1

Social Networks

Imran RasulBREAD Summer School 2008

2

Social NetworksGiven incomplete markets, income volatility, informal institutions of contract enforcement, potentially important role for social networks in developing country contexts

This lecture we look at two examples:education [Progresa]technology adoption in agriculture

Will later see another example on network effects (Kremer and Miguel, 2004, health intervention), and in migration (Munshi 2003)

3

Example 1: EducationDe Giorgi et al (2007) Family Networks and School Enrolment: Evidence From a Randomized Social Experiment

4

Research Questions

5

Why Family?

6

Related Literature

Informalinsurance

Peer effects

7

The Progresa Program

8

The Evaluation Data

9

Surnames in Mexico

10

11

Heads have more familyties present than their wives

Because women more mobile atthe time of marriage

(Rosenzweig and Stark 1989)

Similar family structures within the household (under the same roof)

12

Isolated Households

13

Empirical Method

14

Empirical Method

15

Similar family structures outside the household and within the village, across eligible and non-eligibles

Intra-generational links are more common than Inter-generational links

16

Family Network Descriptives

Lots of variation within the same village

Family networks do not span more than three generations

17

Outcome: Change in Secondary School Enrolment

18

Estimating Equation

Pre

viousl

y es

tablis

hed

res

ults

last

lec

ture

(Shultz

2004)

20No response among eligible but isolated households, neither for boys nor girls

Similar enrolment rates at baseline

21

Similar enrolment rates at baseline

Such high enrolment rates at baseline imply that conditional cash transfers for primary enrolment act as de facto unconditional pure income transfers

22

Econometric Concerns

23

Econometric Concerns

24

Only eligible households thatare connected respond

The causal effect of beingembedded in a family network

No spillover effectsonto non-eligibles

25

26

Understanding Why Families Matter

actual transfers received are endogenouslydetermined by household behavior

hence use potentialtransfer in own household and average potential transfer in local family network, measured in October 1998

similar distributions of transfers households are eligible for among connected and isolated households [balanced on demographic characteristics of children at baseline]

correlation in potential transfers among ownhousehold and local family network is -.01

28

No responseamong isolated

No response if noresources available

to redistribute

29

Start to respond as moreresources in network

available to redistribute

30

Results

31

Which Family Members Matter?

32

Intra-generational Family Links

33

34

Policy Implication

35

Survey Design

36

Example 2: Social LearningLocal entrepreneurs may engage in –

Learning by doingLearning from others

The two types of learning have dramatically different implications for policy and the characteristics of growth

When individuals learn from their own experiences, they undertake an investment that yields uncertain returns

When they learn from each other, not only is there a risky investment, but that investment generates an informational spillover

These learning spillovers underlie some models of endogenous growth, and also imply a role for government intervention

37

These two types of learning form the process of “social learning”

Knowledge generated by experimentation by one individual or firm, increases the individual’s future profits and generates an informational externality that benefits other individuals or firms

Now present a standard target input model in which forward looking farmers use Bayesian updating to learn about the parameters of a new crop technology

The target input model has been developed by Prescott (1972), Wilson (1975), Jovanovic and Nyarko (1994) and applied to learning in agriculture by Foster and Rosenzweig (1995)

Alternative models of learning from networks include those basedon informational cascades (Banerjee 1992, Bickchandani et al1992), or rule-of-thumb learning (Ellison and Fudenberg 1995)

38

A Model of Social Learning (Foster and Rosenzweig 1995)

The parameters of the new crop vary across farmers, because of varying soil conditions sayWhile the form of the underlying production technology is known with certainty, one parameter - the target - is unknownFor farmer i in period t, output, qit, declines in the square of the distance between the input used, kit, and the uncertain target, κit-

39

The target input level, κit, is not known at the time the input is chosenAfter the input has been applied and output realized, the farmerupdates his beliefs on what the target input level isEach time the farmer uses the input can be thought of as a trialthat yields information on the distribution of the target parameterEach period is the length of time over which the input choice affects output, taken to be far shorter than one growing season

can be as short as a week if the input is the amount the new crop has to be watered and the output is plant height

Learning is the process of gathering information to better estimate what, on average, is the optimal target, κ* Farmers learn by doing, gaining information about the optimal target from their own past trialsThey may also learn from others, by observing the trials of their social network

