Do Politicians Change Public Attitudes? Magnus Carlsson Linnaeus University Gordon Dahl UC San Diego May 9, 2014 Dan-Olof Rooth Linnaeus University Introduction A large literature asks: Can voters aect political parties / politicians? But what about the reverse question: Can political parties / politicians aect voters? In this paper, we study whether political representation aects public attitudes on important policy questions The Power of Political Representation By being elected to oce, politicians have an opportunity to shape public attitudes Platform to express ideas Media attention Implementation of policies Shaping attitudes can help (or hurt) with re-election Politicians care about both re-election probabilities and advancing their preferred policy position Setting Local City Council Elections in Sweden Lots of elections (∼290 per election year) Many seats to ll at the local level (∼45) Many parties vying for seats (∼8) Small, single issue parties which attracted voters on two key issues 1 Green Party: anti-nuclear energy (1988-1998) 2 Swedish Democrats: anti-immigration (2002-2012) Advantages Single issue parties: It is clear what attitudes might be aected Small parties: The jumps in the seat share are larger (good for identication) Single Issue Party #1: Green Party and Nuclear Energy Historical timeline 1970's: Nuclear plants built 1979: Three Mile Island accident 1980: Referendum on nuclear power 1981: Formation of Green Party 1986: Chernobyl accident (radioactive exposure in Sweden) Nuclear Plants close to major cities Gothenburg North of Ringhals Malmö South of Barsebäck Stockholm btwn Forsmark / Oskarshamn Single Issue Party #1: Green Party and Nuclear Energy 1) 1988 Green Party Platform Concrete aims 1. ... Phase out nuclear power within 3 years ... 2) 1994 survey: 3 most important topics for each party Most parties had employment or the economy in their top 3 Green Party had environment rst, and was only party with nuclear energy mentioned Single Issue Party #2: Swedish Democrats and Immigration Our most important issues Responsible immigration policy Identication Idea (details later) Take advantage of large non-linearities in the seat assignment rule Idea: thresholds in the rule generate jumps in whether a seat is awarded in otherwise similar elections Simplied example: One city council and two elections Election 1: Party A gets 2.09% of votes and assigned 0 seats Election 2: Party A gets 2.10% and assigned 1 seat Elections 1 and 2 approximately the same except for # seats Compare post election attitudes in Election 1 versus 2 In reality the seat assignment rule is more involved with thresholds varying based on the vote shares of all other parties Preview of Results: Nuclear Energy Evidence that politicians aect public attitudes ... and are (modestly) rewarded at the ballot box the next election Green Party (anti-nuclear) politicians who get elected Cause more negative attitudes towards nuclear energy 1 more seat increases negative attitudes by ∼2.6% Cause higher votes for Green Party in next local election in 2 out of 3 elections Preview of Results: Immigration Evidence that politicians aect public attitudes but not in the party's intended direction ... and no incumbency eect in the next election Swedish Democrat (anti-immigration) politicians who get elected Cause less negative attitudes towards immigration 1 more seat reduces negative attitudes by ∼4% Do not cause higher votes for SD Party in next election once the party reaches a critical mass Preview of Results: Newspaper Coverage and Content Media coverage in local newspapers Focus on SD party and immigration attitudes, since data not available for earlier years SD mentioned more often in local newspapers 1 more seat increases articles mentioning SD by ∼35% Immigration and integration policy mentioned more often in articles, even without mentioning SD SD associated with words such as racism and xenophobia Four Strands of Related Research 1) Policy aected by election results Downs (1957), Lee, Moretti & Butler (QJE 2004), Petterson-Lidbom (JEEA 2010), Besley & Case (JEL 2003), Levitt (AER 1996), Folke (JEEA 2013, similar Swedish data for an earlier time period) 2) Incumbency (person or party) Lee (JOE 2008) plus many poli sci papers 3) Politics and media DellaVigna & Kaplan (QJE 2007), Gerber, Kaplan & Berger (AEJ-Applied 2009), Gentzkow (QJE 2006), Strömberg (QJE 2004), Strömberg & Shapiro (JPE 2010) 4) Attitudes/behavior aected (by religious leaders) Bassi & Rasul (WP 2012), Stroebel & Benthem (WP 2012) Swedish Political Parties Today, there are eight major parties in Sweden 1 Moderate Party (M) 2 Christian Democrats (KD) 3 Centre Party (C) 4 People's Party (FP) 5 Social Democratic Party (S) 6 Left Party (V) 7 Green Party (MP) 8 Swedish Democrats (SD) (New Democracy (NyD) got seats in parliament in 1991) Parties often form coalitions. Coalitions can vary over time and over dierent policy areas (education, healthcare etc.) Swedish Election System General elections occur every 4 years (previously 3) at the national (N=1), the county (N=21), and the city council (N=290) level We focus on local elections to have enough observations Sweden has a multi-party proportional election system There is no single winner in an election Instead each party is allocated seats in the City Council approximately in proportion to the party´s vote share Swedish city councils have a large autonomy Levy taxes (around 30 percent of earnings) Largest expenditures are childcare, education, elderly care Arrange for local provision of electricity and public safety Decide on refugee placement and immigrant integration plans Seat Assignment Rule Council seats allocated according to modied Sainte-Laguë method AKA the highest quotient , or odd number rule After the votes, vp , for each party have been tallied, successive quotients, qp , are calculated for each party based on the number of seats, sp , each party has received so far: (v qp (vp , sp ) = p 1.4 p sp v 2s +1 p sp =0 ≥1 Example #1 with 3 parties and 5 seats Quotient Party Votes % /1.4 /3 /5 /7 Party A 56,250 ~57% 40,178 (1) 18,750 (3) 11,250 (4) 8,035 Party B 28,950 ~29% 20,679 (2) 9,650 (5) 5,790 4,136 Party C 13,496 ~14% 9,640 4,499 2,699 1,928 Discontinuities in Seats There are several ways large discontinuities can arise, i.e., where vote shares move negligibly but result in discrete jumps in seats Example #2: 28 new people decide to vote for Party C, no change in votes for other parties Quotient Party Votes % /1.4 /3 /5 /7 Party A 56,250 ~57% 40,178 (1) 18,750 (3) 11,250 (4) 8,035 Party B 28,950 Party C 13,496 ~29% 20,679 (2) 9,650 (5) 5,790 4,136 ~14% 9,640 4,499 2,699 1,928 Baseline 28 new people now decide to vote for Party C Party A 56,250 ~57% 40,178 (1) 18,750 (3) 11,250 (4) 8,035 Party B 28,950 ~29% 20,679 (2) 9,650 5,790 4,136 Party C 13,524 ~14% 9,660 (5) 4,508 2,705 1,932 Discontinuities in Seats, cont. Even if the votes for Party C remain unchanged, the number of seats for Party C can change. Example #3: 60 voters switch from voting for Party B to voting for Party A Quotient Party Votes % /1.4 /3 /5 /7 Party A 56,250 ~57% 40,178 (1) 18,750 (3) 11,250 (4) 8,035 Party B 28,950 Party C 13,496 ~29% 20,679 (2) 9,650 (5) 5,790 4,136 ~14% 9,640 4,499 2,699 1,928 Baseline Party C unchanged and 60 voters switch from Party B to A Party A 56,310 ~57% 40,221 (1) 18,770 (3) 11,262 (4) 8,044 Party B 28,890 ~29% 20,635 (2) 9,630 5,778 4,127 Party C 13,496 ~14% 9,640 (5) 4,499 2,699 1,928 Many Local Random Experiments Many parties and elections over time The same rule applies over time and across city councils Cut-os for being awarded a seat vary from election to election by vote shares for the various parties the number of seats A large number of local random experiments a proportional election system generates more close elections compared to a majoritarian system many seats: On average 45 seats in a city council small parties do not need many votes to get a seat (∼2%) Regression Model Causal relationship we want to estimate: yjt = αj + θf (zj ,t −1 ) + ujt (1) yjt measures attitudes, j indexes municipality, t indexes time f f (.) is share of seats the SD party gets modied Sainte-Laguë seat allocation method determines this (.) is a known function of inputs zj ,t −1 : number of available seats and votes (or equivalently, vote shares) for all parties Estimate of θ likely to suer from omitted variable bias Votes in prior election are a function of attitudes and attitudes are correlated over time Identication Identication strategy: Only use non-linear threshold variation Augment outcome equation with a exible (but continuous and smooth) function of zj ,t −1 : yjt = αj + θf (zj ,t −1 ) + g (zj ,t −1 ) + ejt Adding g (.) ensures we identify θ (2) solely from