Urban Legend Propagation Ian Dennis Miller 2018-11-08 Ian Dennis - - PowerPoint PPT Presentation

Urban Legend Propagation

Ian Dennis Miller 2018-11-08

Ian Dennis Miller Urban Legend Propagation 2018-11-08 1 / 88

Introduction

Ian Dennis Miller Urban Legend Propagation 2018-11-08 2 / 88

Executive Summary

Urban Legends propagate farther when they are disgusting

phenomenon known as emotional selection We replicate this finding with a simulation study We conduct a new study to simulate urban legend propagation in a larger network

The current work offers two major contributions:

Applied research on the topic of urban legend propagation Fundamental science of computational social modeling

Experimental control is inversely proportional to ecological validity.

The more controlled an experiment is, the less is resembles real life. Lack of ecological validity is a frequent criticism in the psychological sciences. We want both: control and validity. Perhaps these are separate, complementary steps in a research epistemology.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 3 / 88

Overview

Background

Urban Legends A Study of Urban Legends Computational Modeling

Methods Results Discussion Conclusion

Figure 1: https://commons.wikimedia.org/wiki/File: Power_8_mandelbulb_fractal_overview.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 4 / 88

Background: Urban Legends

Ian Dennis Miller Urban Legend Propagation 2018-11-08 5 / 88

Background: Urban Legends

Serial Reproduction Task Communication Networks Contagion and Epidemiology Memes Social Transmission Cascades Urban Legends Emotional Selection

Figure 2: https://commons.wikimedia.org/wiki/File: Harpy.PNG

Ian Dennis Miller Urban Legend Propagation 2018-11-08 6 / 88

Serial Reproduction Task

Remembering: A study in experimental and social psychology.

(Bartlett, 1932) first scientific study of “word of mouth” chains also tested images, separately

kids game: “telephone operator”

A tells B, who tells C, who tells D, . . .

serial reproduction is a useful research method

Figure 3: https://commons.wikimedia.org/wiki/File: Telephone_operators,_1952.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 7 / 88

Communication

vocabulary for discussing communication (Shannon, 1948)

transmitter receiver message channel

information theory

signal-to-noise ratio uncertainty (entropy)

Figure 4: https://commons.wikimedia.org/wiki/File: Shannon_communication_system.svg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 8 / 88

Networks

Graph theory and networks literature Serial reproduction task is:

a communication network represented as a graph

network topology:

directed, acyclic graph with k nodes each node:

in-degree of 1

• ut-degree of 1

except first and last nodes Figure 5: https://commons.wikimedia.org/wiki/File: Hierarchical_Communication_Network.png

Ian Dennis Miller Urban Legend Propagation 2018-11-08 9 / 88

Contagion and Epidemiology

contagion: transmission by contact

not choice

pathogenic contagion

epidemiology, public health formal analysis; modeling

emotional contagion

“emotional state-matching of a subject with an object” (De Waal, 2008) basis for complex social processes

social contagion

social influence (Asch, 1956) false memories (Roediger, et al, 2001)

Figure 6: De Waal. (2008). Putting the Altruism Back into Altruism

Ian Dennis Miller Urban Legend Propagation 2018-11-08 10 / 88

Memes

The Selfish Gene, ch. 7 (Dawkins, 1976) meme: unit of culture

repeatable can be mutated

viral social network phenomena

information cascades meme epidemics, contagions? image macro: image with embedded word overlay

memes are what is transmitted

the meme is the message content/data, for modeling purposes

Figure 7: https://commons.wikimedia.org/wiki/File: I_can_has_cheezburger.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 11 / 88

Social Transmission

transmitted via network “word of mouth”

• nline social media

personal communications

• verheard on the street

broadcast media influences

previous research has looked at:

Emotional Selection (Heath, 2001) Arousal (Berger, 2009) User Generated Content (Miller, 2012)

Figure 8: https://commons.wikimedia.org/wiki/File: NetworkSociality1.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 12 / 88

Cascades

cascade: a time series of sharing events

term used by social network scientists

track a media object (meme) as it propagates

content-oriented, as opposed to people-oriented

examples of cascades:

dancing baby i can haz cheezburger Arab Spring (on Twitter) Gangnam Style music video ice bucket challenge

