Information, accountability, and public goods:Evidence from three linked field experiments in Liberia

PRELIMINARY AND INCOMPLETE. PLEASE DO NOT CIRCULATE.

Mauricio Romero∗ Justin Sandefur† Wayne Aaron Sandholtz∗

December 19, 2017‡

Abstract

Prevailing theories linking democracy with growth require that voters hold politicians accountable for theprovision of public goods. While much empirical work has examined clientelism and corruption, verylittle is known about the more basic relationship between democratic accountability and public goodprovision. Furthermore, previous work has focused on either voters or politicians in a given context,without measuring how both groups respond to one another dynamically. Our paper’s contribution isto follow the impact of improved information on public good provision through the democratic process,using three linked field experiments carried out in an election year in Liberia. First, we leveraged the ran-domized rollout of a controversial school reform to measure the policy’s causal effect on voters’ politicalattitudes and voting patterns. We find that the reform increased household satisfaction with governmenteducation and led to higher voteshare for the presidential candidate of the party that created the pol-icy (in competitive districts). Next, we randomly varied the provision of credible information about thepolicy’s impact and popularity to 681 politicians. The information made them more likely to tell theirvoters they support the policy’s expansion, but only in competitive districts. Finally, we randomly variedthe provision of information about candidates’ positions on the school policy to subset of households.The information made households more likely to support candidates who favor the policy expansion,especially in competitive districts. Overall, we show that improved information can improve democraticaccountability for public good provision through the choices of both politicians and voters.

Keywords: Evidence-Based Policy; Public-Private Partnership; Public Service Delivery; Political Economy;Randomized Controlled Trial;JEL Codes: O10, C93, D72, P16, H41

∗University of California, San Diego†Center for Global Development‡Corresponding author: Wayne Sandholtz. E-mail: wasandholtz@ucsd.edu. We are grateful to the Minister of Education, George

K. Werner, Deputy Minister Romelle Horton, Binta Massaquoi, Nisha Makan, Kammi Sheeler and the PSL team. Thanks to Al-pha Simpson and his team at Q&A, and Kou Gbaintor-Johnson and her team at Center for Action Research and Training, whocollected the data. Thanks to Arja Dayal, Dackermue Dolo, and their team at Innovations for Poverty Action who led the datacollection for the original PSL evaluation, and Avi Ahuja, Dev Patel, and Benjamin Tan, who provided excellent research assistance.For comments which improved the design and analysis, we thank Eli Berman, Prashant Bharadwaj, Jeremy Bowles, Daniel But-ler, Michael Callen, Brian Dillon, Mitch Downey, Horacio Larreguy, Craig McIntosh, Karthik Muralidharan, Simeon Nichter andseminar participants at UC San Diego and WGAPE NYU Abu Dhabi. A randomized controlled trial registry entry is available at:https://www.socialscienceregistry.org/trials/2506 as well as the pre-analysis plan. IRB approval was received from UCSD (proto-cols # 171543 and #171544) and the University of Liberia. The evaluation was supported by Center for Global Development and UCSan Diego department of economics. Romero, Sandefur, and Sandholtz acknowledge financial support from the Central Bank ofColombia through the Lauchlin Currie scholarship and the Research on Improving Systems of Education (RISE) program, and theInstitute for Humane Studies, respectively. The views expressed here are ours, and not those of the Ministry of Education of Liberiaor our funders. All errors are our own.

1 Introduction

A well-developed body of theory holds that political competition leads to growth by allowing voters to

hold governments accountable for providing public goods (Stigler, 1972; Besley, Persson, & Sturm, 2010;

Acemoglu, Naidu, Restrepo, & Robinson, 2014). While existing empirical work documents an association

between political competition and government responsiveness (Besley & Burgess, 2002), there is surpris-

ingly little direct evidence on how these mechanisms of political competition operate in practice. The idea

that voters reward politicians at the ballot box for championing their preferred policies, while consistent

with the existing quasi-experimental evidence, has largely been taken as an article of faith in the literature

until very recently.

One reason for this dearth of evidence is that exogenous policy shocks delivered at an electorally

meaningful geographic level, and at an electorally salient time, are rare. Public good provision is often

targeted to maximize parties’ perceived incumbency advantage, e.g. by redistributing based on factors

such as co-ethnicity or party affinity (Kramon, Posner, et al., 2016). A large body of work establishes that

voters reward politicians for direct redistribution (Manacorda, Miguel, & Vigorito, 2011; Pop-Eleches &

Pop-Eleches, n.d.) and clientelistic transfers (Wantchekon, 2003). There are comparatively few studies on

the electoral returns to programmatic policy thought to bring broad-based growth, though a few notable

exceptions bear mentioning. Larreguy, Marshall, and Trucco (2015) find that Mexican voters reward federal

politicians land titling reforms, and Harding (2015) finds that incumbents in Ghana benefit electorally

where roads improve. However, the picture admits significant nuance. de Kadt and Lieberman (2017) find

that improvements in service provision decrease support for local incumbents in South Africa. Bursztyn

(2016) finds that in poor Brazilian municipalities, increased public education spending is associated with

worse reelection outcomes for incumbents.

If the literature provides little conclusive direct evidence on electoral returns to public good provision,

another way to measure political accountability would be to test whether politicians behave as if voters

hold them responsible for implementing good policies – i.e., do they react to information about their

constituents’ opinions? However, working with elected officials as experimental subjects is difficult. In the

US, Butler and Nickerson (2011) find that New Mexican state legislators who receive information about

their constituents’ opinions are much more likely to vote accordingly. But it’s not clear lawmakers in

younger democracies would react similarly, and we know of no studies in the developing world which

have measured this.

1

Finally, it should similarly be possible to test whether voters behave as if they can hold politicians

accountable by seeing how information on public good provision affects their political actions. Many

studies provide information to voters, but nearly all focus on exposing corruption (A. V. Banerjee, Green,

McManus, & Pande, 2014; De et al., 2011; Ferraz & Finan, 2008) or measuring performance (A. Banerjee,

Kumar, Pande, & Su, 2011; Grossman & Michelitch, 2016). These are important avenues for accountability,

but we know of little research studying the more direct channel of providing information related to the

public goods politicians are supposed to produce. (Bursztyn (2016) is a notable exception here too, show-

ing somewhat surprisingly that poor Brazilian respondents rate their local government more negatively

upon learning that it has reallocated spending from cash transfers to education.)

Not all information is relevant to political action: for example, (Lieberman, Posner, & Tsai, 2014) find

zero effect of providing childrens’ school information on parents’ political action in Kenya. They give some

examples of when information might reasonably be expected to drive political action: “ ... for information

to generate citizen action it must be understood; it must cause people to update their prior beliefs in some

manner; and it must speak to an issue that people prioritize and also believe is their responsibility to

address. In addition, the people at whom the information is directed must know what actions to take and

possess the skills for taking these actions; they must believe that authorities will respond to their actions;

and, to the extent that the outcome in question requires collective action, they must believe that others in

the community will act as well.”

