Behavioral Implications of Causal Misperceptions Part III
Ran Spiegler (TAU & UCL)
ES Winter School, Delhi
December 2019
β’ So far we have taken agentsβ causal model as given.
β’ Let us now endogenize it.
β’ Various forces that generate the agentβs causal model:
β Extrapolating from observations
β Selecting between models offered (opportunistically) by others, according to their anticipatory value
Endogenous Causal Models
β’ Political beliefs are shaped by narratives.
β’ Battles over public opinion involve βcompeting narrativesβ.
β’ A policy is popular when backed by an attractive narrative.
β’ People are attracted to hopeful narratives (stories with a βhappy endingββ¦).
Political Narratives
"Barack Obama was a writer before he became a politician, and he saw his Presidency as a struggle over narrative.β
The New Yorker, 2018
βThere can be little doubt then that people think narratives are important and that crafting, manipulating, or influencing
them likely shapes public policy β¦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.β
LSE Impact blog, 2018
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
β’ ππ = {0,1}ππ, ππ > 2
β’ π₯π₯1 (also denoted ππ) represents an action.
β’ π₯π₯ππ (also denoted π¦π¦) represents a consequence.
β’ ππ β β(ππ) is a joint distribution over π₯π₯1, β¦ ,π₯π₯ππ.
The Model
β’ ππ ππ = 1 = πΌπΌ
β Historical action frequency, endogenized later.
β’ ππ π¦π¦ = 1 = ππ β (0,1), exogenously
β Exogenous and independent of ππ.
β’ (ππ π₯π₯2, β¦ ,π₯π₯ππβ1 ππ, π¦π¦ ) is exogenous with full support for any ππ, π¦π¦.
β’ A narrative is a DAG (ππ, π π ), where:
β ππ β {1, β¦ , ππ} is the set of nodes, ππ β€ ππ β€ ππ.
The Model
Narratives as Causal Models
β’ The set of feasible narratives is a set of DAGs β that include the nodes 1 and ππ, such that 1 is an ancestral node.
β’ ππ is an exogenous limit on DAG size.
β’ When 2 < ππ < ππ, the narrative selects variables (in addition to the action and the consequence).
β’ I identify the DAG with π π unless specified otherwise.
β’ ππ β [ππ, 1 β ππ] is a policy.
β Proposed frequency of playing ππ = 1
β’ A representative agentβs utility is π¦π¦ β πΆπΆ(ππ β ππβ).
β’ ππβ β [0.5,1 β ππ) is the agentβs intrinsically ideal policy.
β’ πΆπΆ is a symmetric, convex cost function, with πΆπΆ 0 = πΆπΆβ² 0 = 0.
Policies
β’ The narrative-policy pair (π π , ππ) induces a gross anticipatory expected payoff (perceived probability of a good outcome):
ππ π π , ππ πΌπΌ = ππ οΏ½ πππ π π¦π¦ = 1 ππ = 1 + (1 β ππ) οΏ½ πππ π π¦π¦ = 1 ππ = 0
β’ The notation highlights that πππ π π¦π¦|ππ - and therefore ππ - may depend on the steady-state frequency πΌπΌ.
β’ How does the representative agent choose between conflicting narrative-policy pairs?
Anticipatory Payoff
Definition: An action frequency πΌπΌ and a distribution ππ over narrative-policy pairs π π , ππ form an equilibrium if:
1. ππππππππ ππ β πππππππππππ₯π₯ π π ,ππ ββΓ[0,1] ππ π π , ππ πΌπΌ β πΆπΆ ππ β ππβ 2. πΌπΌ = β(π π ,ππ) ππ(π π , ππ) οΏ½ ππ
Proposition: An equilibrium exists. (Proof method: Construct an auxiliary game that has a Nash equilibrium.)
Equilibrium
β’ Suppose β only consists of the DAG π π : ππ β π¦π¦.
β’ Then, πππ π π¦π¦ = 1 ππ = ππ for all ππ (rational expectations).
β ππ π π , ππ πΌπΌ = ππ for every feasible π π , ππ .
β In equilibrium, ππ assigns probability one to ππ = ππβ.
Rational-Expectations Benchmark
β’ ππ = 3
β’ The three variables are ππ, π¦π¦, π π :
β ππ = 1 Hawkish policy toward rival country β π¦π¦ = 1 Weak regime in rival country
β π π = 1 Strong nationalistic sentiment in rival country
An Example: Foreign-Policy Narratives
β’ Long-run distribution ππ (with full support) over the variables β ππ ππ = 1 = πΌπΌ
β ππ π¦π¦ = 1 = 12 , independently of ππ
Foreign Policy
Foreign Policy
ππ π π = 1 ππ, π¦π¦ = ππ+1βπ¦π¦2
β’ Hawkish policy tends to strengthen nationalism in the rival country.
