporate real
jumps. Within the class of noncatastrophe models, it is the closest
to the catastrophe models. The logistic model is defined as Y =
λ+σ/(1+exp(α)), where α = a0 +a1 Z1 +a2 Z2 +. . . an Zn . A direct
comparison of log-likelihoods is not feasible since these models are
not nested. The stochastic cusp model uses more parameters (#P ),
and comparison should control for this. As measures of goodness
of fit, Hartelman’s program uses the Akaike information criterion
(AIC) and Bayes information criterion (BIC). The AIC (Akaike 1974)
penalizes the log-likelihood (LL) for the number of parameters (–2
LL + #P * 2). In the BIC (Schwarz 1978), the penalty is a function
of the number of subjects (N) and the number of parameters (–2 LL
+ #P * ln (N)). The model with the lowest AIC and BIC is selected.
In case the AIC and BIC contradict each other, differences between
models are usually too small to make clear-cut decisions. If models
are nested (e.g., restricted vs. unrestricted cusp models), they can
be compared by the likelihood ratio. Under certain mild regularity
conditions, −2(lnLunrestricted −lnLrestricted ) is known to be chi-square
distributed with the difference in the number of estimated parameters
as degrees of freedom.
Hartelman (1997) reports simulations demonstrating accurate
estimation of the parameters for sample sizes larger than 200. The
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 139

distribution of the estimates is also reasonably normal. Simulations

also show that the AIC and BIC favor the simpler logistic model when
data are fitted that are generated with the cusp model but lie outside the
bifurcation set (i.e., the area that gives rise to the characteristic flags).
The last important extension of Cobb’s technique in the program
of Hartelman (1997) is the possibility of introducing restrictions to
parameters to test specific hypotheses (by fixing some of the parame-
ters λ, σ, a0 . . . an and b0 . . . bn to certain values, e.g., 0). In the most
strict cusp catastrophe model for attitudes, the informational variable
only loads on the normal factor, while involvement only loads on the
splitting factor. In this case, the two variables have an independent
joint effect on the attitudinal variable. This model can be tested by
restricting some of the a0 . . . an and b0 . . . bn parameters to 0 and can
be compared (by AIC and BIC) with the unrestricted cusp model. In
the next section, we present two examples to illustrate the usefulness
of Hartelman’s (1997) extension of Cobb’s technique (for another
recent application involving parameter restrictions, see Ploeger, van
der Maas, and Hartelman 2002).

EXAMPLES

In the first example we reanalyze the data used by Latané and Nowak
(1994) to demonstrate that higher levels of involvement, as measured
by intensity of feeling (Study 1) and judged importance (Studies 2
and 3), increase the modality of the behavior or attitude toward an
issue (Study 1) and favorableness toward political statements (Studies
2 and 3). These data sets do not include measures of normal variables.
There is only one independent control variable that should load only
on the splitting factor of the cusp (the loading on the normal factor is
fixed to 0). Hence, the cusp model can only be partly tested with these
data. This test (focusing on divergence) is in itself very informative
since a good fit of the model strongly supports the claims of Latané
and Nowak.
In the second example, we analyze data of a study on social cul-
tural developments in the Netherlands (Felling, Peters, and Schreuder
1985). These data include attitudinal (dependent) variables, vari-
ables that can be interpreted as involvement (splitting) factors, and a
140 SOCIOLOGICAL METHODS & RESEARCH

variable that might operate as an information (normal) factor. This

last factor is political orientation. The idea is that a position on this
dimension reflects the kind of information one has at one’s disposal
about politics. This assumption is made because it is not possible to
assess the information directly from the respondents. Even though
the normal factor does not involve a direct measure of information,
the complete cusp model can still be fitted on these data.

