• Tidak ada hasil yang ditemukan

Behavioral Implications of Causal Misperceptions Part III

N/A
N/A
Protected

Academic year: 2024

Membagikan "Behavioral Implications of Causal Misperceptions Part III"

Copied!
30
0
0

Teks penuh

(1)

Behavioral Implications of Causal Misperceptions Part III

Ran Spiegler (TAU & UCL)

ES Winter School, Delhi

December 2019

(2)

β€’ So far we have taken agents’ causal model as given.

β€’ Let us now endogenize it.

β€’ Various forces that generate the agent’s causal model:

– Extrapolating from observations

– Selecting between models offered (opportunistically) by others, according to their anticipatory value

Endogenous Causal Models

(3)

β€’ Political beliefs are shaped by narratives.

β€’ Battles over public opinion involve β€œcompeting narratives”.

β€’ A policy is popular when backed by an attractive narrative.

β€’ People are attracted to hopeful narratives (stories with a β€œhappy ending”…).

Political Narratives

(4)

"Barack Obama was a writer before he became a politician, and he saw his Presidency as a struggle over narrative.”

The New Yorker, 2018

β€œThere can be little doubt then that people think narratives are important and that crafting, manipulating, or influencing

them likely shapes public policy …by telling a story that includes assertions about what causes what, who the victims are,

who is causing the harm, and what should be done.”

LSE Impact blog, 2018

(5)

β€’ Eliaz-Spiegler 2018: Conceptualizing narratives as causal models that map policies to public consequences.

β€’ An equilibrium model of how narrative-policy pairs compete for public popularity.

Narratives as Causal Models

(6)

β€’ 𝑋𝑋 = {0,1}π‘šπ‘š, π‘šπ‘š > 2

β€’ π‘₯π‘₯1 (also denoted π‘Žπ‘Ž) represents an action.

β€’ π‘₯π‘₯π‘šπ‘š (also denoted 𝑦𝑦) represents a consequence.

β€’ 𝑝𝑝 ∈ βˆ†(𝑋𝑋) is a joint distribution over π‘₯π‘₯1, … ,π‘₯π‘₯π‘šπ‘š.

The Model

(7)

β€’ 𝑝𝑝 π‘Žπ‘Ž = 1 = 𝛼𝛼

– Historical action frequency, endogenized later.

β€’ 𝑝𝑝 𝑦𝑦 = 1 = πœ‡πœ‡ ∈ (0,1), exogenously

– Exogenous and independent of π‘Žπ‘Ž.

β€’ (𝑝𝑝 π‘₯π‘₯2, … ,π‘₯π‘₯π‘šπ‘šβˆ’1 π‘Žπ‘Ž, 𝑦𝑦 ) is exogenous with full support for any π‘Žπ‘Ž, 𝑦𝑦.

β€’ A narrative is a DAG (𝑁𝑁, 𝑅𝑅), where:

– 𝑁𝑁 βŠ† {1, … , π‘šπ‘š} is the set of nodes, 𝑁𝑁 ≀ 𝑛𝑛 ≀ π‘šπ‘š.

The Model

(8)

Narratives as Causal Models

β€’ The set of feasible narratives is a set of DAGs β„› that include the nodes 1 and π‘šπ‘š, such that 1 is an ancestral node.

β€’ 𝑛𝑛 is an exogenous limit on DAG size.

β€’ When 2 < 𝑛𝑛 < π‘šπ‘š, the narrative selects variables (in addition to the action and the consequence).

β€’ I identify the DAG with 𝑅𝑅 unless specified otherwise.

(9)

β€’ 𝑑𝑑 ∈ [πœ€πœ€, 1 βˆ’ πœ€πœ€] is a policy.

– Proposed frequency of playing π‘Žπ‘Ž = 1

β€’ A representative agent’s utility is 𝑦𝑦 βˆ’ 𝐢𝐢(𝑑𝑑 βˆ’ π‘‘π‘‘βˆ—).

β€’ π‘‘π‘‘βˆ— ∈ [0.5,1 βˆ’ πœ€πœ€) is the agent’s intrinsically ideal policy.

