EITM Institutions Week: Information & Experiments John Aldrich Arthur Lupia
Jan 12, 2016
EITM Institutions Week: Information
& ExperimentsJohn Aldrich
Arthur Lupia
Aldrich Lupia EITM 2007
Can We Trust the Voters?The proposition that [the people] are
the best keeper of their liberties is not true. They are the worst conceivable, they are no keepers at all. They can neither act, judge, think, or will . John Adams, 1788.
“Overall, close to a third of Americans can be categorized as “know-nothings” who are almost completely ignorant of relevant political information – which is not, by any means, to suggest that the other two-thirds are well informed….” Critical Review 1999
Aldrich Lupia EITM 2007
Challenge Voters lack details.
Inferences: Common: Voter incompetence. Recent: Voters adapt.
When can people who lack information vote with competence?
Aldrich Lupia EITM 2007
Insurance ReformQ: Can badly informed voters use elite endorsements
to emulate the behavior of better-informed voters?
Background High and fast increasing rates. Industry anti-trust exempt. Legislative stalemate. Five competing initiatives. Over $80 million spent. Ralph Nader involved.
Aldrich Lupia EITM 2007
Result: Emulation A.Lupia. Am.Pol.Sci.Rev. 88: 63-76. 1994
0 20 40 60 80
Prop 100
Prop 101
Prop 103
Prop 104
Prop 106
LO Info 27 15 26 12.5 45
LO Info +Endorsement
53 5 73 17 12.5
HI Info 53 8 72.5 17 11
Prop 100
Prop 101
Prop 103
Prop 104
Prop 106
Aldrich Lupia EITM 2007
Implications Lack of information lack of competence.
To come:
People choose “short cuts” in predictable ways.
Under what conditions do voters use short cuts effectively?
Aldrich Lupia EITM 2007
The Democratic DilemmaThe Democratic Dilemma
Can individuals who are not “well informed about political affairs” make the same choices they would have made had they been “well informed?”
Aldrich Lupia EITM 2007
Who Believes Whom? Common explanations:
People are sheep. Talk is cheap. Certain attributes required.
E.g., heuristics E.g., reputation, repetition
Common assumptions People know each other. Stimulus/response paradigm. Personal character is the key. External forces do not matter.
Aldrich Lupia EITM 2007
Task and ToolsTask: Clarify the political consequences of limited information.
Tools: An exit poll. Game-theoretic models of communication & choice.* Laboratory experiments. Survey experiments. A comparison to other decision makers.
Key Unresolved Question
What if credibility is endogenous?
Aldrich Lupia EITM 2007
One way to model uncertainty
Aldrich Lupia EITM 2007
The “Standard” Way of Modeling Uncertainty
Is mathematically convenient.
It makes sense if You’ve never cracked the binding of a
introduction to psychology textbook You’ve never attempted to collect data about the
nature of people’s beliefs
Aldrich Lupia EITM 2007
A Modeling Tradeoff Continuous message spaces and very
restrictive assumptions about the nature of uncertainty.
Discrete message spaces and far less restrictive assumptions about uncertainty.
Aldrich Lupia EITM 2007
Theoretical Premises An uncertain voter makes a binary choice. (1d-wlog)
A speaker says “better” or “worse.” He can lie.
The voter is uncertain about the speaker’s interests and knowledge.
External forces may be present. Verification, Penalties for lying, Observable costly effort.
Aldrich Lupia EITM 2007
Model Intuitions Absent external forces, persuasion requires
perceived common interests and perceived speaker knowledge.
The model clarifies how external forces substitute for speaker attributes.
Absent sufficient prior information, competence requires that the perceptions be correct. Institutions can help.
Aldrich Lupia EITM 2007
Basic Model Two players: speaker and receiver.
The receiver chooses x or y. Her choice affects both players’ utility.
Sources of uncertainty. Is x or y better for the receiver?
“better” prior: b[0,1]. Do the speaker and receiver have common interests?
“common” prior: c[0,1] Does the speaker have private information about x?
“knows” prior: k[0,1]
Aldrich Lupia EITM 2007
Aldrich Lupia EITM 2007
Extended Model Penalties for Lying
The speaker pays a penalty for making false statements.