The Target Input Model

40

The Optimal TargetThe optimal target fluctuates around κ* as follows –

Where μit~ i.i.d. N(0, σu²)These μit are transitory shocks to the optimal target inputWhile κ* is, on average, the optimal input levelIn period t farmer i has beliefs about κ* which are distributed as N(κit*, σκit²)

Two simplifying assumptions –σu² is knowninput is costless so farmer i's profit is his output multiplied by the constant price of output, p, normalized to one throughout

41

As Et(μit)=0, to maximize expected output, farmer i uses as his input level his expected optimal target level, so –

Hence expected output (profit) is –

Intuitively, expected output rises as the farmer has less uncertain beliefs on the optimal target, and as the variance of transitory fluctuations in the optimal target falls

42

Learning by DoingSuppose farmer i learns in isolation from othersEach period an input level is chosen, output observed, and an inference made on κit which is used to update beliefs about κ*In period t-1 the variance of i's prior belief about κ* is σκi,t-1²In period t, after observing κit, the farmer applies Bayes's rule to update his belief about the variance of κ*, yielding a posterior belief -

See Greene (1997)for details of Bayesian

updating

43

ρo=(1/(σu²)) is the precision of the information generated by i's own trialρi0=(1/(σκi0²)) is the precision of i's initial beliefs about the true value of κ*After repeated backwards substitution, we find can show that –

where It-1 is the number of trials i has with the new crop on his own farm between periods 0 and t-1Substituting this into the expected profit function –

Hence as expected, learning by doing increases expected output –

Weight each of these two factors

44

Learning from OthersIf farmers share the same distribution of the input target, theycan potentially learn from each others' trials

Define the network of farmers that share information with i as n(i)Suppose farmer i can costlessly observe the input choice of farmer j in n(i), κjt

In period t, farmer i uses Bayes's rule to update his prior belief on the variance of κ* using information both from the number of his own past trials, It-1, and the trials made by people in his network, n(i)t-1Hence his posterior belief on the variance of κ* is -

Initial beliefs Learning by doing Learning from others

45

Expected output now also depends on the number of trials of the network –

As one farmer uses the new technology, this creates an informational externality for all farmers within the network, increasing their expected output -

46

Implication 1

Information acquired by learning by doing and learning from others are substitutes for each other -

47

Implication 2

The learning externality is smaller for farmers that have more precise initial information on the true target level –

Hence farmers that are better informed to begin with, are less responsive to the choices of their network

48

Implication 3Farmers that obtain more precise information from their own trials, need to observe fewer trials among their network to have the same change in expected output –

Hence the learning externality varies across networks if in some networks, farmers are better able to observe the input choices of their network members

may be because some networks are more geographically concentrated than others so that soil conditions are more similar for farmers within the network

49

DataHouseholds panel data set from rural India (NCEAR, ACIS)Starting year: 1968-9 (soon after onset of the green revolution when HYV seeds first became available)Follow 4118 households till 1970-1, in 250 villagesInformation on area planted with new HYVs(for wheat and rice), assets (pre 1968)

Build village level aggregates using sampling weights

Descriptive: among

50

Descriptives

Level and intensity effects (experimentation in early years?)

51

Baseline Results

Own experience

Village experience

Own experience coefficients increase with inclusion of village experienceConsistent with notion that certain characteristics have opposite effects onown and neighbours usage…

52

The Adoption DecisionWe now consider each farmer's initial decision to adopt the new technologySuppose farmers have access to some traditional crop with riskless return qLet ait=1 if farmer i uses the new crop in period t, and ait=0 otherwiseIf farmers learn from others, a farmer's initial adoption decision depends on the adoption choices made by members of his social networkTo see this note that the value of the future stream of profits to i from period t to T is…

value of the future stream ofprofits to i from period t to T

Is-1=∑ais is the total number of trials i has conducted up to and including period sn(i)s-1 is the total number of trials of his social network over the same periodδ is the discount rateAs expected output increases in the number of own trials, once a farmer switches to the new technology, he never switches out of itAdoption is thus an absorbing state