In this paper, we bring together evidence from three experiments to address many of those conditions

simultaneously, studying how politicians and voters respond to information about each other in the con-

text of public good provision in Liberia. Liberia is a young democracy. Its first free and fair nationwide

election was held in 2005, after two decades of coups and civil war. Its 2017 election offered the chance for

the first democractic peaceful transfer of power in the country’s history (after Ellen Johnson Sirleaf was

elected to two terms as president in 2005 and 2011). Like many nascent democracies, Liberia’s politics

are broadly thought to be characterized by cults of personality rather than consistent policy platforms.

Clientelism, according to anecdote, is rife, and Maffioli (2017) found suggestive evidence that the govern-

ment manipulated its response to the Ebola crisis for electoral gain. The question of whether democratic

channels allow for voters to choose politicians who will invest in public goods is a vital one for Liberia.

We measure political responses to a public school policy change, carried out and evaluated via RCT in

an election year in Liberia. The Partnership Schools for Liberia (PSL) program, implemented by president

Ellen Johnson Sirleaf’s Ministry of Education, transfered management of 93 randomly selected public

2

primary schools to private school organizations (and provided additional resources), starting in the 2016-

2017 school year. The program was controversial, prompting frequent media coverage in the national

and international press, leading to unusually high salience for a school policy. Romero, Sandefur, and

Sandholtz (2017) show that the program increased test scores by 60% (.19 σ), teacher attendance by 50%,

and satisfaction of both students and parents by about 10%. These results were initially reported publicly

in September 2017, about one month before nationwide elections for the presidency and the House of

Representatives.

We leveraged the random assignment of the program to measure its impact on political outcomes,

including survey responses at the household level and, eventually, official vote totals at the precinct level.

This allows us to directly measure voters’ response to the school policy change. We then conducted an

information experiment among candidates for Liberia’s House of Representatives, varying the provision of

school policy evidence (including the effect on test scores and the effect on households’ political attitudes).

This allows us to measure whether candidates behave as if voters will hold them accountable for their

positions on this school policy, as proxied by the candidates’ public positions on the policy. Finally,

we conducted a separate information experiment among households whose political attitudes we had

surveyed before. We now varied whether the households received information about candidates’ positions

on the school policy, recording their voting intentions and approval ratings of the politicians. This allows

us to measure whether these voters behave as if they can hold politicians accountable for this school policy.

Our results show that better information, coupled with competitive elections, can lead candidates to

adopt the policies voters favor, and can lead voters to elect them. The school policy itself made households

about 7% more likely to approve of the government’s performance on education relative to the control

group, 11% more likely to say education is their top priority for government spending. It also made both

households and teachers to have a more hopeful view of the country’s future, and decreased teachers’

satisfaction with the teacher union by 21%.

Information provided to candidates led to increased support for the policy – but only in places where

certain conditions were met. Our main outcome variable was whether candidates asked us to describe

them as favoring the expansion of the school policy in our voter outreach activity. We identify at least

two groups of candidates that seem to have been sensitive to our information: candidates in competitive

districts (where the margin of victory was less than five percentage points) and candidates who did not

have their own children in primary school. For example, policy information made candidates about 32%

more likely to want us to tell their voters they supported expanding the policy if they were running

3

in a district where the margin of victory ended up being less than five percentage points (these districts

correlate closely with places where turnout was high in the previous Representative elections, in 2011). The

content of the information varied among candidates – some received information about the policy’s impact

on learning, whlie other just received information on the policy’s popularity. While we are underpowered

to confidently distinguish effect sizes from these two arms, the results suggest that effects were driven by

the popularity information.

Informing households of candidates’ positions about the policy made them 17% more likely to report

intending to vote for a presidential candidate from parties associated with the policy, and 44% less likely

to approve of a presidential candidate from parties not associated with the policy. It also made them more

likely to vote for legislative candidates who support the policy – but, again, only in competitive districts.

In districts with an eventual margin of victory of less than five percentage points, candidate information

more than doubled the likelihood households said they planned to vote for a candidate we identified as

favoring the expansion of the school policy.

Overall, we find no average effects of the policy on voting patterns, but in competitive districts, the

policy increased vote share for the presidential candidate from the outgoing president’s party and de-

creased vote shares for incumbent representatives. In places where the policy’s impacts on test scores

were strongest, the policy increased vote share for representative candidates from the outgoing presi-

dent’s party and decreased incumbent representatives’ vote share but also depressed turnout.

We show that policies relating to the provision of public goods, information about the same, can be

electorally consequential in a young democracy. Our study relates to three strands of literature: electoral

incentives for public goods; better voter information to politicians; and better politician information to

voters. In this context, information about public goods was able to improve electoral accountability: voters

are happier in places where the school policy took effect; candidates support the policy when they know

voters favor it; voters support the candidates who support the policy they like. We stress that these are

just three steps in the process of democratic accountability, perhaps necessary but probably not sufficient

to ensure the improvement of public good provision.

4

2 Experimental design

2.1 Background

Education is often seen as one of the main public goods which it is a government’s role to provide.

However, in many developing countries, the public school system fails to provide the level of education

deemed basic by international institutions. Liberia’s moribund public education system exemplifies this

failure. The civil war of 1999-2003 and the Ebola epidemic of 2014 left the Ministry of Education with little

capacity to run a national school system. An effort to clean thousands of ghost teachers from Ministry

payrolls was cut short (New York Times, 2016), and while systematic data is scarce, teacher absenteeism

appears common (Mulkeen, 2009). Nearly two-thirds of primary aged children are not in school, including

over 80% of children in the poorest quintile, placing Liberia in the lowest percentile of net enrollment rates

in the world, and at the 7th percentile in youth (15-24) literacy (EPDC, 2014). Demographic and Health

Surveys show that among adult women who did not go to secondary school, only six percent can read

a complete sentence. In 2013, after all 25,000 high school graduates who sat the University of Liberia’s

college entrance exam failed, President Ellen Johnson Sirleaf said the education system was “a mess.”

2.1.1 The policy

In response, the Liberian Ministry of Education announced a pilot program – “Partnership Schools for

Liberia” (PSL) – which contracts the management of 93 government primary schools to one of eight

private school operators in a Public-Private Partnership (PPP). The government (and donors) provide

these operators with funding on a per-pupil level. The operators were given responsibility for (though not

ownership of) the resources the government normally uses to provide education – schools, classrooms,

materials, and teachers, and a grant equal to USD$50 per pupil on top of that. In exchange, operators

are responsible for the daily management of the schools, and can be held accountable for results. The

operators include high profile, for-profit chains with investors like Bill Gates, Mark Zuckerberg, the World

Bank, and DFID. Other operators are non-profit NGOs based in Liberia and abroad, and one operator is a

respected Liberian religious institution.

2.1.2 The evaluation

Based on criteria established by the evaluation team, the Ministry of Education, and operators, 185 PSL-

eligible schools across 12 of Liberia’s 15 counties were identified. These schools are not a representative

5

sample of public schools in the country – they have better facilities, internet access, and road access than

the average school in the country. The eligible schools were split into pairs matched on administrative

data, and treatment was assigned randomly within matched pairs.

