• Lots of talk about “echo chambers”
  • Surely this is a continuous phenomenon
  • Segregation a clearer concept
• Segregation in this context = different media diets by partisanship
• Related to but distinct from selective exposure theory
  • Interest in how much segregation, not why there is segregation
• How do you measure this?

• Using network analysis to approach this problem
  • Residential (racial) segregation
• Focusing on people as unit of analysis
  • Rather than sources or audiences
• Goal: a number that could be plugged into a regression model

A measure that…

• Can describe people
  • Person X has Y level of segregation
• Accommodates multi-party systems
• Is comparable across contexts
  • parties, countries, non-media contexts
• Requires no more complex designs than usual

Selective exposure research

• Control for partisanship, ideology, congenial media use
• No preference/positive preference for out-party media
• Not a measure of the information environment!

Tracking data (e.g., Bakshy, Messing, & Adamic, 2015)

• Define sources as liberal/neutral/conservative based on who tends to click/share
• How much liberal media on conservatives’ news feeds, vice versa?
  • About 25%

Tracking data

• Gentzkow & Shapiro (2011) have individual-level web and TV tracking data
  • Dichotomize sources
  • True ideology unknown for many respondents
  • Minimal “segregation” (= .08 on a measure bounded at 0 and 1)
  • Similar design and results from Flaxman & Goel (2016)

• Developed to measure residential segregation (Echenique & Fryer, 2007)
• Does not impose binary classification
• Produces measurement at individual, group (party), and system levels
• Can use for people or sources
• Single number bounded approx. at -1 (active avoidance of in-group) and 1 (perfect segregation)

• Requires a network/graph representation

For people:

• Nodes are people, edges reflect co-consumption
• Weight of edges = # of shared sources

For sources:

• Nodes are sources, edges reflect co-consumers
• Weight of edges = # of pairs of people who both use source

Pew Research Center’s American Trends Panel, Wave 1 (2014)

• N = 3,308
• Represenative of adults in USA
• 66 political media sources
  • Mix of TV, internet, print, radio (some overlapping)
  • Did you get news from in the past week? (Yes/No)
  • About 5 sources per respondent
• Sources coded as left/Democrat-favoring, right/Republican-favoring, or non-partisan
• Respondents self-report partisanship

Level of expected segregation depends on group size

• In scenario with random choosing, larger groups more segregated
• Using bootstrapping to create null distribution
  • Randomly assign sources to each R
  • Hold # sources constant for each R
  • Weight probability of selecting source by population distribution
• Compare observed SSI to null

• SSI shows interesting within- and between-party variation in segregation
• Can be used as individual-level variable for theory testing
• Data requirements:
  • List of sources used by respondents (more is better)
  • Party ID
• Inference procedure probably not needed
  • Null SSI \(\approx\) group’s population proportion
• Can be directly compared to:
  • interpersonal (social) network segregation
  • other countries, media systems, etc.