Network Structure and Public Good Provision

Transcription

1 Network Structure and Publc Good Provson Evdence from Vllage Knshp Networks n the Gamba Gft Tontarawongsa Duke Unversty June 16 rh, 2014

2 Overvew Free-rdng problem n publc goods. Informal cooperaton s prevalent especally n developng countres. Does socal network actually play a role n sustanng ths cooperaton? Theory lterature suggests: Socal captal helps mtgate ths market falure. Game theoretcal models of cooperatve games n networks. Lack of causal evdence due to lmted data + endogenety problem Ths paper provdes feld-based evdence of how networks affect decsons to contrbute to publc goods.

3 Research Questons How does socal network poston affect publc good contrbuton? Do better-connected agents n a network contrbute more? At the larger level, does a better-connected network facltate a hgher level of contrbuton? Estmate the effect of household network poston on publc good contrbuton usng knshp network data from rural vllages n the Gamba Address the endogenous network formaton problem usng the arguably exogenous varaton n vllage ethnc composton as an IV.

4 Vllage Network Example

5 Theoretcal motvaton Game theory lterature on cooperatve games n networks No consensus on the predctons of the relatonshp Why mght household network poston affect publc good contrbuton? Facng hgher socal pressure, threat of sancton Beng montored more, easer to get caught shrkng Poston serves as a proxy for household valuaton of publc goods. Prvate nformaton about valuaton s better known n the communty Do not dstngush between these mechansms n ths paper

6 Network centralty measures To capture the poston n networks and relate t to publc good contrbuton, Degree Centralty Number of lnks Egenvector centralty Weghs n centralty of the connected nodes. Idea: beng connected to a well-connected person makes me more central. Correspond to an egenvector of a matrx representaton of a network.

7 Data Overvew Analyss sample: 2,110 households from 48 rural vllages n the Gamba A baselne household survey for the mpact evaluaton of the Gamba Communty-Drven Development Project (CDDP) n Feb-May of Shared wth me by Dany Jamovch of Goethe Unversty Frankfurt

8 Publc Goods n Rural Gamba Dd anyone n your household partcpate n the followng actvtes for the vllage n the past one year? bnary outcomes Publc good contrbuton Fracton of households contrbutng Std. Dev. Road constructon 58% (0.49) Communal buldng constructon 71% (0.46) Water well constructon 48% (0.50) Publc health 67% (0.47) Publc educaton 48% (0.50) Total contrbuton (sum all above) 3.1 (1.29) Form of contrbuton n-knd contrbuton 99% and monetary contrbuton 1%

9 Publc Goods n Rural Gamba Pattern of publc goods Hgh correlaton wthn the vllage Cases where contrbuton to a partcular publc good s 0 wthn the vllage. Government and NGO nvolvement A school buldng project Focus group ntervew report Unversal access at the vllage level

10 Vllage Knshp Networks Household-level relatonshps Kn network blood relatves of the household head and wves. Varable Mean Std. Dev. Mn Max Degree centralty, Kn 3.99 (3.31) 0 15 Egenvector centralty, Kn 0.76 (0.70) Dst.

11 Endogenety of network centralty Observed relatonshp could be drven by underlyng endogenous network formaton (Mansk, 1993) Households may sort based on unobserved characterstcs. Other potental sources of bas Omtted varables - preferences for publc goods. Reverse causalty

12 Identfcaton Strategy Havng more households of the same ethnc group makes t easer to form knshp networks because of the patrlneal culture and arranged marrages. IV for household centralty: household ethnc fracton The fracton of households wthn the vllage that belong to the same ethnc group as a gven household. A C B D E F G ID Ethnc fracton A 3/7 B 3/7 C 3/7 D 1/7 E 2/7 F 2/7 G 1/7

13 Identfcaton Strategy The Gamba s hghly dverse. Explot the arguably exogenous varaton n vllage ethnc composton determned by hstorcal accdent. Ethnc dentty s not salent (Mwakkagle (2010)). kn lnks (17% across ethnc groups) Not a determnant of economc exchange relatonshps (Arcand and Jamovch, 2013) Ethncty % of sample % 2003 census Mandnka Fula Wolof Jola 7 11 Serahule 2 8 Serer 5 3 Manjago 1 2 Others 5 4

14 Identfcaton Strategy Wthn the same ethncty, comparng households n vllages wth varaton n ethnc fracton. Concerns: 1. Varaton s not random and dfferent types of households sort nto vllages based on ths varaton. 2. Ethnc fracton affects publc good contrbuton drectly.

