The Politics of News Personalization Lin Hu 1 Anqi Li 2 Ilya Segal 3 - - PowerPoint PPT Presentation

The Politics of News Personalization

Lin Hu 1 Anqi Li 2 Ilya Segal 3

1Australian National University 2Washington University in St. Louis 3Stanford University

October 2019

Changing News Landscape

Increasing online news consumption via social media and mobile devices: In 2016, 40% Americans frequently consulted online news sources, 62% got news on social media and 18% did so often In 2017, 85% U.S. adults got news on mobile devices In 2018, social media outpaced print newspapers as a news source

Rise of News Aggregators

Aggregator sites, social media feeds, mobile news apps: Gather tons of users’ personal data (demographic attributes, digital footprints, social network positions) Personalized news aggregation in exchange for user attention Use and impact: Google News aggregated contents from more than 25,000 publishers in 2013 The top 3 popular news websites in 2019: Yahoo! News, Google News and Huffington Post, are aggregators Social media feeds in 2016 U.S. presidential election

Potential Impact on Politics

For too many of us, it’s become safer to retreat into our own bubbles, ...especially our social media feeds, surrounded by people who look like us and share the same political outlook and never challenge our assumptions... And increasingly, we become so secure in our bubbles that we start accepting only information, whether it’s true or not, that fits our opinions, instead of basing

• ur opinions on the evidence that is out there.

—Barack Obama, farewell address, January 10, 2017

Research Questions

What kind of personalized news is aggregated for and consumed by rational inattentive voters in equilibrium? How does news personalization affect policy polarization in a model of electoral competition?

Agenda

• 1. Model

News aggregation Electoral competition

• 2. Extensions
• 3. Literature

Agenda

• 1. Model

News aggregation

Setup Optimal news signal

Electoral competition

• 2. Extensions
• 3. Literature

Political Players

Two candidates L and R: Office-motivated Policy space: A = [−a, a] Policy profile: a = aL, aR, fixed to any −a, a, a ≥ 0 for now A unit mass of voters: Types: K = {−1, 0, 1} Population function: q : K → R+, q (−k) = q (k) Valuation of policies: u (a, k) = −|t(k) − a|, t : K → R is strictly increasing and t (k) = −t (−k)

Expressive Voting

Utility difference from choosing candidate R over candidate L: v (a, k) + ω where v (a, k) = u (aR, k) − u (aL, k) ω: valence state about fitness for office:

E.g., whether the state favors experience with the use of hard

• r soft power

Equal ±1 with prob. .5

News Aggregation

A monopolistic infomediary partitions K into market segments using segmentation technology S: Broadcast news: b = {K} Personalized news: p = {{k} : k ∈ K} Aggregates ω into |S| news signals, one for each market segment A news signal Π : Ω → ∆ (Z) is a finite signal structure: Z: set of news realizations Π (· | ω): probability distribution over Z conditional on the state being ω

News Consumption

Each voter can either consume the news signal offered to him or abstain Consume news = absorb the information contained in the news signal: Potential gain from improved expressive voting Attention cost: λ · I (Π) Infomediary’s gross profit = total amount of attention paid by voters

Model Discussion: News Signal

Under signal structure Π : Ω → ∆ (Z), πz: prob. that the news realization is z µz: posterior mean of the state given news realization z

Strictly prefer candidate R to L iff v (a, k) + µz > 0 —————– candidate L to R iff v (a, k) + µz < 0

Bayes’ plausibility:

• z∈Z

πz · µz = 0 The infomediary can commit to any signal structure

Model Discussion: Attention Cost

Assumption 1.

The needed attention level for consuming Π : Ω → ∆ (Z) is I (Π) =

• z∈Z

πz · h (µz) , where h : [−1, 1] → R+ satisfies the following properties: (i) h (0) = 0 and strict convexity; (ii) continuity on [−1, 1] and twice differentiability on (−1, 1); (iii) symmetry around zero. E.g., h (µ) = µ2; h(µ) = H

• 1+µ

2

• , H = binary entropy function

Model Discussion: Miscellaneous

Voter’s inflexibility Attention-based business model Ability to personalize

Agenda

• 1. Model

News aggregation

Setup Optimal news signal

Electoral competition

• 2. Extensions
• 3. Literature

Optimal News Signal

Expected utility gain from news consumption: V (Π; a, k) =

      

• z∈Z

πz [v (a, k) + µz]+ if k ≤ 0 −

• z∈Z

πz [v (a, k) + µz]− if k > 0 Under segmentation technology S, any optimal news signal of market segment s ∈ S solves max

Π I (Π) ·

 

• k∈K:V (Π;a,k)≥λ·I(Π)

q (k, s)

 

(s)

Binary Recommendations and Strict Obedience

For binary news signals, write Z = {L, R} and assume w.l.o.g. that µL < 0 < µR A binary news signal induces strict obedience if the following holds among its consumers: v (a, k) + µL < 0 < v (a, k) + µR (SOB)

Binary Recommendations and Strict Obedience (Cont’d)

Lemma 1.

