Dishonesty in Advice and Promises

• Ernan Haruvy� �Desautels Faculty of Management�McGill University

​

1

Purpose of this talk

• Offer a Perspective
• Inform
• Entertain
• Plant ideas
• Co-create

​

​

2

About my background

• Experimental
• Economics – Rationality, Learning, Other-regarding, DMFO

​

• Finance – Portfolio Selection, Trader Types in Asset Market (Asset Valuation), Governance, Sustainable Preference, Advice

​

• Supply Chain and Operations – B2B Auctions, Bargaining
• Marketing – Pricing, Auctions, B2B, Choice, and Advice

​

• Accounting?

​

3

Common Theme

• 1. Market Design
• 2. Beliefs and Implicit Promise-- Think Trust Game.
• 3. Mediation/Advice – Think Manager, Auditor

​

4

Perspectives of experimental economics

5

Create self-contained and controlled economies

Completely map the economic theory to the experiment

In the experiment itself: Abstraction to the max

Attention to the instructions and to testing comprehension

Attention to incentives and the exact delivery of incentives

Recognition and embrace of bounded rationality in design

Attention to beliefs

Attention to the applied environment and its feature – Never(!) a cover story

Careful attempt to link to the applied environment

Market Design

6

Auctions + Matching

Economist as Engineer

A “cover story” is unacceptable– Need to know the minute details, including first-hand testimony

Need a route to apply

Evolution of Internal Consistency and External Validity in Exp Economics

7

Study 1: A Strategically Lying Adviser

• The point of the work is to show that Subjects Learn To Fully Delegate Instead of Soliciting Advice when the Advisor’s Interests are not fully aligned.

8

Example 1: A Strategically Lying Adviser

9

Equilibrium is for Advisor to be honest 100% of the time when seeing beta (recommend B). Then A recommendation will be followed 100% of the time. The advisor will mix 50:50 when seeing alpha. That makes investors indifferent between accept and reject when seeing B (which they will see 2/3 of the time).

10

11

Equilibrium and Outcome

• Equilibrium is for Advisor to be honest 100% of the time when seeing beta (recommend B). Then A recommendation will be followed 100% of the time. The advisor will mix 50:50 when seeing alpha. That makes investors indifferent between accept and reject when seeing B (which they will see 2/3 of the time).

​

• But the rule being learned is to be honest in both states.
• 89% in State Alpha; 85% in State Beta.

​

​

12

Study 2: Learning to Prefer Bad Advice by a Strategically Lying Advisor

Similar to Study 1, but here the Advisee is not given a profile of Truth Telling for each State but rather one profile only to condition on.

​

13

Optimal Advisor Behavior

• Here the Advisor can safely lie when the incentives are very high to do so and build a truthful reputation when the incentives to lie are weak.

​

14

Setting

1. There are two “states of the world”: Low Variance and High Variance States
2. High variance is rate (p=0.1)

​

15

Π_A^L -- This is the earnings for the retailer if he chooses A and the state is Low Variance

Π_B^L -- This is the earnings for the retailer if he chooses B and the state is Low Variance

Π_A^H -- This is the earnings for the retailer if he chooses A and the state is High Variance

Π_B^H -- This is the earnings for the retailer if he chooses B and the state is High Variance

Π_A- This is the average earning, across states, of the retailer if he chooses A. It is computed as �.9 EΠAL+.1 EΠAH

Π_B- This is the average earning, across states, of the retailer if he chooses B. It is computed as �.9 EΠ_B^L+.1 EΠ_B^H

Advisee

​

• Each subject gets a recommendation – which is bad or good
• Subject can follow or not

16

What is meant by bad/good recommendation

​

• The bad recommendation is one where the low stake tells truth and high stake lies
• The good recommendation is the reverse

17

Conditions Study 1

18

Feedback Screen

19

Results: Advisee follows bad advice�

20

Advisor

• 158 subjects in the role of advisor
• 81 subjects in ASY and 77 for SYM.
• Each round choose recommendation
• As long as truthfulness >= 50% advice is followed
• Random choice otherwise.