Figure 9: https://en.wikipedia.org/wiki/File: DancingBaby.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 13 / 88

Urban Legends

stories passed via word of mouth

• ften with a moral or safety message

Phantom Anesthetist (Johnson, 1945)

examined “mental epidemic” “hysterical” contagious laughter

Best & Horiuchi (1985)

no evidence for “Halloween sadism”

Big Foot Legend transmission not serial

in “real world,” degree is much higher than one

Figure 10: https://commons.wikimedia.org/wiki/File: Urban_Legends_City_logo.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 14 / 88

Emotional Selection

When content elicits an emotion in the reader

affects the reader’s behavior likelihood to share that content

Heath et al (2001): disgusting urban legends

more widespread across urban legends websites elicit more willingness in participants to share

Figure 11: https://commons.wikimedia.org/wiki/File: Fragile_Emotion_cropped.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 15 / 88

Recap: Urban Legends

serial reproduction task: a research paradigm for studying word of mouth communication: the process of transmitting a message from sender to receiver networks of communication: graph theory applied to communication contagion: transmission by contact; epidemiology models memes: ever-changing content that is copied and transmitted social transmission: the human behavioral ecosystem of meme sharing cascade: the “life history” of a viral meme urban legend: stories re-told via word of mouth emotional selection: story propagation is affected by emotion altogether, this provides a precise vocabulary to describe

Eriksson & Coultas (2014) the current work

Ian Dennis Miller Urban Legend Propagation 2018-11-08 16 / 88

Reviewing a study on Urban Legends

Ian Dennis Miller Urban Legend Propagation 2018-11-08 17 / 88

Reviewing a study on Urban Legends

Eriksson & Coultas (2014): Emotional selection in urban legend propagation.

study 2: rate disgusting urban legends study 3: read stories and share to a partner

Setting the stage for the current work

Figure 12: https://commons.wikimedia.org/wiki/File: Bigfoot!_(19720317769).jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 18 / 88

Disgusting Stories

Eriksson & Coultas (2014) break down transmission into 3 stages:

receive, encode-retrieve, transmit

used Serial Reproduction paradigm

• ne person tells the next, in sequence

“fizzles” in the lab; “viral” cascades cannot manifest

paradox: in “real world,” urban legends survive.

I was going to run a very similar in 2014

finding this study was a time-saver!

Ian Dennis Miller Urban Legend Propagation 2018-11-08 19 / 88

Eriksson & Coultas Study 2: choose-to-transmit (1/2)

does disgust affect how funny a story is? Yes 80 participants from mturk provided ratings of urban legends

each participant rated 4 stories high/low disgust assigned with Latin square reported disgust, amusement, “pass-along”

Figure 13: Story topics are manipulated to create high/low disgust versions.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 20 / 88

Eriksson & Coultas Study 2: choose-to-transmit (2/2)

predicting “pass-along ratings”

how likely they are to share the story disgust is a significant driver of the effect

replicates Heath et al (2001) NB: created model of urban legend transmission

maximum likelihood estimates are interpreted as slopes

• n average, sharing is unlikely

(intercept is -0.86) but disgusting stories can counteract much of this (0.79)

Figure 14: Eriksson & Coultas (2014); p. 14

Ian Dennis Miller Urban Legend Propagation 2018-11-08 21 / 88

Eriksson & Coultas Study 3: Receive and Transmit (1/2)

emotional selection during reading? yes study 3 limited to: choose-to-receive, choose-to-transmit

stories are on paper; encode-retrieve not part of it

80 mturk participants

40 “chains” consisting of 2 “generations” 2 steps per generation: read and transmit read 4 stories; choose to share what they read

Figure 15: Flowchart depiction of research method used in Eriksson & Coultas Study 3.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 22 / 88

Eriksson & Coultas Study 3: Receive and Transmit (2/2)

track the number of stories retained after 2 generations

at each stage and each generation, fewer are retained

high disgust stories are retained at a higher rate, retained longer by end of step 4, no low-disgust stories retained emotional selection during both read and transmit

Figure 16: Eriksson & Coultas (2014); p. 17

Ian Dennis Miller Urban Legend Propagation 2018-11-08 23 / 88

The current work

pick up where Eriksson & Coultas left off

transmission chains “fizzle” paradox: in “real world,” urban legends survive. so let’s bring the model out of the lab

current work re-contextualizes transmission model

build empirically calibrated Agent Based Model (ABM) run with different network conditions test for epidemic sharing outcomes in urban legends

