Why social networks are important in development

  Stockholm Doctoral Course Program in Economics Development Economics II: Lecture 3

  Social Networks Masayuki Kudamatsu

  IIES, Stockholm University

13 November, 2013

  Why social networks are important in development

  • Impact of network characteristics

  Peer effects on welfare / behavior

  • (e.g.

  class size on test score) • (e.g.

  

Impact of network behavior

technology adoption)

  • Informal insurance

  Help contract enforcement

  • Group lending
  • Relational contract •
This lecture

  1. Identify the effect of network (Munshi 2003)

  characteristics

  2. Identify the effect of network

  behavior

  • (Manski 1993; Conley & Udry

  How to overcome the reflection

  problem 2010)

  3. Network characteristics that

  enhance contract enforcement

1. Impact of Network

  

Characteristics

  Empirical challenge: omitted

  • variable bias People in the same network share
  • many things in common

  Geography

  • Assortative matching
Digression

  Positive assortative matching: suppose Two types of agents (high and low

  • ability): u i ∈ {H, L}, (H > L) i’s payoff from forming a pair w/ k :
  • V i (u i , u k ), increasing w/ u i & u k
Digression (cont.)

  Ability of each pair: positively

  • correlated if V i (H, H)−V i (H, L) > V i (L, H)−V i (L, L)

  ⇐ High ability type can outbid low

  ability type for forming a pair with high ability type Digression (cont.)

  Dating back to Gary Becker’s

  • analysis on marriage Applications in development:
  • Ghatak (2000) for microfinance; Ackerberg & Botticini (2002) for sharecropping Legros & Newman (2007) for most
  • general treatment of the issue

1. Impact of Network

  

Characteristics (cont.)

  Need an exogenous variation in network characteristics Munshi (2003) uses rainfall shock in

  • Mexico as an IV for the migrant network size in US

1.1 Research question

  Does a larger network size of Mexican migrants in US increase the probability of employment of its members?

  Why important?

  • What’s original?
  • >Is it feasi

1.1 Research question (cont.)

  Why important? # of people living outside country of

  • birth: 175,000,000 in 2000

  2.9% of world population (2.2% in

  • 1965)

  Remittances: huge & growing

  • cf. See Yang (2008)’s lecture note for

  the literature on migration

  (1991-2005) Remittances vs. ODA, FDI Source: World Development Indicators 2007. Data are in current US$. 3

1.1 Research question (cont.)

  What’s original? Many have studied this question

  • But identification in these studies
  • not credible Innovation: Use rainfall as an
  • instrument for network size

1.1 Research question (cont.)

  Is it feasible? Mexican Migration Project: a

  • cross-sectional survey with recall data

  A cross-sectional survey of Mexican

  • communities (survey year differs) 200 hh heads in each community
  • surveyed for retrospective history of

    migration and employment

  ⇒ Panel data

1.2 Background

  Network members for migrants: Those from the same origin community Where to migrate varies a lot across

  • origin communities (Table II.B) 70% obtained a job via referrals
  • from relatives & paisanos (p. 562)

  ⇒ i’s Network = those from the same community

1.3 Theory

  Role of network: overcome information asymmetry Worker’s ability: unobservable to

  • firms Incumbent workers have better info.
  • on the ability of their network members

  ⇒ Firms rely on referrals from

1.3 Theory (cont.)

  ⇒ Incentive to refer high ability workers from their network cf. Anecdotal evidence (sec II.D)

  Established migrants play key role

  • More likely to be employed
  • Have been at destination for longer
  • More to lose if they are fired
  • Have developed firm-specific skills

  ⇒ # of established migrants, not all

1.3 Theory (cont.)

  Who benefits more from referrals Those w/ unfavorable observed

  • characteristics:

  women, older men, less educated

  • New migr

1.4 Empirical specification

  Sample: person-years located in US

  Y ict

  = αMN

  • βME
  • µ
  • η
  • ε

  ct

  ct

  i

  t

  ict

  Y ict : Dummy for being employed / for non-agricultural job

  • For migrant i from community c in year t
  • Nonagricultural job: higher-paying

1.4 Empirical specification (cont.)

  Y ict = α MN ct + β ME ct + µ i + η t + ε ict MN

  • ct : Ratio of surveyed members

  of community c located in US for 1-3 yrs by year t ME : Same as MN but for 4+ yrs

  • ct ct
  • α = 0, β > 0 • β: larger for new migrants, women, • older men

  Prediction:

  1.5a Identification issue (1)

  People from some origin

  • communities may be better skilled for US jobs

  ⇒ More people migrate ⇒ Network size bigger & each

  community member gets employed because of their skill, not network

  1.5a Identification issue (1) (cont.)