Non adoption Adoption

54

The farmer adopts the new crop in period 0 if –

Note that adoption of the new crop may occur even when the traditional crop is currently more profitable. To see this note that since adoption is an absorbing state –

Hence the new technology may be adopted even if in the current period, the old technology is more profitable so that -

55

Farmer i's adoption decision in period 0 depends on the information available from his networkThe derivative of the net gains from adopting in period 0, with respect to n(i)₀ is -

Strategic delay effect (negative)Information externality effect (positive)

56

Two Countervailing ForcesThere are two opposing effects on i's incentive to adopt as the number of adopters among his network increases

First, his incentives increase because use of the new crop by network members creates a learning externality for i, which increases his current profitability if he adopts

Second, i has an incentive to delay adoption because the value of information he receives from his own adoption is lower as more network members adoptHence the gain in future profitability from one additional trial by i with the new technology is decreasing in the number of trials of the networkIf more network members adopt, less additional information is gained by the farmer himself adopting the new crop, giving rise to incentives to strategically delayadoption

57

Heterogeneity Across FarmersThe net gains from adopting in period 0 can therefore be increasing or decreasing in the number of adopters among the network

Intuitively, if a farmer is more myopic, he has less incentive to strategically delay, and the net gains are more likely to be increasing in n(i)₀

This creates a positive correlation between the propensity of farmer i to adopt and the adoption choices of his network, n(i)₀

Other things equal, for more forward looking farmers, the adoption choice is more likely to be negativelycorrelated to those of the network

58

Further ImplicationsThe effect of n(i)₀ on the propensity of farmer i to adopt is non linear as both the informational externality effect and the strategic delay effect are concave in the number of adopters in the network

The marginal effect of n(i)₀ on the net present value of the benefit of adoption is decreasing in ρi0

although both the informational externality effect and strategic delay effects are still present, the more precise information the farmer has to begin with, the less sensitive his adoption decision is to the number of adopters among his network

59

Empirical Evidence (Bandiera and Rasul 2006)

Why Should Networks Matter for Adoption?information is an important barrier to adoptionthis information can relate to the parameters of the new crop, or its profitabilitysocial networks help overcome these informational constraintsfarmers in the same network expect to share information, learn from each others' crop specific knowledge and experiences, and solve problems togetherenforcement of social normsrisk sharing Not related to information

60

OverviewThis paper asks –

is the anticipated sharing of information strong enough to lead to initial adoption decisions to be correlated within a network?

Related literature - Foster and Rosenzweig 1995, Munshi 2000, Conley and Udry 2003

Present evidence on the adoption decisions of 200 farmers across 9 villages, in the Zambezia province of northern Mozambique

Data were collected in conjunction with the NGO Movimondo, who introduced sunflower seeds in the area as part of a broader project promoting sustainable food security

61

62

Measuring Social NetworksMeasure the number of adopters known to each farmer among his family and friends

Social network is defined at a level far smaller than the village

allows us to distinguish the effect of private networks from correlated unobservables at the village or cohort levelevaluation of the importance of networks based on close contact, relative to those based on geographic, demographic or cultural proximity

63

Background and DataMovimondo started distributing sunflower seeds in March 2000Movimondo initially organizes an open village meetingFarmers receive the quantity of seeds they desire (typically ≤ 5kg)A contact farmer is appointed within each village

Seeds are distributed for freeReceiving seeds is not conditional on participation in other projectsDemonstration plots set up, extension workers offer ongoing advice (common to all farmers)

64

Survey Instruments

Household questionnaire to 198 household heads across 9 villages

primary household decision maker with regards to agriculture

Village questionnaire to village leaders –

a village elder, political leader and contact farmer

65

Benefits from Adoption

Sunflower seeds can be pressed into oil, clear nutritional benefitsPotential for commercialization of sunflower seeds

66

Barriers to AdoptionLand –

imperfect marketarea required to produce 1 liter of oil is less than one hectare

Labor –evidence that adopters produce less cashew

Credit –seeds are freely distributed, same input requirement sets for existing crops

Information –uncertainty over cultivation techniques

68

Social NetworksEach farmer, irrespective of their own adoption decision, was asked –

"how many of the people you know are planting the new crop?“"how many of these belong to your family?“"how many of these are your neighbors and friends?“

Measures the number of different sources of available informationCorresponds to why we believe networks matter for adoption

The question –relates to the subset of all individuals known, that were known to be adopting sunflowerdefined at the time of adoption

If correctly interpreted, this measure picks up the set of individuals the farmer both knows and are adopting sunflowerIt does not measure endogenously formed networks after the farmer's own adoption decision

Despite living in small villages, information is not a public goodNumber of farmers known adopting among family and friends is orders of magnitudes smaller than the number of adopters in the village

Catholics lessentrepreneurial?