Verification A fourth source of uncertainty emerges. With probability v[0,1], Nature
replaces the speaker’s signal with the true signal.
Costly effort The speaker must pay a positive cost to say anything.
Each exogenous force’s impact on communication strategies and outcomes is determined endogenously.
Aldrich Lupia EITM 2007
Prior BeliefsPrior Beliefs
Aldrich Lupia EITM 2007
The Effect of a Penalty for LyingThe Effect of a Penalty for Lying
Aldrich Lupia EITM 2007
The Effect of Costly EffortThe Effect of Costly Effort
Aldrich Lupia EITM 2007
Aldrich Lupia EITM 2007
Lab Experiments A subject is a voter or advisor.
The voter predicts coin tosses Earns $1/correct prediction.
Advisor: “heads” or “tails.” we vary perceptions:
hidden die rolls determine speaker interests & knowledge.
we vary “institutions.” penalties for lying, costly effort, verification present in selected
trials.
Persuasion Enlightenment Reasoned Choice
When the Conditions for Persuasion and Enlightenment Were Satisfied
Observed 89 1136/1280
93 1057/1136
83 1067/1280
P redicted 100 100 100 When the Conditions for Pers uas ion Were Not S atis fied Obs erved 58
668/1161 57
383/668 52
600/1161 P redicted 50 50 50
When the Conditions for Pers uas ion Were S atis fied and the Conditions for Enlightenment Were Not
Obs erved 90 104/116
50 52/104
53 62/116
P redicted 100 N/A 50
Summary of Results
From Arthur Lupia and Mathew D. McCubbins. The Democratic Dilemma: Can Citizens Learn What They Need to Know? Ch 7.New York: Cambridge University Press.
50 60 70 80 90 100
NO
PFL
Verif.
Persuasion Reasoned Choice
With sufficient penalty for lying or verification, we expect persuasion and reasoned choice. Otherwise, we do not.
From Arthur Lupia and Mathew D. McCubbins. Elements of Reason: Cognition, Choice and the Bounds of Rationality. Ch. 3.New York: Cambridge University Press.
50 60 70 80 90 100
All
None
P only
Persuasion Reasoned Choice
All: Trials where model predicts persuasion and reasoned choice. None: Model predicts none of the above. P only: Model predicts persuasion only.
Aldrich Lupia EITM 2007
CATI Experiment N=1464 “... talk show host [SENDER] [POSITION] spending money to build
more prisons. What do you think? Is spending money to build prisons a good idea or a bad idea?”
“How much would you say that [SENDER] knows about what will happen if this country spends money to build more prisons -- a lot, some, a little, or nothing?”
“On most political issues would you say that you and [SENDER] agree all of the time, most of the time, only some of the time, or never?”
Random Selection
No Speaker20%
CONTEXT 1Limbaugh
20%
CONTEXT 2Donahue
20%
Message Content: Supports
CONTEXT 1Limbaugh
20%
CONTEXT 2Donahue
20%
Message Content: Opposes
RESPONDENT
From Arthur Lupia and Mathew D. McCubbins. The Democratic Dilemma: Can Citizens Learn What They Need to Know? Ch 8.New York: Cambridge University Press.
Metric: (%Yes|Heard Supports) - (%No|Heard Supports)
L to R: declining perceptions of trust and knowledge (K>k); we expect declining effects.
-40
-30
-20
-10
0
10
20
30
40
AK aK Mid ~ak
LimbaughEffect ofTreatment
Donahue Effectof Treatment
Just Ideology?
Factors such as ideology depend on perceived agreement and knowledge. The converse is not true.
Perceived knowledge & trust are the fundamental source effects.
-30-20
-10
010
20
3040
50
AK aK Other
SameIdeology
DifferentIdeology
Aldrich Lupia EITM 2007
Implications Citizens process political information in systematic ways.
Strategic considerations matter.
People can choose competently despite lacking details. True in a range of experimental environments. Would electoral outcomes be different today?
Results imply different solutions. Circulate endorsement information. Make it easier to “follow the money.” Complexity increases requirements.