  Controlling for individual FE ability ( µ

  i

  ) solves this problem

  • US jobs are as low-skilled as those in Mexico (Tables I.B & II.A)

  ⇒ Individual skills likely to be

  time-invariant

  1.5b Identification issue (2)

  Business environment at

  • destination affects BOTH network size & employment probability ( ε )

  ict ict More people migrate if ε ↑

  • With serial correlation in business
  • environment at destination, even # of

    established migrants (ME )

    ct correlated with ε ict ict High ε causes return migration
  • (ME ↓) ct

  1.5b Identification issue (2) (cont.)

  ⇒ Use rainfall at origin as an IV for network size Rainfall ↓ ⇒ # of migrants ↑

  • Would-be migrants: mostly work in
  • rain-fed agriculture in Mexico (Table

I.B-C)

  1.5b Identification issue (2) (cont.)

  Rainfall in Mexico doesn’t affect

  • employment opportunities in US

  Corr. coeff. btw. rainfall in Mexico &

  • US: 0.01 (p. 570) Origin community is too small to affect
  • labor market conditions at the destination

  

⇒ No impact on US labor market

conditions

  1.5b Identification issue (3)

  Negative rainfall shock in the past

  • ⇒ Among migrants in US today, # of those staying long ↑ ⇒ Mechanically, they’re more likely to be employed (more opportunities to find a job)

  ⇒ Restrict sample to new migrants

  (Tables V(3), VI(4), IX(2))

1.6 IV estimation

  MN & ME : instrumented by

  ct ct

  Mean rainfall in community c over

  • years t to t − 2 Mean rainfall in community c over
  • years t − 3 to t − 6
    • t − 4) (Tables V(2), VI(3))

  

Results robust to a different cut-off (btw. t

− 3 &

1.7 Reduced-form / 1st stage

  results (Table V)

  Digression: reduced-form / 1st stage

  • Reader can check if instruments are

  Always show 1st stage results

  • not weak

  Better to report reduced-form

  • results as well

  

Reduced-form coefficients:

  • proportional to IV coefficients Maybe more relevant than LATE
  • (Deaton 2010)

  

1.8 Results: OLS vs IV (Table VI)

  dep. var.: indicator of being employed at destination

  Similar finding for having

  • non-agricultural job as dep. var. (Table IX)

1.8 Results: OLS vs IV (cont.)

  Why | ˆ β

  IV | > | ˆ β |? OLS

  Endogeneity due to return migration

  • Attenuation bias due to
  • measurement error

  Size of network based on random

  • sample of individuals from community

  Heterogeneous treatment effect (fn

  • 33)

  Rainfall affects low-ability individuals’

  • migration decision

1.9 Heterogenous treatment

  effects

  Network size effect: larger for New migrants (arrive in t − 1 or t)

  • (VI(4)) Women for employment (VI(6))
  • But not for occupation (IX(7))
  • Older men (above 45 yrs old) for
  • employment (VI(7))

  But not for occupation (IX(8))

  • Less educated (<10 yr

2. Impact of network behavior

  Important to distinguish the impact

  • of network behavior from the one of network characteristics

  If network behavior matters

  • ⇒ changing a few people’s behavior is

  enough to induce many more in the

network to change their behavior.

If only characteristics matters

  • ⇒ there will be no such spillover effect of

  a policy

2. Impact of network behavior

  (cont.)

  Early studies regress each person’s

  • behavior on the average behavior of their network members Manski (1993): this methodology is
  • wrong due to the reflection problem

  Does the mirror image cause the

  • person’s movement or reflect them?

