Networks, Information, and Vote Buying Ral Duarte Frederico Finan - - PowerPoint PPT Presentation

Networks, Information, and Vote Buying

Raúl Duarte Frederico Finan Horacio Larreguy Laura Schechter

Motivation

◮ Despite the secret ballot, vote buying remains pervasive

throughout the developing world.

◮ Political brokers are thought to:

◮ enforce the exchange of targeted benefits for votes; ◮ exploit their social connections to sustain vote-buying

exchanges;

◮ and acquire politically-relevant information about voters

through their social networks.

◮ Question: Do social networks diffuse information about

voters that brokers leverage to sustain vote buying?

Usual data challenges

◮ Lack of comprehensive data on social networks.

◮ Social network data from villages in which brokers and voters

are embedded.

◮ Lack of data on brokers’ vote-buying decisions.

◮ Data on vote-buying targeting reported by multiple political

brokers about overlapping voters in each village.

◮ Confoundedness of network measures.

◮ Voter characteristics as reported by both brokers and voters.

Our approach

◮ Compute a broker-voter network measure called hearing. This

is the extent to which a given broker might hear information about a specific voter in his network.

◮ Use broker and voter fixed effects to deal with broker- and

voter-specific confounders.

◮ Assess whether information about voters is diffused to brokers

through the network.

◮ Asses whether this information is used for targeting.

Main results

◮ Hearing significantly predicts:

◮ how much each political broker knows about each voter; ◮ who each political broker targets with vote buying; and ◮ whether a voter claims to support the broker’s party after the

election.

Robustness

◮ Broker- and voter-level confounding variables.

◮ Include broker and voter fixed effects.

◮ Homophily - brokers might both target and be closer to voters

who are similar to them.

◮ Control for the similarity between a voter and a broker in terms

• f their network position and their socio-demographic

characteristics.

◮ Results are not driven by parametric choices when computing

hearing, nor by potential bias due to partial network sampling.

Mechanisms

◮ The networks’ information-diffusion role:

◮ Hearing predicts how much the broker knows about the voter.

◮ Whether brokers target specific voters should not simply be a function of

how much they hear about them, but also what they hear about them.

◮ Brokers are more likely to target voters who are reciprocal and who

are not registered to their party, but only when the network allows brokers to hear information about them.

◮ Rule out that the network’s effect is explained by its enforcement role.

◮ Control for broker-voter network measures that the literature

suggests increase people’s ability to enforce informal transactions.

◮ Rule out that the network’s effect is explained by brokers targeting voters

who are good at convincing others.

◮ Control for voter fixed effects, and for network measures of how

well-connected a voter is interacted with hearing.

Literature on targeting of vote-buying

Who should be targeted?

◮ Voters with weak ideological attachment (Lindbeck and

Weibull, 1987; Dixit and Londregan, 1995).

◮ Core supporters (Cox and McCubbins, 1986; Nichter, 2008). ◮ Reciprocal voters (Finan and Schechter, 2012; Lawson and

Greene, 2014).

◮ Opinion formers (Schaffer and Baker, 2015).

This paper extends Finan and Schechter (2012) in at least two important ways.

◮ How do brokers know which voters are reciprocal? Because

social networks diffuse that information to them.

◮ Brokers are only more likely to target reciprocal voters about

whom they can learn information through the network, and more so when these are not registered to their party.

Literature on social networks in political economy

◮ Political economy more generally

◮ Voter level - Voters spread information about unemployment,

electoral violence, and elections. This affects their voting

• behavior. (Alt et al, 2019; Fafchamps and Vicente, 2013;

Fafchamps et al, 2019)

◮ Vote buying

◮ Candidate level - Well-connected candidates get more votes.

(Cruz et al, 2017)

◮ Voter level - Well-connected voters are more likely to (admit to)

being targeted. (Calvo and Murillo, 2013; Cruz, 2019; Fafchamps and Labonne, 2019; Schaffer and Baker, 2015) They are more likely to be targeted. (Ravanilla et al, 2017)

◮ Broker level - Brokers’ central position in non-political networks

explains their ability to influence vote choice. (Szwarcberg, 2012)

◮ We create a broker-voter level measure of connection, add

two-way fixed effects, and show networks diffuse information from voters to brokers who use it to target vote buying.

