Cheap Talk Messages for Market Design Theory and Evidence from a Labor Market with Directed Search

John Horton

MIT

Ramesh Johari

Stanford

Philipp Kircher

Cornell

https://john-joseph-horton.com/papers/hot_towel.pdf

Search design: facilitating job search through cheap talk

​

• People search for jobs

​

• Websites try to make this easier (search design)

​

• One avenue: providing information
• but "hard" information is difficult.
• so they could simply ask for it (cheap talk)
• why could this be more efficient than “free text”?
• prominent
• easy to classify
• “searchable”

​

​

​

Example

Consider an experienced programmer.

She only wants to compete for the jobs when the firm values her skills and pays appropriately.

Can it help to simply ask firms about this?

Lots of computer-mediated matching markets where platform can create signaling opportunities

Worker

Firm

Application

Expert Worker

Novice Worker

High "quality";

commands a high wage.

Low "quality";

commands a low wage.

H

L

"High"

Which type of worker does the firm want?

"Low"

L

H

Suppose firms differ in willingness to pay (WTP) for "quality"

High WTP for quality

Low WTP for quality

H

L

Workers apply to the right "type" of employer

H

H

L

L

In practice, "types" unknown

?

?

H

L

Can "we" get firms to reveal their WTP for quality, or their type (High, Low)?

If firms reveal type:

I'm a High type

I'm a Low type

H

L

Strategic considerations of stating "vertical" preferences in a matching market

• Stating High might get you:
• Pro: More and/or better applicants
• Con: $$$$$ - Higher bids, as sellers condition on WTP
• Stating Low might get you:
• Pro: $ - Lower bids, as sellers condition on WTP
• Con: Worse and/or fewer applicants

Research Questions - Theory

• Is truthful cheap talk an equilibrium outcome?
• How does it affect the market
1. It increases sorting: more high ability at high messages, less at low messages
2. Applications go down at low message, go up at high message
3. It affects the wages: wages conditional on ability decrease at low message, and increase at high message
4. Efficiency increases

​

​

Research Questions - Evidence

• Is truthful cheap talk an equilibrium outcome? (√)
• How does it affect the market
1. It increases sorting: more high ability at high messages, less at low messages (√)
2. Applications go down at low message (√), go up at high message
3. It affects the wages: wages conditional on ability decrease at low message, and incr at high message (√)
4. Efficiency increases (√)

Literature

Cheap talk in search markets: Menzio JPE2007, Kim and Kircher ECTR2015

pure theory; no complementarities

​

Information provision in actual labor search markets: Belot, Kircher, Muller REStud2019; Altmann et al JPubE2018, Barach and Horton 2017, Horton MS2019

no cheap talk

​

Verifiable information provision about quality in field experiments in consumer product markets: Tadelis and Zettelmeyer AER2015, Luca 2016; Jin and Leslie QJE2003

no cheap talk

​

Cheap talk about horizontal preferences: dating (Lee and Niederle, EE2015); economic junior job market (Coles et al., AEJ-Micro2013, Kushnir GEB2013)

no vertical preferences

​

Theory of sorting in search markets: e.g., random search: Shimer and Smith ECTR2000), Teulings and Gautier REStud2004 & JEEA2006, Eeckhout and Kircher ECTR2010, Bagger and Lentz REStud2019; directed search: Shi JET2001 & REStud2002, Shimer JPE2005, Eeckhout and Kircher ECTR2011, Cai et al (2021)

no cheap talk

​

General trend to look at online (labor) markets: Kuhn and Skuterud AER2004; Kuhn and Mansour EJ2014; Varian AER2010; Marinescu and Wolthoff AEJ2016

​

Big "so what" of this paper:

When a market is computer-mediated, what new tools can the platform use to improve market function?

This talk:

(1) Can a platform help "direct" search?

(2) Should it?

Empirical context

• Large online labor market in which workers apply to job openings proposing an hourly wage bid
• Platform wanted to create a new signaling opportunity for would-be employers

Employer's selection of vertical preference tiers when posting a job opening

"Low"

"High"

Preference shown to would-be applicants

How signal would work market-wide:

1. Employer picks preference; knows it will be shown

2. Would-be applicants shown employer preference

Motivating idea for this feature

• High-type firms will give up some bargaining power / commit to higher wages in exchange for a more appropriate pool of applicants

The feature worked differently during an experimental phase.