Romero et al. (2017) showed that the program increased test scores by 60%, teacher attendance by 50%,

and satisfaction of both students and parents by about 10%. The evaluation also showed that at least some

of the critiques of program detractors were well-founded: one operator chose to enforce class size limits,

forcing hundreds of students to leave their regular school. The same operator also requested reassignment

of 74% of the teachers in the schools it operated. The official report of these results became public about

one month before nationwide elections for the presidency and the House of Representatives.

2.1.3 The political context

The program has enjoyed a relatively high profile, garnering attention from local and international news

outlets, and receiving a special condemnation from a UN Special Rapporteur, who wrote that Liberia was

abrogating its responsibilities under international law. The National Teacher Association of Liberia staged

a strike, calling for the resignation of the Minister of Education. In response, students blocked the main

highway to the country’s international airport, demanding that the government and the teachers’ union

send the teachers back to class.

The policy has been championed by the incumbent president, Ellen Johnson Sirleaf of the Unity

Party (UP). At the time of the policy’s implementation, Sirleaf was nearing the end of her second (and

constitutionally-mandated final) term as president. Her vice president, Joseph N. Boakai, was the UP’s

presidential candidate in the 2017 election. Reforming the education sector has been a priority for the

administration. Both the president and the vice president have championed PSL and gone out of their

way to associate themselves with the program. (President Sirleaf attended a Flag Day celebration at one

of the private operator’s schools, and Vice President Boakai spoke at the graduation ceremony for the

same operator’s teacher training course.) However, as the election neared, the politically powerful Na-

tional Teachers Association of Liberia (NTAL), which opposes the program, became more vocal in their

opposition. Neither Boakai nor any of the other presidential candidates have given this policy a strong

role in their campaign.

A few factors make it less than straightforward to predict how voters will react to this school pol-

icy. On the one hand, the program brought more resources into communities and improved test scores,

something which many parents likely perceived. On the other hand, it could alienate voters who see the

6

privatization of education provision as a dereliction of the government’s duty. Voters may be inclined

to reward politicians for acknowledging the inadequacy of public provision and offering other options.

Or they may punish politicians for abdicating their fundamental responsibility to provide education to

private entities with limited accountability, many foreign and some for-profit.

2.2 Experimental design: Overview and timeline

In this paper, we present results from three different experiments. We devote one subsection below

to the design of each experiment, including sampling, balance, and summary statistics. The first ex-

periment uses the random assignment of schools to the policy in order to measure the policy’s im-

pact on political attitudes and outcomes. We collected this data at the end of the first school year in

which the policy was in place: May/June 2017. The second experiment varies the provision of RCT

evidence on the policy to political candidates and measures how it affects their support for the pol-

icy. It was carried out between 26 September and 9 October, 2017. The third experiment varies the

provision of information on politicians’ support of the policy to households and measures how it af-

fects their voting intentions. It was carried out between 2 October and 9 October, 2017. The elec-

tion was held on 10 October, 2017. The experiments described here were pre-registered along with

pre-analysis plans at https://www.socialscienceregistry.org/trials/1501 (policy impact on political atti-

tudes) and https://www.socialscienceregistry.org/trials/2506 (information experiments for candidates

and households).

2.3 Policy experiment: Design and Sampling

We examine outcomes in two sets of data using the random variation from the policy evaluation itself:

household survey data and official electoral results. The household data were collected from a survey

of the households of a random subset of students in the evaluation of the policy. More details on this

sampling process can be found in Romero et al. (2017).

Electoral data at the voting precinct level were collected from the website of the National Elections

Commission (NEC) of Liberia (http://www.necliberia.org). Precinct coordinates were obtained from NEC.

In all our analyses involving the policy’s impact on election results, there are various ways to link voting

precincts to treatment and control schools. While the randomized evaluation included only 3% of the

country’s public primary schools, 61% of the 2,080 polling places for the 2017 election were within 10km

of one of these RCT schools. We consider a 10km radius as the catchment area for a school, and exclude

7

all polling centers more than 10km from any RCT school. We then define treated polling places as those

within 10km of at least on treated school, and control polling places as those within 10km of no treated

school.1 Widening or narrowing this radius does not qualitatively change the results, nor does assigning

treatment based on a set of Voronoi polygons drawn around all RCT schools (see future Appendix).

2.3.1 Policy experiment: Balance and summary statistics

We check that our treatment definition provides a set of polling places which are balanced on 2011 election

outcomes.1Due to poor data and logistical constraints, the list of schools which ended up being operated privately differed slightly from the

list of schools originally assigned to treatment – Romero et al. (2017) gives a more complete discussion of the treatment assignmentprocess, and uses the originally-assigned intent-to-treat schools in its analyses. In this paper’s analyses, we defined “treated”schools as schools which actually came under private operation. Using the intent-to-treat assignment does not substantively changethe results; these analyses are available upon request.

8

Table 1: Balance – Policy Experiment (2011 Election Results)

Treatment Control Difference Difference(F.E.)

Precinct-level data (N = 1080)

Distance to nearest RCT school 3.75 5.71 -1.96∗∗∗ -1.53∗∗∗

(2.56) (2.94) (0.37) (0.29)Num. public primary schools within 10 km 18.98 9.14 9.83∗∗∗ 3.40∗∗∗

(10.21) (4.35) (2.69) (0.87)Voters registered 2011 1,229.91 916.92 312.98∗∗ 103.66∗∗

(723.25) (612.10) (121.15) (49.01)Invalid pres. vote share 2011 0.06 0.08 -0.01∗ 0.00

(0.04) (0.04) (0.01) (0.00)Voter turnout (pres.) 2011 0.74 0.73 0.00 -0.00

(0.06) (0.08) (0.01) (0.01)UP pres. vote share 2011 0.43 0.41 0.02 -0.01

(0.16) (0.20) (0.04) (0.02)CDC pres. vote share 2011 0.33 0.22 0.10∗∗ 0.02

(0.18) (0.16) (0.04) (0.02)LP pres. vote share 2011 0.04 0.07 -0.03 -0.02

(0.08) (0.16) (0.03) (0.02)Invalid 2011 0.07 0.07 -0.01 -0.00

(0.03) (0.04) (0.01) (0.00)Voter turnout (rep.) 2011 0.73 0.72 0.00 -0.00

(0.06) (0.07) (0.01) (0.01)UP rep. vote share 2011 0.20 0.17 0.02 0.03

(0.17) (0.20) (0.04) (0.02)CDC rep. vote share 2011 0.16 0.11 0.05∗ -0.02

(0.13) (0.15) (0.03) (0.02)Invalid rep. vote share 2011 0.10 0.12 -0.02 -0.03

(0.15) (0.19) (0.03) (0.02)No UP rep. candidate 2011 0.12 0.15 -0.03 -0.03

(0.32) (0.36) (0.03) (0.03)No CDC rep. candidate 2011 0.26 0.36 -0.10 -0.05

(0.44) (0.48) (0.07) (0.05)No LP rep. candidate 2011 0.18 0.17 0.00 -0.04

(0.38) (0.38) (0.04) (0.03)

Sample consists of precincts within 10 km of a school in the Partnership Schools RCT evaluation. Precincts areclassified as treated if they are within 10 km of a school operated as a partnership school and control if not.∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

It is important to note that our treatment definition means that “control” polling places are, almost

mechanically, more remote than “treated” ones, since they are likely to be near fewer schools. They also

had more registered voters in 2011. Therefore all our election regressions include controls for distance to

the nearest school and the number of public primary schools (whether in the RCT or not) within 10km

of the polling place, as well the following 2011 election variables: voters registered, voter turnout, invalid

9

votes, and vote share for the presidential and representative candidates for the three main parties which

took place in both elections: Unity Party (UP), Congress (later Coalition) for Democratic Change (CDC),

and Liberty Party (LP).