  This is an issue of multicollinearity,

Reflection problem

  To answer why an individual tends to behave in a similar way to his/her network members, we want to estimate: y = α + βE(y |x) + E(z|x)

  �

  γ

  �

  • z

  η + u

  • x

  : membership indicators

  • z

  : individual-level determinants of y observed by econometricians

  • u: individual-level determinants of y
Reflection problem (cont.)

  y = α + βE(y |x) + E(z|x)

  �

  γ

  �

  • z

  η + u

  • β: “endogenous effect”

  ie. Impact of network members’ behavior

  • γ

  : “contextual effect”

  

ie. Impact of network characteristics Reflection problem (cont.) �

  Suppose E (u|x, z) = x δ

  • (ie. average u differs across networks)

  ⇒ If δ �= 0, network members behave in

the same way because their

unobservable characteristics that directly affect behavior are the same.

  • Network members share the same

  Two reasons for this:

  • environment (simultaneity bias)

  e.g. Geography, weather, business cycle Reflection problem (cont.) � �

  y = α + βE(y |x) + E(z|x) γ + z η + u Take expectation both sides of the outcome equation conditional on x

  �

  E (y |x) =α + βE(y |x) + E(z|x) γ

  � �

  • E(z |x)η + x δ

  �

  (⇐ E(u|x, z) = x δ) Solving for E (y |x) yields (if β �= 1): Reflection problem (cont.)

  So E (y |x) is a linear function of

  • E (z|x)

  

This is true even if δ = 0 (ie. no

  • omitted variable bias)

  � �

  = α + βE(y |x) + E(z|x) γ + z η + u y

  • cannot be estimated due to multi-collinearity
Reflection problem (cont.)

  Solutions (Manski 2000, p. 129): specify endogenous effect as

  1. Dynamic (ie. lagged mean)

  2. Nonlinear function of mean

  

3. Not mean behavior but, say, median

  behavior

  Conley & Udry (2010) follow these

  • three

  4. Some members affected by

  randomized treatment

  Randomized treatment approach is

  • popular by now But there is a caveat
  • Network may change in response to
  • treatment Carrell, Sacerdote& West (2013): a
  • policy designed by experimental evidence on peer effects may backfire

  

Conley and Udry (2010)

  Detailed data collection by

  • long-term fieldwork (every 6 weeks for 2 years) Knowledge of agriculture (how
  • pineapple grows)

  ⇒ Better identification strategy than

  past studies on social learning for technology adoption

2.1 Research Question

  Do pineapple farmers learn from their friends about the optimal usage of fertilizer?

  • Interesting?
  • If yes, only a few farmers need to be subsidized for universal adoption
  • Original?
  • Overcome the reflection problem
  • Feasible?

2.2 Background & Data

  Panel household surveys (every six

  • week in 1996-98) in 3 villages of southern Ghana Pineapple recently introduced in the
  • study area (Figure 3)

  ⇒ Must be room for learning

2.2 Background & Data (cont.)

  Pineapple takes 5 survey rounds to

  • mature after fertilizer is applied

  ⇒ Once applied, farmer cannot change the use of fertilizer in the same plot until harvest

  Pineapple grows throughout the

  • year

  ⇒ Not everyone plants at the same time ⇒

  Can exploit lagged response to network behavior, to avoid the

2.2 Background & Data (cont.)

  Outcome variable: Changes in

  • amount of fertilizer used Sample: 107 plantings by 47
  • pineapple farmers whose previous planting is also observed (closed circles in Figure 2)

  Other observed plantings are also

  • used for measuring regressors (open circles in Figure 2)

2.2 Background & Data (cont.)

  Each farmer’s network (“information

  neighbors”): obtained by asking

  Among 7 other farmers randomly

  • chosen from the sample, Whom they turn to for advice on their
  • farm

  cf. Previous studies often treat everyone else in the same village as network members

  ⇒ Median # of info neighbors: 2

  Location of all plots: collected by

2.3 Theory

  Basic ideas: Info. neighbors’ behavior per se

  • shouldn’t matter What matters is information each
  • farmer obtains from their info. neighbors Relevant information: expected
  • profits as a function of the amount
  • Farmer i updates E [π

  ⇐ By looking at change, we can control for farmer fixed effect

  ,t

  )?

