Voting Technology, Political Responsiveness, and Infant Health: Evidence from Brazil Thomas Fujiwara Princeton University Place Date
Motivation Why are public services in developing countries so inadequate? Common answer: (disadvantaged) citizens have little influence over politicians. Interventions are proposed to remedy this: Quotas - Pande (2003), Chathopadhyay and Duflo (2004). Participatory budgets - Besley et al. (2005). Plebiscites - Olken (2010).
Motivation Drastic changes in the rules of the game may not be practical in some situations. Less is known about greasing the wheels of democracy. Does removing mundane obstacles to political participation lead to better public services?
Electronic voting in Brazilian elections Paper ballots required reading and writing several error-ridden spoiled (residual) votes, mainly by the less educated. Electronic voting: reduced voter error, increasing the number of valid votes, generated the de facto enfranchisement of the less educated, affected policymaking in local (state) governments.
Summary of results and structure of talk
Literature: democracy, enfranchisement, and redistribution Empirical literature is both voluminous and full of contradictory results (Acemoglu et al., 2013). Country-level: Some find effects of enfranchisement on redistribution (Aidt and Jensen, 2013)......some argue democracy does not matter (Mulligan et al, 2004). Democracy and health: Besley and Kudamatsu (2006) and Kudamatsu (2012). Within-USA: Minorities: Cascio and Washington (2014), Husted and Kenny (1997), Naidu (2012). Women: Kenny and Lott (1999), Miller (2008).
Background on Brazilian Elections State (and federal) elections every 4 years. Same rules and dates for all states: Open-list PR for legislatures: vote for a candidate (not a list). States are the districts (at-large elections). voting becomes complex. In 1998, a voter in São Paulo had to choose one out of: 1265 candidates for state legislature, 661 candidates for federal congress (lower chamber), 10 candidates for governor, 13 candidates for federal senate, 12 candidates for president.
Until 1994, only paper ballots were used
The introduction of electronic voting Mid-1990s: independent electoral authority introduces electronic voting. Main motivation: reducing the time and costs of vote counting. Facilitating voting was a surprising side effect.
The electronic voting device
Voting interface: state legislature election
A vote for candidate Monteiro Lobato n. 92111 x
A residual (spoiled) vote - no candidate has n. 88888
The phase-in of electronic voting Until the 1994 election: only paper ballots were used. 1998 election: municipalities with more than 40,500 registered voters used electronic technology (rest used paper). Why? Limited supply of devices (production capacity of manufacturer). 2002 election: only electronic voting was used.
The 1998 regression discontinuity design Municipalities just above and just below the cutoff are, on average, similar. No manipulation: Forcing variable: number of voters in 1996. Discontinuity announced in 1998. Focus on state legislature elections. Outcome of interest: valid votes. Data sources: Federal Electoral Authority for electoral results, 1991 Census for demographics.
(Zero) Effects on turnout & registration.6.7.8.9 1 0 20000 40000 60000 80000 100000 Number of Registered Voters - 1996 Registered Voters/Total Population Turnout/Registered Voters
Effects on valid votes.6.7.8.9 1 0 20000 40000 60000 80000 100000 Number of Registered Voters - 1996 Valid Votes/Turnout - 1994 Election (Paper Only) Valid Votes/Turnout - 1998 Election (Discontinuity) Valid Votes/Turnout - 2002 Election (Electronic Only)
Estimation framework v m number of voters in municipality m. y m outcome of interest. Treatment effect of change from paper to electronic voting: TE = lim E[y m v m ] lim E[y m v m ] v m 40,500 v m 40,500 Non-parametric estimation: local linear regression. With a narrow bandwidth, TE = β: y m = α + β1{v m > 40, 500} + γv m + δv m1{v m > 40, 500} + ɛ m Imbens-Kalyanaraman (2012) optimal bandwidth.
Results not shown here TE is zero for: Valid votes in 1994 and 2002. Covariates: income, inequality, education, latitude, longitude. Tests of strategic manipulation (continuity in the distribution of municipalities)..
Effects on policy outcomes Discontinuous assignment at the municipal level. But elections are for officials at state governments. policies decided by state governments switch to state level data.
What policies would be affected by enfranchising the less educated? Brazil has a 2-tier health care system. Free universal public health care. Parallel private system. The uneducated have stronger preferences for health care spending (compared to the more educated) political economy models predict that enfranchising the less educated raises health care spending.
The less educated rely more on public health care Q: How do you Cover Your Health Expenses? Answer by Highest Grade Completed Answer: With Private Insurance 0.2.4.6.8 <4th Grade 4th Grade 8th Grade High School College + Source: 1998 Latinobarometro
Why health care (and not other policies)? Other plausible policies are not decided by state governments. Education is supplied by municipalities. Income taxation and redistribution programs are federal. Health care provision is politically salient: Survey: improving health care services is the most common government priority. 51% for low income (0-2 min. wages). 40% for high income (10+ min. wages). Has quick implementation and effects: Suppressed demand: 57% of women claim wait times and difficulties of being served is the main problem with public health care. Ferraz and Finan (2011), Rocha and Soares (2010), Brollo and Troiano (2014), Correa and Madeira (2014).