Figure 17: https://commons.wikimedia.org/wiki/File: Model_train_union_station_066094.JPG

Ian Dennis Miller Urban Legend Propagation 2018-11-08 24 / 88

Background: Computational Modeling

Ian Dennis Miller Urban Legend Propagation 2018-11-08 25 / 88

Background: Computational Modeling

Epistemological approaches to systems modeling Agent-Based Modeling Individual-Based Modeling Pattern-oriented Modeling Patterns of Human Behavior Modeling Psychological Phenomena

Figure 18: https://commons.wikimedia.org/wiki/File: Analog_Computing_Machine_GPN-2000-000354.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 26 / 88

Epistemological approaches to systems modeling

Formal: differential equations

economics, game theory: equilibria

Agent-based: logical, algorithmic (theoretical)

computer science: finite state machines, if/then multi-agent systems: robotics, manufacturing

Individual-based: observed patterns (empirical)

ecological: regression, algorithms

Figure 19: https://commons.wikimedia.org/wiki/File: The_flock_of_starlings_acting_as_a_swarm._- geograph.org.uk-_124593.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 27 / 88

Agent-Based Modeling

ABM: Agent-Based Modeling A method for studying systems

for example: people, animals, cars,

• products. . .

Dublin and Lotke (1925): US population growth Cellular Automata (von Neumann) Conway (1970) Schelling (1971) Epstein & Axtell (1996) Wilensky (1999)

Figure 20: https://commons.wikimedia.org/wiki/File: Glider.svg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 28 / 88

Individual-Based Modeling

Railsback and Grimm (2005; 2011) These ecologists are pragmatic.

Railsback: “Not too simple; not too complex”

Describe system by describing individuals

study how system affects individuals study how individuals affect system

Figure 21: https://commons.wikimedia.org/wiki/File: School_of_fish_02709.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 29 / 88

Strong Inference

epistemological goals:

discover truth falsify incorrect hypotheses

Railsback and Grimm (2011, p. 245)

Strong Inference (Platt, 1964) apply to ABM

HowTo: Strong Inference with ABMs (Railsback & Grimm, 2011)

1 identify alternative traits (hypotheses) for behavior 2 implement the alternative traits in the ABM, testing the software carefully to get “clean

results”

3 test and contrast the alternatives by seeing how well the model reproduces the

characteristic patterns, falsifying traits that cannot reproduce patterns

4 repeat the cycle as needed: revise the behavior traits, look for (or generate, from

experiments on the real system) additional patterns that better resolve differences among alternative traits, and repeat tests until a trait is found that adequately reproduces the characteristic patterns.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 30 / 88

Pattern-oriented Modeling: Scientific Workflow

HowTo: Pattern-oriented Modeling

1 empirical research/literature 2 identify characteristic patterns of

emergent behavior

3 propose theories for a key individual

behavior

4 produce individual-based model 5 how well does IBM reproduce observed

patterns?

Figure 22: Railsback & Grimm (2011, p. 245)

Ian Dennis Miller Urban Legend Propagation 2018-11-08 31 / 88

Pattern-oriented Modeling: Model Building

ODD Protocol: Overview, Design Concepts, Details

HowTo: Construct Pattern-Oriented Modeling Study

1 formulate model using ODD protocol 2 identify characteristic patterns of

emergent behavior

3 define criteria for pattern-matching 4 review model formulation

Figure 23: Grimm, et al. (2005). Science.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 32 / 88

Patterns of Human Behavior

Psychology

empirical science predicated on

• bservation

behavior, cognition, personality,

• development. . .

phenomena are commonly modeled using regression

based on multiple observations, statistics

regression and SEM models for everything!

behavior, cognition, personality,

• development. . .

