The Affective Tipping Point: Do Motivated Reasoners
Ever “Get It”?pops_772 563..594
David P. Redlawsk
Rutgers University
Andrew J. W. Civettini
Knox College
Karen M. Emmerson
Lake Research Partners
In order to update candidate evaluations voters must acquire information and determine
whether that new information supports or opposes their candidate expectations. Normatively,
new negative information about a preferred candidate should result in a downward
adjustment of an existing evaluation. However, recent studies show exactly the opposite;
voters become more supportive of a preferred candidate in the face of negatively
valenced information. Motivated reasoning is advanced as the explanation, arguing that
people are psychologically motivated to maintain and support existing evaluations. Yet it
seems unlikely that voters do this ad infinitum. To do so would suggest continued motivated
reasoning even in the face of extensive disconfirming information. In this study we
consider whether motivated reasoning processes can be overcome simply by continuing to
encounter information incongruent with expectations. If so, voters must reach a tipping
point after which they begin more accurately updating their evaluations. We show experimental
evidence that such an affective tipping point does in fact exist. We also show that
as this tipping point is reached, anxiety increases, suggesting that the mechanism that
generates the tipping point and leads to more accurate updating may be related to the
theory of affective intelligence. The existence of a tipping point suggests that voters are
not immune to disconfirming information after all, even when initially acting as motivated
reasoners.
KEY WORDS: Affective intelligence, Motivated reasoning, Process-tracing, Voting, Candidate
evaluation
Political Psychology, Vol. 31, No. 4, 2010
doi: 10.1111/j.1467-9221.2010.00772.x
563
0162-895X © 2010 International Society of Political Psychology
Published by Wiley Periodicals, Inc., 350 Main Street, Malden, MA 02148, USA, 9600 Garsington Road, Oxford, OX4 2DQ,
and PO Box 378 Carlton South, 3053 Victoria Australia
What does it mean for a voter to be rational? At a minimum, rational voters
know their own preferences, update those preferences accurately upon receipt of
new information, and choose the candidate that best represents their interests. Such
voters should be predictable in the sense that when they encounter new information
about a candidate, evaluations of that candidate will be adjusted up or down
as appropriate. Yet, real people are not nearly the predictable “cool calculators”
rational models seem to require (Redlawsk, 2002). Recent research has convincingly
shown that emotions play an important part in most decision-making realms.
While classical political thought drew distinctions between reason and emotion,
reconceptualizations from neuroscience (Damasio, 1999) to political science
(Lodge & Taber, 2000, 2005; Marcus, Newman, & MacKuen, 2000) demonstrate
that emotions must be an integral part of political decision-making processes.
Existing affective evaluations color how people think about candidates (Redlawsk,
Civettini, & Lau, 2007) and issues (Lodge & Taber, 2000, 2005) and how
new information is processed as it is learned during a campaign. Rational updating
requires that negative information lower the evaluation of a candidate, while
positive information must do the opposite (Green & Gerber, 1999). But what if the
candidate is one a voter already likes; a candidate whom the voter has already
decided is “good”? What happens when negative information is encountered about
that candidate? A developing body of research shows that voters may operate as
motivated reasoners, attempting to hold to their existing positive evaluation by
using any one of a number of processes to explain away new incongruent information
(Kunda, 1990; Lodge & Taber, 2000; Taber & Lodge, 2006; Redlawsk,
2002). In other words, existing affect may interfere with accurate updating.
We argue that existing affect towards an already known candidate is an
important factor in determining the extent to which new information is accurately
perceived and evaluations correctly updated. We present evidence that voters
ignore significant amounts of negative information about positively evaluated
candidates. In fact, voters may become even more positive about a candidate they
like after learning something negative about that candidate. This tendency, so
contrary to classical notions of rational updating, is consistent with theories of
motivated reasoning (Taber & Lodge, 2006), research on a “conservation bias”
(Steenbergen, 2001), and the concept of cognitive dissonance (Festinger, 1957).
Yet we do not believe this can go on without end. At some point even the most
strongly held positive evaluation should flag in the face of repeated negative
information relevant to the evaluation. In this paper, we demonstrate that there is,
in fact, a point at which voters stop reinforcing their preferences, abandon motivated
reasoning, and begin “rational” updating. We call this the affective tipping
point.1
1
While the concept of a “tipping point” is not especially new, Gladwell’s (2000) book The Tipping
Point has brought the idea into the public imagination. It seems a very effective way to describe the
point at which things change, in a sense the straw that breaks the camel’s back.
564 Redlawsk et al.
Theoretical Perspective
Standard cognitive/rational models do not appropriately account for affect in
the evaluation and decision-making process. Green and Gerber’s (1999) description
of a Bayesian updating process in which new information updates prior beliefs
is typical. New information in agreement with an existing candidate evaluation is
assumed to strengthen that evaluation. Information to the contrary does the opposite.
Thus a voter’s early positive impressions of a candidate will be strengthened
and made more positive by learning something good about that candidate. But,
when the voter learns something disagreeable about that same candidate, the prior
evaluation will be updated in a negative direction to account for this new negative
information. While the updating process itself need not be linear, nowhere in this
model is there any suggestion that existing affect towards the candidate might
actually impede attitude change. Further, there is no serious consideration given to
the possibility of asymmetric effects, that negative information might be weighted
differently than positive information in the updating process (Redlawsk, 2007).
Holbrook, Krosnick, Visser, Gardner, and Cacioppo (2001) attempt to address
this nonlinear asymmetric possibility and in doing so propose a more nuanced, but
still “rational” updating process. In their Asymmetric Nonlinear Model (ANM)
positive information, while carrying more weight at the beginning of an updating
process, decreases in importance more rapidly than negative information. Thus, in
order to update, a voter must assess the direction of the information before
incorporating it into an evaluation. This modification of the updating model, while
taking into account the asymmetric nature of positive and negative information,
proceeds in the same vacuum that other cognitive approaches inhabit; the existing
evaluation of the candidate serves merely to anchor the updated evaluation, but
does not directly condition the updating process itself.
Enter Affect
Since Festinger’s (1957) description of cognitive dissonance and Heider’s
(1958) development of balance theory, psychological studies of affect and updating
have regularly suggested that cognitive processes are not so straightforward,
and certainly do not proceed in a vacuum. The hot cognition thesis (Abelson,
1963) argues that affect and cognition are inexorably linked; for every concept or
piece of information in memory there is an associated affective evaluation that is
activated whenever the concept is accessed (Lodge & Taber, 2000, 2005; Taber &
Lodge, 2006). Whether positive or negative, affect cannot be separated from the
underlying information, so theories that focus only on cognitive updating can only
tell part of the story. Indeed, Zajonc (1984) convincingly argues for the primacy of
affect; that affective responses occur before conscious processing. There are both
cognitive and affective systems wired in the human brain, and these may well work
in parallel and independently.
565The Affective Tipping Point
Marcus and his colleagues (Marcus & MacKuen, 1993; Marcus et al., 2000)
propose a theory of affective intelligence to describe the direct effects of the
affective systems on cognitive processes. Affective responses are the result of a
dual process emotional system: a behavioral inhibition system and a behavioral
approach system (Marcus & MacKuen, 1993). The first system compares a new
stimulus to existing expectations, and if a stimulus is found to be incongruent
with those expectations, attention is shifted to it. The new stimulus is thus a
potential threat, and this perceived threat generates negative affect like anxiety,
interrupting normal (essentially below consciousness) processing. This interruption
leads to active processing, where both attention to the new information and
the time taken to process it increases. Thus, negative affect directly motivates the
individual to learn more about the stimulus and the environment in general. If
affective intelligence is right, we would expect that during an election campaign
anxious voters would be more attentive, more informed, and more likely to make
good choices (Lau & Redlawsk, 1997; Lau, Anderson, & Redlawsk, 2008).
Marcus and his colleagues provide evidence that anxious voters show more
learning and attention to campaigns, while calm voters pay less attention
(Marcus et al., 2000). And while they do not test it directly, the logical conclusion
of their theory is that affectively intelligent voters should look more like
rational updaters.
But other research on affect and its effects on cognition somewhat muddy
these waters. Holbrook (2005) used political ads to generate positive, negative,
and neutral affect in subjects. Highly anxious individuals were more responsive
to new information, but overall they were less able to accurately recall information
after the fact. On the other hand, Brader (2005), also using political ads
to generate affect, found that anxious subjects were more likely to recall information
related to the issue in the ads, but failed to seek out more information.
Isen (2000) argues that positive affect improves cognitive processing; contrary to
Marcus et al., affectively positive individuals are more likely to ignore incongruent
information, rather than pay special attention to it. Finally, other evidence
supports the notion that anxious voters pay more attention and accurately
process more information, but this effect may be limited to conditions where
there is a great deal of incongruent information in the environment (Redlawsk,
Civettini, & Lau, 2007). Feeling just a little bit anxious might not be enough to
trigger careful attention.
