biased processing and opinion polarisation

y+ "

Biased processing and opinion polarisation experimental re!nement of argument communication

theory in the context of the energy debate

Opinion Dynamics and Cultural Con!ict

in European Spaces

H2020 – FETPROACT-2016 - 732942

www.odycceus.eu

Sven Banisch (MPI Leipzig) and Hawal Shamon (FZ Jülich) y+ "y+ "y+ "

!"#$#%$&'($)*#+,!"#!$!%&'(!)*$)!(&+&',-#!)*&,.&)"/$'!0,(&'#!,1!/,''&/)"+&!,-"2",2!-.,/&##&#!),!32(&.#)$2(!)*&!0&/*$2"#0#!4&*"2(!)*&!&0&.5&2/&!,1!/,2#&2#3#6!

-,'$."#$)",2!$2(!/,27"/)8

Opinion Dynamics

The Puzzle of Polarisation

! How does a population with moderate initial opinions diverge into groups of agents that strongly support opposing views?

– + – +

?

– + – +?

Argument Communication Theory! Models of collective deliberation by social exchange of pro and

con arguments regarding a certain issue (attitude object)

• Agents communicate arguments and update them in memory

• Attitudes/opinions as number of pro versus con arguments are updated after peer exposure to an argument

• Homophily: Opinion similarity drives interaction

} }

1

1 100 1

00 0

attitude

arguments

evaluativeassociations

from weak to strong

homophily

Argument Communication Theory! ACT provides a promising setting. Research within the frame-

work is advancing towards applied opinion dynamics!! Applications:

! Limitations:

• What is the role of cultural diversity and demogra-phic faultlines in group decision and team processes? (Michael Mäs and others)

• How to opinions across multiple related topics align along ideological lines? (our JASSS paper)

• Lack of a realistic model of argument processing.• Selective information processing is a robust cogni-

tive mechanism which must be included to arrive at better models of group deliberation!

Argument Communication TheoryOnline Demos

www.universecity.de/index.php?site=demos

Overview

! Experiment

! Cognitive model

! Expected attitude changes (under »computerized« experiment)

! How good is the model given the data?

! Collective deliberation with biased processing

! Impact of homophily and selective exposure

The Psychological Puzzle• Humans process information in a biased way:

attitude-congruent arguments are favoured over incongruent ones!

• A lot of experimental research has been invested on the question whether biased processing im-plies attitude polarisation when subjects are exposed to con!icting arguments.

• Attitude polarisation refers to an intra-individual e!ect of becoming more extreme after exposure

• Empirical evidence is mixed!

– +

– +

attitude moderationversus

attitude polarisation

individuals(attitude change)

(argument persuasion experiments)

Experiment! Hawal Shamon carefully designed an experi-

ment in which subjects received a balanced set of pro and con arguments regarding six tech-nologies for energy production.

Experiment! Attitudes measured before

and after exposure to 7 pro and 7 con arguments

! Attitude change data on 6 di!erent energy sources

Cognitive Model of Argument Processing! Attitude structure borrowed from ACT

attitudeon technology

(coal, gas, onshore, o!shore, photovoltaic, biomass)

proarguments

contraarguments

} }

3

0 101 0

11 1

} }

4

1 011 1

00 0

} }

1

1 100 1

00 0

} }

0

1 000 1

10 1

} }

-4 -3 -2 -1 0 1 2 3 4

• Positive evaluation of an argument (a) which is coherent with the opinion (o)

• Attitude strength (|o|) determines the size of this e!ect

• Evaluation V(a) becomes a linear functi-on of the initial opinion

• Matches with experimental data on evaluation bias for a pro argument

Cognitive Model of Biased Processing! Argument evaluation aligned to experimental data and for-

malised in terms of cognitive coherence

Cognitive Model of Argument Processing! Argument adoption is assumed to depend on the evaluation

V(a) and the strength of biased processing ! by

• We are not rational optimizers of cognitive coherence

• The free parameter ! accounts for the strength of biased processing

• ! = 0 means unbiased adoption and if ! is large only coherent arguments are adopted

!