  j ,s−5

  (x

  j ,s

  respond to π

  ,t p

  − x i

  < s ≤ t where t p is the period of i’s previous planting

  i ,t

  )

  ,s−5

  (x j

  ,s

  )] by observing neighbor j’s profit π j

  i ,t−5

  (x

  • t p
  • How does ∆x it ≡ x i

  Implications 1 & 2

  • When x j

  ,s

  = x

  

i

,t

p

  = 0 if good news on x j,s (ie. π j,s (x

j,s −5

) ≥ E i,t pj,s (x j,s −5 )])

  • ∆x it
  • ∆x it
  • When x
  • ∆x it
  • ∆x it

  

�= 0 if bad news

  j ,s

  �= x

  

i

,t

p

  �= 0 if good news on x j,s (ie. π j,s (x

j,s −5

) ≥ E i,t pj,s (x j,s −5 )])

  

= 0 if bad news

⇒ If bad news, i will take different

  behavior from j’s

  • Good news on x > x

  − x

  i ,t p

  − x

  i ,t

  ⇒ x

  i ,t p

  > 0

  i ,t p

  i ,t

  ⇒ x

  i ,t p

  , theory also predicts the direction of behavior change:

  i ,t p

  �= x

  ,s

  For good news on x j

  Implication 3

  • Good news on x < x

  < 0 Implication 4

  All of these effects should diminish

  • with farmer i’s experience

2.4 Measuring good (bad) news

  • j ,s j ,s−5

  For i’s expectation on π (x ),

  (for t < s ≤ t) p

  use median of π (x ) where

  k ,τ k ,τ −5

k : plots within 1km radius of i

  • τ ∈ {s − 3, s − 2, s − 1, s}
  • k ,τ −5 j,s −5

  

1 (x > 0) = 1(x > 0)

  • If π
  • j (x j ) exceeds this, it is a ,s ,s−5

  good news on x = x j ; otherwise

  ,s−5

  bad news

  So theory tells us that farmer i’s

  • behavior is a highly non-linear function of i’s network member behavior

  ⇒ Avoid the reflection problem

2.5 Testing implications 1-2

  1 s (good, x = x it p

  )

  Use logit estimation: Pr (∆x it �= 0) = Λ

  α

  • α
  • α

  2 s (good, x �= x it p

  )

  3 s

  (bad, x = x

  )

  • α

  4 s

  (bad, x �= x

  it p

  )

  • α
    • s (good, x = x

  5 ˜Γ it + zit

  α 6

  it p

  ): share of good

  it p

2.5 Testing implications 1-2

  s (good, x = x

  Use logit estimation: Pr (∆x

  )

  it p

  1

  • α
  • α

  �= 0) = Λ α

  4

  α 6

  � it

  5 ˜Γ it

  )

  it p

  s (bad, x �= x

  )

  it

  it p

  s (bad, x = x

  3

  it p

  s (good, x �= x

  2

  )

  • α
  • α
  • z
    • Theoretical predictions:

2.5 Testing implications 1-2

  Use logit estimation: Pr (∆x it �= 0) = Λ

  )

  1 s (good, x = x it p

  α

  • α

  2 s (good, x �= x it p

  • α

  3 s (bad, x = x it p

  )

  • α

  4 s (bad, x �= x it p

  )

  • α

  ˜Γ

  • z

  5

  it

  � it

  α 6

  )

  • S.E.: Conley (1999)’s spatial GMM
Changes in growing conditions

close

  ˜Γ it ≡ |x − x it |

  it p close

  x : Average of x ks where:

  it

  k : plots within 1km of plot i

  • ∈ {t − 3, t − 2, t − 1, t} s
  • ⇒ Control for common shocks faced by

  

it

Other controls z

  Wealth

  • Soil characteristics
  • Dummies for
  • Clan
  • Village
  • Survey round
  • Novice farmer (adopted in 1994 or
  • later)

2.5 Testing implications 1-2

  (cont.)

