1 Network Effects on Social NetworksQuoc-Anh Do (Sciences Po, Paris) Institutional and Organizational Economics Academy 22-26 May 2017

2 Social Networks and Economics“Social networks” in economics? Economic sociology: Embeddedness of economic activities/interactions in non-economic relationships, even in market economies (Mark Granovetter) Modern economics: structure of pairwise relations that conditions all individual choices. Examples: trade costs due to physical and non-physical distances (global, local levels), trust and organization of the firm (mergers/acquisitions), friends/relatives/compatriots and labor market choices Recent boom in economic research on social networks. Empirical economists rely heavily on the persuasiveness of the empirical methodology used to test hypotheses on data

3 Definition of social networks in economicsMathematical definition of a social network: a graph A set of nodes/vertices: V A set of edges/links/connections: E. Edges may be directional or undirectional. They may carry weights (distance). There can be different types of edges, corresponding to different dimensions of a network (e.g. links between family members, between creditors-debtors, between co-workers) Some simple examples: Trade between countries (total/by industry) Group structure: classroom, ethnic group, etc.

4 Network of U.S. senators, 1999

5 Network of U.S. senators, 2013

6 Andris C, Lee D, Hamilton MJ, Martino M, Gunning CE, et alAndris C, Lee D, Hamilton MJ, Martino M, Gunning CE, et al. (2015) The Rise of Partisanship and Super-Cooperators in the U.S. House of Representatives. PLOS ONE 10(4): e https://doi.org/ /journal.pone

7 Empirical questions on social networksWhat determines the network structure? Testing theories of network formation Individual/pairwise/full-network determinants of links Does the network matter? To what extent? Links versus non-links Variation among peers/connected nodes Network position Testing different theories based on given network structure, such as information diffusion, social learning, games on network, etc. In reality: usually both stages

8 A typical example from the literature12/10/2017 Hogwarts: Assignment into groups by “magic hat”. Same-group students interact a lot Friendship forms more among same-group students … and less across different groups A typical example from the literature, similar to most cases studied before Social Networks and Beliefs Quoc-Anh DO

9 Same-group and friendship influencesPeers in the same group may affect each other But true friends have much more influence

10 Empirical challenges The “peer effect” individual equation:Link individual outcomes to some averages of peers/connected nodes: 𝑦 𝑖 =𝛼+𝛽 𝑦 𝑁(𝑖) + 𝛾𝑥 𝑖 + 𝛿 𝑥 𝑁(𝑖) + 𝜖 𝑖 𝑁(𝑖) denotes the relevant peer group (𝑖’s neighbors) Two problems: 1. Technical identification: Can we estimate separately 𝛽? The reflection problem (Manski 1993); conditions to pass it (Bramoullé et al. 2009) 2. Causal interpretation: Is 𝛽 evidence of causal peer effect? Angrist 2014: it may have resulted from correlated shocks.

11 Difficulties of estimating beta𝑦 𝑖 =𝛼+𝛽 𝑦 𝑁(𝑖) + 𝛾𝑥 𝑖 + 𝛿 𝑥 𝑁(𝑖) + 𝜖 𝑖 A social multiplier effect: effects on 𝑦 𝑖 get “echoed” When peer groups are a partition: multiplier is 1 1−𝛽 (𝛽<1). Not possible to estimate 𝛽 separately: Manski’s (1993) reflection problem When peer groups are not exactly a partition (there are overlaps): Bramoullé et al. (2009) shows valid instruments can be constructed from friends of friends, friends of friends of friends, and so on.

12 Difficulties of causal interpretation𝑦 𝑖 =𝛼+𝛽 𝑦 𝑁(𝑖) + 𝛾𝑥 𝑖 + 𝛿 𝑥 𝑁(𝑖) + 𝜖 𝑖 Even if the equation is identified: Angrist 2014: common shocks that affect 𝑥 𝑖 or 𝑦 𝑖 would produce a positive estimate of 𝛽. Only correlation! The set of neighboring nodes 𝑁(𝑖) may result from individual choices (depending on the network formation process). Correlated selection into peer groups 𝑁(𝑖) may produce a positive estimate of 𝛽. Endogeneity bias due to homophily: “Birds of a feather flock together”

13 Elements of solution Angrist’s suggestion: Separate “receiver” and “source” of peer effects; with exogenous shocks from those sources Example: Kling et al.’s (2007) study of Moving To Opportunities Treated households are randomly given vouchers to move to better-off neighborhoods Estimate the effect of the neighborhood on treated households To address homophily: randomize connections, or find an exogenous instrument variable for connections Implementation of an IV for connection is difficult in a framework of individual-level equation

14 Alternative specificationThe dyadic regression equation: 𝑦 𝑖𝑗 =𝛼+𝛽 𝐿 𝑖𝑗 + 𝛾𝑥 𝑖𝑗 + 𝜖 𝑖𝑗 Each observation is a dyad (pair). Variables need to be constructed by pair (such as differences in certain dimension) 𝐿 𝑖𝑗 denotes the link between 𝑖 and 𝑗. It can be instrumented for. The coefficient 𝛽 could signify convergence in the outcome due to the links. Individual-level interpretation of 𝛽 (e.g. who is converging to whom) is a bit more difficult.