sharp nonlinearities in the voting algorithm, and not from the zj ,t −1 variables themselves f (.), the share of seats awarded to the SD party, does not need to be estimated; it is determined by the voting algorithm g (.) must be estimated and captures how votes / vote shares aect attitudes We will include municipality xed eects Note: f (.) has no time or region subscripts by denition; for our baseline regressions, we impose this on the g (.) function too Control Function Interpretation Estimation strategy can be interpreted as a control function The control function is g (.) Identifying assumption is g (.) = E [ujt |αj , zj ,t −1 ] This implies the error term in the augmented regression is conditionally mean zero: E [e |α , f (z , −1 ), g (z , −1 )] = 0 jt j j t j t Tradeo: need to estimate g (.) exibly, but this comes at the cost of precision We estimate with various order polynomials Identication derives from the large nonlinear jumps in seats Assuming a xed order polynomial, consistent θ̂ is identied and Comparison to Regression Discontinuity yjt = αj + θf (zj ,t −1 ) + g (zj ,t −1 ) + ejt Similar to RD, with a few key dierences The cutos f g (.) is high dimensional g (.) is a global function, assumed to be the same for all (.) vary with each election elections regardless of the cuto f (.) Cannot do local linear regression and no natural way to weight observations relative to distance from changing cutos With many running variables it is dicult to draw a graph Attitudes Attitudes measured at the municipality level 1) Attitudes towards nuclear energy SOM attitude survey 1988-1998 random sample of 16,392 respondents 2) Attitudes on immigration FSI attitude survey 2002-2013 random sample of 21,499 respondents Attitude Variable #1: Nuclear Energy Nuclear energy question, 1988-1998: In 1980 we had a referendum on nuclear power in Sweden. After the referendum, parliament decided to phase out nuclear power by 2010. What is your opinion about nuclear energy use in Sweden? A) Stop nuclear power immediately. B) Stop nuclear power earlier than 2010. C) Phase out nuclear power by 2010. D) Phase out nuclear power, but after 2010. E) Do not phase out nuclear power at all. Keep using it. F) No opinion. Dummy variable for negative attitudes: alternative A or B Negative Attitudes towards Nuclear Energy Distribution of negative attitudes towards nuclear energy across municipalities in the three post-election periods City Council Vote Shares, 1988-1994 Green Party (anti-nuclear) in blue: 5.5% in 1988, 3.4% in 1991, 5.0% in 1994, 4.5% in 1998 Attitude Variable #2: Immigration Immigration question, 2002-2012: Should Sweden continue accepting immigrants to the same extent as now? A) To a greater extent B) To the same extent C) To a lesser extent D) Do not know Dummy variable for negative attitudes: alternative C Negative Attitudes towards Immigration Distribution of negative attitudes towards immigraton across municipalities in the three post-election periods City Council Vote Shares, 2002-2010 SD Party (anti-immigration) in blue: 0.4% in 2002, 2.9% in 2006, 5.7% in 2010 Estimating Equation yijt = αj + λt + β xi + θf (zj ,t −1 ) + g (zj ,t −1 ) + eijt yjt is attitude variable, measured in years after election include municipality f.e.'s include xi to control for individual characteristics f (zj ,t −1 ) is seat share obtained in last election g (zj ,t −1 ) is control function αj and survey year f.e.'s λt has same inputs as seat share algorithm these inputs consist of vote shares for all parties, total number votes in municipality, and total number seats we model g (.) with high order polynomials and interactions Polynomial Control Functions g (.) = g (vs1 , vs2 , vs3 , vs4 , vs5 , vs6 , vs7 , vs8 , vs9, seats , votes ) A) Vote share for main party of interest (1 term): B) 1st order (10 terms) γ1 vs1 γ1 vs1 + γ2 vs2 + γ3 vs3 + γ4 vs4 + γ5 vs5 + γ6 vs6 + γ7 vs7 + γ8 vs8 + γ9 seats + γ10 votes C) 2st order, no interactions (21 terms) γ1 vs1 + γ11 vs12 + γ2 vs2 + γ22 vs22 + γ3 vs3 + γ33 vs32 + ... Similarly for D) 3rd order, no interactions (32 terms) E) 4th order, no interactions (43 terms) F) 4th order, with 2nd order interactions (85 terms) E plus γ12 vs1 vs2 + γ13 vs1 vs3 + γ23 vs2 vs3 + ... G) F + 3rd order interactions for party of interest (121 terms) F plus γ112 vs12 vs2 + γ113 vs12 vs3 + γ123 vs1 vs2 vs3 + ... OLS Estimates of Green Party (MP) Representation on Nuclear Attitudes Dep var: Negative attitude on nuclear energy MP seat share×100 .0013 .0009 -.0018 -.0017 (.0015) (.0015) (.0015) (.0016) Secondary education -.0311** (relative to compulsory) -.0301** (.0074) (.0073) College education -.0197** -.0143** (.0089) (.0099) Female .0991** .1017** (.0053) (.0053) Age -.0077** -.0079** (.0012) (.0013) .0059** .0061** (.0013) (.0013) Age squared×100 Municipality f.e.'s R-squared .0100 .0430 X X .0347 .0682 Notes: Overall mean: .15, obs: 16,392, muncipalities: 290, s.e.'s clustered by muni. CF Estimates of Green Party (MP) Representation on Nuclear Attitudes MP seat share×100 Control fn (# terms) Dep var: Negative attitude on nuclear energy .0066* .0065 .0080** .0089** .0099** .0130*** .0113** (.0039) (.0040) (.0041) (.0042) (.0042) (.0048) (.0052) A) MP vote share only 1 B) 1st order C) 2nd, no interactions D) 3rd, no interations E) 4th, no interactions F) 4th, with 2nd interactions G) F + 3rd order MP terms 10 21 32 43 86 Muni & year f.e.'s X X X X X X R-squared .0685 .0688 .0696 .0699 .0709 .0739 Notes: Overall mean: .15, obs: 16,392, municipalities: 290, s.e.'s clustered by muni. 121 X .0763 OLS Estimates of Swedish Democrat (SD) Representation on Immigration Attitudes Dep var: Negative attitude on immigration SD seat share×100 .0155** .0133** -.0035 -.0046* (.0027) (.0017) (.0027) (.0027) Immigrant Secondary education (relative to compulsory) -.0437** -.0471** (.0092) (.0095) -.0369** -.0339** (.0086) (.0088) College education -.2689** -.2563** (.0085) (.0097) Female -.0348** -.0343** (.0060) (.0061) Age .0037** .0032** (.0013) (.0013) Age squared×100 -.0021 -.0016 (.0014) (.0014) Municipality f.e.'s R-squared .0102 .0760 X X .0394 .0960 Notes: Overall mean: .61, obs: 21,499, municipalities: 290, s.e.'s clustered by muni. CF Estimates of Swedish Democrat (SD) Representation on Immigration Attitudes SD seat share×100 Control fn (# terms) Dep var: Negative attitude on immigration -.0115** -.0132** -.0149** -.0152** -.0172** -.01894** -.0193** (.0058) (.0050) (.0052) (.0053) (.0055) (.0054) (.0056) A) SD vote share 1 B) 1st order C) 2nd, no interactions D) 3rd, no interations E) 4th, no interactions F) 4th, with 2nd interactions G) F + 3rd order SD terms 10 21 32 43 86 Muni & year f.e.'s X X X X X X R-squared .0961 .0967 .0979 .0985 .0991 .1011 Notes: Overall mean: .61, obs: 21,499, municipalities: 290, s.e.'s clustered by muni. 121 X .1027 Robustness Alternative specications Non-linearities in eect of seat share: # of seats eect is almost linear so seat shares is a good approximation Allow g (.) to vary by election year: No change Allow g (.) to vary by region: No change Dropping attitudes collected in election years: No change Dropping municipality xed eects: Estimates somewhat smaller but still signicant SD: Not aected by unstable seats MP: Not simply the fear of living close to a nuclear plant Incumbency Eects and Attitudes for Green Party (MP) MP seat share×100 × 1988 election year × 1991 election year × 1994 election year # muni's Log votes next election Neg. attitudes w/ party Naive OLS Control Fn after election representation .3143** (.0305) .3552** (.0395) .2569** (.0345) .0401* (.0226) .0490** (.0223) -.0130 (.0246) .0085 (.0064) .0160** (.0063) .0160** (.0062) 259 / 284 229 / 286 242 / 288 R-squared .3384 .9079 .0735 Obs. 858 858 16,392 Notes: Standard errors clustered by municipality. Specications include xed eects for municipality and year. Control function includes 4th order polynomials with 2nd order interactions (86 terms). Incumbency Eects and Attitudes for Swedish Democrats (SD) SD seat share×100 × 2002 election year × 2006 election year × 2010 election year # muni's Log votes next election Neg. attitudes w/ party Naive OLS Control Fn after election representation .4424** (.0563) .1887** (.0238) .0764* (.0410) -.0117 (.0297) -.0010 (.0100) -.0191** (.0059) -.0229** (.0058) 29 / 290 144 / 290 246 / 290 R-squared .2231 .8716 .0844 Obs. 580 580 21,499 Notes: Standard errors clustered by municipality. Specications include xed eects for municipality and year. Control function includes 4th order polynomials with 2nd order interactions (86 terms). Newspaper Coverage and Content Does newspaper coverage increases after a party wins a seat? Do newspapers give more attention to winning party? Do newspapers devote more attention to party's key issues? Is coverage