Figure 24: https://commons.wikimedia.org/wiki/File: Mourning_angel.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 33 / 88

Modeling Psychological Phenomena

existing efforts to incorporate decision-making

ODD - D Belief-Desire-Intention (BDI) not psychologically useful (or plausible) behavioral economics and psychology: (e.g. Jackson, 2017)

I propose Individual-Based Modeling

• ffers a better way forward.

epistemology of computational modeling derives from ecology, not computer science

Figure 25: https://commons.wikimedia.org/wiki/File: Extensive_form_game_2.svg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 34 / 88

Recap: Computational Modeling

Numerous ways to think about modeling

multiple epistemologies ecology literature is interesting

Pattern-oriented Modeling: ABM approach

characterize living systems

Psychology: understanding people

patterns of human behavior build models of psychological phenomena

I propose to build computational models

• f psychological phenomena

perform inference with models

Figure 26: https://commons.wikimedia.org/wiki/File: CTA_Loop_Junction_Detail.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 35 / 88

Methods

Ian Dennis Miller Urban Legend Propagation 2018-11-08 36 / 88

Methods

Studies Model (ODD protocol) Simulations

Figure 27: https://commons.wikimedia.org/wiki/File: WomanFactory1940s.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 37 / 88

The Current Work

Extension of Eriksson & Coultas (2014)

Model of social transmission behavior as serial reproduction task Create empirical, psychological ABM with new network topology emergence of viral sharing (not seen in lab study)

Methodological approach:

study 1: Replicate dynamics in original article study 2: demonstrate lack of scaling with serial topology study 3: use new topology and observe viral emergence

Figure 28: https://commons.wikimedia.org/wiki/File: Model_Train_at_Old_Depot_Museum.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 38 / 88

Study 1: Network

Goal: build ABM that replicates Eriksson & Coultas, Study 3 Network layout:

each serial reproduction chain consists

• f 2 agents (i.e. has chain-length of 2)

Generation 1 agent has one directed edge to Generation 2 agent 40 chains; total n = 80

Figure 29: Serial reproduction chains; length = 2

Ian Dennis Miller Urban Legend Propagation 2018-11-08 39 / 88

Study 1: Simulation Overview

first agent in each chain receives 4 stories agents decide whether to read any stories they receive agents may decide to share any stories they read any shared stories are passed to the next agent in the chain there are 2 agents in the chain:

the simulation is done when time = 2 ticks

the number of story shares is logged to a CSV for analysis

Figure 30: Study 1; NetLogo Interface. Agents are displayed in the box on the left. Buttons appearing on the right are used to interact with the simulation.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 40 / 88

Study 2: Scaling Up

Goal: re-run study 1, but increase scale

because it’s difficult

Rationale for increasing scale:

at each step in the chain, fewer stories are shared

simply increase the number of

• pportunities to share

. . . see if that affects total shares

instead of 40 chains, we will run 400 chains

total n = 800

in all other ways, this is identical to study 1

Figure 31: Study 2; NetLogo Interface. There are so many agents and connections that they all blend together.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 41 / 88

Study 3: New Network Topology

Goal: use a different network topology

preferential attachment “rich get richer”

use same n as study 2: 800 simulation runs longer: time = 5 ticks rationale: network has no beginning or end

cascade always dies in the lab could therefore run indefinitely

epidemic: susceptible-infected-susceptible (SIS)

susceptible: agents willing to receive stories infected: while they have stories to transmit

Figure 32: Study 3; NetLogo Interface with network layout.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 42 / 88

Methods: ODD Protocol

Ian Dennis Miller Urban Legend Propagation 2018-11-08 43 / 88

Methods: ODD Protocol

Overview of Model

Entities, State variables, and Scales Process overview and Scheduling

Design Concepts

Emergence Interaction Stochasticity

Details

Ian Dennis Miller Urban Legend Propagation 2018-11-08 44 / 88

Entities

Agents represent humans Edges between agents represent a communication network

directed edges; feed-forward in a social context, these are friendships

Stories represent the content Agents will read

a separate edge called “reader” is used to represent an Agent reading a Story. this graph implies: “a story has readers”

Figure 33: Close-up of edges that connect agents in the ABM.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 45 / 88

Agent variables: Decision States

agents have “state” variables that influence their behavior in the simulation has_content: whether an agent has read any stories has_decided: whether an agent has already considered sharing stories

prevents agents from acting twice in a single time step.

decision_model: whether an agent uses a null model or the Eriksson & Coultas model

this variable controls the agent’s decision-making model

Figure 34: https://commons.wikimedia.org/wiki/File: TheKnightAtTheCrossroads.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 46 / 88