15 Some examples on firms’ political connectionsFirst example on estimating impacts of political connections Question: Do firms draw unmerited benefits from social connections with politicians? We will go through standard designs + my own work

16 Cross-country evidence: Faccio 2006 AERPolitician Top corporate officers/ Large shareholders 1) Politician Top corporate officers/ Large shareholders Chiefs of State (government, parliament) 2) Politician Firm 3) Politician  Board Officers/Large SHs  Politics Politician Firm Cumulative Abnormal Returns 4)

17 Connections to Parties: Goldman et al. 2009Political Party Board Members 1) Political Party Board Members board members with political positions 2) Change in control of both House and Senate following 1994 midterm election, and 2000 Δ in presidency Politician Firm 3) Politician Firm CAR, Procurement 4)

18 Connections to VP Dick CheneyFirms 1) VP Dick Cheney Firms Past CEO, board members, and Halliburton’s board 2) Politician Firm Heart attacks, Cheney becomes running mate 3) Politician Firm CAR 4)

19 Do et al.’s empirical designElected Congressman E Board member M Market reaction to value Former classmates Firm A Differential Value of Connection RDD of close elections: At the limit when vote margin0, equivalent to randomization between winner and loser Defeated Candidate D We construct the social links between politicians and directors using former classmate networks. Politician E and director M are considered connected if they graduated from the same university program within 1 year of each other. Our identification comes from RDD of close elections to US Congress, and our outcome variable is post-election market reaction to value of connected firms. That is, suppose firm A is connected to winning politician E through director M, and firm B is connected to defeated politician D through director N, our treatment effect measures the differential market reaction to firm A and firm B right after the election, which can be interpreted as the value of the connection to the elected politician. Endogeneity and causality problems Unobserved characteristics of elected/appointed politicians and firms; homophily of links Reverse causality: strong firms connect to strong politicians Events are subject to endogeneity issues Concurrent events; prediction precision of event probabilities, especially for rare events Measurement of social networks of politicians and firms By previous experiences, ownerships: endogenous By campaign contributions: hard to address connections to individual politicians, because firms cannot contribute Connections to individual politicians: Look at alumni networks Board member N Market reaction to value Former classmates Firm B

20 Main results of the paperOn average: Negative effect on firm value when directors’ friends are elected to Congress Effect dominated by politicians from positions in state politics, not by politicians from federal offices (incl. incumbents) Explanation: loss of benefits when politicians leave (corruptible) state politics Different from Shleifer Vishny’s (1994) effect as tested by Bertrand et al. 2008 This loss is stronger for more corrupt states, better-governed firms, possibly because of intense media scrutiny differences between federal and state politics

21 Contributions New finding: Smaller benefits when friends are elected to higher offices Benefits are not just function of power, but also scrutiny Explains Tullock’s (1972) puzzle “Why is there so little money in politics?” (complement Ansolabehere Snyder 2003 JEP) Implications on corporate behaviors, institutional design of elections, and checks and balances Results focus on a well-defined type of connections: former university classmates RDD avoids reverse causation, endogeneity of connections, of events; RDD is robust to market’s prediction errors of events Our paper contributes a new finding that in the context of US Congress election, higher offices may entail smaller benefits, and benefits depend not only on power but also on scrutiny. This novel finding helps explain Tullock’s puzzle “…”. This provides implications on “…”. In addition, by implementing RDD, we also greatly improve the quality of the evidence, compared to previous works based on event studies. Finally, we look at a different type of connection, social links instead of direct links. Simple point: scrutiny matters, because favoritism in a democracy can lead to legal problems (unlike nondemocracies) Ansolabehere Snyder 2003 JEP: campaign money is a form of political participation, because low impact on politicians’ decision. “This finding helps to explain Tullock’s (1972) puzzle. Money has little leverage because it is only a small part of the political calculation that a re-election oriented legislator makes. And interest group contributors—the “investors” in the political arena—have little leverage because politicians can raise sufficient funds from individual contributors. It is true that when economic interest groups give, they usually appear to act as rational investors (for example, Snyder, 1990, 1992, 1993; Grier and Munger, 1991; Romer and Snyder, 1994; Kroszner and Stratmann, 1998, 2000; Ansolabehere and Snyder, 1999, 2000a). However, this “investor” money from organized groups accounts for only a small fraction of overall campaign funds. Since interest groups can get only a little from their contributions, they give only a little. As a result, interest group contributions account for at most a small amount of the variation in voting behavior. In fact, after controlling for legislator ideology, these contributions have no detectable effects on the behavior of legislators.” “Since interest groups can get only a little from their contributions, they

22 RDD implementation 𝐶𝐴 𝑅 𝑖𝑓𝑡 =𝛽𝑊𝑖𝑛𝐿𝑜𝑠 𝑒 𝑖𝑡 + 𝑓 𝑊 𝑉𝑆 𝑖𝑡 |𝑊𝑖𝑛 + 𝑓 𝐿 𝑉 𝑆 𝑖𝑡 𝐿𝑜𝑠𝑒 + 𝜀 𝑖𝑓𝑡 Obs. unit: Connection (Pol-Dir) x Firm x Election Year 𝐶𝐴 𝑅 𝑖𝑓𝑡 : market’s cumulative abnormal returns to firm’s stock value 𝑊𝑖𝑛𝐿𝑜𝑠 𝑒 𝑖𝑡 : dummy indicating election success 𝑉 𝑆 𝑖𝑡 : vote share (centered around 50% threshold) Robustness checks: alternative outcomes, additional controls & FEs, different clustering schemes, different polynomial orders, different bandwidths, different kernel functions Our empirical strategy relies on sharp RDD implementation. …

23 Economic interpretationTreatment effect of a connection to an elected Congressman, compared with a connection to a defeated candidate Positive effect: Congress connection more valuable vs. alternative Negative effect: Congress connection less valuable vs. alternative Weighted average treatment effect among firms connected to politicians, weighted by politician’s probability of running a close election. The treatment effect measures the value of a connection to an elected Congressman, compared to that of a connection to a defeated candidate. Our estimate is a WATE among connected firms, weighted by the politician’s probability of running a close election.