Fix any symmetric policy profile −a, a, a ≥ 0 and assume Assumption 1. Then, (i) any optimal broadcast news signal is either degenerate or binary; (ii) any optimal personalized news signal of any type of voters is either degenerate or binary; (iii) any optimal news signal, if binary, induces strict obedience.

Uniqueness

Lemma 2.

Fix any symmetric policy profile −a, a, a ≥ 0 and assume Assumption 1. Then, (i) in the broadcast case, if it is optimal to induce consumption from all voters, then the optimal news signal is unique; (ii) the optimal personalized news signal of any type of voters is unique.

Regularity Condition

Assumption 2.

Under any symmetric policy profile −a, a, a ≥ 0, (i) any optimal news signal is nondegenerate, and the posterior means of the state conditional on its realizations belong to the

• pen interval (−1, 1);

(ii) it is optimal to induce consumption from all voters in the broadcast case.

Notations

Under segmentation technology S: ΠS (a, k): optimal news signal consumed by type k voters µS

z (a, k): the posterior mean of the state given news

realization z ∈ {L, R} πS (a, k) = −

µS

L (a,k)

µS

R (a,k)−µS L (a,k): prob. that candidate R is

endorsed Suppress the notation of k if S = b

Own-Party Bias and Occasional Big Surprise

△ □ ◇ ☒ △ News signal (B) □ News signal for k<0 (P) ◇ News signal for k=0 (P) ☒ News signal for k>0 (P)

• 1
• 0.8
• 0.6
• 0.4
• 0.2

0.2 0.4 0.6 0.8 1 μL μR

Figure 1: Optimal news signals

Own-Party Bias and Occasional Big Surprise (Cont’d)

Theorem 1.

Fix any symmetric policy profile −a, a, a ≥ 0 and assume Assumptions 1 and 2. Then, (i) πb (a) = 1/2 and µb

L (a) + µb R (a) = 0;

(ii) ∀k ∈ K, µp

L (a, −k) + µp R (a, k) = 0, and

(a) πp (a, k) < 1/2 and µp

L (a, k) + µp R (a, k) > 0 if k < 0;

(b) πp (a, k) = 1/2 and µp

L (a, k) + µp R (a, k) = 0 if k = 0;

(c) πp (a, k) > 1/2 and µp

L (a, k) + µp R (a, k) < 0 if k > 0;

(iii) I (Πp (a, k)) > I

• Πb (a)
• ∀k ∈ K.

Agenda

• 1. Model

News personalization Electoral Competition

• 2. Extensions
• 3. Literature

Game Sequence

• 1. The infomediary commits to news signals

2.

a Voters decide whether to consume news or not b Candidates propose policies

• 3. State is realized
• 4. Voters observe signal realizations and policies and vote

expressively; winner is determined by simple majority rule with even tie-breaking

Equilibrium

Under segmentation technology S, a policy profile −a, a and news profile µ can be attained in a PBE if

• µ is a |S|-dimensional random variable, where the marginal

distribution of each dimension s ∈ S solves problem (s), taking −a, a as given a maximizes candidate R’s winning probability, taking µ, candidate L’s policy −a and voters’ behaviors in stages 2(a) and 4 of the game as given

Remark 1.

Assume for now that news signals are conditionally independent across market segments.