​

21

Theory – Expected Value to the supplier-- Symmetric

• Expected Profit (always recommend A) = 250
• Expected Profit (always recommend B) = 250
• Expected Profit (always truthful) = Worst possible payoff for advisor = -15*0.9 + -200*0.1 = -26.75 + …
• Expected Profit (strategically lie when high stakes) = -15*0.9*0.5+200*0.1 = +13.25 + …
• Expected Profit (strategically lie when high stake and when at more than 50%) = [-15*0.5*0.5 + 15*0.5*0.4]*0.9+200*0.1 = -15*0.5*0.1 + 200*0.1 = +19.25 + …

​

22

Advisor’s screen

23

Advisor’s feedback screen

24

Finding

• Suppliers lie when stakes are large and tell the truth when stakes are small. That is, they appear to build costly reputation primarily when the costs are low.

25

Finding

• Finding. When suppliers have sunk low to where they are no longer believed, paradoxically their cost of telling the truth is zero. At that point, their truth telling for high stakes exceed their truth telling for small stakes (in SYM only).

​

26

Conclusions

• Bad advice is believed more often when it implies correct recommendations most of the time.
• Advisees learn to follow bad advice
• Advisors learn to give bad advice

​

• When suppliers are in a state where they are not believed, they may revert to good advice until they are believed again.

27

The Trust Game

• Based on the investment game of Berg, Dickhaut, & McCabe (1995), later known as Trust Game.
• Investor is given $0-$10, which she can sent any amount to Receiver. Any amount the investor sends is tripled, and the Receiver can return any amount.
• A “simple” trust game is any truncation of the investment game—with any multiplier—often a binary choice per player role (Cox and Deck, 2005; Charness and Rabin, 2002)
• The First Mover—the Trustor—“trusts” or opts out.
• If he trusts, the second mover—the Trustee—can reciprocate by returning at least the amount invested plus a “fair” return. Or the Trustee can violate the trust.

_____________________________________________________

• It is easier to add controls (with dictator and ultimatum).
• Easier to add treatments with narrow focus (manipulations).
• Easier to fit to different applications

​

28

Variations of the Trust Game in Accounting/Finance

• King, R. R. (2002). An experimental investigation of self‐serving biases in an auditing trust game: The effect of group affiliation. The Accounting Review, 77(2), 265-284.
• The Auditor is the Trustor; The manager is the Trustee.
• The auditor’s best response is to select an audit level to match the manager's fraud level
• The manager may "promise" his auditor-counterpart that he intends to cooperate by committing low levels of fraud, but not implement the promise.
• The authors find that promises can make auditors have unwarranted trust

29

Other variations with promises– many-to-one

• Manager-investor interaction:
• The manager is making a choice of an accounting system –truthfully disclosing or private. (Lunawat, 2013, European Accounting Review)

​

• Other manager-investor examples:
• Earning Guidance
• Dividend Announcements

​

30

Research Questions

• We ask the following questions:

​

1. If trust increases welfare and a promise increases trust, does it follow that enabling promises increase welfare?
2. Why do people keep unenforceable promise? Is it due to guilt, a preference for honesty, or something else?
3. Does the possibility of plausible deniability (due to chance) increase default—by how much?

31

Example

• For example, in Quebec, there is a possibility of a warranty on the real estate transactions, and most transactions are without this warranty as it presents a legal headache. Does this increase trust and trustworthiness relative to not putting in this feature?

​

32

The First Trust + Promise Game

• Charness & Dufwenberg (2006) conducted the seminal study on the effect of cheap-talk promise– with chance-- on trust formation.

​

33

The First Trust+Promise Game

• They find that promise increases trust by 68% and trustworthiness by 74%.
• The increase of trustworthiness was attributed to the guilt aversion.
• They tested the theory by measuring the probability that the Trustees believe the Trustors are going to choose In.
• That is, trustees are more likely to choose Roll (Be Trustworthy) if Don’t Roll results in too much deviation from trustors’ expected payoff.