24 Politician-director connectionPoliticians: Hand-collected for all politicians involved in 5%-vote margin elections for US Congress, Directors: BoardEx covering past education, employment, other activities of all boards of major public firms, Classmate networks: connect politician-director if they finished the same university program within a 1 year gap (Cohen et al. JPE 2008) Missing data: 2 year gap birth dates University program: undergraduate, general graduate, professional (business, law, medical), precise to campus Other measures: 2-3 year gap Alumni network: no condition on year gap

25 Outcome variable CAR Outcome variables: Cumulative Abnormal Returns (CAR) of connected-firms’ stocks (cumulated daily changes in stock prices, minus predicted changes estimated by standard market model) Different windows of events: 1 week before to 1 month after; Benchmark: 1 day before to 1 week after (-1, +5) Robustness checks with different market models (using different market factors), raw returns, different windows Standard market model: regressions of firm returns on market returns using daily data from one year to 2 months before the event date, separately for each firm. Standard practice in the finance literature. FF 3 factor: additional D(small & lager caps), D(high & lower market to book) A Z-shaped graph of CAR as function of vote margin

26 Cross-sectional identification by RDDWhen election result is partly predicted, CAR only reflects the unpredicted part (pitfall of event studies looking at CAR) Suppose that the market’s prior for a politician’s win is 90%. The return to his actual winning reflects only 10% of the value. Same for his opponent. With close election RDD, pre-election probability does not matter Election result almost randomized at threshold  controls, FEs, market factors used in CAR construction have no expected impact on estimate, but can help reduce noises Some examples of bad market predictions of probabilities: Brexit, Trumpit Justin Wolfers’ calculation of market value of Trump’s presidency based on future market’s predicted probabilities: off by a large margin

27 Data summary Number of close elections 101 Number of connected firms 1,268 # of Senate elections 20 median market cap ($ million) 595 # of Congress elections 81 median employment 1,480 election average margin 2.57% average CAR(-1,+5) 0.00 % of wins by Democrats 57% avg. # of connected politicians 1.3 Number of politicians 170 Number of connected directors 1,171 # coming from state politics 64 1.1 # coming from federal offices 89 avg. # of firms % of Democrats 54% Number of universities 117 avg. # of connected directors 7.7 % of undergraduate connections 88% avg. # of connected firms 9.8 % of graduate connections 12% Final sample includes 1,312 Pol x Dir connections, resulting in 1,819 observations at (Pol-Dir) x Firm x Election Year level.

28 Covariate tests for selected politician, director, and firm characteristics(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Politician characteristics Director characteristics Firm characteristics (1-Year Lag) DEPVAR Gender Age Grad. Degree State Politics Log (Emp.) In-state Activi-ties Gov. Con-tracts Hldgs. by Inst. Owners Board Size Win/Lose 0.100 2.673 0.024 -0.173 -0.030 2.685 -0.016 0.309 45.96 0.010 -0.113 [0.119] [2.098] [0.238] [0.196] [0.041] [2.053] [0.080] [0.269] [0.050] [28.64] [0.020] [0.547] Obs. 1,817 1,797 1,819 1,722 1,484 1,809 1,717 1,087 1,244 No significant differences between winners and losers, between directors connected to winners and losers, or between firms connected to winners and losers.

29 Connections to elected Congresspersons bring negative valuesCAR(-1,+5)  (1) (2) (3) (4) (5) (6) (7) Sample Full sample State Politics Federal Offices State Politics & Local Firms State Politics & Non Local Firm Congress (Incum-bent) Non-Congress Federal Offices Win/Lose [0.0085]*** [0.0131]*** [0.0138] [0.0199]*** [0.0170] [0.0141] [0.0253] Difference [0.0181]** Observations 1,819 594 742 352 242 598 144 Significant 4.8% difference in CAR between firms connected to losers and winners coming from state politics; even stronger effect among local firms. Effect robust to additional controls & FEs, different clustering schemes, polynomial orders, bandwidths, CAR constructions, and non-parametric procedure.

30 The negative discontinuity effect

31 Effects vary by firm size

32 Some anecdotal cases Senator George Allen: Senator Bob CorkerClosely elected in 2000 Previous Governor of VA Graduate from UVA Connected to Theodore Coburn of Measurement Instruments Inc. CAR suffered a big loss the week after election Senator Bob Corker Closely elected in 2006 Previous mayor of Chattanooga, TN (with controversies) Graduate from U of TN Connected firms such as First Acceptance Corp, Wright Medical Group in TN lost significant value

33 What explains the negative effect?Value decrease from State to Federal State politics is more corrupt (Glaeser Saks 2006)  connection to State politicians creates value for firms Federal politics is more scrutinized (Fisman et al. 2012)  previous value of connection is lost Also: “Somebody” in state becomes “nobody” in Congress Also: Longer distance leads to less frequent interactions Alternative view: Politicians force firms to over-employ (Shleifer Vishny 1994, Bertrand et al. 2008) Exploratory tests for value from connection to state politics: Post-election firm’s real outcomes Effect by firm’s corporate governance quality Effect by state’s corruption level & institution quality Media scrutiny & effect by distance to DC

34 Reduction in government contracts(1) (2) (3) (4) (5) (6) (7) (8) 2-Year Change in: No. of Obligated Transactions Having Obligated Transactions No. of Procurement Contracts Having Procurement Contracts Sample State Politics Federal Offices Win/Lose -67.57 2.49 -0.143 0.024 -32.20 12.70 -0.109 -0.063 [30.56]** [41.82] [0.051]*** [0.058] [18.68]* [35.70] [0.057]* [0.074] Observations 561 690 The first piece evidence of the channel is reduction in number of government contracts within 2 years of the election. The decrease happens in both the number of actual transactions (obligated transactions), and the number of contracts (procurement contracts), and most importantly, only among firms connected to state-level politics. Data source: Federal Procurement Data System Columns (1)-(2): 2-year changes in number of obligated transactions Columns (3)-(4): changes in whether the firm is awarded at least one obligated transaction Columns (5)-(6): changes in number of all procurement contracts Columns (7)-(8): changes in whether the firm is awarded at least one procurement contract Number of government contracts goes down among firms connected to winning politicians with coming from state politics.