Balance tables for the households in the original policy evaluation can be found in Romero et al. (2017).

2.4 Candidate experiment: Design and Sampling

Our candidate information experiment was conducted through a phone survey in which we attempted to

call all 992 candidates running for seats in the House of Representatives. Our sample consists of the 681

candidates we reached (69%). We did not exclusively reach inconsequential candidates: of the 66 House

incumbents standing for reelection our sample includes 22, as well as 3 of the 7 retiring candidates. We

also reached 32 of the 73 candidates who eventually won, and 13% of our sample ended up as the winner

or the runner-up in their district. Figure 1 plots the density of vote shares for candidates we did and did

not survey.

Figure 1: Distribution of vote shares for surveyed and unsurveyed candidates

05

1015

Den

sity

0 .2 .4 .6 .8Vote share

Surveyed (666) Not surveyed (290)

Note: Figure excludes about 20 surveyed candidates from two districts for which election results were still pending as of the time

of writing

10

Conditional on reaching a candidate by phone, we randomized them into one of four treatment arms.

The “control” condition included basic information about the school policy, including what it entailed,

and one sentence about what supporters and opponents of the policy said about it:

In the Partnership Schools program, 93 government primary schools became Partnership Schools, man-

aged by one of eight private and NGO school providers. [Sentence describing which providers operated

in the candidate’s county] Teachers in Partnership Schools remain on government payroll, and buildings

remain the property of the government and free to students. These schools also received extra resources

from foreign donors: 50 US per student.

Supporters of Partnership Schools believe that private management can bring innovation and improve-

ment to Liberia’s schools. Opponents of Partnership Schools believe that the resources would be better

spent within the public system, without private contractors.

The “impact information” condition included the control language plus a summary of the findings of

Romero et al. (2017):

The Ministry of Education commissioned an independent scientific evaluation of Partnership Schools

using state-of-the-art methodology. The study was carried out by academics at institutions based in the

United States: The Center for Global Development and the University of California.

The evaluation showed how the outcomes for students and teachers were different in Partnership Schools.

The children in Partnership Schools learned 60% more math and English than children in the traditional

public schools. That means that children in a Partnership School learned more in 6 months than children

in a traditional public school learned in a whole school year. The evaluation also found that teachers in

Partnership Schools were twice as likely to attend school.2 The evaluation also identified some problems:

in some schools run by Bridge International Academies, some students were kicked out and had to

transfer to different schools, and over half of the teachers were removed. The program is also expensive:

the partnership schools cost at least twice as much to run as government schools, and in some cases

much more.

The “popularity information” condition included the control language plus a summary of the findings

on the policy’s impact on political attitudes (presented later in Table 4):

2This was an inadvertent error. Teachers were in fact 50% more likely to attend, not twice as likely.

11

The Ministry of Education commissioned an independent scientific evaluation of Partnership Schools

using state-of-the-art methodology. The study was carried out by academics at institutions based in the

United States: The Center for Global Development and the University of California.

The researchers interviewed voters whose children went to Partnership Schools and traditional govern-

ment schools, as well as teachers in these schools.

They found that voters whose children went to Partnership Schools were: 10% MORE satisfied with

their children’s education, 7% MORE likely to say the government’s performance on education was

good, 11% MORE likely to say education is their top priority for government, and 14% MORE likely

to say Liberia is moving forward .

Teachers in Partnership Schools were: 21% LESS likely to be satisfied with the teachers’ union (NTAL),

and 7% MORE likely to say Liberia is moving forward.

Finally, a fourth condition contained the control language, the impact information, and the popularity

information.

2.4.1 Candidate experiment: Balance and summary statistics

We check the balance on candidates’ characteristics and pre-treatment survey responses among various

survey conditions here. In most subsequent analyses, we will pool all information treatments into a single

“any information” treatment.

12

Table 2: Balance – Candidate experiment (N = 681)

Variable ControlPopularity

InfoEffectiveness

InfoDifferencePop - Ctrl

DifferenceEffect - Ctrl

Incumbent 0.024 0.063 0.023 0.039 -0.001(0.154) (0.244) (0.151) (0.023)* (0.017)

Eventual winner or runner-up 0.131 0.143 0.102 0.012 -0.029(0.339) (0.351) (0.303) (0.039) (0.036)

UP (incumbent pres. party) 0.036 0.063 0.052 0.027 0.016(0.187) (0.244) (0.223) (0.024) (0.022)

CDC (main opposition) 0.072 0.038 0.058 -0.034 -0.014(0.260) (0.192) (0.235) (0.025) (0.027)

Number of attemptsnecessary to interview 2.554 2.424 2.512 -0.130 -0.043

(1.949) (1.728) (2.001) (0.205) (0.215)Has own children in primary 0.601 0.605 0.636 0.004 0.035

(0.491) (0.490) (0.483) (0.056) (0.054)Candidate has Univ. degree 0.717 0.665 0.744 -0.052 0.027

(0.452) (0.474) (0.438) (0.051) (0.048)It’s good for gov’t to work w/

private companies to provide edu. 0.931 0.866 0.934 -0.065 0.002(0.254) (0.341) (0.249) (0.034)* (0.028)

Heard of PSL 0.596 0.677 0.620 0.081 0.023(0.492) (0.469) (0.487) (0.053) (0.053)

Heard of any PSL operator 0.892 0.873 0.866 -0.018 -0.025(0.312) (0.334) (0.341) (0.036) (0.036)

’Strongly’ or ’Somewhat’approve of teachers’ union 0.981 0.974 0.970 -0.007 -0.012

(0.136) (0.159) (0.172) (0.017) (0.017)Believes voters hold exec.

branch responsible for education 0.842 0.899 0.818 0.056 -0.025(0.365) (0.303) (0.387) (0.037) (0.041)

Believes voters hold exec.branch responsible for PSL 0.722 0.738 0.687 0.016 -0.035

(0.450) (0.442) (0.465) (0.054) (0.055)Believes more than ’a few’ voters

have heard of PSL 0.193 0.152 0.186 -0.041 -0.007(0.396) (0.360) (0.390) (0.042) (0.043)

Observations 166 158 172 324 338

Candidates were randomized to receive information about PSL’s popularity, its effectiveness, or both. These columns excludethe group which received both treatments, for simplicity, but comparisons among all interactions are available upon request.∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

A few summary statistics are worth pointing out here. Most candidates have heard of the school

policy’s name (“PSL”), and nearly all have heard of at least one of the operators associated with it. Nearly

all support the idea of public-private partnerships in education, but nearly all also approve of the teachers’

union. Most think that their voters (correctly) attribute responsibility for the school policy to the executive

13

branch, but only about 19% think that more than ‘a few’ of their voters have heard of the policy.