Agent variables: Share Threshold

share_threshold: determines whether an agent will share or not Eriksson & Coultas Study 2

provides linear model of sharing intentions when intention is high enough, agents will share share_threshold is the trigger

share threshold is like a degree of freedom

calibrate against results of E&C study 3

Figure 35: https://commons.wikimedia.org/wiki/File: Synapse_diag5.png

Ian Dennis Miller Urban Legend Propagation 2018-11-08 47 / 88

Urban Legend Stories

Eriksson & Coultas participants scored stories:

how disgusting how amusing

Stories in the model have the following state variables

disgusting: 0 for not disgusting, 1 for disgusting theme: 0 for cake, 1 for dog, 2 for Nepal, 3 for pizza amusing-mean: average amusing rating for this story amusing-sd: standard deviation of amusing rating for story

Figure 36: https://commons.wikimedia.org/wiki/File: Dog_(26345685425).jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 48 / 88

Scales

scales: the types of dimensions in this model Time is discrete increments of 1 hour

the duration of an experimental session.

as if 2-link serial reproduction conducted:

2 separate, 1-hour sessions

• ther scales (e.g. Likert scales) are

centered and standardized

Figure 37: https://commons.wikimedia.org/wiki/File: Golden_Ratio.png

Ian Dennis Miller Urban Legend Propagation 2018-11-08 49 / 88

Process overview and scheduling

Simulation asks agents to act once per time tick (step) agents are updated in the same order each time

all agents are included in fixed update schedule special considerations:

generation 1 agents act only at time=1 - not time=2 generation 2 agents act only at time=2

serial operation: one agent completes its actions before the next agent starts

Figure 38: https://commons.wikimedia.org/wiki/File: TurtlesAllTheWayDown2.png

Ian Dennis Miller Urban Legend Propagation 2018-11-08 50 / 88

Interaction

entity types interact in several ways:

communication: agents interact with

• ne another by transmitting stories

receive: agents interact with stories by choosing whether to read them transmit: agents interact with stories by choosing whether to share them

network implications

receive and transmit decisions create network edges

interactions occur at each major step in the flowchart, at right

Figure 39: Flowchart depiction of research method used in Eriksson & Coultas Study 3.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 51 / 88

Stochasticity

Gaussian distribution

how many standard deviations away from the mean for a given dimension

Poisson distribution is used to model events over time

emotion selection: effect of disgust upon reading and transmitting

the preferential attachment network algorithm is stochastic

similar but different topologies each time

Figure 40: https://commons.wikimedia.org/wiki/File: Gaussian_distribution_thick_lines.svg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 52 / 88

Emergence

• bserving stories retained over time

measured per-chain, per-individual, and per-story

expecting eventual story extinction with serial chain topologies

prolific sharing with preferential attachment network topology

expect high-disgust stories to be retained at a higher rate

Figure 41: https://commons.wikimedia.org/wiki/File: Order_and_Chaos.tif

Ian Dennis Miller Urban Legend Propagation 2018-11-08 53 / 88

Observation

the model provides metrics about the total number of shares that occur

NetLogo provides tools to monitor simulation

the model is visualized using a GUI model results are exported as CSV

Figure 42: https://commons.wikimedia.org/wiki/File: Percival_Lowell_observing_Venus_from_the_ Lowell_Observatory_in_1914.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 54 / 88

Initialization

create agents, connect into network

study 1, 2: use a serial network topology study 3: add edges using preferential attachment

simulations start with the first agents in their transmission chain Experimenter provide stories to the starting agents

Figure 43: https://commons.wikimedia.org/wiki/File: Fiat_Lux,_Sather_Gate_detail.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 55 / 88

Submodel: step

“step” is executed at each time tick; it is how time moves forward while an agent has no stories, that agent does nothing for each story the agent “has”:

decide whether to read that story (choose-to-receive) if it was read, then decide whether to share (choose-to-transmit) ask the next agent to read all the stories this agent is sharing

Figure 44: Flowchart depiction of ABM “step” submodel.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 56 / 88

Submodel: choose-to-receive

When the story is disgusting, it is more likely to be read.