How do state legislators increase health care spending? Main activity of legislators: amending budgets to deliver public services to their constituents (Ames, 2001). 34% of São Paulo state legislators cite health as area of expertise. e.g., the Family Health Program, cited in 107 budget amendments cite the program in one year (from 94 legislators). Cannot disentangle economic mechanism : change in legislator identity vs. repositioning of legislators ( citizen-candidate ) vs. ( median-voter ) Change in identity of voters and legislators. Legislator-level behavior is not observed.
The phase-in of electronic voting 1994: only paper ballots ( paper-only election ). 1998: municipalities above the cutoff used electronic technology ( discontinuity election ). 2002: only electronic voting ( electronic-only election ).
Identification strategy Explore variation across states in: S i share of voters in municipalities above cutoff (in 1998) Different S i imply different patterns of electronic voting adoption through time.
Geographic distribution of S i (share above cutoff) (.7631133,1] (.5240926,.7631133] (.4614621,.5240926] (.4481768,.4614621] (.3886635,.4481768] (.3648933,.3886635] (.3345228,.3648933] (.2953183,.3345228] [.1472582,.2953183]
Intuition: 2 states, with S i = H and S i = L, and H>L
Identification strategy: intuition
Identification strategy: intuition
Identification strategy: intuition
Placebo tests
Graphical Representation Panel A: Paper (1994) to Discont. (1998) Panel B: Discont. (1998) to Electr. (2002) Panel C: Pooled 1994-2002 Change in Valid Votes/Turnout, 94-98 (Residuals) -.05 0.05.1 -.4 -.2 0.2.4.6 % of Voters Above Cutoff (Residuals) Change in Valid Votes/Turnout, 98-02 (Residuals) -.06 -.04 -.02 0.02.04 -.4 -.2 0.2.4.6 % of Voters Above Cutoff (Residuals) Change in Valid Votes/Turnout (Residuals) -.05 0.05.1 -.5 0.5 Change in Use of Electronic Voting (Residuals) Change in Spending Share in Health, 94-98 (Residuals) -.04 -.02 0.02.04 Panel A: Paper (1994) to Discont. (1998) -.4 -.2 0.2.4.6 % of Voters Above Cutoff (Residuals) Change in Spending Share in Health, 98-02 (Residuals) -.05 0.05 Panel B: Discont. (1998) to Electr. (2002) -.4 -.2 0.2.4.6 % of Voters Above Cutoff (Residuals) Change in Spending Share in Health, 94-98 (Residuals) -.05 0.05 Panel C: Pooled 1994-2002 -.5 0.5 Change in Use of Electronic Voting (Residuals)
Effects by Category of Spending Parameter: θ 98 θ 02 (θ 98 θ 02 )/2 Sample (terms): 1994-1998 1998-2002 1994-2002 (Paper-Disc.) (Disc.-Electr.) Sample Avg. (1) (2) (3) log(total Spending) - -0.004-0.257 0.127 (0.093) (0.156) (0.097) {0.946} {0.274} {0.254} Share of Spending 0.099 0.039-0.029 0.034 in Health Care (0.017) (0.013) (0.008) {0.104} {0.044} {0.000} Administration 0.181-0.072 0.121-0.097 and Planning (0.044) (0.083) (0.043) {0.126} {0.192} {0.084} Social Assistance 0.108-0.016-0.053 0.018 (0.017) (0.024) (0.015) {0.316} {0.074} {0.274} Education 0.176-0.005 0.016-0.010 (0.014) (0.015) (0.011) {0.708} {0.324} {0.626}
Effects by Category of Spending - continued Parameter: θ 98 θ 02 (θ 98 θ 02 )/2 Sample (terms): 1994-1998 1998-2002 1994-2002 (Paper-Disc.) (Disc.-Electr.) Sample Avg. (1) (2) (3) Judiciary 0.063 0.008-0.024 0.016 (0.020) (0.011) (0.012) {0.726} {0.048} {0.326} Legislative 0.036-0.002 0.015-0.009 (0.008) (0.011) (0.006) {0.878} {0.318} {0.266} Public Safety 0.074 0.002-0.017 0.009 (0.015) (0.024) (0.010) {0.922} {0.564} {0.348} Transportation 0.053 0.010-0.004 0.007 (0.017) (0.036) (0.017) {0.606} {0.910} {0.658} Other Categories 0.210 0.036-0.025 0.030 (0.031) (0.036) (0.031) {0.316} {0.536} {0.426}
Assessing the empirical strategy Five sets of additional evidence: 1 Placebo tests: no effects on variables that should not be affected. 2 Pre- and post- trends. 3 Sharp timing: year-by-year event study. 4 Controlling for interactions of time dummies and baseline variables. GDP, Gini, Illiteracy, Poverty, Population, Area, #Municipalities. 5 Instrument using distribution closer to the cutoff.