EXAMPLE 1: THE LATANÉ AND NOWAK STUDIES

(USING ONE CONTROL VARIABLE)

To test the cusp model of attitudes, Latané and Nowak (1994)

report three studies. They show that variance in attitude increases
as involvement increases (Study 1) and that perceived importance is
related to attitude extremity (Studies 2 and 3). The increased variance
is explained by the emerging bimodality of the distribution of the
responses. By fitting a constrained cusp model, this hypothesis can be
tested more directly. Following the description of the three original
studies, the results of the cusp fits are presented. Within each study,
two model fits are performed: one with and one without restrictions.
In the restricted model, involvement loads only on the splitting factor.
In Study 1, Latané and Nowak (1994) used a data set from
Stouffer et al. (1950). In this study, U.S. soldiers were asked their
opinion about three issues: the point system for demobilization, post-
war conscription, and the Women’s Army Corps. Attitudes were
scaled by counting the number of questions on each issue to which
a soldier gave a favorable response and ranged from 0 (unfavorable)
to 6 (favorable). In addition, respondents were asked to indicate how
strongly they felt about their answer (from intensity 0 to intensity 5).
A total number of 6,817 attitude scores were assessed. In Figure 3, the
combined results of the attitudinal statements are shown in relation
to intensity. At intensity 0, a normal distribution describes the data
best: Most people have a neutral attitude toward the three statements.
At intensity 5, however, a bimodal distribution of attitudinal values
emerges: a small number of people have a neutral attitude, while most
have a pro or contra attitudinal position. Variance increases regularly
and significantly (Latané and Nowak 1994).
Because the program Cuspfit cannot handle more than 5,000
subjects, we took only 10% of the data simply by dividing the
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 141

frequency of each score by 10 (N = 685).2 The data set contained one

control variable (intensity of feeling) and one dependent variable
(the attitude).
In the second study (Latané and Nowak 1994), 100 undergradu-
ate psychology students (62% females) were given 16 political state-
ments to evaluate. Attitudes toward each statement ranged from –2
(oppose) to +2 (favor). Next they were asked to rate the importance
of the statements (using a 5-point scale from 0 = unimportant to 4 =
important). In addition, respondents had to predict the responses of
their best friend, a male classmate, and a female classmate. Results of
this study resemble those of Study 1: At importance level 0, a more
or less normal distribution is observed because most people have a
neutral attitude toward the statements. At importance level 4, visual
inspection suggests the emergence of a bimodal distribution. The
overall correlation between importance and extremity of attitudes is
significant: r = .42, p < .001.
In their last study (Latané and Nowak 1994), 62 members of
University of North Carolina fraternities and sororities (65% females)
served as respondents. The main difference between this and the orig-
inal study is the expansion of the attitude scale from 5 to 7 points
(−3 = oppose to +3 = favor). Visual inspection of the results sug-
gests the same pattern as both former studies: a shift from a normal
to a bimodal distribution as importance increases.

Results of Example 1

The data sets of Latané and Nowak (1994) were reanalyzed with
the program Cuspfit. According to the AIC and BIC, both the unre-
stricted and restricted cusp models fitted much better than the linear
and logistic models. This finding was obtained for all three cases (see
Table 1). In Study 1, the restricted cusp model did not show a worse
fit than the unrestricted cusp model, χ 2 (1) = −2(−876.8−876.9) =
.2, p > .05. Both the AIC and BIC are lower for the restricted model.
In Studies 2 and 3, the unrestricted cusp model has a better fit than
the restricted cusp, χ 2 (1) = 52, p < .01 and χ 2 (1) = 34, p < .01,
respectively. Table 2 shows the standardized parameter estimates.
In Study 1, the effect of intensity on the normal factor is negligible
compared to its effect on the splitting factor (a1 = −.002;
142 SOCIOLOGICAL METHODS & RESEARCH

TABLE 1: Fit Statistics in the Latané and Nowak Studies: One Control Variable

Log-Likelihood AIC BIC #P

Study 1
Linear –972.0 1,950 1,964 3
Logistic –971.9 1,952 1,970 4
Cusp –876.8 1,766 1,793 6
Cusp restricted –876.9 1,764 1,787 5

Study 2
Linear –2,374 4,754 4,770 3
Logistic –2,141 4,291 4,312 4
Cusp –1,691 3,395 3,427 6
Cusp restricted –1,717 3,444 3,471 5

Study 3
Linear –1,451 2,907 2,922 3
Logistic –1,359 2,726 2,745 4
Cusp –1,052 2,116 2,146 6
Cusp restricted –1,069 2,149 2,173 5