Roadmap

◮ Background ◮ Data ◮ Empirical Strategy ◮ Results ◮ Conclusion

Roadmap

◮ Background ◮ Data ◮ Empirical Strategy ◮ Results ◮ Conclusion

Brief political history of Paraguay

◮ Paraguay was a dictatorship under the rule of Alfredo

Stroessner of the Colorado party from 1954 to 1989.

◮ In 2008, an independent bishop won the presidency ending 61

years of Colorado rule.

◮ Paraguay remains largely a two-party country.

◮ 2006 elections: 66% Colorado mayors, 30% Liberal mayors.

◮ Political parties in Paraguay are not strongly ideological.

◮ “Policy has played little part in the campaigning for Paraguay’s

top job." (The Telegraph, 2008)

◮ “Competition among candidates is very personalized and

ideological differences are unclear." (Rizova, 2007)

Vote buying in Paraguay

Given the small ideological differences across parties, vote buying can be effective.

◮ “Elections in Paraguay are decided by the voters who are

mobilized with money. A very small percentage of the voters are loyal. The incentivized voters define [the election].” (A broker of the Liberal party in General Morínigo) Vote buying is becoming increasingly important to win elections.

◮ “There are three groups of voters: the captive, the thinkers,

and those that can be bought. Relative to previous elections the captive voters have declined and the voters that can be bought have increased.” (A broker of the Colorado party in General Aquino)

Political brokers

Political brokers (operadores políticos) act as intermediaries between candidates and voters, exchanging money and favors for promises to vote accordingly.

◮ “Political brokers are fundamental since they know their zone well.”

(Liberal politician in San Lorenzo) Brokers’ central positions in the network allow them to learn about voters.

◮ “[Brokers] know who [their] party supporters are.” (Liberal official in

Asunción)

◮ “[Brokers know] which Colorado and Liberal voters would sell their vote.”

(Liberal official in Coronel Oviedo)

◮ “It’s all about ñe’embegue (gossip).” (Colorado broker in General Aquino)

Brokers suggest that the voters who they target are likely to reciprocate with their vote:

◮

“While some voters take the money and vote for another candidate, the number of voters like that is small.” (Liberal broker in General Morínigo)

◮

“[The voters they target] always thank favors.” (Colorado broker in General Aquino).

Roadmap

◮ Background ◮ Data ◮ Empirical Strategy ◮ Results ◮ Conclusion

Sample

◮ Combine vote-buying data collected for Finan and Schechter

(2012) with social network data collected for Ligon and Schechter (2012).

◮ Panel data collected in ten villages across two departments in

Paraguay.

◮ 2002 - incentivized experiments measuring reciprocity. ◮ 2007 - added more households and collected social network

data.

◮ 2010 - interviewed political brokers.

2007 social network data

◮ Collected from 30 to 48 households in each of 10 villages.

Direct sampling rate between 12 and 91%, with a cross-village mean of 47%.

◮ Reach between 54 and 100% of all households directly or

indirectly (with a cross-village mean of 88%).

◮ Social connections include:

• 1. One hhd provided assistance when a member of the other hhd

fell sick.

• 2. One hhd provided monetary or in kind transfers to the other

hhd.

• 3. One hhd lent money to the other hhd.
• 4. One hhd would ask to borrow from the other in times of need.
• 5. Any two members of the hhds belong to the same family (i.e.,

parents, children, siblings).

• 6. Any two members of the hhds are “compadres.”

Social network example

◮ 257 links btwn 81 hhds (39 directly surveyed) ◮ Brokers are #s 9 (direct, 14 links) and 73 (indirect, 10 links).

Hearing measure

◮ Hearing, Hib, is the expected number of times broker b would hear

information originating from voter i (Banerjee et al. 2019).

◮ In t = 1, those directly connected to i find out the information with

probability p.

◮ In t = 2, those who received information in the first period transmit

it to nodes with which they are directly connected with probability p.

◮ In t = 3, those who received information in the second period

transmit it to nodes with which they are directly connected with probability np, where n is the number of nodes from whom they received the information.

◮ Process goes on T periods.

◮ Hib is the ibth entry of the matrix H = T

t=1(pg)t, where g is the

adjacency matrix (an element is 1 if the row and column households are socially connected and 0 if they are not).

◮ Set T equal to 7, the largest social distance between any voter and

broker in our sample.