2.4 high-disgust of a maximum 4 stories are retained on average

• therwise, 1.2 of the maximum (4)

modeling number of reads as random Poisson; mean either 1.2 or 2.4

divide by 4 to yield a “retain-ratio” between 0 an 1

function randomly yields “yes” decision in proportion to retain ratio Eriksson & Coultas Study 3 “A related-samples Wilcoxon signed rank tests showed a difference in retention between high disgust (M=2.40, SD=1.30) and low disgust stories (M=1.20, SD=1.43) after the first step of the chain, p=.002, as pre- dicted by emotional selection operat- ing in the choose-to-receive phase.”

Ian Dennis Miller Urban Legend Propagation 2018-11-08 57 / 88

Submodel: choose-to-transmit (1/2)

There are two aspects to the share behavior:

how much does an agent wish to share? is that amount above threshold?

The model for determining share score is a zero-centered, standardized regression model. Coefficients come from Table 2 If share score produced by model is above share-threshold, then agent will share

Figure 45: Eriksson & Coultas (2014); p. 14

Ian Dennis Miller Urban Legend Propagation 2018-11-08 58 / 88

Submodel: choose-to-transmit (2/2)

HowTo: Compute share_score

1 select value at random from Gaussian

distribution for each of the following: disgust, social, amusing, interesting, plausible, surprising

2 substitute values for variables in the

following equation: share_score = -0.86 + (0.79 * disgust) + (0.08 * social) + (0.22 * amusing) + (0.35 * interesting) + (0.25 * plausible) + (0.18 * surprising)

Figure 46: Eriksson & Coultas (2014); p. 14

Ian Dennis Miller Urban Legend Propagation 2018-11-08 59 / 88

Null Model of Behavior

What happens when agents behave randomly?

replace the model from the article we flip a coin for the result applies to both receive and transmit

null model:

permits comparisons to Eriksson & Coultas model strong inference: reject theories that are bad does the Eriksson & Coultas model produce better outcomes?

Figure 47: https://commons.wikimedia.org/wiki/File: Coin_Toss_(3635981474).jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 60 / 88

Methods: Simulations

Ian Dennis Miller Urban Legend Propagation 2018-11-08 61 / 88

NetLogo

NetLogo (Wilensky, 1999) is an ABM environment

NetLogo is the interpreter and the language also provides tools for interacting with models

Alternatives include: MASON, RePast, AnyLogic

it shouldn’t matter which environment is used just need a match between domain and toolkit

NetLogo models are easy to share

Figure 48: https://commons.wikimedia.org/wiki/File: Netlogo-ui.png

Ian Dennis Miller Urban Legend Propagation 2018-11-08 62 / 88

BehaviorSpace

NetLogo includes the BehaviorSpace framework

can run simulations for you

automates process of model parameter search

can also run the same parameters repeatedly

runs the simulation repeatedly and saves the results as CSV used to create distributions of model behavior CSV results are imported with R

Figure 49: R Studio CSV data import script.

Ian Dennis Miller Urban Legend Propagation 2018-11-08 63 / 88

Results

Ian Dennis Miller Urban Legend Propagation 2018-11-08 64 / 88

Study 1: Replication

Pattern of results

high-disgust stories propagate at a higher rate than low-disgust fewer stories are read as time proceeds at each time point, fewer stories are shared than are read by the end of 2 time points, virtually no low-disgust stories are shared

Figure 50: Eriksson & Coultas (2014); p. 17

Ian Dennis Miller Urban Legend Propagation 2018-11-08 65 / 88

Stories retained over time: Side-by-side comparison

Figure 51: Eriksson & Coultas (2014); p. 17 Figure 52: Computational Model Study 1 Results

Ian Dennis Miller Urban Legend Propagation 2018-11-08 66 / 88

Study 1: Replication Results

Average stories received and transmitted per step Summary of results:

high-disgust stories propagate at a higher rate than low-disgust fewer stories are read as time proceeds at each time point, fewer stories are shared than are read by the end of 2 time points, virtually no low-disgust stories are shared

error bars

refer to the distribution of results from multiple simulation runs not the same as the error bars around the estimate of the parameter more in the discussion

Figure 53: Computational Model Study 1 Results

Ian Dennis Miller Urban Legend Propagation 2018-11-08 67 / 88

Calibrate share_threshold parameter

Study 1 results obtained with a share_threshold of -0.5 share_threshold: standard deviation units