34

Challenges to CD

• Ellingsen & Johannesson (2004) and Vanberg (2008) called CD's conclusions to question.
• They argued that changes in payoff expectations do not account for the effects of promise.
• That is, people have an intrinsic preference for promise-keeping.
• The commitment-based theory

35

Guilt-aversion vs. Commitment-based

• Guilt-aversion
• Battigalli & Dufwenberg (2007)
• Bellemare et al. (2011)
• Charness & Dufwenberg (2006)
• Chen & Houser (2019)
• Dufwenberg & Gneezy (2000)
• Ederer & Stremitzer (2017)
• Geanakoplos et al. (1989)
• Guerra & Zizzo (2004)
• Khalmetski et al. (2015)
• Regner & Harth (2014)
• Reuben et al. (2009)
• Schwartz et al. (2019)

​

• Commitment-based
• Bhattacharya &Sengupta (unpublished)
• Ellingsen & Johannesson (2004)
• Gibson et al. (2013)
• Ismayilov & Potters (2015)
• Lundquist et al. (2009)
• Miettinen & Suetens (2008)
• Vanberg (2008)

36

Di Bartolomeo et al. (2019) use a clever no-delivery design to draw equivalence to no-communication. They find that when ‘silence’ has been delivered trustees are less trustworthy. But efficiency is driven by trustors– not trustees. �A was switched or not switched to show promise-mechanism impact

37

Our design vs. CD

• The CD design has two components which are critical to the guilt-aversion theory and which they do not vary in either CD (2006) or CD (2010).
1. The Chance Event.
2. The possibility of Promise itself.
• In this work, we vary both.

​

​

38

Reproduced from CD (2006)

Investigation

• Varying the two elements allows us to answer these questions as well as gauge whether or not, or under what conditions, allowing a promise is expected to improve welfare, and likewise under what conditions removing or reducing chance can drastically improve welfare

39

Experiment Overview

• n = 153
• Recruited from the Cloudresearch online recruiting platform (CONNECT – No MTURK subjects at all)
• SIX variants of CD Trust Games: 3x2

​

{Chance moves AFTER the trustee’s decision node (closest to CD (2006)), Chance moves BEFORE the trustee’s decision node, No Chance}

• X
• {Promise or Silence is possible, Promise is Not Possible}.

40

41

All six variants can be expressed as the following normal form game.

closest to CD

Results

42

Graphic for CD: Trust

43

Trustworthiness (CD)

44

Finding summary

• We show that trust overall decreases when there is a possibility of promise.

​

• Trust– not trustworthiness– is what determines efficiency.

​

45

Chance reduces trustworthiness and trust.

• But not in the way guilt theory prescribes
• The chance event is in fact the guilt alleviating device.
• Tadelis shows evidence for both. Without the chance event, he shows that Trust increases by 85% and Trustworthiness increases by 88%. That is, the chance event hurts efficiency.
• We confirm the results of Tadelis, with the caveat that trust depends on whether the Chance event occurs before or after the Trustee’s decision node in a Trust Game, which has no theoretical bearings on the alleviation of guilt.

​

46

Results

• When we compare the Trust Game variants, we find that trust increases by 15% when the Chance move precedes (vs. follows) the Trustee’s decision node. No meaningful difference was observed in overall Trustworthiness.

47

Results

• Moreover, trust increases by 66% when the Chance move precedes (vs. follows) the Trustee’s decision node under the Unpromisable treatment.
• Although the position of Chance does not alter the strategic equivalence, it affect trust, regardless of the possibility of Promise.

48

Results

• The increase in trust is moderated by the possibility of promise. Taking out chance when there is promise, increases trust by 33%. But taking out chance when no promise is possible increases trust by 125%.

49

Conclusion

• The possibility of making Promises is good with complete attribution but terrible with plausible deniability, even though, and especially because people believe promise.

​

• Likewise, removing plausible deniability is most effective when promises are not possible; only slightly effective when promises are possible.

​

​

50