◮ Set p equal to the inverse of the largest eigenvalue of the adjacency

matrix for each village’s social network. (Between 0.09 and 0.14.)

Hearing in period 0

Figure: Hib(0) = 0

Hearing in period 1

Figure: Hib(1) = 0

Hearing in period 2

Figure: Hib(2) = 0

Hearing in period 3

Figure: Hib(3) = 2 ∗ 0.363 = 0.09

Hearing in period 4

Figure: Hib(4) = 2 ∗ 0.363 + 2 ∗ 0.364 = 0.12

Hearing in period 5

Figure: Hib(5) = 2 ∗ 0.363 + 2 ∗ 0.364 + 19 ∗ 0.365 = 0.23

Additional network measures

◮ Individual-level characteristics:

• 1. degree centrality (number of connections);
• 2. diffusion centrality (average hearing);
• 3. clustering coefficient (share of connections that are connected

to one another);

• 4. betweenness centrality (share of shortest paths between any

two nodes in the network that go through a node); and

• 5. eigenvector centrality (recursive, a node is important if it is

linked to other important nodes).

◮ Broker-voter-level characteristics:

• 1. social proximity (inverse of distance);
• 2. existence of a support pair (having a friend in common,

Jackson et al. 2012);

• 3. number of support pairs; and
• 4. an indicator of prior non-political informal financial

transactions.

Questions asked to brokers

Ask them questions about all 30 individuals in our random sample.

◮ How well do they know the voter?

◮ Know spouse’s name? ◮ Know hectares land owned? ◮ Know years of education? etc...

◮ What were their political interactions with the voter in the

run-up to the 2006 municipal elections?

◮ Did they talk with the voter about politics before the election?

(48% average)

◮ Did they offer the voter vote-buying (in cash or in kind)?

(27% average)

◮ Other political outcome measure:

◮ Indicator for whether, a few months after the election, the

voter reports that he supports the broker’s party (46% average).

◮ Final sample - 296 voters and 932 broker-voter pair

• bservations.

Information outcome variables

◮ Complement 2010 broker data with 2007 voter data to

construct indices of brokers’ knowledge about voters.

◮ Demographics index combines whether the broker knows:

• 1. the voter (89%);
• 2. the voter’s spouse’s name (77%);
• 3. the voter’s years of education (81%); and
• 4. the amount of land the voter owns (42%).

◮ Political index combines whether the broker knows:

• 1. the partisanship of the voter (59%); and
• 2. if the voter voted in the 2006 election (63%).

◮ Social preferences index combines whether the broker knows:

• 1. the extent to which the voter would retaliate wrongdoing

(59%); and

• 2. if the voter trusts at least half the people in his village (66%).

Mediating variables to explain which voters are targeted

◮ Not registered to the same party using official electoral roll

data (59% average).

◮ Reciprocity - use the experimental measure from Finan and

Schechter (2012).

2002 incentivized experiment

◮ Trust game. Each player given 8000 Guaranies. Chooses to

send 0, 2000, 4000, 6000, or 8000 Guaranies to anonymous

• trustee. I tripled the amount sent, and the trustee decided

how much to keep and how much to return. Average of 3745 Guaranies sent. Returned on average 43% of tripled amount.

◮ Strategy method. ◮ All players played both roles.

Measuring reciprocity

◮ In the spirit of Cox et al (2008), we define reciprocity from the

2002 trust game.

◮ Reciprocity: the average share returned when receiving 12, 18,

• r 24 thousand Gs minus the share returned when receiving 6

thousand Gs. Censor this below 0.

◮ Mean is 0.043 with standard deviation of 0.076. ◮ Separates altruism from reciprocity. ◮ Only available for 85 voters and 271 broker-voter observations.

Who are the brokers and why do they talk to us?

Panel survey in 91, 94, 99, 02, and 07; with much continuity of survey team, high levels of trust. Asked households in ten villages to identify political brokers. They identified 43 brokers. We interviewed 38.

◮ 34 live in that village. ◮ 20 were from hhds in original survey (trust us). ◮ 20 Colorado, 17 Liberal, 1 UNACE. ◮ 36 men and 2 women. ◮ Average 51 years old, 8 years of education. ◮ 24 lived in the village all their life, other 18 lived there average

• f 20 years, minimum of 6 years.

◮ Worked as a broker between 5 and 48 years; average of 18

years.