35 Reduction in in-state activities, employment, director’s tenure(1) (2) (3) (4) (5) (6) DEPVAR Post-Election Change in Local Newspaper Mentions Post-Election Change in log(Employment) Director’s RemainingYears Sample State Politics Federal Offices Win/Lose -2.832 -0.572 [0.0044]** [0.0042] [0.0354]* [0.0486] [1.484]* [1.109] Observations 593 741 480 623 594 742 Connected-firms reduce (in-state) activities and employment after politician wins. Connected-directors remain with the firm for shorter period after politician wins. Noisy but qualitatively similar results for other firm real outcomes. Columns (1)-(2): Firm activities are measured in a given state in a given year as "Firms Reported In Local Newspapers" (FRILN), namely the number of search hits for the firm's name in local newspapers normalized by the number of search hits for the neutral keyword "September” (Saiz and Simonsohn’s (2008) idea of “downloading wisdom”).

36 Better-governed firms are more affectedCAR(-1,+5)  (1) (2) (3) (4) (5) (6) (7) (8) Sample Politicians Coming from State Politics Larger Hldgs. by Inst. Block Owners Smaller Hldgs. by Inst. Block Owners More Inst. Block Owners Fewer Inst. Block Owners Smaller Board Size Larger Board Size Higher Shares of Indepen-dent Dir. Lower Shares of Indepen-dent Dir. Win/Lose -0.099 0.023 -0.112 -0.002 -0.082 -0.032 -0.059 -0.057 [0.028]*** [0.018] [0.039]*** [0.023] [0.031]** [0.034] [0.031]* [0.033]* Difference -0.122 -0.109 -0.050 -0.001 [0.037]*** [0.052]** [0.054] [0.052] Observations 182 169 162 279 244 145 218 171 If it were Schleifer & Vishny story (i.e. loss of value by being forced to do favors for winning politicians), we would expect better governed firms to suffer less of a value loss. Effect is stronger under stronger corporate governance: loss of value due to loss of state-political connections after the politicians’ successful elections

37 Effect is stronger in states with higher corruption level (1/2)CAR(-1,+5)  (1) (2) (3) (4) (5) (6) Sample Politicians Coming from State Politics More Corrupt Main City Less Corrupt Main City More Corrupt State Less Corrupt State More Corrupt Conviction Less Corrupt Conviction Win/Lose [0.0264]*** [0.0186] [0.0228]*** [0.0211]* [0.0406]* [0.0183]** Difference [0.0321]* [0.0308] [0.0442] Observations 318 276 341 253 258 336 Columns (1) & (2): use Exalead.com 2009 search hits for the word “corruption” close to name of main city, normalized by hits for name of main city Columns (3) & (4): use Newslibrary.com 2009 all newspapers search hits for “corruption” close to name of state, normalized by hits for name of state Columns (5) & (6): Glaeser and Saks (2006) -- extract actual conviction data from the Department of Justice’s “Report to Congress on the Activities and Operations of the Public Integrity Section” to form a measure of the ratio of convicted corruption cases by population size, averaged from 1976 to 2002 to remove periodical noises Effect is stronger in states/main city with more mentions for corruption in the news, and in states with more convictions of corruption case.

38 Effect is stronger in states with higher corruption level (2/2)CAR(-1,+5)  (7) (8) (9) (10) (11) (12) Sample Politicians Coming from State Politics More Regulations Fewer Regulations Higher ALD 1970 Lower ALD 1970 Higher Generalized Trust Lower Generalized Trust Win/Lose [0.0148]*** [0.0180] [0.0238]** [0.0173]** [0.0181]*** [0.0117]** Difference [0.0231]*** [0.0292] [0.0214] Observations 364 230 269 325 306 288 Colums (9) & (10): ALD as a proxy for quality of governance, measure least likely to be affected by reverse causation. Columns (7) & (8): The index of regulation by state is measured for 1999 in Clemson University’s Report on Economic Freedom, This report combines information on labor and environmental regulations and regulations in specific industries such as insurance. Effect is stronger in states with more regulations passed, more isolated capital city (measured by Average Log Distance), and higher generalized trust.

39 Change in media scrutiny post-election(1) (2) (3) (4) VAR Change in Normalized Local Newspaper Mentions Sample Winners from State Politics Losers from Winners from Federal Offices Federal Offices Mean 0.0464 0.0217 [0.0141]*** [ ]* [0.0155] [0.0223]*** Observations 30 34 48 41 Winners get mentioned more in local newspaper after their successful elections (compared to before the elections).

40 Winners are more scrutinized, while distance to DC does not matter(1) (2) (3) (4) DEPVAR Change in Normalized Local Newspaper Mentions CAR(-1, +5) Sample State Politics Federal Offices Shorter distance to DC Longer distance Win/Lose 0.0527 0.160 [0.0293]* [0.0544]*** [0.0218]* [0.0180]** Observations 64 89 355 239 Winners get mentioned more in local newspaper after their successful elections (compared to losers). Effect does not seem to vary with distance between firm’s headquarter and DC.

