Why Globalized Communication may increase Cultural Polarization

1
Why Globalized Communication may increase
Cultural Polarization

• Paper presented at
• 2005 International Workshop  Games, Networks, and
  Cascades
• Cornell Club (NYC), October 7-9, 2005

Andreas Flache, University of Groningen,
ICS Collaborators on general project Michael W.
Macy, Cornell University James A. Kitts,
University of Washington
2
Cultural diversity and global communication

• Two positions
• Increasingly global communication homogenizes
  cultures
• E.g. Hamelink 1983
• Increasingly global communication makes cultural
  differences and cross-cultural conflict more
  pronounced
• E.g. Huntington 1996

3
Concepts

• What is culture?
• Anderson culture provides a set of ideas,
  values and beliefs that function to provide a
  basis for interaction and understanding among a
  collection of people
• Axelrod culture is set of individual attributes
  that are subject to social influence
• Globalizing communication
• broader range of interaction beyond an
  individuals immediate locale and across cultural
  groups (Greig, 2002)
• Qualitative jump through the internet

4
Computational models of culture formation

• Models proposed by Carley, Axelrod, Mark, Latane
• Homophily the greater the similarity, the more
  likely the interaction (relational dynamic).
• Influence the greater the interaction, the more
  similar become the interactants (opinion
  dynamic).
• Axelrod influence is restricted to local
  neighbors
• Dynamics
• Minimal initial similarity increases probability
  of interaction
• which then increases similarity
• leading to uniformity, not diversity
• Why can there be stable diversity?

5
Axelrods solution interaction thresholds

• Influence stops when individuals are too
  different
• preservation of diverse, isolated subcultures
• Local regions become homogenous over time ?
• Differentiation from neighboring regions ?
• No more mutual influence ?
• Stable diversity

Example of equilibrium 5 features, 15 traits
per feature 20x20 world,
6
Implications of Axelrods model for globalizing
communication

• Modeling globalization
• Inreasing geographical range of communication
• Axelrod (1997)
• Increasing range ? less diversity
• Diversity distinct cultures in equilibrium
• Initial distribution more similar across
  neighborhoods (random)
• more overlap, i.e. smaller chance of isolation of
  local regions
• Follow-up studies
• E.g. Shibani (2001), Greig (2002)
• Global mass media and larger range of interaction
  allow local minorities to find support against
  local conformity pressures ?
• Globalized communication may also increase
  diversity

7
What is missing(1)

• Continuous opinion space
• Axelrod etc assume nominal opinion space
• Many issues are not nominal
• how much money should we spend on?
• Many traditional models of opinion formation use
  continuous space
• e.g. French, Harary, Abelson, Friedkin,
  Hegselmann Krause.
• These models produce unanimity, not stable
  diversity, under a large range of conditions.

8
What is missing(2)

• There is no negative influence
• Axelrod etc assume that agents never change
  opinions to decrease similarity
• Empirically we know people often have a tendency
  to distance oneself from negative referents,
  profiling
• Adding negative influence in a continuous opinion
  space may profoundly change influence dynamics
• ? Macy et al (2002) from uniformity to
  polarization (not just diversity)

9
A Hopfield Model of Dynamic Attraction Modeling
negative influence and continuous opinions

• Nowak Vallacher, 1997
• Node i has or opinion on K dimensions
• (-1 sik 1)
• Nodes i and j are tied by positive or negative
  weights (-1wij1)
• Opinion of j can attract or repel opinion of i,
  depending on wij

10
Influence depends on relations

• Effect of sj on si depends on the connection
  between i and j
• Positive weights opinions become more similar
• Negative weights opinions become less similar
• Change in position of i with regard to issue s is
  weighted average of distances sj-si modified by
  moderation m
• Moderation degree to which actors weigh small
  differences in opinion relatively less (m gt1
  moderate or tolerant)

N size of neighborhood j ? neighborhood
11
And relations depend on influence

• Weight wij increases with agreement in the K
  opinions of i and j
• To be precise weight is adapted gradually to
  match level of (dis) agreement.

K number of opinions j ? neighborhood ?
learning rate
12
More details

• Correction necessary to keep opinions within
  bounds
• Asynchronous updating
• Agent is selected at random
• either weights or states are updated with equal
  probability

13
Access structure channels influence

• Mutual influence only for local neighbors
• For example
• Agents are arranged on a circle
• Parameter range (r)
• of population to which agent has access
• Access is symmetrical

r20
r50
r10
Examples for N20
14
Experiment 1 Does continuous opinion space
reduce diversity?