  Results (Table 4)

  • it p

  1SD ⇑ in share of bad news on x

  ⇒ Prob. of fertilizer use change ⇑ by 15%pt For bad news on x �= x , ⇓ by 9%pt

  • it p

  Mean prob. of fertilizer use change: • 13%)

  • it

  Robust to how to measure ∆x �= 0

  (columns B-C)

2.6 Testing implication 3

  Implication 3 says:

  it

  = x

  − x

  • Good news ⇒ ∆x

  j ,s−5

  i ,t p

  

it

  = 0

  • Otherwise, ∆x

2.6 Testing implication 3 (cont.)

  Therefore, define GoodNews (x ) × (x − x )

  j j i ,s−5 ,s−5 ,t p

  M i ≡

  ,t

  Experience

  it

  • j

    ,s−5

  GoodNews (x ): dummy for

  π (x ) above i’s expectation

  j j ,s ,s−5

  • it

  Experience : How many plantings i

2.6 Testing implication 3 (cont.)

  OLS estimation of

  � 3

  ∆x = β M + β Γ + z β + ν

  it 1 it 2 it it it close

  • it it it p

  Γ ≡ x − x : Changes in

  growing conditions for farmer i at time t

  z

  • it : same as before, plus Γ it defined

  from financial neighbors

2.6 Testing implication 3 (cont.)

  1 M it + β 2 Γ it + z � it

  ∆x it = β

  β 3 + ν it

  • S.E.: Calculated by Conley (1999)’s Spatial GMM

2.6 Testing implication 3 (cont.)

  Results (Table 5)

2.6 Testing implication 3 (cont.)

  • 1SD ⇑ in M i

  ,t

  ⇒ x i

  ,t

  ⇑ by 4 cedis per plant, larger than median level

  • Effect: bigger for novice pineapple farmers
  • Consistent w/ Implication 4
  • • Suggests external validity (Late

  

adopters do not respond less to

learning)

2.7 Robustness Checks

  Endogeneous network formation

  • drives the result? it

  Info. shocks: uncorrelated with z • , conditional on growing conditions

  (page 54) Info. neighbors: measured at t = 0

  • Predicted info neighborhood: same
  • result (Table 6 D)

2.8 Additional findings

  Own learning effect: equally

  • important (Table 6 A) Impact on labor use: similar result
  • for pineapple while no learning for maize-cassava (Table 7 A-B) Good news in geographic
  • neighborhood: misleading results (Table 7 C)

  

⇒ Measuring the ACTUAL network:

2.9 Future research

  If learning is important, info network

  • must be endogenous

  

⇒ Who will connect to early adopters?

Those who value info high

  • Those who incur low cost to link up
  • with them (gender, wealth, religion, etc.)

  What type of farmers should be

  • targeted to maximize technological
Banerjee et al. (2013)

  • How diffusion of new technology (joining microfinance) depends on who was first informed about it?
  • Estimate the model in which
  • If informed, joining is a function of

    own characteristics & fraction of

    informed neighbors joining
  • Information diffusion probability differs

  

by whether informed join or not

  • This model does not contain any
Banerjee et al. (2013) (cont.)

  Fraction of informed neighbors

  • joining does not matter Informed pass info even if not
  • joining Propose “diffusion centrality” that
  • can inform NGOs of the person to target to maximize diffusion
BenYishay and Mobarak (2013)

  To maximize technology adoption,

  • targeting average farmers and giving them an incentive to spread information is the most effective.

3. Network Structure &

  Enforcement

  In developing countries, legal

  • institutions are weak

  ⇒ Contracts cannot be enforced by third party

  Social networks play a role of

  • enforcement

  Grief (1993): Mediterranean traders in

  • 11th century McMillan & Woodruff (1999): firms in •
Early studies treat network

  • structure as a black box Some recent attempts to unpack
  • the black box

  Bloch, Genicot, & Ray (2008):

  • informal insurance Jackson, Rodriguez, & Tan (2012):
  • favor exchange

  

cf. Yves Zenou’s course on network in

  Q3

3.1 Research Question

  Jackson, Rodriguez, & Tan (2012) ask

  • What kind of network structure facilitates favor exchanges?
Why important?

  • (very relevant to developing countries)

  Without external enforcement

  , favor exchanges to achieve higher

  (e.g. informal insurance, credit)

  welfare require repeated interactions between a pair of individuals But each pair may not repeat

  • interactions frequently enough to
Why important? (cont.)