Placebo effects Parameter: θ 98 θ 02 (θ 98 θ 02 )/2 Sample (terms): 1994-1998 1998-2002 1994-2002 (Paper-Disc.) (Disc.-Electr.) Sample Avg. (1) (2) (3) Share of Spending 0.096-0.011-0.016 0.003 in Health Care - (0.021) (0.014) (0.015) Municipalities {0.714} {0.296} {0.932} log(population) - 0.064 0.059 0.002 (0.032) (0.026) (0.015) {0.158} {0.088} {0.886} N (state-terms) 54 54 81 Standard errors clustered at the state level in parenthesis. p-values based on Cameron et al. (2008) cluster-robust wild-bootstrap in curly brackets. State, time, and region-time effects in included.
Placebo effects: periods w/o change in voting technology Pre- and Post Trend Analysis Parameter: θ 94 θ 06 Sample (terms): 1990-1994 2002-2006 (Paper-Paper) (Electr.-Electr.) (1) (2) Share of Spending -0.011-0.016 in Health Care (0.021) (0.014) {0.714} {0.296} N (state-terms) 54 54 Standard errors clustered at the state level in parenthesis. p-values based on Cameron et al. (2008) cluster-robust wild-bootstrap in curly brackets. State, time, and region-time effects in included.
Effects on health services utilization and outcomes Data source: vital statistics covering the universe of Brazilian birth records. Measure of utilization: share of mothers with 7+ pre-natal visits. Measure of infant health: share of low-weight births (<2.5kg). Predicts adult health and economic outcomes. Can be affected by pre-natal visits (mainly through changes in maternal behavior). Separately observe educated and uneducated mothers (with and without primary schooling).
Effects on infant health: uneducated mothers Effects on Health Outcomes: Mothers without Primary Schooling Parameter: θ 98 θ 02 (θ 98 θ 02 )/2 Sample (terms): 1994-1998 1998-2002 1994-2002 (Paper-Disc.) (Disc.-Electr.) Sample Avg. (1) (2) (3) Share with 0.362 0.122-0.023 0.069 7+ Visits (0.065) (0.033) (0.040) {0.154} {0.558} {0.182} Share with LW 7.721-0.370 0.528-0.529 Births (x100) (0.304) (0.269) (0.246) {0.266} {0.104} {0.044} N (state-terms) 54 54 81 Standard errors clustered at the state level in parenthesis. p-values based on Cameron et al. (2008) cluster-robust wild-bootstrap in curly brackets. State, time, and region-time effects in included.
Placebo test: educated mothers Effects on Health Outcomes: Mothers with Primary Schooling Parameter: θ 98 θ 02 (θ 98 θ 02 )/2 Sample (terms): 1994-1998 1998-2002 1994-2002 (Paper-Disc.) (Disc.-Electr.) Sample Avg. (1) (2) (3) Share with 0.569 0.062-0.009 0.031 7+ Visits (0.036) (0.022) (0.019) {0.152} {0.742} {0.134} Share with LW 6.261 0.391-0.023 0.196 Births (x100) (0.474) (0.550) (0.502) {0.398} {0.900} {0.626} N (state-terms) 54 54 81 Standard errors clustered at the state level in parenthesis. p-values based on Cameron et al. (2008) cluster-robust wild-bootstrap in curly brackets. State, time, and region-time effects in included.
Conclusion
Graphical representation Panel A: Paper (1994) to Discont. (1998) Panel B: Discont. (1998) to Electr. (2002) Panel C: Pooled 1994-2002 Share with 7+ Prenatal Visits, 94-98 (Residuals) -.1 0.1.2 -.4 -.2 0.2.4.6 % of Voters Above Cutoff (Residuals) Share with 7+ Prenatal Visits, 98-02 (Residuals) -.1 -.05 0.05 -.4 -.2 0.2.4.6 % of Voters Above Cutoff (Residuals) Share with 7+ Prenatal Visits (Residuals) -.1 0.1.2 -.5 0.5 % of Voters Above Cutoff (Residuals) Share of Low-Weight Births, 94-98 (Residuals) -.01 -.005 0.005.01 Panel A: Paper (1994) to Discont. (1998) Share of Low-Weight Births, 98-02 (Residuals) -.01 -.005 0.005.01 Panel B: Discont. (1998) to Electr. (2002) Share of Low-Weight Births (Residuals) -.01 -.005 0.005.01 Panel C: Pooled 1994-2002 -.4 -.2 0.2.4.6 % of Voters Above Cutoff (Residuals) -.4 -.2 0.2.4.6 % of Voters Above Cutoff (Residuals) -.5 0.5 Change in Use of Electronic Voting (Residuals)
Example of Print Advertising.