Brokers’ unique network location

◮ Use individual-level data on all brokers and voters. ◮ How do brokers differ from non-brokers? Degree Betweenness Diffusion Eigenvector Centrality Centrality Centrality Centrality (1) (2) (3) (4) Broker 0.3844*** 0.1046 0.5739*** 0.6021*** (0.1233) (0.1538) (0.1457) (0.1428) Observations 1,032 1,032 1,032 1,032 Village FE X X X X Directly Surveyed FE X X X X

Roadmap

◮ Background ◮ Data ◮ Empirical Strategy ◮ Results ◮ Conclusion

Basic empirical strategy

◮ Baseline regression:

yibv = α + βHibv + ηbv + ǫibv,

◮ where yibv is an outcome for voter i and broker b in village v; ◮ Hibv is the hearing measure; and ◮ ηbv is a broker b fixed effect.

◮ Preferred specification also includes voter fixed effects θiv. ◮ Two-way clustering of standard errors, at the broker and voter

levels.

◮ Add interactions of controls with Hibv to test which

politically-relevant information is used to target voters.

Extra control variables

◮ Broker- and voter- fixed effects – control for person-specific

factors (broker’s centrality, voter’s social preferences, etc).

◮ Additional broker-voter-level controls in some specifications

◮ to get at homophily:

• 1. absolute age difference between broker and voter;
• 2. indicator for them being same sex;
• 3. geographic distance between their residences;
• 4. absolute difference in their individual centrality measures; and
• 5. broker fixed effects interacted with the voter centrality

measures;

◮ or to get at confounding with hearing:

• 1. social proximity;
• 2. the existence “support pairs” between them;
• 3. the number of “support pairs”; and
• 4. whether they engaged in an informal non-political transaction

in the past year.

Roadmap

◮ Background ◮ Data ◮ Empirical Strategy ◮ Results ◮ Conclusion

Results: vote buying

Broker Broker Support the Vote-buying targeting

• ffered

approached same index voter voter party (1) (2) (3) (4) (5) Hearing 0.2114*** 0.3215*** 0.0720** 0.1833*** 0.1174*** (0.0436) (0.0574) (0.0270) (0.0309) (0.0387) Dep variable mean 0.0000 0.0000 0.2725 0.4775 0.4614 Broker FE X X X X X Voter FE X X X X Observations 932 932 932 932 932 ◮ Column (2) - One sd increase in hearing → 0.32 sd increase

in vote-buying index.

◮ Column (5) - One sd increase in hearing → 12 pp (28%)

increase in likelihood that voter supports broker’s party.

Robustness

◮ Social desirability bias unlikely to cause bias here.

◮ Broker and voter fixed effects.

◮ Links might be formed among co-partisans - Freeman’s

segregation index (FSI):

◮ From 0 (randomly-generated network) to 1 (fully-segregated

network).

◮ Our villages have mean 0.13 and sd 0.15. ◮ Benchmark of Twitter networks in the 2009/2010 US Congress

is 0.59 (Sparks, 2010).

◮ Brokers might target voters with similar traits, and similarity

in traits might correlate with hearing.

◮ Add broker-voter measures of homophily.

◮ Social network’s use to enforce informal transactions might

also confound our results.

◮ Add broker-voter measures of network’s enforcement potential.

Results: robustness

Vote-buying targeting index Hearing 0.3341*** 0.2474* 0.3456*** 0.3656*** 0.3208*** 0.3729** (0.0869) (0.1241) (0.0586) (0.0740) (0.0568) (0.1490) Social proximity 0.0607 0.0352 (0.0961) (0.1023) Existence of support pair

• 0.0989
• 0.1499

(0.1066) (0.1395) Number of support pairs

• 0.0761
• 0.0582

(0.0853) (0.1216) Previous transaction 0.0048

• 0.1720

(0.1335) (0.1683) Broker FE X X X X X X Voter FE X X X X X X Broker-voter controls X X Observations 932 932 932 932 932 932

◮ Point estimate in column (1) unchanged. (Controls: a) same gender; b)

distance in km; c) difference in age, degree, clustering coefficient, betweenness cent, diffusion cent, and eigenvector cent; and d) broker fe interacted with voter’s network measures.)

◮ None of the enforcement measures predicts targeting.