  Favor exchanges can still be

  • sustained if failing to do a favor to someone leads to no opportunity to receive favors from other people in the future Previous studies tend to assume
  • that everyone in the community punishes the deviator
Why important? (cont.)

  • Some people may prefer not

  This is unlikely.

  • punishing based on their own interest. The deviator may not ask a favor to • everybody else in the future.

  So we want to know which pair of

  • individuals in a community needs to be connected to sustain favor exchanges
Why original?

  • Provide a new concept of network structure (“support”) that is key to sustain favor exchanges
  • This concept is derived from game theory analysis

  cf. Sociologists asked the same question

and came up with a concept of “clustering” (how likely two of your friends know each other) without no

formal theoretical justification Why feasible? Mathematical skills...

  • Unique data on favor exchanges in
  • a network

3.2 Model

  n players

  • Discrete time (t ∈ {1, 2, ...})
  • Linked players can do favors for
  • each other The need for a favor arise randomly
  • over time
Model (cont.)

  Prob. that i needs a favor from j in

  • period t (if i & j are linked): p Value of a favor: v
  • Cost of a favor: c
  • Assume v > c > 0 (ie. Doing a
  • favor is socially optimal)

  δ ∈ (0, 1) Discount factor:

Model (cont.)

  Agents choose to keep or delete a

  • link at each period Keep a link ⇒ Doing a favor to the
  • linked agent when called upon Cannot add or rebuild links
  • >For tractability of analysis
In each period t

  • Agents decide whether to keep their links
  • Links are retained if mutually agreed.
  • At most one agent i t is called upon to do a favor for j t

  ∈ N

  i (g t )

  : Network in period t

  • g t

  (g t ): Set of agents linked to i (“neighbors”) in network g t

  • N i

  t

  decides whether to do a favor

  • i
  • • If not doing a favor, the link ij is

  deleted

3.3 Analysis

  Consider a case of n = 2 Expected value of a relationship per

  • period pv − pc

  Present value of keeping a

  • relationship p (v − c) 1 − δ

  ⇒ Doing a favor is preferred if

  δp(v − c) 1 − δ > c Analysis: n = 3

  Consider 3 agents connected to

  • each other (a “triad”) If failing to do a favor causes
  • ostracism, doing a favor is preferred if

  δp(v − c) 2 · > c 1 − δ

  ⇒ Network promotes favor exchange Analysis: Equilibrium refinements

  Consider smallest m that satisfies

  • δp(v − c) m · > c 1 − δ

  Any network where all agents have

  • at least m links is sustainable as a SPE

  If any favor is ever refused, all agents

  • delete all links
Once a favor is refused, some other

  • agents may want to deviate from the punishment strategy, in order to obtain higher payoff The punishment strategy (or
  • off-the-equilibrium strategy) should be renegotiation-proof

  

Definition: renegotiation-proof

  A network g is renegotiation-proof if g is sustainable as a SPE

  • No continuation equilibrium is
  • Pareto-dominated by another renegotiation-proof equilibrium

  

Definition: Renegotiation-proof

(cont.)

  Illustrate this concept for the case

  • where n = 4 and m = 2 Consider the following net

  Suppose player 1 fails to do a favor

  to player 2.

  

⇒ Players 1 & 2 no longer do favors to

  Then 3 & 4 no longer do favors to

  • Everyone prefers the original network to the empty network that would result from failing to do a favor

  Now consider this network

  Suppose player 1 needs to do a

  • favor to 3. 1 does favor if everyone else

  But the following network

  Pareto-dominates the empty network (and is renegotiation-proof as we just saw)

  So this network is NOT

  • renegotiation-proof.
One more equilibrium selection criterion

  Some people may fail to give a

  • favor just by chance If this leads to a huge change in
  • network structure, we should not observe the original network so often in reality Focus on networks that are stable
  • (or “robust”) to such random failure
Definition: Robust network

  A network is robust against social contagion if It is renegotiation-proof

  • For any network sustained in some
  • continuation, if i deletes a link ij, then only i or neighbors of i will delete their links.

  = 2, the left network is not With m

Definition: Supported links

  A link ij ∈ g is supported if There exists k such that ik ∈ g and

  • jk ∈ g
Theorem 3

  If No pair of players could sustain

  • favor exchange in isolation A network is robust against social
  • contagion then

  All of its links in the network are

  • supported
Proof of Theorem 3

  • (We will show this Suppose otherwise.

  implies that the network is not robust.)