First questions

1. Do vertical preferences vary?
2. Do employers condition their signal on knowledge about revelation?

Explicit arm: Employer picks preference knowing whether choice will or will not be revealed

"Would-be applicants will see your choice"

"Would-be applicants will not see your choice"

Randomization

ShownPref = 1

ShownPref = 0

%

employers selecting

Category of work (set ex ante):

Vertical preference choices

%

employers selecting

Vertical preference choices

Δ in fraction by ShownPref

(Standard Error)

Employers did not condition tier choice on whether choice would be revealed

Substantial between-category variation

Substantial within-category variation

First questions

• Do vertical preferences vary? YES.
• Do employers condition their signal on knowledge about revelation? NO.

Research questions on "sorting"

1. Do workers sort in the expected direction (High to High, Low to Low)?
2. Do workers condition their wage bids on willingness to pay?
3. Is the number of applications reduced?

Research questions on "sorting"

• Do workers sort in the expected direction (High to High, Low to Low)?
• Do workers condition their wage bids on willingness to pay?
• Is the number of applications reduced?

Ambiguous arm of the experiment: Employer picks preference not knowing whether choice will be revealed

"Would-be applicants might see your choice"

Ambiguous arm of the experiment: Employer picks preference not knowing whether choice will be revealed

"Would-be applicants might see your choice"

ShownPref = 1

ShownPref = 0

Research questions on "sorting"

• Do workers sort in the expected direction (High to High, Low to Low)?
• Do workers condition their wage bids on willingness to pay?
• Is the number of applications reduced?

We need to characterize worker "type" at the moment they applied

On-platform

experience

Outcome:

An applicant's prior earnings at the time they applied.

Whether vertical preferences were

shown (ShownPref)

Substantial sorting even when preference not shown

(ShownPref = 0)

High

Low

Medium

Employers who had "High" revealed got ~7% more experienced applicants.

Employers who had "Low" revealed got ~18% less experienced applicants.

Employers who had "Medium" preferences revealed got ~5% more experienced applicants

Research questions on "sorting"

• Do workers sort in the expected direction (High to High, Low to Low)? Yes.
• Do workers condition their wage bids on willingness to pay?
• Is the number of applications reduced?

Effect of revealing preferences on applicant wage bids

Effect of revealing preferences on applicant wage bids

10% higher wage bids

from

revelation

in high tier

Effect of revealing preferences on applicant wage bids

5% higher wage bids

from

revelation

in med tier

Effect of revealing preferences on applicant wage bids

15% lower wage bids

from

revelation

in low tier

Research questions on "sorting"

• Do workers sort in the expected direction (High to High, Low to Low)? Yes.
• Do workers condition their wage bids on willingness to pay? Maybe? Couldn't this be composition?
• Is the number of applications reduced?

Workers send

many apps

H

H

L

L

ShownPref = 1

ShownPref = 0

ShownPref = 1

ShownPref = 0

Within-worker approach to detecting conditioning

Log wag bid of worker i to opening j

Within-worker approach to detecting conditioning

worker i specific fixed-effect

Within-worker approach to detecting conditioning

Treatment assignment and tier of job opening j and interactions

Do workers condition wage bids?

Do workers condition wage bids?

Bid 10% lower when they know they face a "Low" employer

Do workers condition wage bids?

Bid 7% higher when they know they face a high-tier employer

Same results, graphically

Some conditioning on tier even when not revealed

Way more conditioning when revealed

Research questions on "sorting"

• Do workers sort in the expected direction (High to High, Low to Low)? Yes.
• Do workers condition their wage bids on willingness to pay? Yes.
• Is the number of applications reduced?

~5% reduction overall from revelation

Effects highly concentrated in the low tier (~15% reduction)

"Sorting" questions

• Do workers sort in the expected direction (High to High, Low to Low)? Yes.
• Do workers condition their wage bids on willingness to pay? Yes.
• Are overall application counts reduced? Yes in the low tier; maybe overall.

Research Questions on matches formed

1. Are compositional/wage bidding changes "passed through" to the workers that are hired?
2. Is the new equilibrium "good"?