This latter point may be crucial to understanding the apparent contradictions
in the research on anxiety and updating. The affective intelligence thesis was tested
by Marcus in the aggregate, using American National Election Studies data,
collected during the particularly rich political environment of a presidential campaign.
Most of the other studies have examined negative affect in a more limited
form in a laboratory with subjects exposed to relatively little information. If
negative affect has its greatest effect when there is a great deal to be anxious
about—for example, when there are significant amounts of affectively incongruent
566 Redlawsk et al.
information which increases the threat to existing evaluations—we might not find
consistent effects unless the information environment becomes especially threatening
to existing beliefs. Why not? Because other processes, such as motivating
reasoning, may be working at cross purposes with affective intelligence.
Where affective intelligence argues that negative affect may produce better
decisions, motivated reasoning suggests that this is not quite the whole story
(Kunda, 1990; Taber & Lodge, 2006; Redlawsk, 2002, 2006). Motivated reasoners
make an immediate evaluation of new information and use it to update an online
tally that summarizes their evaluative affect (Hastie & Park, 1986; Lodge,
McGraw, & Stroh, 1989; Lodge, Steenbergen, & Brau, 1995; Redlawsk, 2001).
Newly encountered information carries with it an affective value. Given an existing
evaluation (represented by the online tally), these affective components interact
so that the online tally directly influences how the new information is evaluated
before it is used to update the tally. This is the key insight missing from both the
cognitive approaches and affective intelligence. Even anxious voters presumably
motivated to learn more and make more accurate assessments may well be subject
to the processing biases of motivated reasoning as they affectively evaluate before
they begin to cognitively process new information.
While a negative emotional response may be generated by incongruence
between expectations (existing affect as summarized by the online tally) and new
information, motivated reasoning suggests that this incongruence does not necessarily
lead to greater accuracy in evaluation or greater information search. Instead
voters committed to a candidate may be motivated to discount incongruent information;
they may mentally argue against it, bolstering their existing evaluation by
recalling all the good things about a liked candidate even in the face of something
negative. Motivated reasoning describes an interaction between existing affective
evaluations and new information, but unlike affective intelligence, the effect of
affect may lead to less accurate updating, rather than more.
Inaccurate updating might be of different kinds. One possibility is that a voter
updates her beliefs in the correct direction, but not to the appropriate magnitude.
That is, instead of becoming less positive by a factor of “X” in the face of negative
information, the evaluation might become less positive by something less than
“X,” in effect a conservatism bias (Steenbergen, 2001). Such updating failures
might have relatively limited consequences, as long as they are directionally
accurate. On the other hand, attitude strengthening effects (also called polarization
effects, Lord, Ross, & Lepper, 1979) where updating is in the wrong direction have
been demonstrated. Redlawsk (2002) finds that voters with an existing positive
evaluation of a candidate become more positive about that candidate when encountering
negative information. Lodge and Taber (2000, 2005; Taber & Lodge, 2006)
show a similar effect in studying issue preferences, and Edwards and Smith (1996)
find that individuals confronted with an argument in conflict with their prior
beliefs judge that argument to be weak, spend longer scrutinizing it, and generate
a list of relevant thoughts and arguments that tend to refute the argument rather
567The Affective Tipping Point
than support it, a process consistent with attitude strengthening. In the context of
a campaign, a voter learning something negative about a favorite candidate might
first doubt the validity of the information, spend time reviewing and trying to
comprehend it, and in the process create a list of relevant thoughts, most of which
argue that the information is either false or unimportant. This thought listing, in
refuting the new piece of information, could call to mind many of the reasons for
the initial support of the candidate and leave a better feeling about the candidate
even after encountering negative information.
In the end we consider that there may be two processes with contradictory
effects. One, motivated reasoning, suggests that small amounts of incongruent
information can be countered in the service of maintaining existing evaluations.
The other, affective intelligence, might not come into play until there is a significant
threat to expectations, driving up anxiety, and overcoming the motivation to
maintain evaluations. It is the existence of this hybrid process that we test for in the
remainder of this paper.
Hypotheses
We know that voters update their evaluations based on new information
encountered during the campaign. We also know that affect influences this process
in many ways. Voters with positive affect towards a candidate may want to hold on
to that evaluation and resist changing their opinion.2
The question here is whether
there is a point at which the positive affect motivated reasoners try to maintain is
overwhelmed by a growing threat to the existing evaluation as more incongruent
information is encountered, and thus leading to more accurate updating.3
And if
there is such a point, what drives it? We know that updating is not a linear process,
but it may be that it is also a hybrid process, with voters updating one way when
encountering just a bit of unexpected (incongruent) information, but another way
when the threat represented by the new information grows large. To test this
empirically we must do several things. First, an initial candidate evaluation must
be established—that is, the voter must have time to learn something about the
candidates. Second, we must challenge the voter’s positive evaluation of the
2
It is also possible that a similar effect occurs with a disliked candidate—that is, that voters who
develop negative affect towards a candidate may be unwilling to positively update their evaluations,
at least at first. While the design of our study allows us to examine this possibility, for reasons of both
space and theoretical clarity our focus in this paper will be on positively evaluated candidates.
3
Another way to think of this in more Bayesian terms is to consider whether as more negative
information is encountered, a voter’s positive “priors” become less and less important in the calculation
of the posterior evaluation. Gill (2007) notes in a different context that as the N of new
information points goes to infinity, the “data” eventually win. In other words, is there a point at which
the priors no longer exert influence on the calculation of a revised evaluation given new information?
Of course, we would argue in the present case that the N need not go to infinity at all, but that there
is a finite tipping point at which evaluation begins to update more accurately.
568 Redlawsk et al.
preferred candidate by providing negative information about that candidate. Third,
we must assess whether the updated evaluation shows evidence of motivated
reasoning or some more accurate form of updating, and finally, if both processes
are at work, we must find the point at which the impact of new negative information
on evaluation changes.
In summary, we believe the updating process works something like this: (1)
A voter develops an initial positive evaluation of a candidate through the early
information that is learned; (2) if a small amount of negative information is
encountered, rather than adjusting downward this initial evaluation becomes more
positive, showing the motivated reasoning attitude strengthening effect; (3) if
enough negative information is encountered to heighten the voter’s anxiety about
the preferred candidate, affective intelligence suggests that he or she will become
more careful in processing additional new information; (4) This increased anxiety
and careful processing may lead to an affective tipping point where additional
negative information begins to generate downward adjustments to the evaluation.
We suggest that a voter’s evaluation of a positively evaluated candidate should
follow a pattern such as the one seen in Figure 1 as greater amounts of negative
information are encountered.4
Putting this into a more structured hypothesis, we expect that:
4
It is certainly possible that a voter might never actually encounter negative information about a liked
candidate. There is evidence that given a choice, voters may well work to confirm their evaluations by
avoiding information that might challenge them (Taber & Lodge, 2006). This particular study does
not address the issue of avoiding incongruent information; instead most of the subjects in the
experiment to be described did in fact encounter such information, though a small number did not and
thus provided a useful control group.
Amount of Incongruent Information
Evaluation
NegativeNeutralPositive
Figure 1. Expected Effects of the Amount of Incongruent Information on Evaluation of a
Preferred Candidate.
569The Affective Tipping Point
Hypothesis 1: Updating for positively evaluated candidates in the face of
incongruent information will begin as a motivated reasoning process,
showing attitude strengthening effects. Given enough incongruent information,
an affective tipping point exists where increasingly anxious
voters will begin to update candidate evaluations more accurately in the
face of increasing incongruent information.
Hypothesis 1 suggests that if we find attitude strengthening effects for small
amounts of incongruent information then these effects are the result of motivated
reasoning. Of course, we cannot see motivated reasoning as it happens, only its
results. But one means by which motivated reasoners might support their existing
evaluations is by bolstering, that is, bringing to mind positive information already
known to offset the new negative information that has been encountered. The
results of such a process may then be visible in the memories people report about
candidates after the election. In particular:
Hypothesis 2: Voters encountering small amounts of negative information
about a liked candidate will report more positive memories about that
candidate than will those encountering no negative information or those
encountering large amounts of negative information.
Finally, increasing anxiety on the part of the voter who is learning increasingly
negative information about a positively evaluated candidate may be the
mechanism by which motivated reasoning is overcome and attitude strengthening
ends. In following Marcus et al. (2000) on this point, we argue that at a high
enough level of incongruency—unexpectedly negative information about a liked
candidate—the environment in which these voters are operating will become more
threatening, increasing anxiety resulting in more accurate updating. Another way
of thinking about this is that voters should also become less certain that they made
the “right” choice as their anxiety grows. Thus:
Hypothesis 3a: As increasing incongruency drives up anxiety about a
positively evaluated candidate, voters will become less certain that they
have made the right choice when called upon to cast a vote.