NOTE: AIC is the Akaike information criterion, BIC is the Bayes information criterion, and
#P is number of estimated parameters. In Study 1, the restricted cusp model has the lowest
AIC and BIC; in Studies 2 and 3, the unrestricted cusp model has the lowest AIC and BIC.

b1 = 1.002). Hence, the restricted model (a1 = .0) has lower AIC
and BIC than the unrestricted model. In Studies 2 and 3, the effect of
importance on the normal factor is larger but is still small compared
to its effect on the splitting factor (Study 2: a1 = .205, b1 = .999;
Study 3: a1 = .166, b1 = .655). Figure 3 gives a graphical display of
the data and the fitted model of Study 1. Results clearly support the
hypothesis of divergence of Latané and Nowak (1994).

EXAMPLE 2: USING TWO CONTROL VARIABLES

In this example, we used data from a study on social-cultural

developments in the Netherlands (Felling et al. 1985). In this study,
3,003 respondents answered a questionnaire containing 800 attitude
items about issues such as religion, value systems, work, politics,
ecology, health, and ethnocentrism. We focused on politics because
a well-defined normal factor could be extracted. This variable was
operationalized by the request to indicate a political position on a
10-point scale (1 = left wing and 10 = right wing). The second control
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 143

TABLE 2: Standardized Parameter Estimates of the Restricted and Unrestricted Cusp

Models of Studies 1, 2, and 3

Normal Factor (a) Splitting Factor (b) Location (λ) Scale (σ )

Study 1

Unrestricted model
0 Constant –0.154 0.619 0.016 1.006
1 Intensity –0.002 1.002

Restricted model
0 Constant –0.017 0.619 0.017 1.006
1 Intensity 0.000a 1.002

Study 2

Unrestricted model
0 Constant 0.365 1.777 –0.539 0.852
1 Importance 0.205 0.999

Restricted model
0 Constant 0.451 1.730 –0.589 0.870
1 Importance 0.000a 0.115

Study 3

Unrestricted model
0 Constant 0.275 2.128 –0.417 0.783
1 Importance 0.166 0.655

Restricted model
0 Constant 0.319 2.119 –0.440 0.789
1 Importance 0.000a 0.749

a. Fixed

variable (involvement) was represented in this study by 12 items. A

principal component analysis was carried out on the 12 involvement
items. The following involvement measures load on the same factor:
“interest in politics,” “no idea about political future,” “read about
politics in newspapers,” “important to engage in politics,” “inquiries
about politics,” “ideas about political future,” “little knowledge about
politics,” and “discuss politics with other people.” The factor scores
of this factor are used.
We selected 24 different attitudinal statements as possible behav-
ioral variables. The statements are selected on the basis of the corre-
lation with the political orientation variable, as well as the increase
 144 SOCIOLOGICAL METHODS & RESEARCH

6
attitude
5

0
0 1 2 3 4 5
intensity of feeling

Figure 3: Fit of the Cusp Model to Data of Latané and Nowak

NOTE: For five intensities of feeling, the frequencies of the attitude scores are dis-
played. The dotted line displays the predicted values of the cusp model, indicating
an increase of bimodality with higher values of intensity of feeling.

of bimodality of the distribution of the responses when involvement

was increased. The fit of the cusp catastrophe model was better than
that of the linear and logistic models in all 24 cases. However, not
all analyses showed a clear result: In most cases, involvement and
political orientation were not completely independent. The left-wing
position was associated with more involvement. In the 24 analyses,
the correlation between political orientation and involvement varied
between .017 and .101 (in the latter case, p < .05), with left-wing
respondents being more involved. In 8 cases, involvement could be
interpreted as the sole splitting variable, and political orientation could
be interpreted as the sole normal variable. Below, we report analysis
of a typical (imperfect) case in more detail. The chosen attitudinal
statement is as follows: “The government must force companies to let
their workers benefit from the profit as much as the shareholders do”
(1 = totally agree, 5 = totally disagree). For illustrative purposes, all
16 possible cusp models are fitted—for example, all 16 possibilities
of restricting the loadings of the two control factors (42 , one model
without restrictions included).
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 145