Were Bush Tax Cut Supporters “Simply Ignorant?” A Second
Look at Conservatives and Liberals in “Homer Gets a Tax
Cut”
Arthur Lupia, Adam Seth Levine, Jesse O. Menning, Gisela Sin
Perspectives on Politics - Nov 2008
Aldrich Lupia EITM 2007
“Homer Gets a Tax Cut” “[T]he strong plurality support for Bush’s tax cut…
is entirely attributable to simple ignorance.” Bartels (2005:24)
Using the same data, we show that this claim is untrue.
Aldrich Lupia EITM 2007
One View of the Impact of Knowledge
All respondents
0.7 0.69 0.67
0
0.2
0.4
0.6
0.8
1
Very Low toAverage (345)
Fairly High(321)
Very High (208)
Interviewer Information Rating (post)
Proportion supporting
Dependent variable: Support for/Opposition to the Tax Cut
Aldrich Lupia EITM 2007
And Another
Republicans
0.893 0.909 0.962
0
0.2
0.4
0.6
0.8
1
Very Low toAverage (150)
Fairly High(175)
Very High (107)
Interviewer Information Rating (post)
Proportion supporting
Over 96% of the most informed Republicans supported the tax cut.
Democrats
0.542
0.3930.348
0
0.2
0.4
0.6
0.8
1
Very Low toAverage (168)
Fairly High(135)
Very High (89)
Interviewer Information Rating (post)
Proportion supporting
Aldrich Lupia EITM 2007
Multivariate Analysis: “Homer” vs. Partisans
Variable “Homer” Dems Reps
Political Information (0 to 1)
-.907(.353)
-1.567(.492)
.102(.461)
Republican Party ID (-1 to 1)
.759(.055)
Family Income (in 1000s)
.0002(.001)
.002(.002)
.000(.001)
“President Bush” wording
-.080(.049)
-.072(.086)
-.032(.051)
Constant .873(.153)
1.013(.308)
.648(.316)
Obs 896 387 418
Dependent variable: Support for/Opposition to the Tax Cut
Aldrich Lupia EITM 2007
Conclusion The “Homer” result is due to:
Presuming that people with different worldviews react to “more information” in the same way.
The fact that liberal opinions on this issue varied far more than those of other respondents.
Examined several ways, it appears that much of the support for the tax cut is attributable to something other than simple ignorance.
Money, Time and the Ability to Answer Political Knowledge
QuestionsMarkus Prior and Arthur Lupia
AJPS - January 2008
Aldrich Lupia EITM 2007
Research Design A two-by-two design. Factor 1: Time
Control: no incentive Treatement: $1 per correct answer
Factor 2: Money Control: 60 seconds per question Treatment: 24 hours to complete all questions
Aldrich Lupia EITM 2007
Research Design Embedded in representative survey
conducted by Knowledge Networks Oct. 19 – Nov. 1, 2004 N = 1,550 (1,220 completes, 79%)
The questions 7 on political issues, 7 on economic issues about information relevant to the 2004 election
Aldrich Lupia EITM 2007
Prior/Lupia resultsNo incentive $1/correct
answer
<60s/Q control 11% increase
<24h/14 Qs 18% increase 24% increase
Aldrich Lupia EITM 2007
Why, and for what class of problems, are Nash-based concepts
satisfactory?
Aldrich Lupia EITM 2007
Question 1
Q: Do people, such as voters, think about politics in ways that popular equilibrium concepts imply.
A: Many do not.
Aldrich Lupia EITM 2007
Question 2 Q: Should differences in reasoning affect game-
theoretic political science?
A: It depends. - In some cases, the “as if” assumption is difficult to justify.
What is the cognitive basis of game-theoretic political science?
Aldrich Lupia EITM 2007
Nash Equilibrium
All players share the same counterfactual assessment along the equilibrium path. Two players cannot disagree about the play of a third.
Gibbons (1992: 8-9)
Aldrich Lupia EITM 2007
SCE Components A finite number of players (iI) have types () and take
actions (a A). i:i(Ai), a probability distribution over actions.
Common knowledge: players know their own utility functions & the extensive form. They need not know others’ strategies or the distribution of
Nature’s type.
Aldrich Lupia EITM 2007
SCE Requirements Condition i states that each player has correct
beliefs about her own type.