Direct evidence of greater anxiety as voters encounter more incongruency will also
provide some evidence of this otherwise unseen process:
Hypothesis 3b: Encountering incongruent information generates anxiety
that grows as more incongruency is encountered. Anxiety will increase
until the voter adjusts to the new information and begins to consider other
candidates.
570 Redlawsk et al.
Methodology
While a decision may be a single choice made at one point in time, evaluation
is a process that occurs over some period of time. To understand a process, we
should observe it as it occurs. Process-tracing experiments have been employed
outside of political science by using information boards that allow subjects to
choose exactly what they would like to learn about a set of alternatives presented
to them (Ford, Schmitt, Schechtman, Hults, & Doherty, 1989; Jacoby, Jaccard,
Kuss, Troutman, & Mazursky, 1987). Within political science similar information
boards have been used to examine voting (Herstein, 1981), political decision
making (Riggle & Johnson, 1996), and information search in political environments
(Huang, 2000; Huang & Price, 1998) among other subjects. However, they
have rarely been used to study candidate evaluation, though it would seem that
process tracing could yield great insights into this subject.
The problem is that the traditional information board is static and allows
constant access to all attributes for all alternatives under consideration. In the
context of an election, this would be as if a voter had access to any piece of
information about a candidate at any time, allowing easy comparison between
candidates across all attributes. In a real election, however, information is much
less organized, somewhat more chaotic, and the time allowed for learning and
information gathering is limited by Election Day. Information comes and goes,
and candidates do not always make it easy for voters to make a comparison or even
get a clear understanding of where they stand on issues. Lau and Redlawsk’s (Lau,
1995; Lau & Redlawsk, 2001, 2006) computer-based dynamic process-tracing
methodology offers a way to model the vagaries of a political campaign in a
controlled experimental environment. The system generates an ever-changing
information environment that mimics the flow of information throughout a campaign
and makes only limited amounts and types of information available at any
point in time, much like in an election campaign in which many issues are “here
today, gone tomorrow.” It can also overwhelm voters with potentially unmanageable
amounts of unorganized information in a way that resembles the media
maelstrom in a real political environment. Yet, the dynamic information board
retains the essential characteristic of process-tracing experiments in that it tracks
the evaluation and decision-making process as it happens and as information is
acquired.
We use this dynamic process-tracing environment to present a simulated
presidential primary election campaign to subjects who learn about four candidates
from within their party, evaluate them, and make a vote choice.5
The
5
A primary election was chosen to limit the direct effects of partisanship in this particular study.
Obviously partisanship is of great import during a general election and undoubtedly plays a critical
role in establishing candidate preference and possibly in resisting change to that preference. However,
it adds a layer of individual difference that negatively impacts experimental control. Thus we settled
on offering a primary election where partisanship would not be a factor for this study. If we can
571The Affective Tipping Point
campaign consists of a wide range of information about each candidate, including
27 issue positions, plus group endorsements, personality traits, and background
information, along with preelection polls. As voters learn about the candidates, the
system collects data on the information they access, how long they spend on each
item, how they feel about each item, and their vote choices and evaluations of each
candidate. These last two measures are obtained multiple times throughout the
campaign in the form of “polls” in which voters are asked to choose a favorite
candidate and rate all candidates. Given all the measures we are able to monitor,
this methodology clearly provides an excellent way of tracking the evaluation and
decision-making process during a campaign and the role that existing evaluative
affect plays in the processing of new information and updating of candidate
evaluations.
Experimental Design
A total of 207 nonstudent subjects were recruited from the Eastern Iowa area
to participate in a mock presidential primary featuring four candidates from one
party.6
Candidates in the primary were fictional but designed to realistically represent
the range of ideologies within their parties. Since the candidates were not
real, subjects clearly had no prior knowledge about any of them, requiring evaluations
to be determined only by the information accessed and inferences made
during the campaign. Subjects registered as either Democrat or Republican before
being exposed to information for the candidates from their party. Subjects were
only allowed to vote in the party for which they had registered. Once the campaign
began, subjects chose what they wished to learn about the candidates from an
ever-changing set of candidate attributes presented over a 25-minute time period.
When subjects arrived they were seated at a computer and given an oral
introduction to the experiment. They then completed an online questionnaire
measuring their political preferences, knowledge, and interests. These questions
allow us to gauge each subject’s placement on the issues that were used in the
simulation, which was necessary to be able to manipulate incongruency through
subject-candidate agreement. Subjects were given a chance to practice with the
dynamic process-tracing environment. They then began the primary campaign
where they had 25 minutes to learn about the four candidates in their party,
following which they voted for one of them. The campaign was interrupted after
about seven minutes by a poll where subjects were asked to report their vote
establish the existence of a tipping point in the first place in this simpler environment, we can move
on in future work to examine the conditions under which the tipping point itself varies. The strength
of partisanship would clearly be one of those conditions worth closer examination.
6
Subjects were recruited in a variety of ways to ensure some level of diversity, specifically in age and
income. We do not claim that the subject pool is representative of any particular population. Subjects
ranged in age from 18 to 88 years and had household incomes ranging from 7.5 to 100 thousand
dollars per year. Fifty-six (56) percent of the subjects were female. Subjects who completed the study
received $20 for their time.
572 Redlawsk et al.
preference and feeling thermometer evaluations of all four candidates. This poll
was repeated two more times, once at about 13 minutes and again at about 20
minutes into the campaign. At the end of the campaign subjects voted and did one
more set of evaluations. They were then asked a number of follow-up questions,
including a memory-listing task asking subjects to record “everything you can
remember” about each candidate, using only the candidate name as a prompt.
Subjects were then prompted to indicate whether each memory made them feel
enthusiastic, anxious, or angry about the candidate. Finally, subjects completed a
cued recall task where they indicated whether they recalled examining each piece
of information they had seen and if so, what they recalled about their affective
response to it. They were then debriefed and dismissed.
The key experimental manipulation embedded in the election simulation
varied the probability of encountering incongruent information during the campaign,
and thus varied the information environment in which subjects operated.
Incongruent information is defined as any candidate attribute at odds with the
subject’s preferences. For example, if a subject was pro-choice, an incongruent
piece of information about a positively evaluated candidate would be that the
candidate was pro-life. In this way a pro-choice subject learning her preferred
candidate was pro-life would clearly have her expectations violated.7
We varied the
probability that subjects would encounter a certain amount of information like
this; unexpectedly bad positions taken by a liked candidate.
As noted above, about seven minutes into the campaign subjects were polled
and asked to indicate which candidate they would vote for if the election were held
7
A second manipulation was also included, but is not central to the analyses here which focus on how
individual pieces of information generate a sense of threat or anxiety as they accumulate. An
“embedded instructions” manipulation was intended to alter the overall emotional state of the subjects
just before the simulation began. Approximately half of the subjects were given special instructions
about the experiment designed to heighten their overall sense of anxiety about their performance in
the study. The instructions told subjects that their performance in the experiment was critical to the
continuation of our research funding and that they were expected to do a good job. The other half of
the subjects did not receive these instructions. An analysis of this intended manipulation showed that
it was not strong enough to generate the expected reaction; as a consequence we do not consider it
further here.
A third manipulation involved asking subjects how they felt about individual pieces of information
that they accessed. One-half of subjects (immediate affect group) were asked immediately after
viewing each item whether or not it made them feel enthusiastic, anxious, and/or angry toward the
candidate. They were also asked to recall this affective response at the end of the study during a cued
recall process. The other half (post affect group) were only asked to recall their affect at the end of
the study, well after the simulation had been completed. This manipulation was designed to test
whether or not subjects could accurately recall the affect attached to information they learned during
the election when they are asked about it later. An initial examination of the data (Civettini &
Redlawsk, 2005) suggests that affect recall is problematic at best. This manipulation is not directly
relevant to our discussion of the affective tipping point since we do not rely here on subjects’ own
assessment of affect. However, because asking immediate affective responses impacted the number of
items that could be examined by taking up time, we use this manipulation as an instrumental variable
to predict the amount of information examined for each liked candidate in Table 2 below, in order to
control for its effects.
573The Affective Tipping Point
at that point.8
Subjects also evaluated each candidate on a 0–100 feeling thermometer,
providing a ranking of candidate preferences. Following this first poll, subjects
were randomly assigned to one of five levels of incongruent information.