15 Valdty of the IV - I Concern: dfferent types of households sort nto vllages wth large or small ethnc representaton. Balance check. Obs = 2,110 Varable: Household Ethnc fracton Log ncome Household sze Head age Female head HH head llterate # sck members # workers lands owned (hectares) Self-reported hgh qualty land Role: vllage chef Role: vllage development commttee Role: tradtonal healer Role: Imam 4.060** Role: Marabout Farmer Muslm Martal status1: unmarred Martal status2: monogamous Martal status3: polygamous Relatve wealth: md low Relatve wealth: md hgh Relatve wealth: hgh Agrcultural Income Number of corrugated huts Mssng number of corrugated huts Number of grass-roof huts Mssng number of grass-roof huts

16 Valdty of the IV - II Lmted mgraton n rural Gamba Mgraton Reports (GBoS, 1993, 2009) Internal mgraton n urban Kanfng and Banjul areas (not n the data) Vllage-level ethnc composton data (GBoS census, 1993, 2003) Ethnc fracton reman stable snce out of the 60 vllages had large changes n a few ethnc groups Close to the Banjul areas No data before 1993, anthropologcal and hstorcal evdence Gamble (1949), Sonko-Godwn (1985, 1988)

17 Valdty of the IV - III Ethnc fracton should affect publc good contrbuton only through networks. Drect ethnc conflct Mwakkagle (2010), Sonko-Godwn (1985, 1988) Arcand and Jamovch (2012) ethncty s not related to econexchange. Dverged preferences over the type of publc goods to be funded. Regressng dfferent publc goods on ethncty not sgnfcant

18 Valdty of the IV - IV Ethnc fracton should affect publc good contrbuton only through networks. contrbuton j 0 1 ELF j X j j Outcome varable Roads Commu. Buldngs Water wells Publc health Publc educaton Total Panel A: OLS estmates wthout controllng for network densty Ethnc dversty ndex (0.17) (0.18) (0.15) (0.17) (0.19) (0.53) Panel B: OLS estmates controllng for network densty Ethnc dversty ndex (0.17) (0.20) (0.15) (0.15) (0.19) (0.49) Network densty 0.801* *** 1.818*** ** (0.47) (0.82) (0.43) (0.55) (0.92) (1.98) Mean outcomes Observatons = 60 vllages. Controllng for: number of households, total populaton, areas, log per capta ncome, ncome Gn coeffcent, llteracy rate, % of households wth no electrcty, ethnc dversty (Herfndahl ndex), relgon dversty (Herndahl ndex), dstrct FE.

19 Results: Frst stage centralty 0 1ethnc_fra cton 2 X j u v Endogenous varable: Degree Cent. Egenvector cent. IV: household ethnc fracton 3.956*** 0.874*** (0.506) (0.090) Constant (1.539) (0.503) Frst-stage F statstcs R-squared Observatons 2,110 Robust standard errors clustered at the vllage level n parenthess. Controllng for: ethnc-group FE and vllage FE. The set of household characterstcs nclude log ncome, household sze, age of household head, whether head s female, llteracy, number of sck household members, number of workers, sze of land owned, whether land s of hgh qualty, status as the vllage chef, tradtonal healer, vllage development commttee, beng a relgous leader, relgon, martal status, number corrugated huts, and number of grass-roof huts. Graph

20 2 nd Stage Specfcaton contrbuton centralty X j v contrbuton = 5 publc good types, total number of contrbuton centralty Degree centralty Egenvector centralty X = household characterstcs j = ethnc group fxed effects v = vllage/network fxed effects

21 Results: Degree centralty contrbut on degree X j v (1) (2) (3) (4) (5) (6) Publc good : Com. Water Publc Publc Road Buldngs Well health educ Total cont. OLS estmates Degree centralty *** ** (0.004) (0.003) (0.004) (0.004) (0.004) (0.010) IV estmates Degree centralty * ** (0.012) (0.013) (0.014) (0.009) (0.011) (0.023) Frst stage F stat Mean of outcomes Sample sze: 2,110, #Vllages = 48 Robust standard errors clustered at the vllage level n parenthess. Controllng for: ethnc-group FE and vllage FE. The set of household characterstcs nclude log ncome, household sze, age of household head, whether head s female, llteracy, number of sck household members, number of workers, sze of land owned, whether land s of hgh qualty, status as the vllage chef, tradtonal healer, vllage development commttee, beng a relgous leader, relgon, martal status, number of corrugated huts, number of grass-roof huts. Probt

22 Results: Egenvector centralty contrbut on α βeg_cent γx δ j θ v ε (1) (2) (3) (4) (5) (6) Publc good : Com. Publc Road Water Well Buldngs health Publc educ Total cont. IV estmates Egenvector centralty * ** (0.056) (0.056) (0.069) (0.040) (0.050) (0.097) Mean outcome Sample sze: 2,110, #Vllages = 48 Robust standard errors clustered at the vllage level n parenthess. Controllng for: ethnc-group FE and vllage FE. The set of household characterstcs nclude log ncome, household sze, age of household head, whether head s female, llteracy, number of sck household members, number of workers, sze of land owned, whether land s of hgh qualty, status as the vllage chef, tradtonal healer, vllage development commttee, beng a relgous leader, relgon, martal status, number of corrugated huts, number of grass-roof huts.