41 Effect diminishes with social distanceCAR(-1,+5)  (1) (2) (3) (4) (5) (6) (7) Sample Within 1Y Within 2Y Within 3Y Within 4Y Alumni Election within 1 Year of Last Reunion Election Not within 1 Year of Last Reunion Win/Lose [0.009]*** [0.008]** [0.007]** [0.006] [0.0228]** [0.0353] Difference [0.0412] Observations 1,819 3,071 4,232 5,370 28,074 295 215 Effect is weaker as strength of politician-director relationship weakens. Effect is much stronger in the year of alumni reunion.

42 How does it work? Requirement: Information is known to a few individuals, who can trade on the related stocks Abnormal Trading Volume (Campbell Wasley 1996): Sampled stocks are traded more around the elections (significant at 1%) (-5,-1): 5.21% more. (-1,+5): 2.22% more. All significant at 1%. Lots of build-up  reason to believe the market is informed and prices pre-election information

43 Summary Key findings: Firms need not benefit from social connections to politicians in higher office, if they are more scrutinized State-level connections matter more to firms, especially in corrupt states, and for well-governed firms Empirical strategy: RDD of close elections using classmates social network Companion paper (Do et al. 2017): close gubernatorial elections. Benefits to connected firms are positive. Implications: Room for rent-seeking from corporate and political perspectives  design of checks and balances at state level

44 Some further thoughts Narrow targeting versus broad-based targeting:Corruption versus lobbying Positive spillover from one firm to similar firms in the industry (i.e. network effects on the same side) Test: control for/estimate the effect on similar firms Result: strong evidence of narrow targeting. Homophily issue: Related to broad-based targeting (but not the same) Difficult to control completely for, unless one finds an exogenous source of variation in the formation of links.

45 Some examples on firms’ political connectionsSecond example on estimating impacts of friendships on political opinion Question: Do people listen to their friends? When people with different beliefs meet and interact in a social network: Do beliefs really converge? How?

46 Targets of the empirical designCausal effect of a friendship link on beliefs between pairs Endogeneity bias due to homophily (towards higher correlation between beliefs and friendship): “Birds of a feather flock together” Average friendship effects may hide important heterogeneity: possibly no effect among common group peers, or divergence “I don't need a friend who changes when I change and who nods when I nod; my shadow does that much better.” – Plutarch Psychology: the need to distinguish oneself among friends Social Networks and Beliefs Quoc-Anh DO

47 Social Networks and Beliefs Quoc-Anh DOEmpirical design Context: Exogenous allocation of study groups at college entry in Sciences Po Design: Survey social networks, opinions and attitudes Use same-group membership as IV for friendship Look at (variation of) friendship’s effects on belief difference Contributions: Friends, not just peers (classmates) Beliefs and (some) actions Separation of friendship influence from homophily bias Social Networks and Beliefs Quoc-Anh DO

48 Setting of tutorial groupsStudents attend few large lectures, then gather in small classes in fixed tutorial groups (triplettes, ~20 students) Tutorial groups are the primary reason to meet other students in 1st year 3 mandatory core classes each semester in 1st year Classroom materials are synchronized No elective classes before 2nd year, except arts and sports Social Networks and Beliefs Quoc-Anh DO

49 Assignment into tutorial groupsTutorial group assignment is essentially exogenous Students mostly choose on course schedule: Students are told to log in separately from home, at a specific time. They see the schedule of one class (out of 3) at a time. All triples of tutorials are balanced with unattractive and attractive timing. Assignments are done within minutes. Enrollment is hectic, and few get their first choices. Very difficult to coordinate. No summary document is available to students. Information is very limited. No information on instructors, on next semester’s schedules. Social Networks and Beliefs Quoc-Anh DO

50 Social Networks and Beliefs Quoc-Anh DOData collection March 2014: Online survey/game for two classes (entry in 2013 and 2009, with focus on 2013) Incentive-compatible questions to elicit social networks (Leider et al 2009) Each student chooses up to 10 friends Friends’ names are from “autosearch” lists Incentives on cross-validation by friends (name, first meeting) High incentives to get high participation rate (Chandrasekhar Lewis 2014) No priming on tutorial groups Questions on political opinions, attitudes (similar to European Value Survey) Retrospective questions on pre-Sciences Po political opinions. Other data obtained from administrative records: tutorial groups, student characteristics, etc. Planned: a second survey in June 2015, further online experiments (trust, cooperation, etc.) in September 2015. Social Networks and Beliefs Quoc-Anh DO

51 Social Networks and Beliefs Quoc-Anh DOSurvey design Social Networks and Beliefs Quoc-Anh DO

52 Social Networks and Beliefs Quoc-Anh DO

53 Social Networks and Beliefs Quoc-Anh DOEmpirical design Dyadic regression: 𝐷𝑌 𝑖𝑗 =𝛼+𝜌 𝐷𝑌 𝑖𝑗 0 +𝛽 𝐿 𝑖𝑗 +𝛾 𝑋 𝑖𝑗 + 𝜖 𝑖𝑗 Unit of observation: any pair of students i and j with available information 𝐿 𝑖𝑗 : undirected friendship link between i and j 𝐷𝑌 𝑖𝑗 : difference in outcomes (opinion, attitude, etc.) 𝐷𝑌 𝑖𝑗 0 : initial difference in opinion 𝛽: coefficient of convergence caused by friendship 𝑋 𝑖𝑗 : controls of initial differences between i and j Dummies for common gender, nationality, admission type, high-school honors, district of high school, profession of parents, residence zip code; differences in tuition fees (parents’ income brackets) Social Networks and Beliefs Quoc-Anh DO