Expected Attitude Change (virtual experiment)

! How would arti!cial agents react to the same experimental treatment? What is their expected attitude change?

! M = 4 pro and con arguments (to match the 9-point attitude scale)

! How many new argu-ments are adopted?

! What is the respective attitude change?

} }1

1 100 1

00 0

1 11 1

1 11 1

4 proarguments

4 conarguments } }

?

1 1?? 1

?? ?

computational implementation (of balanced argument treatment)

T" T#


! probability of adopting k pro-arguments

! probability of adopting k con-arguments

! probability of attitude change k

! expected attitude change

! assuming symmetric distribution over npro and ncon


! Attitude moderation if bias ! = 0. Negative opi-nions become more posi-tive and vice versa.

! This is what current models assume!

! Attitude polarisation if ! is large. Negative opinions become more negative and vice versa.

The psychological puzzle: attitude moderation versus attitude polarisation

– +– +

! Bifurcation from attitude moderation to attitude polarisation if ! crosses a critical value of 1/2

! Biased processing leads to attitude polarisation only if the bias is strong

! This strength may vary across topics that experi-ments have addressed

0 0.2 0.4 0.6 0.8 1 1.2

biased processing (!))

1.2

1.4

1.6

1.8

2

2.2

2.4

2.6

2.8

3

mea

n sq

uare

erro

r

Calibration with Experimental Data" Assuming that agents adapt by argument exchange (as in ACT

models), what is the adequate choice for the processing bias !?

" Least square error of theoretical prediction of mean attitude ch-anges to the data (N = 1078)

" Best match for ! close to the critical point 1/2!

! Clear signature that the re!ned model with moderate biased processing improves ACT

previousmodels

calibratedmodels

0 0.2 0.4 0.6 0.8 1 1.2

biased processing (!))

1.2

1.4

1.6

1.8

2

2.2

2.4

2.6

2.8

3

mea

n sq

uare

erro

r

Calibration with Experimental Data" Assuming that agents adapt by argument exchange (as in ACT

models), what is the adequate choice for the processing bias !?

" Least square error of theoretical prediction of mean attitude ch-anges to the data (N = 1078)

" Best match for ! close to the critical point 1/2!

! Clear signature that the re!ned model with moderate biased processing improves ACT

previousmodels

calibratedmodels

0 0.2 0.4 0.6 0.8 1 1.20.5

1

1.5

2

2.5

3

3.5

biased processing (!)

mea

n sq

uare

err

or

negative initial opinion

positive initial opinion

all initial opinions

consistent results on bothsides of the attitude scale

Di!erences across topics! Least square error of theo-

retical prediction of mean attitude changes to the data (N ! 170 per topic)

! Indication that bias is weak (< 0.5) for new issues (biomass, gas) and stron-ger for long-debated issues (> 0.5)

! Attitude moderation in the former cases, attitude pola-risation in the latter ones0 0.2 0.4 0.6 0.8 1 1.2

biased processing (")

0.5

1

1.5

2

2.5

3

3.5

mea

n sq

uare

err

or

gas biomass

coal onshore

o#shore

solar alltechnologies

CollectiveDynamics

– +

– +

– +

moderateconsensus

versus

choiceshift

versus

collectivepolarisation

!!!"#$%&'()'*%+,"-'

.-'/0).10-'234520('("6+/2

Collective Deliberation with Processing Bias! Argument communication models describe processes of collecti-

ve attitude formation as repeated social exchange of arguments

!"#$#%&'()*#+,-#./#0.1%2324)#51)%36#26#752*)8#53#*5%8.'"

9"#:;)#6)%8)*#5*320&<53)6#5#*5%8.'#5*1&')%3#3.#5#*)0)24)*"

="#:;)#*)0)24)*#58.736#3;26#5*1&')%3#>23;#5#7*.(5(2<2?3@#7#5%8#&7853)6#236#53323&?8)#500.*82%1<@"

} }

4

1 011 1

00 0

} }

1

1 100 1

00 0

&"'("! !"#"$%"!