Those in Group 0 viewed candidates who were assigned issue positions that
remained ideologically consistent throughout the campaign (i.e., the most liberal
Democrat always taking the most liberal position or the moderate Democrat
always taking more conservative positions held by his party). For this group
candidate attributes were not manipulated in any way. In Group 1, a random 10%
of information made available to the subject was manipulated to be incongruent
with the subject’s own preferences, while 90% of the information remained consistent
with the candidate’s established ideology. Group 2 subjects were assigned
a 20% probability that available information would be incongruent with the subject’s
preferences with 80% remaining ideologically consistent, and Groups 3 and
4 were assigned 40% and 80% incongruency, respectively.
The assignment of incongruency occurred without the subjects’ knowledge.
In choosing what information to view about candidates, subjects could not know
before choosing an item whether it would be congruent or incongruent with their
initial candidate evaluation and therefore they could not control their information
environment or the amount of incongruency they encountered.9
Data
Process-tracing methodologies provide data that are both extensive and
complex. Before we could begin analyzing our results several steps were needed to
clean the dataset. To begin with, we dropped eight subjects who either failed to
complete the study or who did not take the study seriously. We then removed 10
additional subjects who looked at fewer than 50 pieces of information (less than two
per minute over the course of the election) or more than 200 pieces of information
(more than eight per minute). While retaining these 10 subjects does not change
either the significance or substance of our results, these subjects were clear outliers
when we examined the distribution of the number of items accessed across all
subjects. We were left with 189 subjects whose data were suitable for analysis.10
It is important to make clear that subjects themselves decided when to click on
information headlines to learn the details about any given candidate. Thus the
8
By the first poll, the average subject had looked at 15–20 pieces of information, which were
generally evenly spread across the four candidates in his or her party.
9
Subjects clicked on boxes that contained “headlines” stating what information was available. These
headlines were generally unsourced and valence-free. Examples include “Singer’s position in Iraq”
and “Rodgers’ political philosophy.”
10
After debriefing the experimenter coded the degree of seriousness with which the subject approached
the study, so this measure is based on observation of the subject’s demeanor at the time of the study.
Subjects examining fewer than 50 or more than 200 items were clear outliers. The mean number of
items examined was 111.43, with a standard deviation of 35.07. Removing these subjects from the
dataset does not substantively change the results and maintains consistency with Civettini and
Redlawsk (2009).
574 Redlawsk et al.
actual effect of the manipulation depends to some extent on the headlines that
subjects chose to examine. Within the experimental groups the actual amount of
incongruent information that was encountered was distributed around the assigned
probability based on two factors. First, because subjects chose from the available
headlines what they wanted to know, and because we were randomly assigning the
proportion of available information that could be incongruent, we could not
control exactly how much incongruent information subjects actually learned.
Second, because unmanipulated information was held ideologically consistent for
the candidate based on his initial position on the liberal-conservative spectrum,
subjects who themselves were not as ideologically consistent as the candidates
would have encountered information that was incongruent relative to their own
preferences but which was not specifically manipulated.
Ultimately for each subject there is a probability of x that any single item of
information about a liked candidate was incongruent with the subject’s preference
and a probably of 1-x that the item was at a set point along the liberal-conservative
spectrum based on the candidate’s assigned ideology, where x is the manipulated
probability: 0, 10, 20, 40, or 80%. If a subject’s initially preferred candidate was
the most liberal, we would expect that the subject would herself generally prefer
the most liberal positions. However, if the subject held some more moderate
preferences in her mix then she would still have the chance of encountering
candidate positions incongruent with her own preferences from among the items
we did not manipulate. The more internally inconsistent a subject’s own issue
preferences the more likely that the subject would encounter incongruent information.
In the end, though, our interest in this study is to examine the extent to
which encountering a more or less incongruent information environment results in
more or less or accurate updating of candidate evaluations; we are not examining
effects of the specific items themselves.
As the dynamic process-tracing experiment progressed, subjects clicked on
valence and source neutral headlines that appeared and disappeared from the
screen; each named an attribute that could be learned for a candidate. By clicking
on the headline, subjects could learn detailed issue positions, candidate traits, and
group endorsements. For each of these items that a subject accessed we recorded
whether it was congruent with the subject’s evaluation of the candidate. For traits
this was fairly simple, since the information hidden behind the headline was
clearly positive or clearly negative—for example, “Martin is considered egotistical
and difficult to work with” would certainly be seen as negative, while “Even
Martin’s opponents consider him an honorable man” would be viewed as a
positive trait.11
Endorsements could also be readily determined to carry positive or
negative valence, based on the subject’s own expressed group preferences (as
11
An initial set of trait statements was independently evaluated by three research assistants, who coded
them as positive or negative. Only statements that were agreed to be positive or negative by all three
coders were actually used in the study.
575The Affective Tipping Point
measured in the preexperiment questionnaire). If a disliked group endorsed a liked
candidate, this was incongruent, while a liked group endorsing the candidate
would be clearly congruent.
Coding positions as congruent or incongruent was a little more complicated.
For each issue in the campaign, eight different positions were created which could
then be assigned to any of the candidates on the fly as the simulation progressed.
Once an issue position was assigned to a candidate that candidate consistently took
that position and the position was not available to any other candidate. We were
then able to calculate the distance between the subject and the candidate on a
standard 7-point liberal-conservative scale for each individual issue. Where a
subject was randomly assigned to receive an incongruent issue item for a liked
candidate, the computer attempted to assign the most distant available issue
position to that candidate.12
As described earlier, subjects in Group 0 where candidate issue positions were
not manipulated could still encounter incongruent information since congruency
was determined by subject-candidate issue agreement. Many voters have inconsistent
ideologies, more liberal or conservative on some issues than others, so these
subjects were bound to disagree with even their favorite candidate at times. And of
course in the four experimental groups where congruency was manipulated, some
unmanipulated items might still have been incongruent with a subject’s own
preferences for the same reason. Thus our calculation of the amount of incongruency
subjects encountered is a combination of the proportion of manipulated items
that were incongruent plus incongruence encountered because of a subject’s own
ideological inconsistency. No matter what the cause of incongruence, it operated
in the same way—subjects would learn unexpectedly negative information about
a liked candidate. To establish the congruency of unmanipulated items, we manually
coded items as congruent or incongruent based on the actual candidate-subject
issue distance. Issue positions closer than 3.5 points on the liberal-conservative
scale were coded as congruent and issues 3.5 or more points away were coded as
incongruent, though tests of other cut points make no significance difference in the
results. Finally, to calculate the percentage of incongruent information each
subject encountered over the entire campaign, we divided the total number of
incongruent items the subject examined by the overall total number of items that
the subject examined, resulting in a measure of total incongruency for each
12
The issue position ratings were obtained by giving the list of positions to several graduate students
and faculty members in the University of Iowa’s Department of Political Science. Each person coded
all the items and the final rating was the average rating across all coders. The subject’s placement was
self-reported during the questionnaire at the beginning of the study.
During the experiment before the first poll was administered and the manipulation began,
positions were assigned to candidates as the subject selected pieces of information to read. Since
these positions could not be changed after the subject had viewed them, the fewer options were
available for the computer to choose from when executing the manipulation. In effect, the strength
of the manipulation was attenuated when only a less distant position was available to assign to the
candidate.
576 Redlawsk et al.
subject.13
This measure is a continuous variable, ranging from 0 to 100 percent,
with a mean of 37.9% and a standard deviation of 30.5.
Results
We have two goals in these analyses. First, we examine our process-tracing
data and candidate evaluations to determine whether the patterns of evaluation
match our theorized process. That is, as more negative information is encountered
about a liked candidate do we see a pattern of initial attitude strengthening, which
reaches some tipping point after which evaluations update in a negative direction?
Second, if the patterns our subjects exhibit match our theory, we must then
examine whether we have evidence to support our argument that both motivated
reasoning and affective intelligence processes are at work.
The Interaction of Affect and Candidate Evaluation
We turn first to candidate evaluation.14
The first part of Hypothesis 1 suggests
an attitude-strengthening effect in the face of a small amount of negative information
about a liked candidate. Voters who process information according to
normative standards should become less positive about a liked candidate with each
piece of negative information they learn. But motivated reasoners will try to
counter the negativity to maintain their existing evaluation and may in the course
of doing so become even more positive about a liked candidate. We begin by
simply looking graphically at the actual evaluations for each of our randomly
assigned groups. Figure 2 presents this information graphing the mean evaluation
of the initially preferred candidate from the first poll (before any manipulation of
information) through the second and third preelection polls, and the final postvote
evaluation (poll 4). The evaluations were made on a 0–100 point feeling thermometer
scale.