In the Low tier

Log wage

Applicant wage bids

In the Low tier

Log wage

Hired worker wage bids

In the Low tier

Log wage

Employers hire even lower wage bid workers

In the Low tier

Log wage

But revelation effects "pass through"

Change in average applicant wage bid lead to approximately same % change in hired worker wage, across tiers

Effects on applicant and hired worker experience

Approximate pass-through on experience measures

Research Questions on matches formed

• Are compositional/wage bidding changes "passed through" to the workers that are hired? Yes.
• Is the new equilibrium "good"?

Is a separating equilibrium good for:

• The platform / social planner?
• Match surplus (wages are just a transfer)
• The employer/firm?
• Surplus from match, minus wages paid
• The worker?
• Wages paid, minus cost of performing

​

One perspective on surplus:

focus on quantity transacted

(wage bill & hours-worked)

Reduction in # applications sent

No evidence of a change on extensive margin (i.e., hires made)

Shifts in wages (but recall large composition shifts)

Overall quantities increase

What about over time?

Concluding thoughts

• Vertical signaling feature was a success for the platform
• Employers are willing to reveal; workers sort as expected
• Increase in hours-worked is consistent with improved match quality
• Largest effects: helping relatively low-skilled workers get hired
• This kind of signaling mechanism is a "design pattern" - can be applied to lots dimensions in lots of different markets

Thank you

Draft:http://www.john-joseph-horton.com/papers/hot_towel.pdf

Backup Slides

What about SUTVA?

Paper Title: "At What Quality and What Price? Eliciting Buyer Preferences as a Platform Design Problem"

Authors: John Horton & Ramesh Johari

Are applicant pools smaller following revelation?

Some evidence of overall reduction in applicant pools ~ on the order of 1-5%

Much clearer evidence of a decline in number of applicants to Low Tier following revelation, ~7%

Research questions

1. Is there variation in vertical preferences not already conveyed by other salient employer/job opening attributes?
2. Will employers reveal an extreme preference (high or low) when they know their preference will be revealed, or will they "pool" on intermediate?
3. Do workers knowing preferences "sort" to employers of the right type?
4. Do workers tailor bids in response to perceived employer willingness to pay?
5. How does preference revelation affect the characteristics of the worker ultimately hired (if any)?
6. How do the resultant contracts turn out?

Market designers as information providers

• Injecting information about goods, services, attributes of counterparties etc., into the market
• Restricting access to the market
• Building & maintaining reputation systems
• Verifying information (particularly about past transactions)
• Creating opportunities for participants to signal

​

Signaling in markets

• Signaling to communicate latent quality/ability (Spence 1973, Milgrom & Roberts 1986)
• More recently, used as a tool to inject seller preference information into designed markets (Coles et al. in JEP 2010; Lee & Niederle 2015)
• Signaling "buyer" preferences
• "Horizontal" lacks a strategic component.
• "Vertical" preference is more strategically fraught

This talk

• Present evidence from the introduction of a new vertical preference signaling mechanism into a large online labor market
• Focus on "engineering" questions---does it "work" and if not, why not?
• Is there something to convey? Will employers convey preferences? Do workers "sort"? Do workers condition their bids? Does it affect who gets hired?

Arms of the experiment

• Two experimental arms:
• "Explicit" Arm: Employers were told ex ante that their preference selection either would be shown to applicants (ShownPref = 1) or not shown (ShownPref = 0).
• "Ambiguous" Arm: Employers were told their preferences might be revealed and then were randomly either shown (ShownPref = 1) or not shown (ShownPref = 0).

Detecting compositional changes in applicant pools

• Each job opening gets a collection of applicants that differ on observable attributes (e.g., past market experience) For each ambiguous arm opening:
• Calculate the qth percentile for some attribute (e.g., "what is the 25th percentile level of experience in the 12 worker applicant pool?")
• Average those quantiles, by revelation and vertical preference tiers
• Compare those averages by revelation

Market-designer perspective

• Signaling as a solution to/consequence of informational problems (Spence, 1973; Crawford & Sobel, 1982) not typically seen as something "designed"
• Exceptions for conveying idiosyncratic "horizontal" preferences (Coles et al., 2010; Lee & Niederle 2015)

​

Quantiles in the pool: Effects of revelation

More experienced applicants in "High Tier" from revelation (~10%)

Some reduction in applications to "low" type employers

Not much on the extensive margin with hiring and interviewing

​

Wages:

High: Go up

Medium: Stay same

Low: Go down

​

Hours:

Up for *every* group

Reveal "High"

Smaller pools?