TABLE 3: Fit Statistics of Example 2

Model Log-Likelihood AIC BIC #P Parameters Fixed at 0

Linear –2,065 4,139 4,160 4

Logistic –1,868 3,746 3,773 5
Cusp 1 –1,908 3,823 3,844 4 a1, a2, b1, b2
Cusp 2 –1,901 3,811 3,838 5 a1, a2, b1
Cusp 3 –1,906 3,822 3,848 5 a1, b1, b2
Cusp 4 –1,900 3,812 3,843 6 a1, b1
Cusp 5 –1,873 3,755 3,781 5 a1, a2, b2
Cusp 6 –1,866 3,743 3,774 6 a1, a2
Cusp 7 –1,868 3,749 3,780 6 a1, b2
Cusp 8 –1,864 3,741 3,778 7 a1
Cusp 9 –1,816 3,642 3,668 5 a2, b1, b2
Cusp 10 –1,809 3,631 3,662 6 a2, b1
Cusp 11 –1,811 3,634 3,666 6 b1, b2
Cusp 12 –1,805 3,625a 3,661a 7 b1
Cusp 13 –1,816 3,644 3,675 6 a2, b2
Cusp 14 –1,809 3,631 3,668 7 a2
Cusp 15 –1,811 3,636 3,672 7 b2
Cusp 16 –1,805 3,626 3,668 8

NOTE: AIC is the Akaike information criterion, BIC is the Bayes information criterion, and
#P is number of estimated parameters.
a. The lowest value for AIC and BIC.

Results of Example 2

Table 3 shows the goodness-of-fit statistics of the 16 cusp models.

The model that shows the best fit is Model 12 (according to both the
likelihood ratio tests and the AIC/BIC). This model has also a lower
AIC and BIC than either the linear or the logistic model. Although
Model 12 has the lowest AIC and BIC, Model 10 has almost the same
BIC as Model 12. Model 10 is the most stringent model to test for
the cusp catastrophe hypothesis (political orientation only loads on
the normal factor, and involvement only loads on the splitting factor).
In comparison, Model 12 is marginally less strict (political orienta-
tion only loads on the normal factor). Table 4 shows the standardized
parameter estimates of Models 10, 12, and 16. Involvement had a
modest impact on the normal factor (Model 12: a2 = −.097; Model
16: a2 = −.093), while political orientation had a modest effect on
the splitting factor (Model 16: b1 = .079). These effects are much
smaller, though, than the estimated effects of involvement on the
146 SOCIOLOGICAL METHODS & RESEARCH

TABLE 4: Standardized Parameter Estimates of Restricted Models 12, 10, and 16

(Example 2)

Normal (a) Splitting (b) Location (l) Scale (s)

Model 10
0 Constant 0.178 0.120 –0.134 1.143
1 Political Orientation 0.450 0.000a
2 Involvement 0.000a –0.250

Model 12
0 Constant 0.152 0.120 –0.120 1.141
1 Political Orientation 0.463 0.000a
2 Involvement –0.097 –0.230

Model 16
0 Constant 0.161 0.128 –0.137 1.140
1 Political Orientation 0.457 0.079
2 Involvement –0.093 –0.239

splitting factor (Models 10, 12, and 16, respectively: b2 = −.250,

−.230, and −.239) and political orientation on the normal factor
(Models 10, 12, and 16, respectively: a1 = .450, .463, .457).
Figure 4 displays the fit of Model 12. The data are plotted in the
control plane (see also Figure 1). It shows that political orientation
only loads on the normal axis and that involvement mainly loads on
the splitting axis. A reasonable portion of the data falls within the
bifurcation set.

DISCUSSION

Catastrophe theory provides a way to model and test transitions in

attitudes. Hypotheses can be formulated and tested by making use
of the catastrophe flags. In doing so, different predictions can be
brought together into one coherent model. Examples of issues that
can be explained by the cusp catastrophe model are polarization and
ambivalence. In our view, existing models of attitude (and stereotype)
change, such as the bookkeeping, conversion, and subtyping models,
can also be incorporated in the cusp model. In addition, this model
includes a less well-studied mechanism of attitude change: the sudden
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 147

Figure 4: Location of the Data from Example 2 in the Cusp Control Plane
NOTE: The figure shows involvement and political orientation in relation to the
splitting and normal factor.