It is immediately clear that updating of evaluations does not proceed linearly
across the groups. The group assigned to no incongruency (that is, where all
information remained ideologically consistent for the candidate) shows a small
uptick in the evaluation of the liked candidate by the end of the campaign. This is
what we would expect since these subjects learned little or nothing that should
13
Nearly every item that was available to subjects could be assessed for congruency with a few
exceptions. In addition to issues, endorsements, and candidate personality traits, subjects could learn
poll results throughout the campaign and could also learn about candidate background and experience.
These items were not assessed as either congruent or incongruent and are ignored in the
analysis.
14
While we manipulated both the highest and lower rated candidates, we focus in this analysis only on
the candidate most preferred (highest rated) in the first poll, after which the manipulation of
congruency began. We would expect some similar effects (though reversed) for a disliked candidate;
however, it may also simply be that voters ignore a disliked candidate once they establish the
evaluation. Available space precludes us from this examination here.
577The Affective Tipping Point
have made them feel negatively about the candidate. But interestingly, so does the
group that was assigned to 10% incongruency. Even those assigned to 20%
negative information about the initially most preferred candidate do not, in the end,
show significant updating of their evaluations of that candidate. It is only within
the groups assigned to encounter 40% and 80% incongruent information that we
see the expected decline in the evaluation over time.
But this does not give us the full picture of the evaluation as a factor of the
actual level of incongruent information encountered, since as detailed above, the
randomly assigned groups provided a probability of encountering incongruent
information, rather than a certainty. The exact level of incongruency was controlled
by both the manipulation and the subject’s own level of ideological constraint.
The more constrained the subject the more likely that unmanipulated
information would be perceived as congruent with expectations and manipulated
information would be perceived as incongruent. We recoded our subjects into
incongruency quartiles, after first removing those subjects who demonstratively
never encountered any incongruency (based on our calculation of actual incongruency
as described earlier). This gives us five groups with differing incongruency
means, encountering from zero to 74.6% incongruent (negative) information
about the initially most preferred candidate. Table 1 describes these observed
incongruency groups and compares them to the original randomly assigned experimental
groups.
55
60
65
70
75
Poll 1 Poll 2 Poll 3 Poll 4
MeanFTRating
0% Incongrent
10% Incongruent
20% Incongruent
40% incongruent
80% incongruent
Figure 2. Evaluation of Preferred Candidate by Assigned Incongruency Groups Over the Course of
the Campaign.
578 Redlawsk et al.
When we graph the mean evaluations over time for each of these observed
groups, the differences seen in Figure 2 become even more pronounced. Now the
group that never actually encountered any incongruent information (Group 0)
actually ends up somewhat less positive about their favorite candidate at the end
than either of the first two quartiles of incongruency (Groups 1 and 2). And those
in the first quartile—averaging about 20% incongruent information—become
consistently more positive about their candidate, even in the face of a nonnegligible
amount of negative information. This finding clearly fits patterns previously
shown by Taber and Lodge (2006) as well as Redlawsk (2002) with attitude
strengthening effects evident in the face of incongruent information.
But it is also clear in Figure 3 that this strengthening effect does not occur for
all levels of incongruent information. While subjects who encountered smaller
amounts of incongruency seem to have their positive impression strengthened over
time, those encountering above the median amount of incongruency (Groups 3 and
4) show a very different pattern. These subjects begin lowering their evaluation of
the preferred candidate immediately, and that evaluation continues to decline over
time, though leveling off towards the end of the campaign. These subjects, then,
do show evidence of more accurate updating of their priors compared to those
encountering less incongruency.
We can calculate the change in evaluation from the first poll to the vote for
each of the randomly assigned and observed groups and test whether the changes
are statistically different from zero and from each other. We use a one-way
ANOVA to predict the mean evaluation for each incongruency group.15
We present
the results graphically in Figure 4, which includes both the random and observed
groups. The evaluation updating process appears to correspond with the theorized
15
Apost hoc LSD analysis of the significance of the difference between each group shows in both cases
(manipulated and observed groups) that the change in evaluation from first poll to the vote for the
two groups with the most incongruent information is statistically different from the other three
though not from each other. Likewise, the change in evaluations for the three groups with the least
incongruency, while different from the other two, are not statistically different from each other.
Table 1. Incongruency Encountered by Observed and Randomly
Assigned Groups
Target Observed Randomly Assigned
Mean SD N Mean SD N
Group 0 0 0.0 0.0 18 15.5 24.9 37
Group 1 10 20.2 18.3 51 24.2 24.4 37
Group 2 20 30.2 17.7 35 34.3 23.8 34
Group 3 40 40.5 19.3 43 39.3 18.2 43
Group 4 80 79.2 19.7 42 74.6 23.8 38
Note. Means in percentages of Incongruent Information.
579The Affective Tipping Point
55
60
65
70
75
Poll 1 Poll 2 Poll 3 Poll 4
MeanFTRating
Group 0
Group 1
Group 2
Group 3
Group 4
Figure 3. Evaluation of Preferred Candidate by Observed Incongruency Groups Over the Course of
the Campaign.
-12
-8
-4
0
4
84210
Groups
EvaluationChange
Random Groups
Observed Groups
Figure 4. Mean Evaluation Change from First Poll to the Vote.
580 Redlawsk et al.
curve presented in Figure 1.16
While the differences between the three groups with
the lower levels of incongruency are attenuated in the curve for the randomly
assigned groups, when we account for the information actually encountered
through our observed groups, the pattern is clearer. But in neither case does the
evaluation updating curve come close to approximating a normatively correct
process where evaluations consistently decline as more negative information is
encountered.
Calculating the Updating Curve and Tipping Point
The graphical presentation of the random and observed group data shows
clear differences in updating by those who encountered small amounts of incongruency
compared to those who encountered greater amounts, suggesting our first
hypothesis finds support. Low incongruency subjects appear to resist negatively
updating the evaluation of a preferred candidate while those encountering larger
amounts of incongruency update as we would expect. We now proceed to calculating
the nature of the evaluation updating curve and examining whether in fact a
tipping point can be identified. Note that Figure 4—change in evaluation over
time—suggests something more complex than the quadratic function described in
Figure 1.
Recall that our observed groups—designed to account for the actual amount
of incongruent information encountered by subjects—deviate from random
assignment, raising the serious question of whether the amount of incongruency
is endogenous to our experimental design and limiting the causal claims that we
can make. One way to address this problem is to use an instrumental variable
approach. Instead of using the randomly assigned groups or our artificially constructed
groups based on observed information search, we will employ a two-stage
least-squares regression with instrumental variables, the first stage of which will
use the randomly assigned groups as instruments for the observed incongruency
levels.
An initial fit of the data without any control variables (not shown) suggests
that a cubic function works better than a quadratic, which is not surprising given
the pattern shown in Figure 4. Our complete model allows for this cubic function
and controls for the initial rating of the liked candidate and amount of information
encountered about the candidate. We control for the initial rating to address ceiling
effects. Subjects who started with a very high rating of their preferred candidate
have less room to move up than those who started lower. We control for the amount
of information viewed because incongruency is measured in terms of the percentage
of information encountered. Subjects varied in how much information they
examined for their liked candidate; controlling for this is necessary to recognize
16
Note that the x-axis scale in Figure 4 describes the relative amounts of incongruency for each group
compared to the other groups,
581The Affective Tipping Point
that some subjects simply learned more information than did others. We use the
randomly assigned targets for information incongruency as instruments in the first
stage to predict the actual amount of incongruent information encountered.
Further, because one of our manipulations slowed half of the subjects down by
asking their affective response after every item examined (see Note 7) those
subjects systematically examined fewer items. Thus we employ a dummy for this
manipulation as an instrument to account for the actual variation in total items
examined for liked candidates.
Table 2 shows the results for the second-stage equation. The dependent variable
is change in the initially preferred candidate’s evaluation at the end of the
election compared to the first poll. The coefficients on the linear, quadratic, and
cubic terms for information incongruency are all significant and in the expected
directions. The coefficients on initial candidate evaluation and total number of
items examined for the liked candidate are not significant, though they are in the
expected direction. Importantly, the results of this analysis confirm the results of
our initial examination of the data using only the randomly assigned groups or our
observed groups. The curve resulting from the equation shows an initial increase
in the final evaluation for candidates when a small amount of incongruent information
is encountered. At some tipping point, the evaluations begin to decline and
updating proceeds in a more normatively correct direction, as more negative
information is encountered about the preferred candidate. For subjects encountering
relatively little incongruency, the predicted change in evaluations is positive
over the course of the campaign, that is, encountering small amounts of negative
Table 2. Two-Stage Least Squares Regression for Change in
Evaluation for Initially Preferred Candidate by Incongruency
Levels: Second Stage Regression Results
Change in Evaluation
B SE Sig.