54 Instrumental Variable strategyIf friendship is randomized: correct treatment effect 𝛽 𝑂𝐿𝑆 Homophily: correlation between 𝐿 𝑖𝑗 and omitted 𝑋 𝑖𝑗 Size of omitted variable bias: 𝛾 𝑜𝑚𝑖𝑡𝑡𝑒𝑑 𝐶𝑜𝑟𝑟( 𝐿 𝑖𝑗 , 𝑋 𝑖𝑗 ) Bias 𝛽 𝑂𝐿𝑆 away from zero, all the more for strong omitted predictors of friendships and differences in outcomes Instrumental variable: random common-group membership 𝐶𝐺 𝑖𝑗 First stage: 𝐿 𝑖𝑗 = 𝛼 1 + 𝜌 1 𝐷𝑌 𝑖𝑗 0 +𝜋 𝐶𝐺 𝑖𝑗 + 𝛾 1 𝑋 𝑖𝑗 + 𝜂 𝑖𝑗 Reduced form: 𝐷𝑌 𝑖𝑗 = 𝛼 0 + 𝜌 0 𝐷𝑌 𝑖𝑗 0 +𝜃 𝐶𝐺 𝑖𝑗 + 𝛾 0 𝑋 𝑖𝑗 + 𝜁 𝑖𝑗 Homophily bias of observables: IV estimates should not differ much when observables are removed Social Networks and Beliefs Quoc-Anh DO

55 Interpretation and InferenceEstimate the LATE among compliers: those who become friends because of the assignment Very strong first stage, exogeneity and monotonicity are likely satisfied Exclusion restriction: common membership does not affect opinions through other channels Possible checks: convergence among same-group non-friends Explore heterogeneity of the effect: Estimate the effect as a nonparametric function of different characteristics: 𝛽(𝑝𝑜𝑙𝑖𝑡𝑖𝑐𝑎𝑙 𝑜𝑝𝑖𝑛𝑖𝑜𝑛𝑠, 𝑛𝑒𝑡𝑤𝑜𝑟𝑘 𝑝𝑜𝑠𝑖𝑡𝑖𝑜𝑛𝑠, 𝑒𝑡𝑐.) with weights by a kernel function Double clustering at group level: allow for arbitrary correlations of error terms across 2 observations with a common group. Social Networks and Beliefs Quoc-Anh DO

56 Social Networks and Beliefs Quoc-Anh DOData description 12/10/2017 Quality of network survey: Participation rate of 66% (completed survey) for the 2013 cohort: in line with Leider et al (2009), Goeree et al (2010) High quality of reported friendship mean VARIABLES (sd) Number of reported friends 8.560 (2.056) Number of reciprocal friends 4.064 (1.998) Correct answer : meeting 0.760 (0.274) Correct answer : time spent 0.521 (0.337) Correct answer : activity 0.461 (0.305) Correct answer : strength of the relationship 0.515 (0.310) Observations 498 Social Networks and Beliefs Quoc-Anh DO Social Networks and Beliefs Quoc-Anh DO

57 Network structure and characteristicsWe use the “OR” undirected network from the network survey (Leider et al 2009, 2010, Goeree et al 2010) Network statistics are similar to the literature Network statistics  Value Mean number of degree per individual 8.8625 Variance number of degree per individual Median number of degree per individual 10 Maximum number of degree per individual 21 Minimum number of degree per individual Diameter of the network 9 Average path length 3.7008 Overall clustering coefficient 0,241 Average clustering coefficient 0,271 Social Networks and Beliefs Quoc-Anh DO

58 Social Networks and Beliefs Quoc-Anh DOOutcome variables Social Networks and Beliefs Quoc-Anh DO

59 Political opinion distributionDistribution of opinions before and after the first year Social Networks and Beliefs Quoc-Anh DO

60 Dyadic control variablesCommon (binary) characteristics, differences in continuous variables Social Networks and Beliefs Quoc-Anh DO

61 Social Networks and Beliefs Quoc-Anh DOBalancing tests Most variables are really balanced Unbalanced in program and some special admission types, because of double degree requirements. Social Networks and Beliefs Quoc-Anh DO

62 Social Networks and Beliefs Quoc-Anh DOResults Strong first stage: common group membership predicts friendship Social Networks and Beliefs Quoc-Anh DO

63 Convergence of political opinionConvergence of about 11% of standard deviation Half-life of political differences: roughly 4 years. Most control variables do not matter to IV IV estimate is bigger than OLS Social Networks and Beliefs Quoc-Anh DO

64 Convergence in other outcomesPolitical and associative participations Determinants of success Social Networks and Beliefs Quoc-Anh DO

65 Social Networks and Beliefs Quoc-Anh DOBenchmark results There is convergence on political opinions, determinants of success, and associative and political participation The estimated bias (IV-OLS) is not in the direction predicted by homophily It is probably due to heterogeneous effects: LATE on compliers (those who become friends because of same-group exposure) Effects are stronger among compliers, compared to always-takers and never-takers Policy driven convergence is strong Social Networks and Beliefs Quoc-Anh DO

66 On the exclusion of common shocksThe design takes care of potential homophily biases in network formation: common group membership uncorrelated with common characteristics/differences in characteristics Another identification concern: within-group common shocks that lead to convergence (exclusion hypothesis) First test: Subsample of non-friend dyads, common group membership does not lead to convergence. Second test: Convergence effect along social distances Social Networks and Beliefs Quoc-Anh DO