• nly free parameter in entire ABM

when share_threshold is:

exactly 0: the average transmit intention score triggers transmission below 0: a lower transmit score triggers (i.e. more sharing) above 0: less sharing

visualize entire share_threshold parameter space

compare to original pattern of results

this is how -0.5 is obtained

Figure 54: share_threshold Parameter Search Results}

Ian Dennis Miller Urban Legend Propagation 2018-11-08 68 / 88

Compare choose-to-receive null model

null behavior model for comparisons

result of a coin flip instead of empirical model experimentally disable choose-to-receive

procedure

replace the behavioral model in ABM re-run simulation

do not reject Eriksson & Coultas computational model

pattern mis-match with choose-to-receive null model low-disgust stories are read too frequently

Figure 55: Null choose-to-receive Results

Ian Dennis Miller Urban Legend Propagation 2018-11-08 69 / 88

Compare choose-to-transmit null model

now replace the transmit decision

disable transmission decision-making now sharing is the result of a 50/50 coin flip

do not reject Eriksson & Coultas computational model

pattern mis-match shares of high-disgust drops off too rapidly by 2nd generation

Figure 56: Null choose-to-transmit Results

Ian Dennis Miller Urban Legend Propagation 2018-11-08 70 / 88

Compare both null models to Computational model

a 2x2 grid

bottom-right quadrant: null read and transmit top-left quadrant: empirical read and transmit top-right and bottom-left: repeats of the previous 2 plots

can now see unique contributions of receive versus transmission do not reject Eriksson & Coultas computational model

Figure 57: Null receive and transmit Results

Ian Dennis Miller Urban Legend Propagation 2018-11-08 71 / 88

Study 2: Receive-Transmit pattern at 10x scale

is it possible that widespread sharing requires more observations?

i.e. more chances for stories to “live”?

re-run simulation with 800 agents

10x more than original study

2 ways to increase scale. either:

longer reproduction chains more reproduction chains

larger sample needed? no

pattern of results is the same irrespective of n

Figure 58: Study 2 Results

Ian Dennis Miller Urban Legend Propagation 2018-11-08 72 / 88

Study 3: Preferential Attachment Network

replace serial topology with preferential attachment (PA)

n is still 800

visualization format

total reads and shares per story not average what would the denominator be?

new outcomes

low-disgust stories rapidly die off but high-disgust stories demonstrate the opposite pattern!

Figure 59: Study 3 Results

Ian Dennis Miller Urban Legend Propagation 2018-11-08 73 / 88

Side-by-side comparison of total stories retained

Figure 60: Study 2, Total Stories Figure 61: Study 3, Total Stories

Ian Dennis Miller Urban Legend Propagation 2018-11-08 74 / 88

Null models with Preferential Attachment topology

2x2 grid

all-null: bottom-right all-empirical: top-left

• bservations about null/null model:

too many low-disgust stories read and shared

• verall positive bias towards sharing

Figure 62: Study 3, Null Comparison

Ian Dennis Miller Urban Legend Propagation 2018-11-08 75 / 88

Summary

Study 1: Replicated the serial topology pattern of results from Eriksson & Coultas

Calibrated sharing threshold Selectively disabled receive/transmit behaviors Did not discard computational models in favor of null models

Study 2: Increased scale 10x to look for scaling properties; found nothing. Study 3: Changed topology to preferential attachment

in contrast to lab, high-disgust stories increase over time

Figure 63: Study 3, Total Stories

Ian Dennis Miller Urban Legend Propagation 2018-11-08 76 / 88

Discussion

Ian Dennis Miller Urban Legend Propagation 2018-11-08 77 / 88

Recap

The Eriksson & Coultas lab study was highly controlled

but lacked ecological validity i.e. lab results don’t look like real life study isn’t flawed; to the contrary, it’s excellent

• ur ABM simulation behaves as expected

(when run with more realistic network) high-disgust stories were shared at increasing rate meanwhile, low-disgust stories “die off”

Figure 64: https://commons.wikimedia.org/wiki/File: CTA_Loop_Junction_Detail.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 78 / 88

What the error bars mean

each model configuration is simulated multiple times

• utcome measures may be summarized

across simulations e.g. story usage (read/transmit) is an

• utcome

summary of model runs is an estimate

• f an emergent parameter

with error

“real life” Eriksson & Coultas error bars are different

estimates of the model parameters error bars related to confidence in the estimate of the mean each point is an individual, not a simulation population