Observed Incongruency Level
Linear1
1.207 .712 .092
Quadratic1
-.055 .026 .036
Cubic1
.0005 .0002 .033
Total Items Examined for Candidate1
.156 .509 .760
Initial Evaluation of Candidate -.064 .134 .635
Model significance: F-test = 2.668, p = .024.
Table entries are un-standardized coefficients and Standard Errors.
Significance reported are z-tests, two-tailed. N = 189.
1
Variable was estimated by instruments in first stage equation.
First stage instruments were 1) the randomly assigned target levels
for incongruency, the square of the target levels, and the cube of
the target levels and 2) a dummy variable indicating the presence
or absence of the immediate affect manipulation.
582 Redlawsk et al.
information does in fact result in a more positive evaluation. Likewise, at relatively
large amounts of incongruency, evaluations no longer continue to decline,
resulting in the cubic function as best fitting. This counterintuitive result most
likely occurs because at high levels of incongruency subjects turn away from the
initially preferred candidate and begin examining more information for other
candidates.
Where does the tipping point occur? That is, how much incongruent information
is necessary to force updating to begin to take “reality” into account?
Calculating this tipping point is simple, since it is the local maximum of the cubic
function within the bounds of 0–100% incongruency. This simple calculation
yields a rounded value of 13.4, suggesting that in these data once about 13% of
information about an initially preferred candidate is incongruent with a subject’s
own preferences, evaluations stop becoming more positive. But evaluations of
candidates do not actually become more negative than for an ideal candidate (no
incongruency) until about 28% incongruent information is encountered. Thus in
our data there is a range of incongruency (I), 0 < I < 28, between which evaluations
of an initially liked candidate are on average higher than for the ideal
candidate, that is, one who takes positions perfectly congruent with a subject’s
own preferences. But the exact tipping point itself is less important than the fact
that our results strongly support both motivated reasoning effects and accurate
updating, at different levels of incongruency.
Tipping Point Mechanisms: Motivated Reasoning and
Affective Intelligence
Having established the presence of a tipping point in candidate evaluation, we
turn now to an explanation for this observed behavior. Recall that we are working
with two different affective theories. Motivated reasoning suggests that evaluators
will become more positive in the face of negative information about a liked person,
as existing positive affect for that person interacts with negative affect for new
information, triggering a (cognitive) effort to make sense of the new information
while striving to maintain the existing (affective) feeling about the person. This
process may lead to attitude strengthening. Somewhat in opposition to this, affective
intelligence theory argues that a threatening environment generates increasing
anxiety, which operates to cause a person to learn more about the environment in
order to prepare a response. Thus increasing anxiety leads to learning, which
normatively should lead to more, rather than less, accurate updating of evaluations.
It is our contention that the updating pattern we have established can be explained
by the operation of both motivated reasoning (at low levels of threat; that is, small
amounts of incongruency) and affective intelligence at higher threat levels. We
turn now to examining whether the pattern we see can be accounted for by these
two affective processes.
583The Affective Tipping Point
Motivated Reasoning
It is well established that information that does not violate expectations is
more readily processed than information that does. Encountering unexpected
information piques our interest and forces us to concentrate more on it, compared
to information that simply confirms expectations. Multiple explanations for this
phenomenon have been advanced, including affective intelligence’s dual affective
systems (Marcus et al., 2000), Petty and Cacioppo’s (1981, 1983) central versus
peripheral routes, and hot cognition’s interaction of existing affect and new information
(Taber & Lodge, 2006; Redlawsk, 2002). Regardless of the mechanism
that might explain information processing differences, the first step here is to show
that such differences actually exist in our data. If they do not, then we can
immediately discount any process—including motivated reasoning and affective
intelligence—that would expect incongruency to take longer to process.
We examined the amount of time subjects took to read incongruent information
for their preferred candidate and compared it to the time they took to read congruent
information for the same candidate. Initially we simply compared mean processing
time for each type of information across our subjects. The results show that
processing time for incongruent information was more than 10% greater than for
congruent information (Mincongruent = 6.99, Mcongruent = 6.35 seconds, t = 2.187, 1797
df, p < .03.)17
However, different information items in the dynamic process-tracing
system may be of different lengths. Obviously the longer the item, the longer it will
take to read. So the analysis must also control for the number of words in the item
and the individual subject’s reading ability, measured as the time it took to read a set
of instructions. A regression analysis controlling for these factors is shown in
Table 3. The result is clear. Even after controls are applied, congruency affects
processing time. On average, incongruent items take more than half a second longer
to process, all else equal. To put this into perspective, we calculated the length of
time to read a 30-word item for someone at the mean reading speed. On average such
an item took 8.56 seconds to read if it was congruent, but 9.20 seconds to read if it
was incongruent, all else equal, an increase of 7.4% in processing time.
The candidate evaluation pattern supports our argument about the form of the
updating process, with an initial increase in evaluations in the face of a small
amount of negative information, followed by more accurate updating as incongruency
builds. And our processing time analysis suggests that incongruent information
takes longer to process, as we would expect. But are we seeing motivated
reasoning as such? And when subjects do reach a tipping point do increasing levels
of anxiety correspond to the downturn in evaluations as we would expect from
affective intelligence?
17
While we are not examining disliked candidates in this paper, we see the same effect of processing time
for those candidates as well. Congruent information (negative information about a disliked candidate)
is processed much faster (6.3 seconds compared to 7.1 seconds for incongruent). The t-test is
significant at p < .02. These results replicate the findings in Redlawsk (2002) with a different dataset.
584 Redlawsk et al.
As Hypothesis 2 suggests, we can look to the memories our voters report as
an indicator of the presence of motivated reasoning. Motivated reasoners attempt
to maintain their affective evaluations in the face of unexpected information. One
way in which they do this is to use bolstering—to recall into active memory factors
which support the existing evaluation and which may then overwhelm the new
incongruent information.18
If this happens, memory for these attitude-supporting
attributes may be enhanced, as repeated access to a concept increases the likelihood
that the concept will be remembered (Fiske & Taylor, 1991). We can examine
the likelihood that our subjects recalled positive (enthusiastic) memories as a
function of the amount of incongruency encountered. If we find that small amounts
of negative information increase positive memory, we will have evidence of a
motivated reasoning process.
Figure 5 plots the mean number of reported memories for each of the instrumentally
defined incongruency groups. The results almost perfectly support the
expectations of Hypothesis 2. Subjects who saw a small amount of negative
information about their most liked candidate (Group 1) report more positive
memories and more overall memory than any other group, including those who
encountered the least amount of incongruent information (Group 0). Further, those
in Group 2 report about the same number of positive memories as Group 0. It is not
until levels of incongruency beyond the tipping point that positive memories for
the initially liked candidate begin to decline. Thus we have evidence of a motivated
reasoning process that operates as would be expected if bolstering were under way
at low levels of incongruency.
18
Let us be clear about this process. We do not suggest that this is cognitively driven at the start.
Bolstering itself is a by-product of the associative nature of memory (Anderson, 1983) and hot
cognition (Lodge & Taber, 2005). In activating the construct stored for the candidate (and the affect
associated with it) other connected memory nodes are also activated. Since we are dealing with
positively evaluated candidates, most of these associations are also positive. As memory nodes are
activated they are more likely to be recalled when later tested. The net result is that more positive
memory traces should be evident when negative information stimulates processing since incongruent
information is processed more carefully than congruent information (Redlawsk, 2002).
Table 3. Processing Time for Congruent and Incongruent
Information Preferred Candidate
Predictor B SE
Item Incongruency (1 = Congruent) -.637*** .228
# Words in Item .154*** .006
Reading Speed -.401*** .012
Constant 4.257*** .343
Adj r2
.487
*p < .1, **p < .05, ***p < .01
Table entries are un-standardized OLS coefficients and Standard
Errors. N = 1587 Items.
585The Affective Tipping Point
Affective Intelligence
What about once evaluations begin to adjust downward in the face of incongruent
information? What causes this tipping point? Hypotheses 3a and 3b test the
claim of affective intelligence that more anxious voters are better voters (Marcus
et al., 2000). If negative affect increases as incongruency increases, it may well be
a mechanism that leads to the more accurate updating of evaluations past the
tipping point. Recall that affective intelligence argues in favor of two emotional
subsystems, one of which is characterized by positive affect and the other by
negative. This negative affective subsystem—the behavioral inhibition system
(BIS)—is responsible for focusing attention on negative stimuli so as to avoid (or
address) potentially dangerous situations. Our expectation then is that as the
information environment for our subjects becomes increasingly at odds with
expectations about a liked candidate, feelings of negative affect—a sense of
anxiety about the original evaluation—will grow.