positive or negative? Data: All articles for all newspapers in Sweden Information only available for 2006-2013 period Given available data, focus on newspaper coverage of Swedish Democrats and immigration for 2 elections 139 newspapers large public subsidies of local newspapers newspapers often cover more than one municipality We search for key words dealing with SD and immigration Eect of SD Representation on Newspaper Coverage Dep var: arcsinh(articles per 100 days) SD seat share×100 Search term: Swedish Democrat .0525** .1349** .1415** .1630** .1603** .2128** (.0215) (.0454) (.0592) (.0678) (.0688) (.0787) Control function (# terms) A) None (OLS) 0 B) SD vote share 1 C) 1st order 10 D) 2nd order, no interactions 21 E) 3rd order, no interactions F) 4th order, no interactions Ave. # articles per paper per year 110 R-squared .9851 .9860 .9869 .9896 Newspaper f.e.'s X X X X Notes: 278 observations, 139 newspapers, s.e.'s clustered by newspaper. 32 .9920 X 43 .9932 X Eect of SD Representation on Newspaper Content Dep var: arcsinh(articles per 100 days) Search terms: Swedish Democrats AND Racism OR Xenophobia OLS SD seat share×100 Control Fn (Racism OR Xenophobia) OLS Control Fn .0072 .1903* .0396 .1873** (.0267) (.1053) (.0243) (.0682) .9755 .9874 .9706 .9878 X X X X Ave # articles per paper per year R-squared Newspaper f.e.'s 90 24 Notes: 278 observations, 139 newspapers, s.e.'s clustered by newspaper. Control function includes 4th order polynomials with no interaction terms (43 terms). Eect of SD Representation on Newspaper Content, continued Dep var: arcsinh(articles per 100 days) Search terms: Swedish Democrats AND Immigrants AND Integration SD seat share×100 Immigrants AND Integration OLS Control Fn OLS Control Fn -.0117 .1603** -.0064** .0687* (.0230) (.0636) (.0156) (.0404) .9720 .9893 .9505 .9708 X X X X Ave # articles per paper per year R-squared Newspaper f.e.'s 21 5 Notes: 278 observations, 139 newspapers, s.e.'s clustered by newspaper. Control function includes 4th order polynomials with no interaction terms (43 terms). Heterogeneous Eects of Political Representation on Attitudes Dependent Variables: Negative Attitudes (toward Nuclear Energy or Immigration) High Low Age Age All Educ Educ Females Males ≤45 >45 A: MP Party (Reminder: Negative nuclear energy attitudes higher for less educated, female, and young) MP seat share .013** .019 .009* ×100 (.005) (.012) (.005) Obs. 16,392 3,402 12,602 .016* (.008) 8,216 .007 (.006) 8,171 .010 (.007) 8,583 .009 (.007) 7,374 B: SD Party (Reminder: Negative immigration attitudes higher for less educated, male, and older) SD seat share -.019** -.006 -.023** -.007 -.024** -.010 -.020** ×100 (.005) (.009) (.007) (.007) (.006) (.008) (.006) Obs. 21,499 7,166 14,160 11,998 11,250 10,032 11,433 Notes: Standard errors clustered by municipality. Specications include xed eects for municipality and year. Control function includes 4th order polynomials with 2nd order interactions (86 terms). Exogeneity Tests Dependent Variables: Municipality Characteristics % % net Tax % % % UE Voting migration rate foreign college > 45 rate MP seat share -.0008 ×100 (.0414) Obs. 848 -.0207 (.0452) 844 A: MP Party (1988, 1991, 1994) .0175 -.0004 -.0267 .0254 -.0218 (.0315) (.0303) (.0366) (.0538) (.0840) 844 844 844 844 843 B: SD Party (2002, 2006, 2010) Lagged attitudes -.0081 (.0057) 16,392 SD seat share -.0279 .0206 .0080 .0073 -.0268 -.0425 .0816 .0076 ×100 (.0447) (.0359) (.0323) (.0354) (.0325) (.0468) (.0671) (.0047) Obs. 864 863 863 863 863 863 863 24,030 Notes: Standard errors clustered by municipality. Specications include xed eects for municipality and year. Control function includes 4th order polynomials with 2nd order interactions (86 terms). Conclusion Can politicians change public attitudes? YES Causal evidence that political representation aects public attitudes for two important policy issues Green Party causes more negative attitudes towards nuclear energy, in line with the party's policy position Rewarded at the ballot box in 2 of 3 ensuing elections Swedish Democrats reduce negative attitudes towards immigration, opposite the party's policy position Experience no incubency benet after reaching critical mass Evidence for media attention as a mechanism Swedish Democrats seem to have increased the debate about immigration policy in local newspapers, but not always in a light favorable for their party