2.5 Household experiment: Design and sampling

Our household information experiment was also conducted as a phone survey. We called households

who had been previously interviewed (in person) as part of the 1-year evaluation of the school policy. We

reached 494 of the 833 households for whom we had at least one unique phone number (59%).

It is important to note that this sample is not nationally representative, in that it consists only of

households with children in (a non-representative set of) public primary schools. Our results should not

be interpreted as informative about the average Liberian voter, but rather a subpopulation for whom a

school policy may be expected to be particularly salient.

The control condition consisted in this brief mention of the three presidential candidates who took part

in a debate:

Thank you. We are near the end of the survey. Now I just want to give you some information about the

candidates.

Liberia’s last presidential debate was on Septermber 26th. The three candidates who attended the debate

were: MacDella Cooper from LRP, Alexander Cummings from ANC, and Mills Jones from MOVEE.

The treatment condition included that prelude as well as the candidates words regarding the school

policy from that debate:

In that debate, each candidate made a statement about Partnership Schools or PSL. I’m going to read

you a part of each candidate’s statement. Please listen:

MacDella Cooper said: “It’s a test project. Maybe at the end of the test, we’ll see . . . Putting the

Liberian public school in the hands of a private organization, I don’t see the benefit yet.”

Alexander Cummings said “We should also be open to different solutions. And we can’t be fixated on

only one traditional way of doing things. We got to be creative. We got to be bold.”

Mills Jones said: “We are not going to do it. It suggests to me that we have given up on our own

capacity to solve our problems and so we must look outside for help. We’re not going to do that.”

The treatment condition also included a list of the representative candidates we had spoken to in the

candidate survey, who had asked us to let their voters know their position on the school policy:

14

Some of the candidates for Representative in YOUR district also have made statements about PSL, which

they wanted us to share with you. Please listen carefully:

These candidates say PSL should be taken into more schools, and supported by the national budget:

[names]

These candidates say PSL needs to be tested more before making a decision: [names]

These candidates say PSL should be stopped immediately, and normal government schools should get

that support: [names]

The median (interquartile range) household in our sample received information about the positions of

4 (2,6) candidates, representing 30 (24,46)% of the candidates on the ballot in their district.

2.5.1 Balance and summary statistics: Households

We check balance of our household subsample on pre-information-treatment characteristics.

15

Table 3: Balance – Household experiment (N = 494)

Variable ControlCandidate

Info Difference NSchool policy treatment 0.543 0.514 -0.028 494

(0.499) (0.501) (0.045)Knows current Representative’s name 0.822 0.866 0.045 494

(0.383) (0.341) (0.033)Satisfied with Representative 0.342 0.308 -0.034 477

(0.475) (0.463) (0.043)Related to a rep. candidate 0.170 0.184 0.014 492

(0.376) (0.388) (0.034)Related to member of teachers’ union 0.333 0.277 -0.056 485

(0.472) (0.448) (0.042)Attended any campaign event 0.358 0.329 -0.029 489

(0.480) (0.471) (0.043)It’s good for gov to work w/

private companies to provide sch 0.962 0.942 -0.021 480(0.190) (0.235) (0.020)

Any candidate has talked about PSL 0.146 0.160 0.014 444(0.354) (0.367) (0.034)

Heard of PSL 0.644 0.615 -0.028 494(0.480) (0.487) (0.044)

Heard of any operator 0.862 0.842 -0.020 494(0.345) (0.365) (0.032)

Legislature created PSL 0.008 0.008 0.000 494(0.090) (0.090) (0.008)

Executive branch created PSL 0.259 0.271 0.012 494(0.439) (0.446) (0.040)

Reps. have authority to change PSL 0.294 0.263 -0.032 359(0.457) (0.441) (0.047)

PSL should be expanded and fundedthrough the national budget. 0.803 0.782 -0.021 478

(0.398) (0.413) (0.037)Children learn more in PSL schools 0.827 0.850 0.022 446

(0.379) (0.358) (0.035)

Notes∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

The assignment of the Candidate Inforation treatment is balanced in terms of the original PSL treat-

ment. Households randomly assigned to receive candidate information also happened to be in districts

for which we had information for slightly fewer candidates.

Households in our sample are reasonably well-informed. Almost all respondents to the household

survey could correctly name their current Representative, but only 34% are satisfied with him or her. 36%

have attended at least one campaign event for a Representative candidate, and 17% are related to someone

running for office. A third are related to a member of the teachers’ union, but almost everyone supports

16

the idea of public-private partnerships in education, and 4 in 5 think children learn more in PSL schools

and that the policy should be expanded. But it’s not clear to which political actors voters attribute the

policy, if any. Nearly everyone (correctly) does not credit the legislature with creating the policy, but 29%

believe representatives have the authority to change it. Only a quarter believed the executive branch had

created it. 15% said they had heard at least one candidate mention PSL, a group we will come back to

later to help define outcomes.

3 Experimental results

In this section we present three sets of results.

1. The impact of the randomized school reform itself on political outcomes

2. The effect of policy evidence on candidates’ positions

3. The effect of information about candidates’ positions on households’ stated voting intentions

3.1 Policy experiment results

In this section we present two subsets of results which exploit the random variation of the school policy

itself: survey outcomes on political attitudes, measured in May/June 2017; and vote tallies from the

October 2017 election at the precinct level.

17

3.1.1 Survey outcomes

Table 4: Selected household and teacher survey outcomes (policy midline)

Treatment Control Difference Difference(F.E.)

Household midline survey (N = 1207)

Satisfied w/ childs edu (HH) 0.743 0.689 0.054∗ 0.069∗∗∗

(0.437) (0.463) (0.028) (0.020)Gov performance on edu is good (HH) 0.566 0.549 0.017 0.034∗

(0.496) (0.498) (0.032) (0.020)Schools top priority for gov spending (HH) 0.811 0.739 0.072∗∗∗ 0.068∗∗∗

(0.392) (0.440) (0.026) (0.020)Liberia is moving forward (HH) 0.577 0.507 0.070∗∗ 0.071∗∗∗

(0.494) (0.500) (0.032) (0.019)

Teacher midline survey (N = 752)

Satisfied w/ teacher union (teachers) 0.394 0.434 -0.040 -0.089∗∗

(0.489) (0.497) (0.046) (0.040)Liberia is moving forward (teachers) 0.745 0.690 0.056 0.053∗

(0.436) (0.464) (0.038) (0.030)

This table presents the mean and standard error of the mean (in parentheses) for the control (Column 1) andtreatment (Column 2) groups, as well as the difference between treatment and control (Column 3), and the differencetaking into account the randomization design (i.e., including “pair” fixed effects) in Column 4. Standard errors areclustered at the school level. The sample is the original (intent-to-treat) treatment and control allocation.∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

Table 4 shows how the school policy itself affected some political attitudes of the households affected

by it. Households became more impressed with the goverment’s peformance on schools and more likely

to say schools were their top priority for government spending. Teachers likewise became more hopeful

abot the country’s direction in general, and became less satisfied with the teachers’ union. It should be

noted, however, that many other political outcomes were tested; most saw no effect (see Appendix A).