Agenda

• 1. Model

News personalization Electoral Competition

Key concepts Main characterization Comparative statics

• 2. Extensions
• 3. Literature

Key Concepts

A deviation a′ by candidate R from −a, a to a′ attracts type k voters if v

−a, a′, k + µs

L (a, k) > 0

and it repels type k voters if v

−a, a′, k + µs

R (a, k) < 0

If a′ does not attract or repel type k voters, then it does not affect the latter’s voting decisions

Key Concepts (Cont’d)

Define the k-proof set by ΞS (k) =

• a ≥ 0 : v (−a, t (k) , k) + µS

L (a, k) ≤ 0

• and type k voters’ policy latitude by

ξS (k) = max ΞS (k)

Key Concepts (Cont’d)

Under segmentation technology S and population function q, Let ES,q denote the set of policy a’s such that the symmetric policy profile −a, a can arise in equilibrium Define aS,q = max ES,q as the degree of policy polarization Type k voters are disciplining if aS,q = ξS (k)

Agenda

• 1. Model

News personalization Electoral Competition

Key concepts Main characterization Comparative statics

• 2. Extensions
• 3. Literature

Main Characterization

Theorem 2.

Assume Assumptions 1 and 2. Then under all segmentation technology S ∈ {b, p} and population function q, ES,q =

• 0, aS,q

and aS,q > 0. In particular, (i) ab,q = ξb (0) ∀q; (ii) ap,q =

  

ξp (0) if q (0) > 1/2, min

k∈K ξp (k)

if q (0) ≤ 1/2.

Proof Sketch: Broadcast News

Since all voters receive the same voting recommendation, a deviation by candidate R is profitable ⇐ ⇒ it attracts a majority of voters ⇐ ⇒ it attracts median voters Thus median voters are always disciplining, i.e., Eb,q = Ξb (0) ∀q

Proof Sketch: Personalized News

In the case where q (0) ≤ 1/2, a deviation is profitable if it attracts any type k voters, holding other things constant Conditional independence implies that the above deviation strictly increases candidate R’s winning probability in the event where type k voters are pivotal Interestingly, a policy profile can be attained in equilibrium if the above deviation is unprofitable. In fact...

Proof Sketch: Personalized News

Lemma 3.

Assume Assumptions 1 and 2. Then the following are equivalent in the case where S = p and q (0) ≤ 1/2: (i) −a, a, a ≥ 0 can be attained in equilibrium; (ii) no unilateral deviation of candidate R to any a′ ∈ [−a, a] attracts any voter whose bliss point lies in [−a, a]. Thus Ep,q = A (0) ∪ A (1), where A (0) = [0, t (1)) ∩ Ξp (0) A (1) = [t (1) , a] ∩

• k∈K

Ξp (k)

Proof Sketch: Final Steps

Strict obedience = ⇒ aS,q > 0 ∀S, q Characterizing policy latitudes establishes the interval property and pins down the disciplining voter

Takeaway

News personalization makes attracting any type of voters—albeit non-majorities—a profitable deviation Voters with the smallest policy latitude are the most susceptible to policy deviations and therefore constitute the disciplining entity for equilibrium polarization Deviations could be more effective in the personalized case than in the broadcast case

Who are Disciplining Under Personalized News?

Lemma 4.

When a is large, (i) if ξb (0) ≥ t (1), then ξb (0) = −µb

L := µb L (t (1));

(ii) ξp (k) = −

t (k) + µp

L (k)

, where µp

L (k) := µp L (|t (k) |, k).

Policy preference vs. belief about fitness: Right-wing (base) voters most prefer candidate R policy-wise but are the most pessimistic when news is unfavorable The opposite is true for left-wing (opposition) voters Base voters have a bigger policy latitude than opposition voters if and only if µp

L (1) + µp R (1) < −2t (1) .

(∗)

Agenda

• 1. Model

News personalization Electoral Competition

Key concepts Main characterization Comparative statics

• 2. Extensions
• 3. Literature

The Politics of News Personalization

Proposition 1.

Fix any population function q, assume Assumptions 1 and 2 and let a be large. Then news personalization strictly increases policy polarization if and only if under personalized news, one of the following conditions hold: (i) median voters are disciplining; (ii) extreme voters are disciplining and have a bigger policy latitude than the median voters hearing broadcast news, i.e., ξb (0) < min {ξp (1) , ξp (−1)} . (∗∗)

The Politics of News Personalization (Cont’d)

Proposition 1 (cont’d).

Condition (∗∗) holds if ξb (0) < t (1). When ξb (0) ≥ t (1), (a) if right-wing voters are disciplining under personalized news, i.e., (∗) is violated, then (∗∗) is equivalent to µp

L (1) − µb L < −t (1) ;

(b) if left-wing voters are disciplining under personalized news, i.e., (∗) is satisfied, then (∗∗) is equivalent to t (−1) < µb

L − µp L (−1) .

Illustrative Example

Example 1.