β€’ 𝐢𝐢 is a symmetric, convex cost function, with 𝐢𝐢 0 = 𝐢𝐢′ 0 = 0.

Policies

(10)

β€’ The narrative-policy pair (𝑅𝑅, 𝑑𝑑) induces a gross anticipatory expected payoff (perceived probability of a good outcome):

𝑉𝑉 𝑅𝑅, 𝑑𝑑 𝛼𝛼 = 𝑑𝑑 οΏ½ 𝑝𝑝𝑅𝑅 𝑦𝑦 = 1 π‘Žπ‘Ž = 1 + (1 βˆ’ 𝑑𝑑) οΏ½ 𝑝𝑝𝑅𝑅 𝑦𝑦 = 1 π‘Žπ‘Ž = 0

β€’ The notation highlights that 𝑝𝑝𝑅𝑅 𝑦𝑦|π‘Žπ‘Ž - and therefore 𝑉𝑉 - may depend on the steady-state frequency 𝛼𝛼.

β€’ How does the representative agent choose between conflicting narrative-policy pairs?

Anticipatory Payoff

(11)

Definition: An action frequency 𝛼𝛼 and a distribution 𝜎𝜎 over narrative-policy pairs 𝑅𝑅, 𝑑𝑑 form an equilibrium if:

1. 𝑆𝑆𝑆𝑆𝑝𝑝𝑝𝑝 𝜎𝜎 βŠ† π‘Žπ‘Žπ‘Žπ‘Žπ‘Žπ‘Žπ‘šπ‘šπ‘Žπ‘Žπ‘₯π‘₯ 𝑅𝑅,𝑑𝑑 βˆˆβ„›Γ—[0,1] 𝑉𝑉 𝑅𝑅, 𝑑𝑑 𝛼𝛼 βˆ’ 𝐢𝐢 𝑑𝑑 βˆ’ π‘‘π‘‘βˆ— 2. 𝛼𝛼 = βˆ‘(𝑅𝑅,𝑑𝑑) 𝜎𝜎(𝑅𝑅, 𝑑𝑑) οΏ½ 𝑑𝑑

Proposition: An equilibrium exists. (Proof method: Construct an auxiliary game that has a Nash equilibrium.)

Equilibrium

(12)

β€’ Suppose β„› only consists of the DAG 𝑅𝑅: π‘Žπ‘Ž β†’ 𝑦𝑦.

β€’ Then, 𝑝𝑝𝑅𝑅 𝑦𝑦 = 1 π‘Žπ‘Ž = πœ‡πœ‡ for all π‘Žπ‘Ž (rational expectations).

– 𝑉𝑉 𝑅𝑅, 𝑑𝑑 𝛼𝛼 = πœ‡πœ‡ for every feasible 𝑅𝑅, 𝑑𝑑 .

– In equilibrium, 𝜎𝜎 assigns probability one to 𝑑𝑑 = π‘‘π‘‘βˆ—.

Rational-Expectations Benchmark

(13)

β€’ π‘šπ‘š = 3

β€’ The three variables are π‘Žπ‘Ž, 𝑦𝑦, 𝑠𝑠:

– π‘Žπ‘Ž = 1 Hawkish policy toward rival country – 𝑦𝑦 = 1 Weak regime in rival country

– 𝑠𝑠 = 1 Strong nationalistic sentiment in rival country

An Example: Foreign-Policy Narratives

(14)

β€’ Long-run distribution 𝑝𝑝 (with full support) over the variables – 𝑝𝑝 π‘Žπ‘Ž = 1 = 𝛼𝛼

– 𝑝𝑝 𝑦𝑦 = 1 = 12 , independently of π‘Žπ‘Ž

Foreign Policy

(15)

Foreign Policy

𝑝𝑝 𝑠𝑠 = 1 π‘Žπ‘Ž, 𝑦𝑦 = π‘Žπ‘Ž+1βˆ’π‘¦π‘¦2

β€’ Hawkish policy tends to strengthen nationalism in the rival country.

β€’ Nationalism and regime weakness are negatively correlated.

– This correlation is not causal; it is due to latent confounding variables.

(16)

Examples of Narratives

β€’ π‘Žπ‘Ž β†’ 𝑠𝑠 β†’ 𝑦𝑦 is a β€œlever narrative” that incorporates 𝑠𝑠 into the story as a mediator.