• Comparison with Axelrod no negative influence
• weights are mapped linearly to 0..1 interval
• ? Zero influence only if maximum difference in
  opinions
• From dichotomous towards continuous opinions
• Discreetize opinion space into g equidistant
  positions
• Gradually increase g and test effect on diversity
  in equilibrium.
• Diversity measured as of different opinion
  vectors surviving.
• We also measured variance of opinions in
  equilibrium
• Conservative scenario
• resembles conditions where Axelrod found high
  diversity
• Opinion space is one-dimensional, k1 (few
  features)
• Strongly local interaction (circle, r2)

• More settings
• N100
• linear influence function (moderation1)
• Fast learning (?1)
• Initial opinion is uniformly distributed in
  -1..1
• initial weight proportional to initial agreement

15
Experiment 1 Results

• Consistent with Axelrod
• more possible opinions increase diversity
  (opinions in equilibrium)
• But inconsistent with Axelrod
• variance of opinions in equilibrium approaches
  zero as g increases
• No diversity at all in continuous opinion space

g number of equidistant opinions g gt 1000 ?
continuous opinion space
16
Experiment 2.Stable diversity in a continuous
opinion space negative influence

• Experiment 1 as baseline
• But now continuous opinion space
• k1, N100,
• Strongly localized interaction (r2)
• Manipulations
• Positive influence only (Axelrod) vs. Positive
  negative influence
• weights 0..1 vs. weights -1..1.
• Results
• With positive influence only, unanimity in
  equilibrium
• With posneg, stable polarization two maximally
  different subgroups
• By and large, this result is robust across a
  large range of conditions, e.g. for larger N, K
  and higher levels of m

17
Experiment 3 What is the effect of globalizing
communication?

• Experiment 2 as baseline
• But now always positive negative influence of
  interaction
• Continuous opinion space, k1, N100,m1,
• Manipulation
• Range of interaction increases gradually from
  2..50
• 10 replications per condition
• Outcome measures (after maximally 1000
  iterations)
• Diversity distinct opinions / N
• Polarization variance of pairwise agreement
• Variance of states
• But first an illustrative scenario k2, r2
• Larger range increases influence range of
  extremists
• ?no more gradual shift of opinions between
  neighbouring regions
• ?agents either move towards or distance
  themselves from extremists
• ?pressure towards polarization

18
A stylized explanation
critical distance
19
A stylized example large interaction range
critical distance
20
A stylized example small interaction range
critical distance
21
Experiment 3 Results

• Consistent with Axelrod
• a larger range of interaction decreases diversity
  (opinions in equilibrium)
• But inconsistent with Axelrod
• Stable diversity with continuous opinions
• Increasing variance of opinions with increasing
  range of interaction
• Increasing polarization with increasing range of
  interaction

Range size of local neighborhood in population
22
Experiment 3 Robustness tests

• Positive effect of range on polarization changes,
  
• When number of issues (k) increases
• Negative ties less likely from random start
• Effect tends to become negative
• When moderation (m) increases
• Large opinion differences weigh relatively more
• Positive effect becomes stronger
• Inverted U-shape effect of range possible
• Range has two opposing effects
• Larger range increases overlap between
  neighboring regions
• ?pressure towards conformity
• ..it also increases influence range of
  extremists ?pressure towards polarization

23
How can range increase polarization?The
diffusion of regional conflicts

• Illustrative scenario isolated caves
• N100, range5, k3, moderation1
• From a random start, homogeneity develops in most
  local regions, but in a small proportion of local
  regions polarization emerges
• When ties between polarized and homogenous
  regions are added, agents in homogenous regions
  either move towards or distance themselves from
  extremists
• Extremism spreads through random ties

24
Robustness tests of effects of range

• Noise
• Qualitative effects robust against small error in
  perception of others influence (/- .5)
• Population size
• Same qualitative effects found for N100, 200,500
• Dimensions of opinion space
• Polarization occurs also with higher k, but only
  with much higher moderation
• Moderation
• The less moderation, the less polarization
• Random access structure
• Qualitative effects remain unchanged

25
Conclusions

• Some previous models suggest cultural diversity
  can persist despite global interaction range,
  others dont
• All rely on nominal opinion space.
• Model with continuous opinion space and negative
  social influences generates tendency towards
  polarization when interaction gets global
• Depending on moderation and issues, effect of
  increasing range of interaction is
• increasing polarization
• decreasing polarization
• Inverted U-shape
• Model suggests that globalized communication may
  promote diffusion of regional conflicts

26
Future research

• Theoretical towards analytical models
• E.g. stochastic stability (Young)
• Empirical
• social influence in experiments / online
  interaction
• Is there influence? Is it negative?
• E.g. world value survey and data on accessibility
  of internet in different countries or social
  strata
• Is there a relationship between cultural
  convergence / divergence and access to the
  internet?