Condition ii states that any action that a player plays with positive probability must maximize her utility given her conjecture about other players and Nature.
Condition iii requires that players’ observations be consistent with their conjectures.
Aldrich Lupia EITM 2007
All important beliefs i() is player i’s prior belief about Nature’s
type. Players need not have correct (i=p) or common (i= j)
priors.
yi=yi(a,) is player i’s private signal about the play of the game. The quality of the private signal is a key variable in SCE
models.
Aldrich Lupia EITM 2007
SCE Definition
True type distribution
True strategies
Conjecture about types
Conjecture about actions
Aldrich Lupia EITM 2007
SCE Properties Observations and conjectures must be
consistent.Incorrect conjectures are allowed.
Two players can disagree about a third (or Nature).
Bayesian updating is not required.
More precise observations imply greater constraints.
Aldrich Lupia EITM 2007
Congress-President Standoff
If C or P “stand off”, gov’t shuts down. If gov’t shuts down, V assigns blame. The outcome (~S, ~S) is a SCE.
Each believes that the voter will blame the other for a standoff. The outcome (~S, ~S) is not a BNE.
Any voter strategy induces at least one other actor to continue the standoff.
Adapted from Fudenberg and Kreps 1995.
Aldrich Lupia EITM 2007
Example: Thinking Differently in Jury Theorems Condorcet (1785): There are N jurors. Each chooses guilt or
innocence. One verdict is correct, If each juror votes correctly with probability >.5, the probability that a majority votes correctly goes to 1 as N.
Austen-Smith and Banks (1996): Each juror receives an independent evidentiary signal. If she uses Bayes Rule and has certain kinds of priors about other jurors, voting in accordance with her signal is not an equilibrium. “If I am pivotal, then…”
Aldrich Lupia EITM 2007
Example: Thinking Differently in Jury Theorems Feddersen and Pesendorfer (1785): Extends Austen-Smith
and Banks to unanimous verdicts. Jurors will vote against “innocent” evidentiary signals. Unanimous verdicts do not minimize false convictions.
Guernaschelli, McKelvey, and Palfrey (2000): Experiments of F&P offer mixed support. “Feddersen and Pesendorfer (1998) imply that large unanimous
juries will convict innocent defendants with fairly high probability… this did not happen in our experiment.”
The gap between the theoretical prediction and the experimental data grew with jury size.
Aldrich Lupia EITM 2007
Empirical Studies Pennington and Hastie “story model”
Each juror attempts to make sense of the evidence by assembling it into a narrative format.
“many jurors tended to construct only one of the possible stories,”
“jurors were surprised to discover that there were other possible stories”
Cacioppo and Petty “need for cognition” While some citizens enjoy dealing with logical
abstractions, others strive to minimize the mental effort devoted to such activities.
Aldrich Lupia EITM 2007
Our model (a variant of F&P) There are with N={1,2,…n} jurors.
Let ={G,I}, where =G means that the defendant committed the crime in question and =I means that he did not.
Each evidentiary signal is an independent Bernoulli random variable, mj{g,i}, which, for each juror j, reveals the true value of with probability p(.5,1).
After observing mj, each juror casts a vote Xj{A,C}, where Xj=A is a vote by juror j to acquit and Xt=C is a vote to convict.
If all n jurors choose C, then the group decision is C; otherwise it is A. All jurors prefer to convict only the guilty and to set only the innocent free.
Aldrich Lupia EITM 2007
Our model allows different Juror Types
Findings Letting high and low NFC
jurors think in different ways drives this same probability to zero very quickly as they jury grows.
To understand how often unanimity rule convicts the innocent, we need to know more about how jurors think.
Aldrich Lupia EITM 2007
Conclusion Choosing an equilibrium concept is equivalent to making
an assumption about how actors reason in equilibrium.
Nash family equilibrium concepts impose a rigorous kind of reasoning.
Concepts, such as SCE, may be more appropriate for actors who have limited ability or incentive to contemplate others’ strategies.
Aldrich Lupia EITM 2007
Challenges The SCE agenda poses an important challenge to
the unseen assumptions about cognition that support Nash-based inferences.