3.1.2 Precinct-level voting outcomes

Using precinct-level election returns, we show how the school policy affected voting patterns. While the

randomized evaluation included only 3% of the country’s public primary schools, 61% of the 2,080 polling

places for the 2017 election were within 10km of one of these RCT schools. We consider a 10km radius

as the catchment area for a school, and exclude all polling centers more than 10km from any RCT school.

We then define treated polling places as those within 10km of at least on treated school, and control

18

polling places as those within 10km of no treated school.3 Widening or narrowing this radius does not

qualitatively change the results (see future Appendix).

It is important to note that this treatment definition means that “control” polling places are, mechan-

ically, more remote than “treated” ones, since they are likely to be near fewer schools. Therefore all our

election regressions include controls for distance to the nearest school and the number of public primary

schools (whether in the RCT or not) within 10km of the polling place.

Table 5: School policy effects on 2017 election

ShareUP (pres.)

ShareUP (rep.)

ShareIncumbent

(rep.)

ShareInvalid(rep.)

Turnout(rep.)

VotersRegistered

Any treated schoolwithin 10 km -0.014 0.005 0.001 -0.005∗ -0.005 56.709

(0.015) (0.039) (0.039) (0.003) (0.009) (56.577)

N 1076 754 923 1058 1058 1076DV Mean 0.286 0.176 0.243 0.050 0.759 1331.093Controls Y Y Y Y Y YStandard errors clustered by school. Sample consists of precincts within 10 km of a school in the PartnershipSchools RCT evaluation. Precincts are classified as treated if they are within 10 km of a school operated as a part-nership school and control if not. Controls: 1st-round presidential vote shares and representative vote shares from2011 for the three main parties which fielded candidates in both 2011 and 2017 (Unity Party, Congress/Coalitionfor Democratic Change, Liberty Party); share of invalid votes in 2011; number of registered voters in 2011; voterturnout in 2011; distance of the voting center to the nearest RCT school; and number of total primary publicschools within 10 km of the voting center.* p<0.10, ** p<0.05, *** p<0.01

On average, we detect little to no impact of the school policy on voting patterns. This is unsurprising

– school policy is just one of many factors that may influence voters’ decisions, and voting is a blunt

instrument for preference expression. But because the evaluation was designed using matched pairs, we

can see whether treatment mattered more in places where the policy had a big effect.

3Due to poor data and logistical constraints, the list of schools which ended up being operated privately differed slightly from thelist of schools originally assigned to treatment– ? (?) for a more complete discussion of the treatment assignment process, and thatpaper uses the originally-assigned intent-to-treat schools in its analyses. In this paper’s analyses, we defined “treated” schools asschools which actually came under private operation. Using the intent-to-treat assignment does not substantively change the results,which are available upon request.

19

3.1.3 Heterogeneity in election outcomes by policy impact

Table 6: School policy effects on 2017 election, interacted with treatment effect

ShareUP (pres.)

ShareUP (rep.)

ShareIncumbent

(rep.)

ShareInvalid(rep.)

Turnout(rep.)

VotersRegistered

Any treated schoolwithin 10 km -0.015 0.018 -0.004 -0.004 -0.003 62.346

(0.015) (0.043) (0.033) (0.003) (0.008) (56.101)

Any treated schoolwithin 10 km × TE -0.019 0.147∗ -0.240∗∗∗ 0.008 -0.057∗∗∗ -141.139

(0.025) (0.084) (0.073) (0.005) (0.018) (97.235)

N 1031 733 883 1013 1013 1031DV Mean 0.287 0.173 0.240 0.050 0.760 1339.400Controls Y Y Y Y Y YStandard errors clustered by school. Sample consists of precincts within 10 km of a school in the Partnership Schools RCTevaluation. Precincts are classified as treated if they are within 10 km of a school operated as a partnership school andcontrol if not. Controls: 1st-round presidential vote shares and representative vote shares from 2011 for the three mainparties which fielded candidates in both 2011 and 2017 (Unity Party, Congress/Coalition for Democratic Change, LibertyParty); share of invalid votes in 2011; number of registered voters in 2011; voter turnout in 2011; distance of the votingcenter to the nearest RCT school; and number of total primary public schools within 10 km of the voting center.* p<0.10, ** p<0.05, *** p<0.01

Table 6 shows the effect of interacting treatment assignment with the size of the local treatment effect – de-

fined as the difference between a matched pair’s treatment and control schools’ average test scores. Places

where the policy had a bigger impact saw decreased vote shares for incumbents (and, perhaps, increased

vote share for representatives from the Unity Party). This is slightly puzzling, as representatives were

not responsible for the school policy’s creation or management – it was a centralized program from the

Ministry of Education. Perhaps voters associated the UP with the policy and hoped UP legislators would

enshrine it in law. Or maybe the result indicates decreased apathy – voters who have seen their public

goods improve may be more willing to consider political newcomers who promise change. However,

strong treatment effects also appear to have depressed turnout. It’s conceivable this reflects a diminished

strength of the vote-mobilizing power of the teacher unions, but that reading is highly speculative at this

point.

Next, we tested whether electoral results differed by political competitiveness, as defined by margin of

victory. (Districts where elections were close are also characterized by high turnout – see future Appendix.)

Table ?? shows the effect of interacting treatment with the difference in voteshare between the top two

representative candidates in the district.

20

3.1.4 Heterogeneity in election outcomes by electoral race tightness

Table 7: School policy effects on 2017 election, interacted with victory margin

ShareUP (pres.)

ShareUP (rep.)

ShareIncumbent

(rep.)

ShareInvalid(rep.)

Turnout(rep.)