change of an attitude or stereotype after a gradual increase of (one-

sided) information.
We emphasized that the often-heard criticism that catastrophe mod-
els lack rigorous statistical tests can now be discarded. The way we
tested the cusp model of attitude change meets the standards used in
most fields of mathematical and statistical modeling in psychology. In
our two examples (distinguished by making use of one or two control
variables), the data are directly fitted in the cusp catastrophe model.
In both examples, the cusp model fits the data better than either the
linear or the logistic model.
148 SOCIOLOGICAL METHODS & RESEARCH

In the first example, the results from the Stouffer et al. (1950) data
set clearly indicated that involvement can be understood as the split-
ting factor. In the two other data sets, involvement contributed rela-
tively little to the normal factor. In the second example, involvement
and a left-wing political orientation were correlated. For this reason,
the involvement measures could not exclusively be taken as split-
ting variables since they also measured some informational content.
In future research, a distinction should be made between different
types of involvement. These types of involvement activate different
types of informational factors (Johnson and Eagly 1990). The results
of the second example suggest that involvement is a splitting factor
in relation to political attitudes, while political orientation works as
the main informational factor. Better results are expected when direct
measures of information and involvement are used.
A possible way to solve the measurement problem is suggested in
van der Pligt et al. (2000), who describe attitudes as an associative
network of attributes, representing positive and negative aspects of
the attitude. How many and which attributes are part of the network
can be determined by questionnaires or protocol analysis. van der
Pligt et al. quantify ambivalence (A) by the following formula: A =
(Np + Nn)/2 − |Np − Nn|, where Np and Nn are the number of
positive and negative attributes. The formula states that ambivalence
is a function of the sum of attributes minus the absolute difference
between the number of positive and negative attributes (weighing of
the importance of attributes might be necessary).
Based on this idea, we suggest defining involvement as the sum
of the number of positive and negative attributes (β = Np + Nn)
and information as the difference between the number of positive
and negative attributes (α = Np − Nn). As argued earlier, in the
cusp model, ambivalence occurs when β is high and α is close to
zero, which corresponds nicely with the formula of van der Pligt et
al. (2000). Polarization (divergence) occurs when α is about zero and
β strongly increases (e.g., when the number of positive and negative
attributes increases equally without distorting the balance between
them). Hysteresis can also be defined in these terms. This more refined
definition of information and involvement might help to solve the
problem with the operationalization of the control variables that was
the major problem in our analyses.
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 149

As argued in this article, the model based on catastrophe theory

can now be tested statistically. This development, in combination
with a more refined measurement of the various factors, offers new
opportunities for our understanding of attitudes and attitude change.

NOTES

1. An anonymous reviewer suggested that the problem with Guastello’s methods is mainly
due to the reverse hierarchical entry method (and to problems relating to ergodicity and sta-
tionarity) and not to the difference equation per se. Dismissal of the reverse hierarchical entry
method would indeed prevent the false conclusion in favor of the cusp for random data. Yet, we
think that the use of the difference equation is also wrong because catastrophe theory deals with
equilibrium behavior, whereas the equation of Guastello (1988) is about transient behavior. Fur-
thermore, Guastello’s method fails to distinguish between minima and maxima (see Figure 2).
2. A reanalysis with 50% of the data produced very similar results.