23 Interpretng the magntudes (1) (2) (3) (4) (5) (6) Publc good : Com. Publc Road Water Well Buldngs health Publc educ Total cont. IV estmates Egenvector centralty * ** (0.056) (0.056) (0.069) (0.040) (0.050) (0.097) Interpret the magntudes 25 percentle percentle percentle percentle Mean outcome Egenvector centralty can be nterpreted as a weghted sum of walks that starts from a gven node, wth longer walks dscounted more. Egenvector centralty, mean = 0.76

24 Robustness check: addtonal controls Bnary varables Exstence of relatonshps outsde of the vllage. Varable Mean Std. Dev. Land exchange 0.14 (0.35) Labor exchange 0.06 (0.25) Input exchange 0.10 (0.31) Credt exchange 0.15 (0.36) Marrage 0.85 (0.35) Enumerator fxed effects Address potental systematc measure errors

25 Robustness check: addtonal controls contrbut on degree X X 2 j v (1) (2) (3) (4) (5) (6) Publc good contrbuton: Com. Water Publc Publc Road Buldngs Well health educ Total cont. Man specfcaton Degree centralty, Kn * ** (0.012) (0.013) (0.014) (0.009) (0.011) (0.023) Robustness check Degree centralty, Kn * ** (0.014) (0.013) (0.015) (0.010) (0.010) (0.023) Mean of outcomes Sample sze: 2,110, #Vllages = 48 Robust standard errors clustered at the vllage level n parenthess. Controllng for: ethnc-group FE and vllage FE. The set of household characterstcs nclude log ncome, household sze, age of household head, whether head s female, llteracy, number of sck household members, number of workers, sze of land owned, whether land s of hgh qualty, status as the vllage chef, tradtonal healer, vllage development commttee, beng a relgous leader, relgon, martal status, number of corrugated huts, number of grass-roof huts.

26 Robustness check: expanded sample (1) (2) (3) (4) (5) (6) Publc good contrbuton: Com. Water Publc Publc Road Buldngs Well health educ Total cont. Stable vllages (Obs = 2,110, vllage = 48) Degree centralty, Kn * ** (0.012) (0.013) (0.014) (0.009) (0.011) (0.023) Mean of outcomes Full sample (Obs = 2,697, Vllage = 60) Degree centralty, Kn * 0.031*** ** 0.058** (0.010) (0.010) (0.011) (0.008) (0.010) (0.024) Mean of outcomes Robust standard errors clustered at the vllage level n parenthess. Controllng for: ethnc-group FE and vllage FE. The set of household characterstcs nclude log ncome, household sze, age of household head, whether head s female, llteracy, number of sck household members, number of workers, sze of land owned, whether land s of hgh qualty, status as the vllage chef, tradtonal healer, vllage development commttee, beng a relgous leader, relgon, martal status, number of corrugated huts, and number of grass-roof huts.

27 Alternatve specfcatons Possble nonlnear relatonshps Bandera and Rasul (2006) nverse U shape Specfcatons: 1. add a squared term of network centralty. 2. I also break down degree centralty nto a set of dummy varables. Lttle evdence of non-lnear relatonshp. Lnear peer effects specfcaton (Mansk, 1993) Postve relatonshp between contrbutons of connected household. Nonlnear Peer

28 Other types of networks Economc networks - four types of economc transactons n the past one year Some postve correlaton but the IV doesn t pass the weak IV test. OLS estmates: small correlatons wth contrbuton. Only sgnfcant for communal buldng constructon. Combned networks both kn and economc networks IV works well, possbly because of the kn networks. IV results: smlar patterns wth larger magntudes than the kn IV results. Interpret wth cauton because of the weak IV problem for the economc networks. Table

29 Concluson Central agents contrbute to publc goods more. Support the models n the game theory lterature Consstent wth punshment n an experment n Carpenter et al. (2012) Networks may explan why ethnc fragmentaton lowers publc goods Unexplaned cases lke Chna, Hat (Esteban et al., 2012) Vllage-level