VotersRegistered

Any treated schoolwithin 10 km =1 0.015 0.012 -0.116∗∗∗ -0.004 0.002 68.542

(0.018) (0.039) (0.033) (0.005) (0.012) (94.860)

margin of victory (rep.) 0.204∗ -0.080 -0.159 0.017 0.039 -0.550(0.113) (0.310) (0.262) (0.019) (0.065) (303.905)

Any treated schoolwithin 10 km =1 ×

margin of victory (rep.) -0.222∗∗ -0.030 0.752∗∗∗ -0.009 -0.050 -88.161(0.111) (0.335) (0.204) (0.017) (0.070) (383.578)

N 1058 754 923 1058 1058 1058DV Mean 0.285 0.176 0.243 0.050 0.759 1332.572Controls Y Y Y Y Y YStandard errors clustered by school. Sample consists of precincts within 10 km of a school in the Partnership SchoolsRCT evaluation. Precincts are classified as treated if they are within 10 km of a school operated as a partnership schooland control if not. Controls: 1st-round presidential vote shares and representative vote shares from 2011 for the threemain parties which fielded candidates in both 2011 and 2017 (Unity Party, Congress/Coalition for Democratic Change,Liberty Party); share of invalid votes in 2011; number of registered voters in 2011; voter turnout in 2011; distance of thevoting center to the nearest RCT school; and number of total primary public schools within 10 km of the voting center.* p<0.10, ** p<0.05, *** p<0.01

In competitive districts, the school policy reduced vote share for the presidential candidate from the

outgoing president’s party (the coefficient on the interaction term is negative because larger margins

of victory correspond to less competitiveness). The program also decreased vote shares for incumbent

representatives in competitive districts.

3.2 Candidate experiment results

Candidates received either information about the policy’s impact on students and schools, or its impact

on households’ and teachers’ political attitudes.

21

Table 8: Policy information’s effect on candidate survey outcomes

ExpandPSL

Fund PSLw/ taxes

Gov shouldwork w/pvt. edu.

Too muchforeigncontrol

PSL⇒↑learning

Voterssupport PSL

Popularity info 0.075 0.008 0.013 -0.039 -0.086∗∗ -0.081(0.056) (0.053) (0.035) (0.057) (0.036) (0.056)

Impact info 0.050 0.035 0.014 -0.051 -0.013 0.003(0.054) (0.053) (0.034) (0.056) (0.035) (0.055)

Both -0.063 -0.015 -0.015 0.068 0.085∗ 0.090(0.077) (0.074) (0.048) (0.078) (0.050) (0.077)

N 679 657 665 655 620 652DV Mean 0.523 0.668 0.893 0.550 0.894 0.600Controls N N N N N NRobust standard errors in parentheses. ’Expand PSL’ means the candidate asked us to tell their voters they supportexpanding the school policy. Other columns are candidate survey outcomes.* p<0.10, ** p<0.05, *** p<0.01

Candidates who received information about the school policy were, on average, no more likely to

tell voters they supported expanding it, nor did other outcomes appear to move much (although, oddly,

popularity information seems to have decreased candidates’ belief that the program boosted test scores).

However, as with the electoral results, these average effects mask important heterogeneity. The remain-

der of the results presented here pool all information treatment arms into a single “any information”

treatment. See Appendix B for (underpowered) analyses which separate the heterogeneous effects of the

treatment arms – the effects appear to be driven mostly by the popularity information.

3.2.1 Heterogeneity in candidate outcomes

While we think these heterogeneity analyses shed light on what factors condition candidates’ repsonses

to information, we take them with a grain of salt as not all of them were prespecified.4

4A randomized controlled trial registry entry and the pre-analysis plan are available at:https://www.socialscienceregistry.org/trials/2506.

22

Table 9: Candidate heterogeneity: Children in primary school

ExpandPSL

Fund PSLw/ taxes

Gov shouldwork w/pvt. edu.

Too muchforeigncontrol

PSL⇒↑learning

Voterssupport PSL

Any Info 0.195∗∗∗ 0.009 0.032 -0.128∗ -0.037 -0.041(0.071) (0.068) (0.044) (0.073) (0.046) (0.072)

Kids in primary 0.243∗∗∗ -0.002 0.047 -0.167∗∗ 0.020 -0.092(0.079) (0.076) (0.050) (0.082) (0.052) (0.080)

Any Info ×Kids in primary -0.213∗∗ 0.047 -0.036 0.165∗ -0.008 0.030

(0.092) (0.088) (0.057) (0.094) (0.060) (0.092)

N 651 633 638 631 597 626DV Mean 0.535 0.678 0.895 0.550 0.894 0.605Controls N N N N N NRobust standard errors in parentheses. ’Expand PSL’ means the candidate asked us to tell their voters they supportexpanding the school policy. Other columns are candidate survey outcomes.* p<0.10, ** p<0.05, *** p<0.01

Among candidates whose own children are not in primary school, information had a significant pos-

itive effect on the likelihood of supporting the policy, which rose to the level of both informed and un-

informed candidates who do have children in primary school. This variable may be a proxy for the

population most lacking in knowledge about the policy – candidates whose own children are in primary

school might already know most of the information we shared in the treatment (although it should be

noted that relatively few of the candidates’ own chlidren are in public schools).

23

Table 10: Candidate heterogeneity: Close races

ExpandPSL

Fund PSLw/ taxes

Gov shouldwork w/pvt. edu.

Too muchforeigncontrol

PSL⇒↑learning

Voterssupport PSL

Any Info -0.003 0.058 -0.020 -0.002 -0.047 -0.008(0.056) (0.054) (0.035) (0.058) (0.036) (0.056)

Margin < 5 pp -0.142∗ 0.058 -0.081 0.094 -0.058 0.055(0.083) (0.080) (0.052) (0.086) (0.054) (0.084)

Any Info ×Margin < 5 pp 0.174∗ -0.065 0.080 -0.094 0.038 -0.045

(0.095) (0.091) (0.060) (0.097) (0.062) (0.096)

N 664 642 650 640 605 637DV Mean 0.518 0.668 0.892 0.547 0.893 0.593Controls N N N N N NRobust standard errors in parentheses. ’Expand PSL’ means the candidate asked us to tell their voters theysupport expanding the school policy. Other columns are candidate survey outcomes.* p<0.10, ** p<0.05, *** p<0.01

These results mirrors the election results, where competitiveness played an important role. In general,

candidates in close races are less likely to support the policy than candidates in uncompetitive races.

We speculate that this could reflect greater involvement of the teachers’ union in close races – or the

candidates’ fear of the teachers’ union involvement. Learning that the policy decreased union support

among teachers may make candidates feel more sanguine about opposing the union. However, further

research is necessary to determine whether this is an important mechanism behind the effect we observe

here.

3.3 Household experiment results

Here we describe the results of providing households with information about candidates’ positions. First,

a word about defining outcomes. Since many of our outcomes consist of asking households about long

lists of candidates, we need a reasonable way to aggregate these outcomes. In order to identify the set

of parties associated with the school policy, we use the household survey question which asked which

candidates, if any, had mentioned the policy. 14% of our respondents reported hearing a candidate talk

about the policy, yielding a list of 65 candidates. Then, for each of the 20 parties represented among all the

candidates, we regressed a party dummy on PSL mentions in the full list of 992 candidates to see which

parties positively and significantly predicted being identified by our households as having mentioned the

24

policy. The coefficients are plotted in Figure 2.