β€’ π‘Žπ‘Ž β†’ 𝑦𝑦 ← 𝑠𝑠 is a β€œthreat/opportunity narrative” that incorporates 𝑠𝑠 as an exogenous force.

(17)

β€’ 𝐢𝐢 ≑ 0

β€’ Therefore, the agent’s anticipatory utility is 𝑉𝑉(𝑅𝑅, π‘Žπ‘Ž;𝛼𝛼) = 𝑝𝑝𝑅𝑅 𝑦𝑦 = 1 π‘Žπ‘Ž

Anticipatory Payoff

(18)

Unique Equilibrium

β€’ 𝛼𝛼 β‰ˆ 0.57

β€’ 𝑆𝑆𝑆𝑆𝑝𝑝𝑝𝑝(𝜎𝜎) consists of two narrative-action pairs:

1. Lever narrative 𝑅𝑅𝐿𝐿: π‘Žπ‘Ž β†’ 𝑠𝑠 β†’ 𝑦𝑦 coupled with the dovish action π‘Žπ‘Ž = 0.

2. Opportunity narrative 𝑅𝑅𝑂𝑂: π‘Žπ‘Ž β†’ 𝑦𝑦 ← 𝑠𝑠 coupled with the hawkish action π‘Žπ‘Ž = 1.

(19)

Why does 𝑅𝑅 𝐿𝐿 Support a Dovish Policy?

𝑝𝑝 𝑠𝑠 = 1 π‘Žπ‘Ž, 𝑦𝑦 = π‘Žπ‘Ž + 1 βˆ’ 𝑦𝑦 2

β€’ (π‘Žπ‘Ž, 𝑠𝑠) are positively correlated

β€’ (𝑠𝑠, 𝑦𝑦) are negatively correlated

β€’ π‘Žπ‘Ž β†’ 𝑠𝑠 β†’ 𝑦𝑦 induces a (perceived, indirect) negative causal effect of π‘Žπ‘Ž on 𝑦𝑦.

(20)

Why does 𝑅𝑅 𝑂𝑂 Support a Hawkish Policy?

𝑝𝑝 𝑠𝑠 = 1 π‘Žπ‘Ž, 𝑦𝑦 = π‘Žπ‘Ž+1βˆ’π‘¦π‘¦2

β€’ 𝑝𝑝 𝑠𝑠 π‘Žπ‘Ž, 𝑦𝑦 is a function of π‘Žπ‘Ž βˆ’ 𝑦𝑦.

β‡’ For any given 𝑠𝑠, 𝑝𝑝 𝑦𝑦 = 1 π‘Žπ‘Ž, 𝑠𝑠 increases with π‘Žπ‘Ž.

β€’ But 𝑝𝑝𝑅𝑅𝑂𝑂 regards 𝑠𝑠 as independent of π‘Žπ‘Ž.

β€’ Therefore, 𝑝𝑝𝑅𝑅𝑂𝑂(𝑦𝑦 = 1|π‘Žπ‘Ž) increases with π‘Žπ‘Ž.

(21)

Why Two Narratives?

β€’ β€œDiminishing returns” to belief manipulation

– 𝑉𝑉(𝐺𝐺,π‘Žπ‘Ž;𝛼𝛼) goes down as 𝛼𝛼 moves toward π‘Žπ‘Ž.

– As the action that one narrative promotes gets implemented more frequently, the competing narrative becomes more appealing.

β€’ Other DAGs induce rational expectations (can’t sell illusions).

1 𝛼𝛼

0 1

2 𝛼𝛼

π‘Žπ‘Ž β†’ 𝑦𝑦 ← 𝑠𝑠 π‘Žπ‘Ž β†’ 𝑠𝑠 β†’ 𝑦𝑦

(22)

Hawkish Bias

β€’ Objective distribution treats π‘Žπ‘Ž = 1 and π‘Žπ‘Ž = 0 symmetrically.

β€’ And yet, 𝛼𝛼 > 12 : Average equilibrium policy is hawkish.

β€’ Why?