β’ Nationalism and regime weakness are negatively correlated.
β This correlation is not causal; it is due to latent confounding variables.
Examples of Narratives
β’ ππ β π π β π¦π¦ is a βlever narrativeβ that incorporates π π into the story as a mediator.
β’ ππ β π¦π¦ β π π is a βthreat/opportunity narrativeβ that incorporates π π as an exogenous force.
β’ πΆπΆ β‘ 0
β’ Therefore, the agentβs anticipatory utility is ππ(π π , ππ;πΌπΌ) = πππ π π¦π¦ = 1 ππ
Anticipatory Payoff
Unique Equilibrium
β’ πΌπΌ β 0.57
β’ ππππππππ(ππ) consists of two narrative-action pairs:
1. Lever narrative π π πΏπΏ: ππ β π π β π¦π¦ coupled with the dovish action ππ = 0.
2. Opportunity narrative π π ππ: ππ β π¦π¦ β π π coupled with the hawkish action ππ = 1.
Why does π π πΏπΏ Support a Dovish Policy?
ππ π π = 1 ππ, π¦π¦ = ππ + 1 β π¦π¦ 2
β’ (ππ, π π ) are positively correlated
β’ (π π , π¦π¦) are negatively correlated
β’ ππ β π π β π¦π¦ induces a (perceived, indirect) negative causal effect of ππ on π¦π¦.
Why does π π ππ Support a Hawkish Policy?
ππ π π = 1 ππ, π¦π¦ = ππ+1βπ¦π¦2
β’ ππ π π ππ, π¦π¦ is a function of ππ β π¦π¦.
β For any given π π , ππ π¦π¦ = 1 ππ, π π increases with ππ.
β’ But πππ π ππ regards π π as independent of ππ.
β’ Therefore, πππ π ππ(π¦π¦ = 1|ππ) increases with ππ.
Why Two Narratives?
β’ βDiminishing returnsβ to belief manipulation
β ππ(πΊπΊ,ππ;πΌπΌ) goes down as πΌπΌ moves toward ππ.
β As the action that one narrative promotes gets implemented more frequently, the competing narrative becomes more appealing.
β’ Other DAGs induce rational expectations (canβt sell illusions).
1 πΌπΌ
0 1
2 πΌπΌ
ππ β π¦π¦ β π π ππ β π π β π¦π¦
Hawkish Bias
β’ Objective distribution treats ππ = 1 and ππ = 0 symmetrically.
β’ And yet, πΌπΌ > 12 : Average equilibrium policy is hawkish.
β’ Why?
β At πΌπΌ = 12 , π π ππ generates higher false correlation between ππ and π¦π¦ than π π πΏπΏ. Then, πΌπΌ needs to go up to restore
indifference between the two narratives.
β’ Tension between βeasy fixβ and rational-expectations narratives
β True narrative: The consequence is determined by a βdeep causeβ
that is impervious to policy
β The βEasy fixβ narrative falsely attributes the consequence to a symptom, which is affected by policy.
β When these are the feasible narratives, both prevail in equilibrium;
the false narrative feeds off the true one.
Another Example: βEasy-Fix Narrativesβ
Proposition: Let β be a set of perfect DAGs that contains a DAG π π such that πππ π (π¦π¦|ππ) is non-constant in ππ.
Then, in any equilibrium (πΌπΌ, ππ), ππ assigns positive probability to exactly two policies: ππ1 β₯ ππβ and ππ0 β€ ππβ.
β’ The proof makes heavy use of the βno distortion of marginalsβ
property of perfect DAGs - ππ π π , πΌπΌ πΌπΌ = ππ
General Result: Equilibrium Polarization
β’ Suppose ππ assigns probability one to πΌπΌ.
β’ ππ π π , πΌπΌ πΌπΌ = ππ for any feasible narrative π π .
β πΌπΌ = ππβ: Deviate to lever narrative with ππ β ππβ.
β πΌπΌ β ππβ: Deviate to ππ β π¦π¦ with ππβ.
β’ Therefore, there must be mixing over policy.
β’ No-marginal-distortion implies πππππ₯π₯π π ππ π π , ππ πΌπΌ only depends on whether ππ > πΌπΌ or ππ < πΌπΌ; this gives two-policy equilibrium.
Proof Strategy
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?
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)
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
β Bayesian-network tools (equivalent DAGs, perfection, d- separation) for analyzing behavioral implications
Conclusion
β’ Purely graphical (and hence non-parametric) characterizations of behavioral features:
β Equilibrium effects in individual choice β Systematically biased estimates
β Multiple prevailing narratives when the criterion for selecting narratives is based on anticipatory utility
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 theoryβ¦)
β’ Complexity considerations in probabilistic inference
β’ Boundedly rational causal inference