In the case where h (µ) = µ2, we have that ξb (0) =

• 1 +
• 1 − 16λt (1)
• / (4λ) > t (1) ,

and that ξp (k) =

      

1/ (2λ) − 3t (1) if k = −1, 1/ (2λ) if k = 0, 1/ (2λ) − t (1) if k = 1. Thus, Left-wing voters have the smallest policy latitude, followed by right-wing voters and then median voters Personalization increases policy polarization if and only if λt (1) > 1/18

Mass Polarization vs. Elite Polarization

Define increasing mass polarization by adding mean-preserving spreads to voters’ type distribution as suggested by Fiorina and Abrams (2008) and Gentzkow (2016)

Proposition 2.

Assume Assumptions 1 and 2. Then ap,q ≥ ap,q′ for all population functions q and q′ such that q (0) > q′ (0), and the inequality is strict if and only if q (0) > 1/2 ≥ q′ (0) and 0 / ∈ arg min

k∈K

ξp (k).

Marginal Attention Cost and Regulatory Implications

Allow perfect competition between infomediaries ⇐ ⇒ increase λ in the personalized case = ⇒ reduce policy polarization

Proposition 3.

Assume Assumption 1 and take any λ′ > λ > 0 that satisfy Assumption 2. Then for all a ≥ 0 and k ∈ K, we have that µp

L (a, k, λ) < µp L (a, k, λ′) and that µp R (a, k, λ) > µp R (a, k, λ′).

Agenda

• 1. Model
• 2. Extensions
• 3. Literature

General Model

In general, the set of policy a’s that can be attained in equilibrium is

• 0,

min

C′s formed under χ,q ξS (C)

• where

C: influential coalition

example

ξS (C): policy latitude of influential coalition C χ: news configuration

example

Thus, Joint news distribution affects polarization through χ, whereas marginal distributions do so through ξS (·) Enriching influential coalitions reduces polarization, holding

• ther things constant

General Model (Cont’d)

In the personalized case, Relaxing conditional independence can only increase polarization mink∈K ξp (k) is the exact lower bound for the polarization that can be attained across all scenarios Skewness is crucial for personalization to increase polarization

Other Extensions

General state distribution Skewness vs. level effect Alternative candidate motive · · ·

Literature

Media bias:

Prat and Str¨

• mberg (2013), Str¨
• mberg (2015), Anderson et al.

(2016) Mullainathan and Shleifer (2005), Bernhardt et al. (2008), Gentzkow and Shapiro (2010), Martin and Yurukoglu (2017) Calvert (1985a), Suen (2004), Burke (2008), Che and Mierendorff (2018)

Own-party bias and occasional big surprise:

Fiorina and Abrams (2008), Barber and McCarty (2015), Gentzkow (2016) Chiang and Knight (2011), Flaxman et al. (2016) DellaVigna and Gentzkow (2010)

Literature (Cont’d)

Rational inattention:

Sims (1998, 2003), Matˇ ejka and Mckay (2015), Caplin (2016), Ma´ ckowiak et al. (2018) Caplin and Dean (2015), Zhong (2017), Denti (2018), Tsakas (2019), Caplin et al. (2019); H´ ebert and Woodford (2017), Morris and Strack (2017); Dean and Nelighz (2019) Matˇ ejka and Tabellini (2016)

Media as flexible and profit-maximizing information channel:

Str¨

• mberg (2004), Chan and Suen (2008), Yuksel and Perego (2018)

Literature (Cont’d)

Economics of news aggregators: Athey and Mobius (2012), Athey

et al. (2017), Chiou and Tucker (2017), Jeon (2018)

Strict obedience vs. continuous signal distribution: Calvert

(1985b), Duggan (2000), Patty (2005), Duggan (2017)

Influential Coalitions

S = b S = p q(0) > 1/2 majorities majorities q(0) < 1/2 majorities 2K − ∅ Table 1: influential coalitions under any symmetric policy profile −a, a, a ≥ 0: baseline model.

back

News Configuration (Cont’d)

When S = b, χ∗ :=

     

1 1 . . . . . . 1

     

News Configuration (Cont’d)

When S = p and signals are conditionally independent, χ∗∗ :=

       

1 · · · 1 · · · · · · 1 1 · · · 1 · · · · · · 1 . . . . . . . . . · · · . . . . . . · · · . . . · · · 1 · · · · · · 1 · · · 1 · · · 1 · · · 1 · · · 1

       

• 2|K| columns

back