REFERENCES

Akaike, Hirotugo. 1974. “A New Look at the Statistical Model Identification.” IEEE Transac-
tions on Automatic Control 19:716-23.
Alexander, Ralph A., Glenn R. Herbert, Richard P. Deshon, and Paul J. Hanges. 1992. “An
Examination of Least Squares Regression Modeling of Catastrophe Theory.” Psychological
Bulletin 111:366-74.
Beek, Peter J., Frans Verschoor, and Scott Kelso. 1997. “Requirements for the Emergence of
a Dynamical Social Psychology.” Psychological Inquiry 8:100-4.
Burlingame, Gary M. and Carolen A. Hope. 1997. “Dynamical Systems and Social Psychology:
The Promise and the Pitfalls.” Psychological Inquiry 8:104-10.
Castrigiano, Domencio P. L. and Sandra A. Hayes. 1993. Catastrophe Theory. Reading, MA:
Addison-Wesley.
Cobb, Loren. 1978. “Stochastic Catastrophe Models and Multimodal Distributions.” Behavioral
Science 23:360-74.
———. 1980. “Estimation Theory for the Cusp Catastrophe Model.” Pp. 772-76 in Proceed-
ings of the Section on Survey Research Methods. Alexandria, VA: American Statistical
Association.
———. 1981. “Parameter Estimation for the Cusp Catastrophe Model.” Behavioral Science
26:75-78.
Cobb, Loren and Bill Watson. 1980. “Statistical Catastrophe Theory: An Overview.” Mathe-
matical Modelling 1:311-17.
Cobb, Loren and Shelemyahu Zacks. 1985. “Applications of Catastrophe Theory for Statistical
Modeling in the Biosciences.” Journal of the American Statistical Association 80:793-802.
150 SOCIOLOGICAL METHODS & RESEARCH

Eagly, Alice H. and Shelly Chaiken. 1993. The Psychology of Attitudes. Fort Worth, TX:
Harcourt Brace.
Eiser, J. Richard, ed. 1994. Attitudes, Chaos & the Connectionist Mind. New York: Blackwell.
Felling, Albert J. A., Jan Peters, and Osmund Schreuder. 1985. Sociaal-culturele Ontwikke-
lingen in Nederland. Vakgroep Methoden van Onderzoek, Katholieke Universiteit van
Nijmegen, SWIDOC Steinmetz Archive Database, NL STAR DATA P1012.
Flay, Brian R. 1978. “Catastrophe Theory in Social Psychology: Some Applications to Attitudes
and Social Behavior.” Behavioral Science 23:335-50.
Gilmore, Robert. 1981. Catastrophe Theory for Scientists and Engineers. New York: John
Wiley.
Guastello, Stephan J. 1988. “Catastrophe Modeling of the Accident Process: Organizational
Subunit Size.” Psychological Bulletin 103:246-55.
———. 1992. “Clash of the Paradigms: A Critique of an Examination of the Polynomial
Regression Technique for Evaluating Catastrophe Theory Hypotheses.” Psychological Bul-
letin 111:375-79.
Hartelman, Pascal A. I. 1997. Stochastic Catastrophe Theory. Dissertatie reeks 1997-2, Facul-
teit Psychologie, Universiteit van Amsterdam.
Hartelman, Pascal A. I., Han L. J. van der Maas, and Peter C. M. Molenaar. 1998. “Detecting
and Modeling Developmental Transitions.” British Journal of Developmental Psychology
16:97-122.
Johnson, Blair T. and Alice H. Eagly. 1990. “Involvement and Persuasion: Types, Traditions,
and the Evidence.” Psychological Bulletin 107:375-84.
Kaplan, Kalman J. 1972. “On the Ambivalence-Indifference Problem in Attitude Theory and
Measurement: A Suggested Modification of the Semantic Differential Technique.” Psycho-
logical Bulletin 77:361-72.
Krosnick, Jon A. and Richard E. Petty. 1995. “Attitude Strength: An Overview.” Pp. 1-24 in
Attitude Strength: Antecedents and Consequences, edited by Richard E. Petty and Jon A.
Krosnick. Mahwah, NJ: Lawrence Erlbaum.
Kruglanski, Arie W., Russell W. Clement, and John T. Jost. 1997. “The New Wave of
Dynamism: Will It Engulf the Field?” Psychological Inquiry 8:132-35.
Lange, Rense, Terence A. Oliva, and Sean R. McDade. 2000. “An Algorithm for Estima-
tion Multivariate Catastrophe Models: GEMCAT II.” Studies in Nonlinear Dynamics and
Econometrics 4:137-68.
Latané, Bibb and Andrzej Nowak. 1994. “Attitudes as Catastrophes: From Dimensions to
Categories With Increasing Involvement.” Pp. 219-249 in Dynamical Systems in Social
Psychology, edited by R. R. Vallacher and A. Nowak. San Diego: Academic Press.
Latané, Bibb, Andrzej Nowak, and James H. Liu. 1994. “Measuring Emergent Social Phe-
nomena: Dynamism, Polarization, and Clustering as Order Parameters of Social Systems.”
Behavioral Science 39:1-24.
McCutcheon, Allen L. 1987. Latent Class Analysis. Beverly Hills, CA: Sage.
McGarty, Craig and S. Alexander Haslam. 1997. The Message of Social Psychology: Perspec-
tives on Mind in Society. Oxford, UK: Blackwell.
Oliva, Terence A., Wayne S. Desarbo, Diana L. Day, and Kamel Jedidi. 1987. “GEMCAT: A
General Multivariate Methodology for Estimating Catastrophe Models.” Behavioral Science
32:121-37.
Pettigrew, Thomas F. 1981. “Extending the Stereotype Concept.” In Cognitive Processes in
Stereotyping and Intergroup Behavior, edited by D. L. Hamilton. Hillsdale, NJ: Lawrence
Erlbaum.
Petty, Richard E., Duane T. Wegener, and Lenadre R. Fabrigar. 1997. “Attitudes and Attitude
Change.” Annual Review of Psychology 48:609-47.
 Mass et al. / SUDDEN TRANSITIONS IN ATTITUDES 151