67 Heterogeneous effectsHeterogeneity of convergence effects with respect to pre-Sciences Po distances (political opinion, common gender, common admission group). The instruments include common group membership and its interactions with those variables. Heterogeneity of convergence effects with respect to predicted propensity to become friends between pairs, excluding the effect of same group membership. Run probit regressions of undirected friendship on control variables. Then predict the propensity to become friends between any pairs of students. Social Networks and Beliefs Quoc-Anh DO

68 Heterogeneous convergenceConvergence is higher among students with more initial differences Social Networks and Beliefs Quoc-Anh DO

69 Propensity score of friendship and convergenceConvergence is strong among unlikely friends Not much difference between OLS and IV. Homophily bias unlikely strong. Effect over propensity score similar to avoidance of double counting. Social Networks and Beliefs Quoc-Anh DO

70 Convergence, divergence, and extremismConvergence of political opinion, by initial political opinion Social Networks and Beliefs Quoc-Anh DO

71 Extremism and divergenceDivergence among extreme leftists in political opinion: Social Networks and Beliefs Quoc-Anh DO

72 Extremism and divergence… and attitude towards immigration Social Networks and Beliefs Quoc-Anh DO

73 Convergence and eigenvector centralityConvergence seems strongest between a star (top 10% centrality) and a non-star (bottom 90% centrality) Social Networks and Beliefs Quoc-Anh DO

74 Path lengths and convergenceUse same-group membership as IV for path length between pairs: convergence from reduction in path lengths Convergence takes place mostly among direct friends Social Networks and Beliefs Quoc-Anh DO

75 Social Networks and Beliefs Quoc-Anh DOSummary Beliefs on politics among friends converge rapidly after one year Overall, homophily matters, but homophily bias seems weak. Convergence is highly heterogeneous: It is strongest among least likely friends, friends with most different predetermined characteristics Homophily reduces beliefs diffusion even more (complementary to Golub Jackson 2012) Divergence among extreme leftists How networks structure matters to convergence: Works mostly through direct friends Convergence towards star Common friends matter Social Networks and Beliefs Quoc-Anh DO

76 Network Effects on Social NetworksThank you! Network Effects on Social Networks

77 Social connections and firm valuesINTRODUCTION Social connections and firm values Important direct links (former/present employment or ownership) to top-level politics may increase firm value In developing countries: Fisman 2001 AER (Suharto), Johnson Mitton 2003 JFE (Mahathir), Khwajia Mian 2005 QJE (Pakistan), Faccio 2006 AER (cross countries) In American politics (strong safeguards against bribery for favor): Fisman et al (Dick Cheney brings no value) vs. Goldman et al RFS (Rep. Party links bring huge value), Acemoglu et al. JFE forth. (Tim Geithner), Roberts 1990 AJPS (dead U.S. senator) Our question: Do social connections in the U.S. between politicians and corporate directors affect firm value? Who make it stronger/weaker? A comparative look into U.S. politics Covering the topic of political connections and firms values, there has been strong evidence in developing countries that firms benefit from direct links to top politics, such as in the case of Indonesia, Malaysia, or Pakistan. In the US, there are stronger safeguards against corruption, and similar works on the US have yielded mixed results. For example, Fisman et al. finds that connection to Dick Cheney brings no value, while Goldman et al. finds that Republican links brings huge value to firms after the 2000 election. Moreover, all of these studies look at direct, explicit links, defined by former/present politicians’ employment or ownership in the firm. Our paper, on the other hand, explores social links (friendships) between politicians and businessmen, and studies the value of these links in the US context. Fisman 2001 AER: Suharto health problems  23% of firm value explained by connection to the Suharto family (provided by a consultancy) Johnson Mitton 2003 JFE: stronger control of capital flow during 1997 crisis benefits firms connected to Malaysian president & key government officials Khwajia Mian 2005 QJE: firms directed by politicians participating in elections (matched based on names) get favorable loans Faccio 2006 AER: worldwide dataset of directly connected firms & effect of announcements of politiciansbusiness and businessmenpolitics Fisman et al 2006: Dick Cheney connections bring no value Goldman Rocholl So 2008 RFS: election 2000 & R. vs. D. connections (former politicians/officials in firms). Another paper: effect on procurement. Acemoglu et al. JFE forth.: firms connected to Tim Geithner (Obama’s 1st Treasury Secretary) experience the value of connection in a crisis. Identification by synthetic control & CARs. Roberts 1990 AJPS: a dead senator (Henry “Scoop” Jackson tells no lie). Returns of firms connected (in different ways) to dead senator. Seniority in Congress pays. Cohen et al 2007 JPE: information transmission among friends from firms & fund managers Knight 2007 JPubE: industries that benefit from economic platform of Bush & Gore. Snowberg et al 2007 QJE: prediction market & bond yields. Bertrand et al 2008: France, test Shleifer Vishny’s idea. Politicians induce firms to hire more in election year elections affect equity prices Nguyen 2008: France, friendship between CEO & board members  CEO remains longer at firm, less likely to be fired

78 Further robustness checksCollapse the data by different levels Using different weighting schemes CAR calculated from raw returns, Fama-French 3-factor and 4-factor models Standard error clustering at different levels Near-randomness of assignment: no pre-election variable predicts election results

79 RDD Implementation (Semi-)Parametric Regression: CARift on Win/LoseitUnit of observation: Political Connection – Firm – time. Clustered standard errors are obtained from regression Non-parametric procedure: Run non-parametric (local linear/polynomial) regressions of CARift on VoteShareit on each side of the cutoff Take the difference at the cutoff & calculate standard errors. Robustness check with different polynomial orders and different bandwidths.