A3;26#26#%.3#3*&)B+5C#D#E-

} }

1

1 100 1

00 0

} }

2

1 000 1

00 0

)45<&532.%F#G+5C-#D#!58.732.%F#7H+G-"!

6.025<#2%3)*5032.% 0.1%2324)#7*.0)662%1

Collective Deliberation with Processing Bias! Argument communication models describe processes of collecti-

ve attitude formation as repeated social exchange of arguments

} }

4

1 011 1

00 0

} }

1

1 100 1

00 0

!"#$"% %"&"'("%

!"#$%&$%&'("&")*+,-./&0&12

} }

1

1 100 1

00 0

} }

2

1 000 1

00 0

+3.4*."$('5&6-./2&0&7.8(9"$('5&9:-62"!

%(;$.4&$'"+).;"$(' ;(<'$"$3+&9)(;+%%$'<

ModelPhenomenology

– +

– +

– +

consensusformation

versus

choiceshift

versus

collectivepolarisation

First Transition! The introduction of weak biased processing speeds up group

decision processes by two orders of magnitude• Groups without processing bias may remain in indecision for a long time

5000 10000 15000 20000 25000 30000 35000 40000

43210

-1-2-3-4

attit

ude

200 400 600 800 1000 1200 1400 1600 1800 2000

43210

-1-2-3-4

attit

ude

dive

rsity

(sta

ndar

d de

viat

ion)

dive

rsity

(sta

ndar

d de

viat

ion)

IIIbI IV

IIIa IV

from no to

weak biased

processing

• Reinterpretation of „extreme con-sensus“ as an e!ective group decision process

• Evolutionary origins of biased processing!?

Second Transition! Strong biased processing may

lead to persistent intra-group con!ict• Biased processing alone is suf-

!cient for collective bi-polari-sation (without any assumptions about social composition)

• Only in the regime of attitude polarisation at individual level

• Strength of processing bias is what matters! (depends on topic)

from weak to

strong biased

processing

0 0.2 0.4 0.6 0.8 1 1.2

1

10

102

103

104

105

prob

abili

ty to

ent

er b

i-pol

arise

d st

ate

time

in b

i-pol

arise

d st

ates

strength of processing bias "0 0.2 0.4 0.6 0.8 1 1.20

0.2

0.4

0.6

0.8

1

– +

– +

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80

0.2

0.4

0.6

0.8

1

strength of processing bias !

prob

abili

ty to

ent

er b

i-pol

arise

d st

ate

strong homophily(exchange when difference less than 4)

biased processing only(strength governed by ћ)

moderate homophily(exchange when difference less than 6)

weak homophily(no exchange between extremes)

Opinion Homophily! The most prevailling assumptions in opinion dynamics is that the

interaction probability depends on the opinion similarity.

! As all previous ACT studies-draw on homophily, it is im-portant to understand the interplay of biased processing and homophily

! Assumption: agents do not interact if opinion di"erence is larger than a threshold! weak form: no exchange bet-

ween extremes (os = -4, or = +4)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80

0.2

0.4

0.6

0.8

1


prob

abili

ty to

ent

er b

i-pol

arise

d st

ate

strong homophily(exchange when difference less than 4)


moderate homophily(exchange when difference less than 6)


Opinion Homophily! One of the most prevailling assumptions in opinion dynamics ...! Homophily signi"cantly shifts

the critical value of polarisation to a weaker processing bias! Even under the mild assumption

that only those at the extreme ends do not talk

! From the perspective of previous models: biased processing ch-anges the picture completely

! We need to check whether pre-vious conclusions still hold!!

