The Formation of Job Referral Networks: Evidence from a Field Experiment in Urban Ethiopia

ETHIOPIAN DEVELOPMENT RESEARCH INSTITUTE

TheFormationofJobReferralNetworksEvidence from a Field Experiment in Urban EthiopiaEvidencefromaFieldExperimentinUrbanEthiopia

A.StefanoCaria1 andIbrahimWorku2IFPRIESSP‐II

EthiopianEconomicAssociationConferenceJuly 18 2013July18,2013AddisAbaba

11UniversityofOxford,CentrefortheStudyofAfricanEconomies2IFPRI‐EthiopiaSupportStrategyProgrammeII

Design Predictions Data Results Conclusions

Social interactions matter for labour market outcomes

• Strong influence on labor market outcomes, throughinformation and referrals (Granovetter 1995, Topa 2011)

• In Ethiopia referrals common in flower sector (Mano et al 2010)and network advice is popular search strategy (Seernels 2007)

• In our sample:• 41pct of workers have first heard of their current job from social ties• 29pct have received a referral

• Exclusion from referral networks is likely to be asubstantial disadvantage in labour market

Design Predictions Data Results Conclusions

Figure 1: The job contact network of a neighborhood in urban Ethiopia

Design Predictions Data Results Conclusions

Empirical degree distribution is quite unequal

Figure 2: Distribution of degree in job contact networks

Design Predictions Data Results Conclusions

• Theory suggests agents have both self-regarding and otherregarding reasons to link with the so far poorly connected

• This prediction does not fit the real data

• Models could be misconstruing the incentives in the field, or thedecision making process. We focus on decision making

1 Would agents include peripheral peers when this maximisesthe chance of getting a referral?

2 Do agents also have other-regarding reasons to includeperipheral peers?

Design Predictions Data Results Conclusions

• We devise an AFE to test for these hypotheses, based onBeaman Magruder (2012)

• We find evidence for self-regarding but not for other regardingmotives to link with peripheral agents

Design Predictions Data Results Conclusions

Outline

1 Design

2 Predictions

3 Data

4 Results

5 Conclusions

Design Predictions Data Results Conclusions

The game

• Subjects add two links to an exogenous undirected network• Specify a partner or ask that one is randomly drawn for them

• The network determines who can refer whom

• A lottery determines whether participants get a lab-job

• Lab-job holders make one referral to a random unemployed tie

Design Predictions Data Results Conclusions

The protocol

1 Network positions are randomly assinged

2 Dictator game

3 Test for understanding

4 Linking decisions

5 Jobs are drawn

6 The network is updated

7 Referrals are given

Design Predictions Data Results Conclusions

Treatments isolate motives for linking behaviour

• In SELF treatments network updated with links of one randomlydrawn unemployed player

• Other regarding concerns switched off• Second order, strategic considerations switched off

• In OTHER treatments we implement the links of one randomlydrawn employed player

• Other-regarding concerns primed, self-regarding switched off

• 2x2 design: we also vary anonymity (decisions remain private)

• 5th treatment checks understanding at the end to limit priming

Design Predictions Data Results Conclusions

The network

A

I F B G

E C

D

H

Figure 3: ID letters

Design Predictions Data Results Conclusions

Jobs are drawn

A

I F B G

E C

D

H

Figure 4: Bold IDs have jobs

Design Predictions Data Results Conclusions

SELF treatment

A

I F B G

E C

D

H

Figure 5: Network augmented with links of one unemployed person

Design Predictions Data Results Conclusions

OTHER treatment

A

I F B G

E C

D

H

Figure 6: Network augmented with links of one employed person

Design Predictions Data Results Conclusions

Outline

1 Design

2 Predictions

3 Data

4 Results

5 Conclusions

Design Predictions Data Results Conclusions

Theory suggests two mechanisms of inclusion

1 Models of strategic network formation posit agents considercosts and benefits of each link (Jackson Wolinsky 1996, Bala Goyal2000)

When people compete for referrals, links with peripheral peopleare very valuable (Calvo Armengol, 2004)

2 Other regarding preferences may also motivate linking choices• If agents are altruistic (efficiency minded or inequity averse) they will

also try to maximise the chance that peers are referred for a job• In our game, this implies linking to the peripheral agents• Directed altruism in non anonynous treatment

Design Predictions Data Results Conclusions

We derive four predictions

1 Subjects in SELF treatments will create new links with peripheralagents

2 Subjects in OTHER treatments will be create new links withperipheral agents

3 DG giving correlated with link decisions in OTHER, but not in SELFtreatments

4 Subjects in OTHERn will be more likely to refer those whom theyknow in real life. Decisions of subjects in SELFn will not be affected

Design Predictions Data Results Conclusions

We analyze the data with the following dyadic regression model:

rij = α + βc2j + γc3j + uij (1)