80 Generalizability of EstimatorDATA & METHODOLOGY: METHODOLOGY – TO SKIP Generalizability of Estimator Heterogeneous effect: The estimator is: This estimates the Weighted Average Treatment Effect on a firm’s cumulative abnormal returns It is not just the effect at the cutoff.

81 Methodology Details Alternative outcome measures:Different windows of event: -7 to -1 to +5 to +20 CAR, CAR normalized by volatility, returns of Buy-and-Hold portfolios Alternative specifications: Narrow down the vote gap: 5% to 2% Non-parametric tests: Estimate a non-parametric function of vote share separately on each side of the 50% threshold, then take the difference Use “hypothetical thresholds” at 48% - 52%: no effect is expected Use quantile regressions instead of OLS to avoid impact of outliers Control of polynomial in Vote Share: Use linear controls in tables. Additional robustness checks: Check different weighting schemes Standard errors clustered in several ways (by politician-election in tables) Check near-randomness of assignment

82 Homophily Concerns Even when election results are almost random, homophily concerns are still relevant Homophily: like-minded directors and politicians chose the same school  issue of of untargetted policy Untargetted policy: A politician promotes policies that benefit a like-minded director’s firm, without favoring that specific network link Address by within-school-election-year and within- industry-election-year comparisons Selection into ‘connected pairs’ because of similar characteristics: a question of external validity Effect conditioned on specific type of social network

83 Internal Validity for Homophily ConcernNETWORK ANALYSIS Internal Validity for Homophily Concern Homophily-induced policies will affect all firms in a single industry, and come from all alumni If some politician-alumni from same school (not necessarily from same year) wins (𝑊 𝐿 𝑠𝑐ℎ𝑜𝑜𝑙,𝑡 ) If someone connected to another firm in the same industry wins (𝑊 𝐿 𝑖𝑛𝑑𝑢𝑠𝑡𝑟𝑦,𝑡 ) 𝑌 𝑖𝑝𝑡 =𝛽𝑊 𝐿 𝑝𝑡 + 𝛿 𝑠𝑐ℎ𝑜𝑜𝑙 𝑊 𝐿 𝑠𝑐ℎ𝑜𝑜𝑙,𝑡 + 𝜅 𝑖𝑛𝑑𝑢𝑠𝑡𝑟𝑦 𝑊 𝐿 𝑖𝑛𝑑𝑢𝑠𝑡𝑟𝑦,𝑡 + 𝑈 𝑖𝑝𝑡 Stronger specification: School FE × Decade ×𝑊 𝐿 𝑠𝑐ℎ𝑜𝑜𝑙,𝑡 Industry FE × Size Quintile ×𝑊 𝐿 𝑖𝑛𝑑𝑢𝑠𝑡𝑟𝑦,𝑡 Election Year FE (which captures 𝑊 𝐿 𝑠𝑐ℎ𝑜𝑜𝑙,𝑡 , 𝑊 𝐿 𝑖𝑛𝑑𝑢𝑠𝑡𝑟𝑦,𝑡 ) Untargetted policy response from homophily: People sharing same values/interests attend same college, and later run companies and promote policies in the same direction Stronger specification: School * Election Year FE or Industry * Election Year FE Let homophily effect vary over time (i.e. time-variant school culture) or firm size

84 Treating Homophily NETWORK ANALYSISCAR (-1,+5) (1) (2) (3) (4) (5) (6) (7) (8) FE included School * Decade * Elec. Year School * Decade * Win(Sch.) Industry * Elec. Year Industry * Win(Ind.) Industry * Size Q. * Elec. Year Industry * Size Q. * WL(Ind.) (2) and (4) (1) and (3) Win/Lose [0.0157]* [0.0120]** [0.0094]** [0.0010]*** [0.0110]** [0.0113]*** [0.0129]*** [ ]*** Vote Shares (Win & Lose) Yes Observations 1,819 1,804 1804 Homophily does not matter to the qualitative conclusion In the most stringent tests: homophily may have biased results towards zero.

85 Relation to other existing literaturesTransmission of Beliefs and Trust: Persistence: Putnam (1993), Guiso et al. (1997) Transmission within families (Bisin-Verdier 2001, Tabellini 2008, Guiso Sapienza Zingales 2008, Benabou Tirole, 2003) and social interactions (Benabou Tirole 2010) Role of education in shaping Beliefs: Capitalist indoctrination: Bourdieu Passeron (1970), Bowles Gintis (1976) Education and Beliefs: Dewey (1944), Helliwell Putnam (2007), Milligan et al. (2004), Heckman (2008) Social networks and Beliefs: Social preferences on existing social networks: Leider et al (2009, 2010), Goeree et al (2010) Belief formation and social networks: Mobius Rosenblatt (2014), DeGroot (1974) Diffusion in networks: Banerjee et al (2013, 2014), Jackson Yariv (2007) Peer effects, and (quasi-)experimental evidence: Sacerdote (2014), Sacerdote (2001), Marmaros Sacerdote (2006), Carrell et al (2013), Shue (2013), Ahern et al (2014), Manski (1993) Identification of social networks effects: Bramoullé et al (2009), De Giorgi et al (2010), Jackson (2008), Durlauf Ioannides (2010), Blume et al (2010) Structural estimation to deal with homophily: Goldsmith-Pinkham Imbens (2013), Christakis et al (2010) Social Networks and Beliefs Quoc-Anh DO