– At 𝛼𝛼 = 12 , 𝑅𝑅𝑂𝑂 generates higher false correlation between π‘Žπ‘Ž and 𝑦𝑦 than 𝑅𝑅𝐿𝐿. Then, 𝛼𝛼 needs to go up to restore

indifference between the two narratives.

(23)

β€’ Tension between β€œeasy fix” and rational-expectations narratives

– True narrative: The consequence is determined by a β€œdeep cause”

that is impervious to policy

– The β€œEasy fix” narrative falsely attributes the consequence to a symptom, which is affected by policy.

– When these are the feasible narratives, both prevail in equilibrium;

the false narrative feeds off the true one.

Another Example: β€œEasy-Fix Narratives”

(24)

Proposition: Let β„› be a set of perfect DAGs that contains a DAG 𝑅𝑅 such that 𝑝𝑝𝑅𝑅(𝑦𝑦|π‘Žπ‘Ž) is non-constant in π‘Žπ‘Ž.

Then, in any equilibrium (𝛼𝛼, 𝜎𝜎), 𝜎𝜎 assigns positive probability to exactly two policies: 𝑑𝑑1 β‰₯ π‘‘π‘‘βˆ— and 𝑑𝑑0 ≀ π‘‘π‘‘βˆ—.

β€’ The proof makes heavy use of the β€œno distortion of marginals”

property of perfect DAGs - 𝑉𝑉 𝑅𝑅, 𝛼𝛼 𝛼𝛼 = πœ‡πœ‡

General Result: Equilibrium Polarization

(25)

β€’ Suppose 𝜎𝜎 assigns probability one to 𝛼𝛼.

β€’ 𝑉𝑉 𝑅𝑅, 𝛼𝛼 𝛼𝛼 = πœ‡πœ‡ for any feasible narrative 𝑅𝑅.

– 𝛼𝛼 = π‘‘π‘‘βˆ—: Deviate to lever narrative with 𝑑𝑑 β‰ˆ π‘‘π‘‘βˆ—.

– 𝛼𝛼 β‰  π‘‘π‘‘βˆ—: Deviate to π‘Žπ‘Ž β†’ 𝑦𝑦 with π‘‘π‘‘βˆ—.

β€’ Therefore, there must be mixing over policy.

β€’ No-marginal-distortion implies π‘šπ‘šπ‘Žπ‘Žπ‘₯π‘₯𝑅𝑅𝑉𝑉 𝑅𝑅, 𝑑𝑑 𝛼𝛼 only depends on whether 𝑑𝑑 > 𝛼𝛼 or 𝑑𝑑 < 𝛼𝛼; this gives two-policy equilibrium.

Proof Strategy

(26)

Competing Narratives: Comments

β€’ Open question: When does the two-narrative result hold when we allow for imperfect DAGs?

β€’ Extension: State-dependent evaluation of narratives (example:

denialist vs. exaggerationist narratives)

β€’ Eliaz-Spiegler-Weiss 2019: What is the maximal predicted correlation that a misspecified DAG can generate?

(27)

Other Relevant Works

β€’ Schumacher-Thysen 2017: Contracting with an agent who has a

wrong model of the mapping from effort to output/cost

β€’ Spiegler 2018, Eliaz-Spiegler-Thysen 2018: A dual approach

(Bayesian-network factorization as a representation of extrapolating a belief from partial statistical data)

(28)

Conclusion

β€’ A framework for equilibrium models with non-rational

expectations, highlighting the role of subjective causal perceptions

β€’ DAGs represent causal models and capture systematic errors of causal attribution

– Bayesian-network tools (equivalent DAGs, perfection, d- separation) for analyzing behavioral implications

(29)

Conclusion

β€’ Purely graphical (and hence non-parametric) characterizations of behavioral features:

– Equilibrium effects in individual choice – Systematically biased estimates

– Multiple prevailing narratives when the criterion for selecting narratives is based on anticipatory utility

(30)

Research Directions

β€’ How do decision makers form causal models?

β€’ Connections to other approaches to decision making under misspecified models

β€’ Applications (macro, IO, political economics, contract theory…)

β€’ Complexity considerations in probabilistic inference

β€’ Boundedly rational causal inference

Referensi

Dokumen terkait