  Then there exists a link ij ∈ g that is not supported. Consider player h ∈ {i, j} /

  • Deleting a link involving h results in
  • a robust network that includes link ij (see next two slides for why)
Proof of Theorem 3 (cont.)

  If h is not connected to either i or j

  • Deleting a link of h leads to a ro
Proof of Theorem 3 (cont.)

  If h is connected to either i or j

  • Deleting a link of h leads to a ro

  ⇐ No pair of players could sustain favor exchange in isolation

  Repeat the deletion of a link of a player other than i and j.

  • Remember robustness requires local

    link deletion for any network

    sustained in some continuation from the original network
  • Only the link ij will remain
  • But this link is not a robust network

3.4 Take it to data

  • Are links in favor networks “supported” in reality?
  • Data from 75 villages in Karnataka, India.

  (downloadable from Esther Duflo’s website)

  • Stratified random sample of HHs (half of the population) based on a full census
  • Each HH: asked to name
  • friends, those they visit/invite, borrow/lend kerosene or rice or money, give/receive advice, ask for medical
HHs i and j are measured as

  having a link if either of i or j mentions the other

  

If HHs outside the sample are

  • mentioned, such information is thrown away

  Measurement error causes

  • underestimation of # of supported links

  Only half of HHs interviewed • HHs may forget mentioning a link

  • For biases caused by measuring a network
A link is measured as supported if a

  • relationship of one type is supported by relationships of any other type

  

⇐ e.g. you repay the debt because you don’t want to lose connections with people who would give you advice on farming, not necessarily because you want to keep borrowing. Calculate the fraction of favor

  • network links (borrow/lend, advice, medical help) that are supported in each of 75 villages

  ⇒ In most villages, the fraction

  exceeds 50% (Figure 6)

  A lot lower fraction of any pair of HHs • have a common “friend” (Figure 7) x-axis : percentile of villages in terms of the fraction of supported favor network links

  Fraction of supported links: higher

  for favor networks than hedonic networks (visiting each other / talk most often) Other supporting evidence A link is more likely to be formed if • supported, even conditional on the geographic distance between the pair (sec

  VI.F) Ratio of fraction of pairs having a common • connected HH btw. linked & unlinked pairs in favor network: higher for pairs w/ different subcaste than w/ same subcaste (Table 4)

4. Further readings

  Recent papers on social networks in development Angelucci et al. (2010): extended

  • family in Mexico & secondary school enrolment Khwaja et al. (2011): business
  • network in Pakistan & credit access Jia et al. (2013): role of
  • connections in promotion of

4. Further readings (cont.)

  Social network can backfire Banerjee & Newman (1998): a

  • theory Munshi & Rosenzweig (2006): men
  • in India trapped in caste network while women benefit from trade liberalization

  Social network may solve poverty trap

  References

[1] Ackerberg, Daniel A., and Maristella Botticini. 2002. “Endogenous Match-

Economy 110(3), pp. 564-92. ing and the Empirical Determinants of Contract Form.” Journal of Political

[2] Manuela Angelucci, Giacomo De Giorgi, Imran Rasul, and Marcos Rangel.

ized social experiment.” Journal of Public Economics, 94, 197-221. 2010. “Family Networks and School Enrollment: evidence from a random-

  

[16] Legros, Patrick, and Andrew Newman. 2007. “Beauty is a Beast, Frog is a

[15] Jackson, Matthew O, Tomas Rodriguez-Barraquer, and Xu Tan. 2012. So-

ican Economic Review 102(5): 18571897. cial Capital and Social Quilts: Network Patterns of Favor Exchange. Amer-

[17] Jia, Ruixue, Masayuki Kudamatsu, and David Seim. 2013. “Complemen-

tary Roles of Connections and Performance in Political Selection in China.” 2007. Prince: Assortative Matching with Nontransferability.” Econometrica, July

  [29] Yang, Dean. 2008. “International Migra- http://dse.univr.it/ssef/documents/material2008/YangMigration.ppt

tion” BREAD Summer School Lecture Note