Ploeger, Annemie, Han L. J. van der Maas, and Pascal Hartelman. 2002. “Catastrophe
Analysis of Switches in the Perception of Apparent Motion.” Psychonomic Bulletin &
Review 9:26-42.
Poston, Tim and Ian Stewart. 1978. Catastrophe Theory and Its Applications. London: Pitman.
Reardon, Richard and Abraham Tesser. 1982. Is Conformity a Catastrophe? Unpublished
manuscript, University of Georgia, Athens.
Rothbart, Myron. 1981. “Memory Processes and Social Beliefs.” Pp. 145-82 in Cognitive
Processes in Stereotyping and Intergroup Behavior, edited by D. L. Hamilton. Hillsdale,
NJ: Lawrence Erlbaum.
Saris, Willem E. 1997. “Public Opinion About the EU Can Easily Be Swayed in Different
Directions.” Acta Politica 32:406-35.
Schwarz, Gideon. 1978. “Estimating the Dimension of a Model.” Annals of Statistics 6:461-64.
Sherif, Carolyn W., Muzafer Sherif, and Roger E. Nebergall. 1965. Attitude and Attitude
Change: The Social Judgment-Involvement Approach. Philadelphia: Saunders.
Stouffer, Samuel A., Louis Guttman, Edward A. Suchman, Paul F. Lazarsfeld, Shirley A.
Star, and John A. Clausen. 1950. Measurement and Prediction. Princeton, NJ: Princeton
University Press.
Sussman, Hector J. and Raphael S. Zahler. 1978. “Catastrophe Theory as Applied to the Social
and Biological Sciences: A Critique.” Synthese 37:117-216.
Taylor, Shelley E. 1981. “A Categorization Approach to Stereotyping.” Pp. 83-113 in Cognitive
Processes in Stereotyping and Intergroup Behavior, edited by D. L. Hamilton. Hillsdale,
NJ: Lawrence Erlbaum.
Teahan, John E. 1975. “A Longitudinal Study of Attitude Shifts Among Black and White Police
Officers.” Journal of Social Issues 31:47-56.
Tesser, Abraham. 1976. “Attitude Polarization as a Function of Thought and Reality Con-
straints.” Journal of Research in Personality 10:183-94.
———. 1980. “When Individual Dispositions and Social Pressure Conflict: A Catastrophe.”
Human Relations 33:393-407.
Tesser, Abraham and John Achee. 1994. “Aggression, Love, Conformity, and Other Social
Psychological Catastrophes.” Pp. 95-109 in Dynamical Systems in Social Psychology, edited
by R. R. Vallacher and A. Nowak. San Diego: Academic Press.
Thom, René. 1972. Stabilité Estructurelle et Morphogénesè. New York: Benjamin.
Thompson, Megan M., Mark P. Zanna, and Dale W. Griffin. 1995. “Let’s Not Be Indifferent
About (Attitudinal) Ambivalence.” Pp. 361-86 in Attitude Strength: Antecedent and Con-
sequence, edited by R. E. Petty and J. A. Krosnick. Mahwah, NJ: Lawrence Erlbaum.
Titterington, D. Mike, Adrian F. M. Smith, and Udi E. Makov. 1985. Statistical Analysis of
Finite Mixture Distributions. Chicester, UK: Wiley.
Tourangeau, Roger 1987. “Attitude Measurement: A Cognitive Perspective.” Pp. 149-62 in
Social Information Processing and Survey Methodology, edited by H. Hippler, N. Schwarz,
and S. Sudman. New York: Springer-Verlag.
Vallacher, Robin R., Andrzej Nowak, and J. Kaufman. 1994. “Intrinsic Dynamics of Social
Judgment.” Journal of Personality and Social Psychology 67:20-34.
van der Maas, Han L. J. 1998. “The Dynamical and Statistical Properties of Cognitive Strategies:
Relations Between Strategies, Attractors, and Latent Classes.” Pp. 161-76 in Applications
of Nonlinear Dynamics to Developmental Process Modeling, edited by K. M. Newell and
P. C. M. Molenaar. Mahwah, NJ: Lawrence Erlbaum.
van der Maas, Han L. J. and Peter C. M. Molenaar. 1992. “Stagewise Cognitive Development:
an Application of Catastrophe Theory.” Psychological Review 99:395-417.
152 SOCIOLOGICAL METHODS & RESEARCH