In order to assess whether affective intelligence processes are at work, we
have three measures of the impact of the primary election on our voters. The
first of these measures is derived from asking subjects after they had voted to
indicate how difficult it was to make a choice between the candidates. The
second comes from asking how confident subjects were that they had chosen the
“right” candidate. Both were asked on a simple scale from 1 to 5 coded so 1 was
0
1
2
3
4
5
6
43210
Groups
Total Memory
Positive Memory
Negative Memory
Figure 5. Incongruency and Memory for a Preferred Candidate.
586 Redlawsk et al.
“Very Easy” or “Not at all Confident” and 5 was “Very Difficult” or “Very
Confident.”
Figure 6 presents a summary of both the difficulty and confidence measures
by incongruency group. Both point in the same direction and support Hypothesis
3a. As levels of incongruent information increase so does reported difficulty while
confidence decreases, but only up to a point. Subjects encountering incongruency
about their most liked candidate that extends beyond the tipping point actually
report an easier decision and greater confidence at the conclusion of the campaign
than those at lower levels of incongruency. This occurs because once past the
evaluative tipping point subjects make peace with the fact that their initially most
liked candidate was just not what they thought he was and they moved on to
another option. In fact, voters in the highest incongruency group were less than
20% likely to vote for the candidate they initially preferred, well below all other
groups.
Our third and most direct measure of the impact of the information environment
on subjects comes from the administration of Watson, Clark, and Tellegen’s
(1988) Positive and Negative Affect Scale (PANAS), which allows us to assess
subjects’ affective states. This measure is explicitly aimed at assessing how a
subject feels at the time the scale is administered. A brief questionnaire was
administered at two points in time—before the experiment began, and again
immediately before the cued recall process. We can compare the level of negative
2
2.5
3
3.5
4
43210
Groups
Difficulty
Confidence
Figure 6. Reported Difficulty and Confidence in Primary Vote Decision by Incongruency Groups.
587The Affective Tipping Point
affect expressed by our subjects after they voted to the level they expressed before
they started and examine the extent to which change in negative affect varies by the
amount of incongruency subjects encountered. We specifically use the PANAS
Negative Affect scale (NA), which consists of ten affect words.19
For each, subjects
were asked to indicate how strongly the affect word described their current
feelings at the time of administration on a 1–5 scale, with 1 indicating that the
word did not describe them at all and 5 indicating it described their current feelings
very strongly. The NA scale in our data has a Cronbach’s Alpha of .884, indicating
high reliability. Previous studies have shown the NA scale itself to be a good
measure of anxious states (Crawford & Henry, 2004; Mehrabian, 1997), and it is
widely used in social psychology. We compare the postexperiment NA score with
the preexperiment NA score and examine the net difference by the instrumentally
defined levels of incongruency in the information environment. Results are in
Figure 7.
Just as with reported difficulty and confidence, this more direct measure of
anxiety responds as we predict in Hypothesis 3b. The group of subjects who
encountered the least incongruency (Group 0) evidences no change at all in their
negative affective state post experiment compared to preexperiment. Subjects
encountering greater amounts of incongruent information about their liked candidate
report a more negative affective state immediately following the election. As
19
These ten words are: distressed, upset, guilty, scared, hostile, irritable, ashamed, nervous, jittery, and
afraid.
0
0.1
0.2
0.3
0.4
43210
Groups
Figure 7. Change in Negative Affect at End of Campaign by Levels of Incongruency.
588 Redlawsk et al.
levels of incongruency climb so too does negative affect. But at the highest levels
of incongruency, the increase in negative affect, while still greater than those
encountering no incongruency, drops substantially. Once past the affective tipping
point subjects become less anxious than they do immediately before it, though the
effect is not instantaneous. Given that subjects at the tipping point are significantly
more positive about their initially liked candidate—despite his growing
negatives—it takes more negative information to bring the rating down below the
initial point before attitude strengthening began. Thus too it appears to take time
for the sense of anxiety evidenced to dissipate.
As negative affect reached high enough levels, our subjects began to reassess
their initial evaluations. While we cannot with these data establish without question
the causal link we would like, we do see patterns we would expect if affective
intelligence was at work. As more negative information about a liked candidate is
encountered, the information environment becomes more threatening, leading to
increasing negative feelings and a sense that the decision is more difficult. This
increasing challenge is resolved at the tipping point, as subjects stop the process
that leads to attitude strengthening and instead begin a process leading to accurate
updating.
Discussion
Motivated reasoners strive to maintain existing evaluative affect, even in the
face of countervailing information. This effect has been well established in the
literature (Kunda, 1990) and replicated here. Thus where we would expect
“rational” voters to approximate normatively correct updating, motivated reasoners
show evidence of attitude strengthening, becoming even more positive about
a liked candidate in the face of negative information about that candidate. And
while the voters in our experiment show heightened negative affect and evidence
that the choice becomes more difficult as incongruency grows, they also show
attitude strengthening effects as motivated reasoning predicts. But we also show
clear evidence that these effects do not necessary continue under all circumstances.
At some point our voters appear to wise up, recognize that they are
possibly wrong, and begin making adjustments. In short, they begin to act as
rational updating processes would require. Why? Our results are consistent with
the idea that as anxiety increases (leading to more difficulty in the decision and
less confidence) voters pay closer attention to the environment (and processing
time increases). They then begin to more carefully consider new information and
potentially override existing affective expectations. Such a process would be
consistent with affective intelligence overriding motivated reasoning.
It is worth keeping in mind the limitations of this study which represents only
a first attempt to show that motivated reasoning does not continue ad infinitum and
to propose a mechanism—affective intelligence—that might interfere with it. The
study is experimental, the environment in which our voters operated, while having
589The Affective Tipping Point
many of the key features of a real world election environment, certainly was not
a real campaign. Experimental control meant our candidates were made up, and
while our subjects told us the candidates seemed quite real, no subject knew
anything about them before the study. Further, we limited ourselves to a primary
election, thus negating any effects of partisan identification. We would expect, for
example, that strong partisans in a general election would have a very high tipping
point, at least compared to non-partisans. Likewise, ideologues, as opposed to
moderates, might also be harder to move off their initial support for a candidate
very close to them. We do not test either of these possibilities here. Yet while our
environment is not a real election, we believe that the psychological processes our
subjects engage in their attempts to learn about and evaluate candidates are really
no different whether in the laboratory or in the midst of a “real” campaign. Voters
learn about candidates, assess whether the new information fits with their expectations,
and in some fashion ultimately create and update evaluations in both
environments.
We argue that the role of affect in candidate evaluation is not a matter of
either/or. Voters are not either motivated reasoners or rational processors. Instead,
as our experiment demonstrates, voters can be both depending on the information
environment in which they are operating. When the amount of incongruency is
relatively small, the heightened negative affect does not necessarily override the
motivation to maintain support for a candidate in which the voter is already
positively invested. But as V. O. Key (1966) noted four decades ago, voters are not
always fools. An affective tipping point exists at which existing positive evaluations
give way to a newly understood reality—the candidate is just not what he or
she seemed to be at first.
If voters are, in fact, somewhat immune to small amounts of negative information
about their favored candidates, what are the implications in the real political
world? Should candidates not worry about minor mess-ups, flip-flops, and
fleeting scandals, or is the atmosphere of the modern campaign already so negative
that most voters are pushed way past the tipping point months before Election
Day? A closer look at the modern campaigning environment might help answer
that question, but the fact remains that for a while at least, a candidate’s early
supporters will probably resist attempts to change their minds. Candidates who
need to win new voters without alienating their bases should be able to lean to the
middle, as long as they don’t lean too far. However, in a real campaign where prior
beliefs about candidates are long standing and based on much more information
than subjects in our study were exposed to early on, reaching the affective tipping
point will require substantially more negative information than our subjects
encountered. It is easy to imagine a long-time fan of a presidential candidate
rejecting virtually all new negative information about him or her and sticking to an
early evaluation. Yet even such a fan might, in the face of overwhelming information
counter to expectations, awake to the changed reality and revise her beliefs
accordingly.
590 Redlawsk et al.
ACKNOWLEDGMENTS
We acknowledge the support of the Social Science Funding Program at the
University of Iowa and thank the group of Iowa graduate and undergraduate
research assistants who helped us carry out this labor-intensive study: Molly
Berkery, Kimberly Briskey, Brian Disarro, Paul Heppner, Jason Humphrey, Jason
Jaffe, Jeff Nieman, Matt Opad, Anthony Ranniger, and Christian Urrutia. Correspondence
concerning this article should be sent to David P. Redlawsk, Department
of Political Science, Rutgers University, New Brunswick, NJ 08901. E-mail:
redlawsk@rutgers.edu
REFERENCES
Abelson, R. P. (1963). Computer simulation of “hot cognitions.” In S. Tomkins & S. Messick (Eds.),
Computer simulation and personality: Frontier of psychological theory (pp. 277–298). New
York: Wiley.
