APPARENT BIAS: WHAT DOES ATTITUDE POLARIZATION SHOW?

• Author: Juan Dubra,Jean‐Pierre Benoît
• Date: 01 November 2019
• Published date: 01 November 2019
• DOI: http://doi.org/10.1111/iere.12400

INTERNATIONAL ECONOMIC REVIEW

Vol. 60, No. 4, November 2019 DOI: 10.1111/iere.12400

APPARENT BIAS: WHAT DOES ATTITUDE POLARIZATION SHOW?∗

BYJEAN-PIERRE BENOˆ

IT AND JUAN DUBRA1

London Business School, U.K.; Universidad de Montevideo, Uruguay

Many, though not all, experiments have found that exposing groups of subjects who disagree to the same

evidence may cause their initial attitudes to strengthen and move further apart, or polarize. Some have concluded

that findings of attitude polarization show that people process information in a biased manner so as to support

their initial views. We argue that, on the contrary, polarization is often what we should expect to find in an

unbiased Bayesian population, in the context of experiments that find polarization.

1. INTRODUCTION

In a classic study, Lord et al. (1979) took two groups of subjects, one of which believed in

the deterrent effect of the death penalty and one of which doubted it, and presented them with

the same mixed set of studies on the issue. Both groups became more convinced of their initial

positions. Numerous, though by no means all, subsequent experiments have also found that

exposing groups of subjects who disagree to the same mixed evidence may cause their initial

attitudes to strengthen and move further apart, or polarize.2Many scholars have concluded that

this polarization shows that people process information in a biased manner, so as to support

their preexisting views. We argue that, on the contrary, polarization is often what we should

expect to find in an unbiased Bayesian population in the context of the experiments that find

polarization.3

There are two aspects to attitude polarization, which we term pairwise polarization and

population polarization. Pairwise polarization occurs when the opinions of a particular pair of

individuals are reinforced and move further apart as a result of a common piece of information.

Population polarization occurs when this divergence is systematic, so that the opinions of the

population on the whole diverge.

The economics literature has taken the view that there is, on the face of it, something puz-

zling about pairwise polarization4and has examined the extent to which pairwise polarization

is consistent with Bayesian updating. This is a crucial step in the understanding of attitude

polarization. At the same time, this literature has not made a clear distinction between pairwise

and population polarization, yet it is population polarization that experimental studies focus

∗Manuscript received July 2017; revised February 2019.

1We thank Juan Pedro Gambetta, Gabriel Illanes, and Oleg Rubanov for outstanding research assistance. We also

thank Lewis Kornhauser, Vijay Krishna, David Levine, Michael Mandler, Frederic Malherbe, Wolfgang Pesendorfer,

Madan Pillutla, Debraj Ray, Jana Rodriguez-Hertz, Andrew Scott, and Stefan Thau for valuable comments. This is a

substantially revised version of a paper that we circulated previously as “A Theory of Rational Attitude Polarization.”

Please address correspondence to: Juan Dubra, Universidad de Montevideo, Prudencio de Pena 2440, Montevideo

11600, Uruguay. E-mail: dubraj@um.edu.uy.

2Experiments on attitude polarization include Darley and Gross (1983), Plous (1991), Miller et al. (1993), Kuhn and

Lao (1996), and Munro and Ditto (1997).

3Although we develop our ideas in a common priors rational setting, our reasoning is not restricted to rational agents.

Rather, full rationality provides a convenient benchmark of unbiased reasoning. Thus, our theory can also be applied

to unbiased subjects who, say, are guilty of base rate neglect.

4For instance, Andreoni and Mylovanov (2012) write, “How can two people see the same information and draw

opposite conclusions?” See Section 3 for a discussion of the economics literature.

1675

C

(2019) by the Economics Department of the University of Pennsylvania and the Osaka University Institute of Social

and Economic Research Association

1676 BENOˆ

IT AND DUBRA

on as being problematic.5Indeed, most attitude polarization experiments explicitly craft the

evidence given to subjects to be sufficiently ambiguous that it can, legitimately, simultaneously

have a positive impact on some individuals and a negative impact on others. Darley and Gross

(1983) include a control treatment that verifies that the evidence it uses induces both positive

and negative beliefs in (different) unbiased subjects. Thus, it is only population polarization

that these experiments cite as evidence of bias.

A clear illustration of the issues involved is provided by Plous’s (1991) nuclear deterrence

study. Plous begins by dividing his subjects into two groups, according to whether they entered

the experiment with a belief that a strategy of nuclear deterrence makes the United States safer

or less safe. He then gives all subjects the same article to read, describing an actual incident

where an erroneous alert caused the United States to enter a heightened state of readiness

for nuclear war with the Soviet Union. The crisis lasted only three minutes, as officials quickly

realized the alert was a false alarm. After reading the article, the beliefs of subjects in each

group move further in the direction of their initial inclinations.

How should unbiased subjects react to the article? As Plous writes, “Given the fact that (a)

the system malfunctioned and (b) the United States did not go to war despite the malfunction,

the question naturally arises as to whether this breakdown indicates that we are safer or less

safe than previously assumed.”

The evidence in the article is equivocal—its implications depend on beliefs about an an-

cillary consideration, to wit, whether it is more important for a system’s safety that it has a

well-functioning primary unit or that it has effective safeguards. It is not clear which one is

more important, and a person could legitimately believe either one is, depending upon his or

her previous information on the matter. A fortiori, the fact that the opinions of two partic-

ular subjects polarize after being given evidence of a rectified malfunction—an opponent of

nuclear deterrence becomes more opposed, whereas a proponent becomes more in favor—is

unproblematic.

However, even if people can legitimately update in different directions, a challenge remains.

Why would it be that, systematically, subjects who are in favor of nuclear deterrence respond

positively to the evidence, whereas those who are opposed respond negatively? Put differently,

why would it be that people who believe in the safety of nuclear deterrence also believe

that safeguards are paramount, whereas people who are skeptical of nuclear deterrence also

believe that primary units are crucial, instead of beliefs in these two dimensions being, say,

uncorrelated? If these beliefs were uncorrelated, although there would be many instances of

pairwise polarization, there would be just as many instances of pairs converging; overall these

instances would cancel each other out and the population would not polarize. It is the fact that

the population polarizes, not just isolated pairs, and that leads Plous to conclude that people

process information in a biased manner to support their initial beliefs.

Is the conclusion of biased reasoning warranted? We now argue that it is not.

Plous tells us that most of the subjects in his experiment knew of the false alarm incident before

entering the experiment6, though, presumably, they did not know all of the details provided

in the article. Suppose that the subjects that arrive with a favorable opinion, despite their

knowledge of a previous malfunction that was caught, are the ones that consider the reliability

of safeguards to be more important than the reliability of the primary unit—hence, their initial

favorable view.7These subjects would naturally tend to increase their beliefs that nuclear

deterrence is safe after being given further descriptions of properly functioning safeguards. On

the other hand, subjects that consider a malfunction of the primary unit to be telling would have

a negative view initially and would tend to revise downward after being given further evidence

5Results purporting to demonstrate bias are reported in group terms. For instance, Lord et al. (1979) contrasts

changes in the mean attitude of pro-deterrence subjects with changes in the mean attitude of anti-deterrence subjects.

6In a variant treatment, which also yields population polarization, subjects are instead provided with a description

of a near-miss incident that is unfamiliar to them.

7In line with this supposition, the comments of pronuclear subjects tend to focus on safeguards, whereas those of

antinuclear subjects focus on breakdowns.