Figure 65: https://commons.wikimedia.org/wiki/File: Beech_forest_Matra_in_winter.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 79 / 88

Explanations for Emergent Network Behavior

Preferential Attachment network emergence

high-disgust stories shared more, not less, at generation 2 more retained at successive generations

network topology:

in-degree above 1: receive stories from n agents

• ut-degree above 1: transmit stories to

n agents

epidemiology dynamics:

preferential attachment implies S-I-S

susceptible, infected, susceptible repeated “infections” with novel stories

serial reproduction task implies S-I-R

susceptible, infected, recovered Figure 66: https://commons.wikimedia.org/wiki/File: Gaussian_primes.png

Ian Dennis Miller Urban Legend Propagation 2018-11-08 80 / 88

Scaling a serial reproduction task

the pattern of results is practically unchanged

irrespective of the behavior models

practically any other network topology is parameterized by scale

serial chains could be unique in this respect particularly well-suited to experimental control?

Figure 67: https://commons.wikimedia.org/wiki/File: Pierpont_exponent_distribution.png

Ian Dennis Miller Urban Legend Propagation 2018-11-08 81 / 88

Messages and Topologies

Eriksson & Coultas: “. . . a link between emotional selection and social network structure might act as a factor that magnifies the cultural differences between different communities, an idea left for future research.” What about messages that change networks? social networks are dynamic

messages can cause us to add or remove connections e.g. twitter, facebook

I have been puzzling over how to obtain friendship dynamics data

but simulation offers a different way forward

Figure 68: https://commons.wikimedia.org/wiki/File: Battle_of_Antietam2.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 82 / 88

Triggering Media

A common expression these days is “being triggered”"

i.e. experience trauma from social media that evokes painful memories refers to a message that induces a reaction

we can construct a literal simulacrum of this phenomenon

the “trigger” metaphor maps onto “threshold” for behavior

triggering and behavioral or emotional contagion

Figure 69: https://commons.wikimedia.org/wiki/File: Trigger_mechanism_bf_1923.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 83 / 88

Statistical Inference

produce error bars of arbitrary size

determined by num simulation runs need tighter distribution? run more simulations.

Andrew Gelman (2009): “p-value can be viewed as a crude measure of sample size” proposed statistical inference process:

contrast 2 or more simulation theories perform repeated simulation runs, measure an emergent result test: num simulation runs required until distributions no longer overlap?

working backwards from Gelman:

number of simulation runs is a crude measure of p-value?

Figure 70: Computational Model Study 1 Results

Ian Dennis Miller Urban Legend Propagation 2018-11-08 84 / 88

A General Method for Psychological Research

ABM methods can be applied to many findings in psychology probably works with many sorts of behavior frameworks

regression is demonstrated in this article how about Bayesian models of behavior? cognitive frameworks like ACT-R?

the class of regression models contains thousands of examples of behavior

many of these can be adapted and calibrated

this approach can be more affordable than big-data observation

Figure 71: https://commons.wikimedia.org/wiki/File: Compound_Microscope_1876.JPG

Ian Dennis Miller Urban Legend Propagation 2018-11-08 85 / 88

Conclusion

Ian Dennis Miller Urban Legend Propagation 2018-11-08 86 / 88

Computational Modeling and Urban Legends

a general method for adapting psychological research for computational exploration

inspired by the ecologist approach to agent-based modeling

developed empirically calibrated model of urban legend transmission

• riginal empirical findings from the

literature results were obtained in the isolation of a lab the current work re-contextualized laboratory results with a realistic network, realistic results were obtained

Figure 72: https://commons.wikimedia.org/wiki/File: Power_8_mandelbulb_fractal_overview.jpg

Ian Dennis Miller Urban Legend Propagation 2018-11-08 87 / 88

Contact

Ian Dennis Miller PhD Candidate Psychology Department University of Toronto @iandennismiller i.miller@utoronto.ca https://www.sisrlab.com http://imiller.utsc.utoronto.ca

Figure 73: Photo Credit: Geoff MacDonald

Ian Dennis Miller Urban Legend Propagation 2018-11-08 88 / 88