(Selective) Media Exposure! Empirical work on politicised science issues as climate change

suggests that selective processing and media exposure rein-force one another (Feldman, 2014)

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80

0.2

0.4

0.6

0.8

1


prob

abili

ty to

ent

er b

i-pol

arise

d st

ate



selective exposure(strength governed by ћ)

combined effect(selective exposure and weak homophily)

neutral exposure(at a rate of 0.5)

! With a chance of 1/2 agents receive an argument from an external source:1) transition sharpens under neut-

ral exposure!2) transition shifts to lower levels of

bias under selective exposure

! Information diversity is not helpful if biased processing is strong.

12

Conclusions

Argument exchange with biased processing matches better with experimental data. 2.Promising approach to measu-re the extend to which subjects engage in biased processing.3.

0 0.2 0.4 0.6 0.8 1 1.2


0.5

1

1.5

2

2.5

3

3.5

mea

n sq

uare

err

or

gas biomass

coal onshore

o"shore


Biased processing may lead to attitude polarisation; but it crucially depends on the strength of the bias.

1.

– +– +

Weak biased processing speeds up group decision processes by orders of magnitude4.Strong biased processing is su!cient for persistent int-ra-group polarisation5.

0 0.2 0.4 0.6 0.8 1 1.2

1

10

102

103

104

105

prob

abili

ty to

ent

er b

i-pol

arise

d st

ate

time

in b

i-pol

arise

d st

ates

strength of processing bias !0 0.2 0.4 0.6 0.8 1 1.2

0

0.2

0.4

0.6

0.8

1

– +

– +

Homophily and selective expo-sure may lead to intra-group con"ict at signi#cantly lower rates of biased processing

6.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80

0.2

0.4

0.6

0.8

1


prob

abili

ty to

ent

er b

i-pol

arise

d st

ate






12

Opinion diversity does not help if biased processing is strong!7.

?Conclusions

Argument exchange with biased processing matches better with experimental data. 2.Promising approach to measu-re the extend to which subjects engage in biased processing.3.

0 0.2 0.4 0.6 0.8 1 1.2


0.5

1

1.5

2

2.5

3

3.5

mea

n sq

uare

err

or

gas biomass

coal onshore

o"shore


Biased processing may lead to attitude polarisation; but it crucially depends on the strength of the bias.

1.

– +– +

Weak biased processing speeds up group decision processes by orders of magnitude4.Strong biased processing is su!cient for persistent int-ra-group polarisation5.

0 0.2 0.4 0.6 0.8 1 1.2

1

10

102

103

104

105

prob

abili

ty to

ent

er b

i-pol

arise

d st

ate

time

in b

i-pol

arise

d st

ates

strength of processing bias !0 0.2 0.4 0.6 0.8 1 1.2

0

0.2

0.4

0.6

0.8

1

– +

– +

Homophily and selective expo-sure may lead to intra-group con"ict at signi#cantly lower rates of biased processing

6.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.80

0.2

0.4

0.6

0.8

1


prob

abili

ty to

ent

er b

i-pol

arise

d st

ate






12

Opinion diversity does not help if biased processing is strong!7.

Thanks!

How Good is the Model Given the Data?

0.0 0.2 0.4 0.6 0.8 1.0 1.2

0.2

0.4

0.6

0.8

1.0

strength of processing bias !!

generaladoptiontendency"

! Maximum likelihood test (log likelihood) indicates:

1. biased argument adoption is relevant (! > 0)

2. a moderate bias 0.4 < " < 0.7 explains observed attitude changes best

3. # is very small (here # = 0.01)

! Clear signature that the cogni-tive agent model provides a valid microfoundation

0.0 0.2 0.4 0.6 0.8 1.0 1.2

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0 1.2

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0 1.2

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0 1.2

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0 1.2

0.2

0.4

0.6

0.8

1.0

0.0 0.2 0.4 0.6 0.8 1.0 1.2

0.2

0.4

0.6

0.8

1.0

biomass

generaladoptiontendency

!generaladoptiontendency

!

strength of processing bias " strength of processing bias "

gas

coal

o#shore

onshore

solar

strength of processing bias "

Proc

essi

ng b

ias a

cros

s top

ics

biased processing and opinion polarisation

Documents