Results: information diffusion

◮ Does hearing predict brokers’ knowledge about voters? Demographics Political Social Preferences Index Index Index (1) (2) (3) (4) (5) (6) Hearing 0.2692*** 0.1703*** 0.3293*** 0.2088*** 0.1621*** 0.1401*** (0.0425) (0.0411) (0.0524) (0.0508) (0.0499) (0.0441) Broker FE X X X X X X Voter FE X X X Observations 932 932 932 932 932 932 ◮ A one-sd increase in hearing is associated with

◮ a 0.17-sd increase in knowledge of demographic characteristics; ◮ a 0.21-sd increase in knowledge of politically-relevant

information; and

◮ a 0.14-sd increase in knowledge about voter social preferences.

Mediating variables

◮ Brokers should not just be more likely to target those about

whom they hear more information;

◮ they should be especially likely to target voters about whom

the information they hear makes them think a vote-buying transfer will be more effective.

◮ We add interactions of hearing with:

• 1. voter reciprocity and
• 2. whether the voter is registered to the broker’s party.

Results: mediating variables

Vote-buying targeting index (1) (2) (3) (4) (5) (6) (7) (8) Hearing 0.241*** 0.284*** 0.235*** 0.266** 0.156** 0.116 0.149** 0.138* (0.055) (0.100) (0.052) (0.107) (0.061) (0.080) (0.057) (0.074) Experimental reciprocity

• 0.064

0.145 (0.066) (0.142) Not reg. to same party

• 0.666***
• 0.773***
• 0.692***
• 0.813***

(0.127) (0.167) (0.137) (0.176) Experimental reciprocity × Hearing 0.127* 0.122

• 0.081
• 0.056

(0.067) (0.138) (0.081) (0.128) Experimental reciprocity × Not reg. to same party

• 0.210
• 0.249

(0.162) (0.221) Not reg. to same party 0.056 0.040 0.052 0.052 × Hearing (0.099) (0.141) (0.091) (0.139) Not reg. to same party × Hearing × Reciprocity 0.287** 0.284* (0.105) (0.144) Broker FE X X X X X X X X Voter FE X X X X Observations 244 244 244 244 244 244 244 244

◮ Political brokers are more likely to target the voters who are

not affiliated with their party if they have learned through the network that the voter is reciprocal.

Alternative explanation

◮ Brokers might target voters who are good at persuading

• thers – voter fixed effects help.

Vote-buying targeting index Hearing 0.357*** 0.360*** 0.335*** 0.332*** (0.062) (0.073) (0.070) (0.056) Voter degree × Hearing

• 0.070**

(0.031) Voter diffusion centrality × Hearing

• 0.056

(0.041) Voter eigenvector centrality × Hearing

• 0.024

(0.046) Voter betweenness centrality × Hearing

• 0.051*

(0.026) Broker FE X X X X Voter FE X X X X Observations 932 932 932 932 ◮ Brokers are not differentially targeting more well-connected

voters in an attempt to purchase more persuasion.

◮ Similar exercise with broker’s centrality looks the same.

Competition between parties

Vote-buying targeting Overall knowledge index index (1) (2) (3) (4) Hearing 0.279*** 0.265*** 0.219*** 0.231*** (0.063) (0.067) (0.049) (0.056) Mean other

• 0.102
• 0.115

0.010 0.021 party hearing (0.091) (0.092) (0.077) (0.084) Hearing × Mean

• 0.002
• 0.003
• ther party hearing

(0.043) (0.051) Broker FE X X X X Voter FE X X X X Observations 932 932 932 932

◮ Possibility of strategic interactions between brokers – but only

a broker’s own hearing matters for whether he targets the voter, the hearing of other brokers does not.

Roadmap

◮ Background ◮ Data ◮ Empirical Strategy ◮ Results ◮ Conclusion

Conclusion

◮ Vote buying is pervasive, and political brokers are key

intermediaries.

◮ We use ‘unique’ network and vote-buying data to show that

the amount of info a broker can hear about a voter through the network

◮ predicts whether the broker targets him and ◮ predicts how much the broker learns about the voter.

◮ Targeting depends not only on how much information the

broker can hear, but also on what he hears.

◮ Among voters not registered to their party and about whom a

broker can hear more information, brokers target the reciprocal.

◮ Information diffusion through social networks can help sustain

an electoral practice widely believed to weaken political accountability and limit public good provision.