The agenda does not fill in the theory of beliefs and belief change that is needed. A place where greater integration of empirical and
theoretical work can open great windows of opportunity.
Aldrich Lupia EITM 2007
Experimental Design:A Close Analogy of our Theory
Our experiments contained a principal and a speaker.
The principal’s job was to predict coin toss outcomes. The speaker made statements.
We paid each principal $1 for each correct prediction.
We varied the speaker’s knowledge, interests, and incentives. Die roll to determine interests, knowledge. Penalties for lying, costly effort, verification.
These variations defined our treatment and control conditions and were the means for evaluating our theory.
Aldrich Lupia EITM 2007
Theoretical Premises and Experimental Analogies
Aldrich Lupia EITM 2007
Other design elements. Quiz on instructions.
Means of communication controlled.
Multiple interactions -- independent payoffs, no added information about prior periods, no reputation effects.
Multiple principals -- independent payoffs for principals, no way for speaker to distinguish principals.
Aldrich Lupia EITM 2007
Lab Experiments A subject is a voter or advisor.
The voter predicts coin tosses Earns $1/correct prediction.
Advisor: “heads” or “tails.” we vary perceptions:
hidden die rolls determine speaker interests & knowledge.
we vary “institutions.” penalties for lying, costly effort, verification present in selected trials.
Aldrich Lupia EITM 2007
Sequence of Events Experimental Series 3
We flip the coin.
We show the results to the reporter.
The reporter makes a statement.
The predictors make their predictions.
If the predictor makes a correct prediction, then both the Predictor and the Reporter earn $.50.
Apart from this, a predictor earns $.05 for predicting heads.
Aldrich Lupia EITM 2007
Sequence of Events Experimental Series 4 We flip the coin.
There is a 70% chance that we show the result to the reporter.
The reporter makes a statement.
The predictors make their predictions.
If the predictor makes a correct prediction, then both the Predictor and the Reporter earn $.50.
Apart from this, a predictor earns $.05 for predicting heads.
Aldrich Lupia EITM 2007
Sequence of Events Experimental Series 5 We flip the coin.
The reporter makes a statement without seeing the coin flip.
The predictors make their predictions.
If the predictor makes a correct prediction, then both the Predictor and the Reporter earn $.50.
Apart from this, a predictor earns $.05 for predicting heads.
Aldrich Lupia EITM 2007
Sequence of Events Experimental Series 6 We flip the coin.
The reporter makes a statement without seeing the coin flip.
The predictors make their predictions.
If the Predictor makes a correct prediction, then the Predictor earns $.50 and the Reporter earn nothing.
If the Predictor makes and incorrect prediction, then the Predictor earns nothing and the Reporter earns $.50.
Apart from this, a predictor earns $.05 for predicting heads.
Aldrich Lupia EITM 2007
Sequence of Events Experimental Series 7 We flip the coin.
We show the result to the reporter.
The reporter makes a statement.
The predictors make their predictions.
If the predictor makes a correct prediction, then both the Predictor and the Reporter earn $.50.
Aldrich Lupia EITM 2007
Sequence of Events Experimental Series 8 We roll the dice. If it lands 1 through 7, the Reporter is
paid $.50 every time a predictor makes a correct prediction. If it lands 8 through 10, then the reporter is paid $.50 every time a predictor makes an incorrect prediction.
We flip the coin.
We show the result to the reporter.
The reporter makes a statement.
The predictors make their predictions.
Aldrich Lupia EITM 2007
Sequence of Events Experimental Series 9 We roll the dice. If it lands 1 or 2, the Reporter is paid $.50
every time a predictor makes a correct prediction. If it lands 3 through 10, then the reporter is paid $.50 every time a predictor makes an incorrect prediction.
We flip the coin.
We show the result to the reporter.
The reporter makes a statement.
The predictors make their predictions.
Aldrich Lupia EITM 2007
Your trials Complete information. + Incomplete information. No advice. ~ Expertise & Common Interests. + Expertise & Conflicting Interests. ~ No expertise & Common Interests. ~ No expertise & Conflicting Interests. ~ Expertise & Common Interests. + Expertise & 70% Common Interests. Persuasive. Expertise & 30% Common Interests. ~