• Unit of observations is all initially unlinked dyads

• Linea probability model

• Standard errors are clustered at session level

• The coefficients on c2j and c3j will provide the basic test forhypotheses 1 and 2

Include interactions for treatments, understanding and DG giving:

rij = α + βc2j + γc3j + δti + θti ∗ c2j + λti ∗ c3j + uij (2)

Design Predictions Data Results Conclusions

Outline

1 Design

2 Predictions

3 Data

4 Results

5 Conclusions

Design Predictions Data Results Conclusions

The experiment

• A 50k town in northern Ethiopia with a growing industrial sector

• Randomly sampled blocks and interviewed all individuals 20-40

• Everyone invited to play game: 447/518 subjects participated

• 10 sessions of SELFa OTHERa OTHERn, 11 sessions of SELFn, 9sessions of SELFawp

1 Covariate balance across assigned network centrality is good

2 Some observable differences (at 10pct s.l) across session treatments

3 Understanding was high and uncorrelated with treatment

Design Predictions Data Results Conclusions

Outline

1 Design

2 Predictions

3 Data

4 Results

5 Conclusions

Design Predictions Data Results Conclusions

Result 1Subjects in SELF treatments are more likely to link with lesscentral peers

Result 2Linking behaviour in SELF treatments is highly correlated withunderstanding, and not correlated with giving in the DG

Design Predictions Data Results Conclusions

Table 1: LPM: SELF treatments

Base Controls Treatments(1) (2) (3)

j centrality = 2 -.167 -.179 -.195(.039)∗∗∗ (.068)∗∗∗ (.060)∗∗∗

j centrality = 3 -.198 -.200 -.340(.047)∗∗∗ (.075)∗∗∗ (.072)∗∗∗

Non anonymous -.023(.048)

Non anonymous X c = 2 .020(.079)

Non anonymous X c = 3 .160(.073)∗∗

No probabilities -.035(.087)

No prob X c = 2 .005(.116)

No prob X c = 3 .038(.136)

Const. .397 .407 .436(.028)∗∗∗ (.046)∗∗∗ (.043)∗∗∗

Obs. 1594 1528 1528

Design Predictions Data Results Conclusions

Table 2: LPM: SELF treatments

Understanding1 Understanding2 OtherRegarding(1) (2) (3)

j centrality = 2 -.016 -.261(.137) (.086)∗∗∗

j centrality = 3 -.010 -.241(.097) (.087)∗∗∗

Understanding .131 .125(.033)∗∗∗ (.054)∗∗

Understand X c = 2 -.199 -.187(.043)∗∗∗ (.110)∗

Understand X c = 3 -.229 -.222(.064)∗∗∗ (.087)∗∗

DG sent -.007(.006)

Sent X c = 2 .012(.009)

Sent X c = 3 .006(.009)

Const. .290 .298 .453(.020)∗∗∗ (.070)∗∗∗ (.058)∗∗∗

Obs. 1517 1517 1528

Design Predictions Data Results Conclusions

Result 3Subjects in OTHER treatments are NOT more likely to link withless central peers

Result 4Linking behaviour in OTHER treatments is uncorrelated withunderstanding or giving in the dictator game

Design Predictions Data Results Conclusions

Table 3: LPM: OTHER treatments

Base Controls Treatments(1) (2) (3)

j centrality = 2 -.065 -.101 -.159(.059) (.081) (.098)

j centrality = 3 .012 -.085 -.129(.076) (.088) (.122)

Non anonymous -.033(.077)

Non anonymous X c = 2 .119(.112)

Non anonymous X c = 3 .089(.144)

Const. .302 .336 .352(.041)∗∗∗ (.052)∗∗∗ (.071)∗∗∗

Obs. 1072 1022 1022

Design Predictions Data Results Conclusions

Result 5In non anonymous treatments, subjects are more likely to link withknown peers

SELFn OTHERn(1) (2)

j centrality = 2 -.177 .002(.052)∗∗∗ (.076)

j centrality = 3 -.165 .058(.044)∗∗∗ (.102)

i knows j .101 .241(.051)∗∗ (.123)∗

Same gender -.055 -.003(.024)∗∗ (.035)

Sum age .002 -.002(.002) (.001)

Diff age -.002 .003(.001)∗∗ (.001)∗∗

Const. .303 .320(.086)∗∗∗ (.096)∗∗∗

Obs. 563 452

Table 4: LPM: Non anonymous treatments

Design Predictions Data Results Conclusions

Outline

1 Design

2 Predictions

3 Data

4 Results

5 Conclusions

Design Predictions Data Results Conclusions

Hiring policies can re-direct network formation

• Individuals may have both self and other regarding motives in theformation of job contact networks

• We find strong evidence in support of self-regarding motives

• We are unable to find evidence of other-regarding motives

• Policy can target incentives in network formation processes

• Employers can be incentivized to ask more referrals from membersof peripheral groups. This would strengthen the latter’s position injob networks