Anderson, J. R. (1983). The architecture of cognition. Cambridge, MA: Harvard University Press.
Brader, T. (2005). Striking a responsive chord: How political ads motivate and persuade voters by
appealing to emotions. American Journal of Political Science, 49(2), 388–405.
Civettini, A. J. W., & Redlawsk, D. P. (2005). A feeling person’s game: Affect and voter information
processing and learning in a campaign. Paper presented at the Annual Meeting of the American
Political Science Association, Washington, DC.
Civettini, A. J. W., & Redlawsk, D. P. (2009). Voters, emotions, and memory. Political Psychology,
30, 125–151.
Crawford, J. R., & Henry, J. D. (2004). The positive and negative affect schedule (PANAS): Construct
validity, measurement properties and normative data in a large non-clinical sample. British
Journal of Clinical Psychology, 43(Pt. 3), 245–265.
Damasio, A. R. (1999). The feeling of what happens: Body and emotion in the making of consciousness.
New York: Harcourt.
Edwards, K., & Smith, E. (1996). A disconfirmation bias in the evaluation of arguments. Journal of
Personality and Social Psychology, 7(1), 5–24.
Fiske, S. T., & Taylor, S. E. (1991). Social cognition (2nd edn.). New York: McGraw Hill.
Festinger, L. (1957). A theory of cognitive dissonance. Stanford, CA: Stanford University.
Ford, J. K., Schmitt, N., Schechtman, S. L., Hults, B. M., & Doherty, M. L. (1989). Process tracing
methods: Contributions, problems, and neglected research questions. Organizational Behavior
and Human Decision Processes, 43, 75–117.
Gill, J. (2007). Bayesian methods: A social and behavioral sciences approach (2nd
Ed). Boca Raton,
FL: CRC Press.
Gladwell, M. (2000). The tipping point: How little things can make a big difference. New York: Little,
Brown, and Company.
Green, D., & Gerber, A. (1999). Misperceptions about perceptual bias. Annual Review of Political
Science, 2, 189–210.
Hastie, R., & Park, B. (1986). The relationship between memory and judgment depends on whether the
task is memory-based or on-line. Psychological Review, 93, 258–268.
Heider, F. (1958). The psychology of interpersonal relations. New York: Wiley.
Herstein, J. A. (1981). Keeping the voter’s limits in mind: A cognitive process analysis of decision
making in voting. Journal of Personality and Social Psychology, 40, 843–861.
591The Affective Tipping Point
Holbrook, R. A. (2005). Candidates for elected office and the causes of political anxiety: Differentiating
between threat and novelty in candidate messages. Paper presented at the annual meeting of
the Midwest Political Science Association.
Holbrook, A. L., Krosnick, J. A., Visser, P. S., Gardner, W. L., & Cacioppo, J. T. (2001). Attitudes
toward presidential candidates and political parties: Initial optimism, inertial first impressions,
and a focus on flaws. American Journal of Political Science, 45(4), 930–950.
Huang, L., & Price, V. (1998). Motivations, information search, and memory structure about political
candidates. Paper presented at ICA ‘98.
Huang, L. (2000). Examining candidate information search processes: The impact of processing goals
and sophistication. Journal of Communication, 50(Winter), 93–114.
Isen, A. M. (2000). Positive affect and decision making. In M. Lewis & J. Haviland-Jones (Eds.).
Handbook of emotions (2nd
ed., pp. 417–435). New York: Guilford.
Jacoby, J., Jaccard, J., Kuss, A., Troutman, T., & Mazursky, D. (1987). New directions in behavioral
process research: Implications for social psychology. Journal of Experimental Social Psychology,
23, 146–175.
Key, V. O., Jr. (1966). The responsible electorate. Cambridge, MA: Belknap Press.
Kunda, Z. (1990). The case for motivated political reasoning. Psychological Bulletin, 108(3), 480–498.
Lau, R. R. (1995). Information search during an election campaign: Introducing a process tracing
methodology for political scientists. In M. Lodge & K. McGraw (Eds.), Political judgment:
Structure and process (pp. 179–206). Ann Arbor: University of Michigan Press.
Lau, R. R., Anderson, D., & Redlawsk, D. P. (2008). An exploration of correct voting in recent U.S.
presidential elections. American Journal of Political Science 52(2), 395–411.
Lau, R. R., & Redlawsk, D. P. (1997). Voting correctly. American Political Science Review, 91(3),
585-598.
Lau, R. R., & Redlawsk, D. P. (2001). An experimental study of information search, memory,
and decision making during a political campaign. In J. Kuklinski (Ed.), Citizens and
politics: Perspective from political psychology (pp. 136–159). New York: Cambridge University
Press.
Lau, R. R., & Redlawsk, D. P. (2006). How voters decide: Information processing in an election
campaign. New York: Cambridge University Press.
Lodge, M., & Taber, C. S. (2000). Three steps toward a theory of motivated political reasoning. In
A. Lupia, M. McCubbins, & S. Popkin (Eds.), Elements of reason: Cognition, choice, and the
bounds of rationality (pp. 182–213). London: Cambridge University Press.
Lodge, M., & Taber, C. S. (2005). The primacy of affect for political candidates, groups, and issues: An
experimental test of the hot cognition hypothesis. Political Psychology, 26, 455–482.
Lodge, M., McGraw, K., & Stroh, P. (1989). An impression-driven model of candidate evaluation.
American Political Science Review, 83, 399–419.
Lodge, M., Steenbergen, M., & Brau, S. (1995). The responsive voter: Campaign information and the
dynamics of candidate evaluation. American Political Science Review, 89, 309–326.
Lord, C. G., Ross, L., & Lepper, M. (1979). Biased assimilation and attitude polarization: The effects
of prior theories on subsequently considered evidence. Journal of Personality and Social Psychology,
37(11), 2098–2109.
Marcus, G. E., & MacKuen, M. (1993). Anxiety, enthusiasm, and the vote: The emotional underpinnings
of learning and involvement during presidential campaigns. American Political Science
Review, 87(September), 672–685.
Marcus, G. E., Newman, W. R., & MacKuen, M. (2000). Affective intelligence and political judgment.
Chicago: University of Chicago Press.
Mehrabian, A. (1997). Comparison of the PAD and PANAS as models for describing emotions and for
differentiating anxiety from depression. Journal of Psychopathology and Behavioral Assessment,
19(4), 331–357.
592 Redlawsk et al.
Petty, R. E., & Cacioppo, J. T. (1981). Attitudes and persuasion: Classic and contemporary
approaches. Dubuque, IA: William C. Brown.
Petty, R. E., & Cacioppo, J. T. (1983). Central and peripheral routes to persuasion: Application to
advertising. In L. Percy & A. Woodside (Eds.), Advertising and consumer psychology (pp. 3–23).
Lexington, MA: Lexington Books.
Redlawsk, D. P. (2001). You must remember this: A test of the on-line voting model. Journal of
Politics, 63, 29–58.
Redlawsk, D. P. (2002). Hot cognition or cool consideration? Testing the effects of motivated reasoning
on political decision making. Journal of Politics, 64, 1021–1044.
Redlawsk, D. P. (2006). Motivated reasoning, affect, and the role of memory in voter decision-making.
In D. P. Redlawsk (Ed.), Feeling politics: Emotion in information processing (pp. 87–108). New
York: Palgrave-Macmillan.
Redlawsk, D. P. (2007). Understanding vs. prediction in candidate evaluation. Paper presented at the
annual meeting of the Midwest Political Science Association and at the annual meeting of the
International Society of Political Psychology, Portland, OR.
Redlawsk, D. P., Civettini, A. J., & Lau, R. R. (2007). Affective intelligence and voting information
processing and learning in a campaign. In A. Crigler, M. MacKuen, G. E. Marcus, & W. R.
Neuman (Eds.), The affect effect: Dynamics of emotion in political thinking and behavior.
Chicago: University of Chicago Press.
Riggle, E. D. B., & Johnson, M. M. S. (1996). Age differences in political decision making: Strategies
for evaluating political candidates. Political Behavior, 18(1), 99–118.
Steenbergen, M. (2001). The Reverend Bayes meets J.Q. Public: Patterns of political belief updating in
citizens. Paper presented at the annual meeting of the International Society of Political Psychology.
Cuernavaca, Mexico.
Taber, C. S., & Lodge, M. (2006). Motivated skepticism in the evaluation of political beliefs. American
Journal of Political Science, 50(3), 755–769
Watson, D., Clark, L. A., & Tellegen, A. (1988). Development and validation of brief measures of
positive and negative affect. Journal of Personality and Social Psychology, 54(6), 1063–1070.
Zajonc, R. B. (1984). On the primacy of affect. American Psychologist, 39(2), 117–123.
593The Affective Tipping Point