van der Maas, Han L. J., Maartje E. J. Raijmakers, Pascal A. I. Hartelman, and Peter
C. M. Molenaar. 1999. “Reaction Time as Recovery Time After Pertubation.” Pp. 159-69
in Nonlinear Analysis of Developmental Processes, edited by G. Savelsbergh, H. L. J. van
der Maas, and P. van Geert. Amsterdam: KNAW.
van der Pligt, Joop, Nanne K. De Vries, Antony S. R. Manstead, and Frenk van Harreveld.
2000. “The Importance of Being Selective: Weighing the Role of Attribute Importance in
Attitudinal Judgment.” Advances in Experimental Social Psychology 32:135-200.
Weber, Renee and Jennifer Crocker. 1983. “Cognitive Processes in the Revision of Stereotypic
Beliefs.” Journal of Personality and Social Psychology 45:961-77.
Zimbardo, Philip G., Ebbe B. Ebbesen, and Cristina Maslach. 1977. Influencing Attitudes and
Changing Behavior: An Introduction to Method, Theory and Applications of Social Control
and Personal Power. 2d ed. Reading, MA: Addison-Wesley.
Zeeman, E. Cristopher. 1976. “Catastrophe Theory.” Scientific American 234:65-83.

Han L. J. van der Maas is a professor of developmental psychology at the University of

Amsterdam. His research interests include cognitive development, catastrophe theory,
statistical methods, phase transitions, and mathematical modeling. Recent publications
include the following: “Evidence for the Phase Transition From Rule I to Rule II on the
Balance Scale Task” in Developmental Review (2001, with B. R. J. Jansen); “Catastro-
phe Analysis of Switches in the Perception of Apparent Motion” in Psychonomic Bul-
letin & Review (2002, with A. Ploeger and P. Hartelman); “A Phase Transition Model
for Mother Child Interaction” in Infant and Child Development (2000, with M. E. J.
Raijmakers); and “Stagewise Cognitive Development: An Application of Catastrophe
Theory” in Psychological Review (1992, with P. C. M. Molenaar).

Rogier Kolstein is a former student in the Psychology Department at the University of

Amsterdam and carried out various parts of the research described in this article.

Joop van der Pligt is a professor of experimental social psychology at the University of
Amsterdam. His research interests include attitudes and attitude change, risk perception,
and decision making under uncertainty. Recent publications include the following: “The
Importance of Being Selective: Weighing the Role of Attribute Importance in Attitudinal
Judgment” in Advances in Experimental Social Psychology (2000, with N. De Vries,
T. Manstead, and F. Van Harreveld); “Attribute Importance and Accessibility” in the
British Journal of Social Psychology (2000, with F. Van Harreveld and N. De Vries);
and “The What, When and How of Affective Influences on Interpersonal Behavior” in
Psychological Inquiry (2002, with T. Manstead).