More entrepreneurial?

Low levels of oilconsumption

71

Empirical MethodMeasure the information available to i on sunflower cultivation from his social network, n(i)0, as the number of adopters among his family and friends

As this network is individual specific and defined within the village, we control for village fixed effects, Zv

These control for village determinants of adoption such as the proximity to markets, land quality, and any component of information on sunflower cultivation that is commonly known to all farmers

Finally, control for i’s individual characteristics (Xi) to capture the precision of his initial beliefs about the parameters of thenew technology and other characteristics that determine the costs and benefit of adoption in period 0

72

Denote the unobservable (to the econometrician) present value of net gains from adoption to farmer i in village v, aiv* -

Dropping time subscripts, n(i) refers to the number of adopters in i’s network in period 0uiv contains unobserved individual and network characteristics that determine the present value of net gains from adoption

Net Gains From Adoption

73

The net gains from adopting may be increasing or decreasing in n(i), depending on whether the positive effect of the contemporaneous information externality prevails over the negative effect of the incentive to delay

In addition the effect of n(i) on aiv* is non linear as both the information sharing (positive) effect and of the strategic delay(negative) effects are concave in n(i)

To account for this, assume the present value of net gains to adoption to be non linear in the number of adopters among the network of friends and family and linear in all other variables -

Some non linear function

74

Actual AdoptionThe actual adoption decision, aiv, is a discrete choiceIt is observed and defined as –

Hence the probability that farmer i is adopts sunflower is;

We experiment with a number of alternative specifications for f (.)

Robust n

on

linear effect

76

Turning point consistentwith parametric specifications

No effects of povertyas expected

Farmer ability

Risk

Land quality?

Religion andentrepreneurship?

78

Do Networks Matter Equally for All Farmers?

Some farmers have good substitutes for information, such as their own abilityThese individuals are more able to solve problems arising from adoption on their own, all else equalIf networks overcome informational constraints, they matter less for those who have better information to start with, relative to the information they receive from their network

79

To see if this hypothesis finds support in the data, we allow the effect of the network to vary according to farmers’characteristicsThese characteristics proxy for the precision of his own initialbeliefs on sunflower cultivation, ρi0The present value of net gains from adoption are –

Where I and U identify farmers that are “informed” and “uninformed” about sunflower cultivation, respectivelyNamely I = 1 if the farmer is informed, 0 otherwise, and U = 1− IEstablish whether the marginal effect of the network is lower for informed farmers, that is if |ψ1 + 2ψ2n(i)| < |β1 + 2β2n(i)|

Networks matter more for the most “uninformed” farmers This is not due to lack of variation in network size amongst uninformed farmers

81

We find that the individual propensity to adopt is inverse-U shaped in the number of adopters among friends and family

This is consistent with social learning if the positive information externality provided by the network prevails in small networks, while the incentive to delay prevails in larger networks

We find heterogeneous effects of the network, as implied by a model of social learning

Summary

82

Other ExplanationsSocial learning is not the only mechanism that links adoption choices. Some explanations predict a positive relationship between adoption decisions in a network

marketing externalities in adoption

Other explanations predict a negative relation between adoption decisions within a network –

if networks provide insurance against idiosyncratic shocks

None of these other explanations by itself would predict an inverse-U pattern of adoption choices, heterogeneous social effects, and that the strength of social ties matters

83

ConclusionsThe result that networks matter differentially across farmer types has both methodological and policy implications that extend past the specific application of technology adoption in rural settings

result suggests that network effects are not symmetric across pairs of individuals

raises the possibility that a given individual may respond heterogeneously to the choices of different members of his social network. Thus, the identity of network members may also be important for social learning and, as a consequence, for optimal policy targeting