Figure 2: Defining PSL-linked parties

ALPANCCDACDCCLPDJPGDPLLFPLINULPLPPLRPLTPMDRMOVEEMPCNLPNONEPUPRDCTWPUPUPPVCPVOLT

1__

0000

03

-.4 -.2 0 .2

Only three party dummies significantly predicted mentions of the policy: the Alternative National

Congress (the vehicle of a reform-minded Coca-Cola executive), the Liberty Party (helmed by perennial

candidate and lawyer Charles Brumskine), and the Unity Party (who was in power when the policy was

created). In subsequent analyses, “PSL-linked” parties refer to these three.

25

Table 11: Candidate information’s effect on household voting intentions and approval: President

Vote any pres.candidate

Vote PSL-linkedParty pres.candidate

Avg approval ofPSL-linked partypres. candidates

Avg approval ofNon-PSL party

pres. candidates

Feels closeto PSL-linked

party

Candidate info -0.016 0.117∗∗ 0.024 -0.080∗∗ 0.197∗∗∗

(0.044) (0.051) (0.028) (0.032) (0.067)

N 494 318 353 348 191DV Mean 0.628 0.695 0.362 0.177 0.670Controls N N N N NRobust standard errors in parentheses. Sample consists of a subset of households originally contacted as part of PSL midlineevaluation, reached by phone for this follow-up survey about one week before the election on 10 October 2017.* p<0.10, ** p<0.05, *** p<0.01

Information about candidates’ positions increased the likelihood that households reported intending to

vote for presidential candidates from PSL-linked parties. We also asked households to rate their approval

of six of the frontrunner candidates. Candidate information reduced approval rates for non-PSL-linked

candidates (both of the presidential candidates whose debate quotes about the policy were negative are

from non-PSL parties).

Table 12: Candidate information’s effect on household voting intentions: Representative

Vote Rep.PSL Yes

Vote Rep.PSL Maybe

Vote Rep.PSL No

Vote Rep.PSL No Info

Vote AnyRep.

Vote Rep.Incumbent

Candidate info 0.002 -0.060∗ 0.013 0.063 -0.053 -0.001(0.039) (0.032) (0.009) (0.051) (0.044) (0.055)

N 317 317 317 317 494 272DV Mean 0.139 0.091 0.006 0.707 0.605 0.279Controls N N N N N NRobust standard errors in parentheses. Sample consists of a subset of households originally contacted as part of PSLmidline evaluation, reached by phone for this follow-up survey about one week before the election on 10 October2017.* p<0.10, ** p<0.05, *** p<0.01

Candidate information has little to no effect on voting patterns for representatives, on average. House-

holds receiving the information were no more likely to vote for representatives who we had told them

supported (or opposed) the expansion of the policy, or for candidates for whom we gave no information.

(They may have been less likely to vote for candidates who were lukewarm on the policy.) This makes

sense; the school policy was implemented by the executive branch, and the path to electoral accountability

over it is more circuitous for representatives. But there is a group for whom the information appears to

have mattered.

26

3.4 Heterogeneity in household outcomes

Table 13: Voting for representatives in tight races

Vote Rep.PSL Yes

Vote Rep.PSL Maybe

Vote Rep.PSL No

Vote Rep.PSL No Info

Vote AnyRep.

Vote Rep.Incumbent

Candidate info -0.057 -0.045 0.014 0.127∗ -0.116∗ 0.033(0.053) (0.045) (0.012) (0.070) (0.059) (0.073)

District won by < 5 pp -0.114∗∗ 0.062 -0.000 0.098 -0.002 -0.088(0.056) (0.047) (0.013) (0.074) (0.065) (0.077)

Candidate info ×District won by < 5 pp 0.164∗∗ -0.045 0.001 -0.183∗ 0.081 -0.047

(0.080) (0.068) (0.019) (0.106) (0.091) (0.111)

N 307 307 307 307 451 264DV Mean 0.140 0.094 0.007 0.704 0.643 0.273Controls N N N N N NRobust standard errors in parentheses. Sample consists of a subset of households originally contacted as part of PSL midlineevaluation, reached by phone for this follow-up survey about one week before the election on 10 October 2017.* p<0.10, ** p<0.05, *** p<0.01

Electoral competitiveness appears to be an important condition for how households react to information.

In competitive districts, candidate information increased households’ likelihood of intending to vote for a

candidate who we had told them supports the policy’s expansion (and less likely to vote for candidates

who we gave no information about). It is possible, of course, that we are simply seeing the follow-on effects

from giving new information to candidates: since competitive districts were the places where information

moved candidates’ positions, they are the places where the information we gave to voters required them

to update what they already knew.

4 Conclusion

It is common to worry that systems of democratic accountability, perhaps especially in the developing

world, are too broken to have any simple fix. These worries are certainly not baseless. Democractic

accountability for public good provision has a lot of moving parts. It seems to require that voters pay

attention to public goods and to what politicians say and do about them, and it seems to require that

politicians believe voters notice these things and act on them. This is probably just the tip of the iceberg

of what democratic accountability requires, but at least as regards this particular school policy in 2017 in

27

the competitive districts of Liberia, meeting those requirements can be facilitated by the wider availability

of more accurate information about voters and politicians.

We show that the policy in question led to increased satisfaction with the government’s performance

on education and greater vote share, in competitive districts, for the presidential candidate of the party

who created the policy. Information about the policy’s effects on learning and on voter satisfaction led

candidates (in competitive districts) to tell their voters they support the policy. Information about which

candidates support the policy, in turn, led households – in competitive districts – to vote more frequently

for the candidates who support the policy.

Some important caveats apply: our study does not involve a representative sample of either politicians

or households. The politicians we studied are those who selected themselves into our study. The voters

are the parents of children in a selected set of government primary schools. It should also be noted that

how the effects of more information play out in equilibrium is far from certain. Policy evidence might be

met with policy counterevidence, and it’s unclear how all the relevant actors would judge the authority

of competing claims. The rent-capturers who hamper the creation of inclusive economic institutions have

incentives to throw every wrench in the gears of democratic accountability – in the words of Acemoglu

(2010), “large-scale shocks and policy interventions will create political economy responses from those

who see their economic or political rents threatened or from those that see new options to increase these

rents.”

Still, exogenous changes to public good provision are rare. The Partnership Schools for Liberia pro-

gram, coupled with its timing just before the October 2017 Liberian presidential election, provides a

valuable opportunity to gain empirical evidence on how voters react to rigorously-evaluated policy in a

young democracy.

References

Acemoglu, D. (2010, aug). Theory, General Equilibrium, and Political Economy in Development Eco-

nomics. Journal of Economic Perspectives, 24(3), 17–32. Retrieved from http://pubs.aeaweb.org/doi/

10.1257/jep.24.3.17 doi: 10.1257/jep.24.3.17

Acemoglu, D., Naidu, S., Restrepo, P., & Robinson, J. A. (2014). Democracy does cause growth (Tech. Rep.).

National Bureau of Economic Research.

28

Banerjee, A., Kumar, S., Pande, R., & Su, F. (2011). Do informed voters make better choices? experimental

evidence from urban india. Unpublished manuscript, Harvard university.

Banerjee, A. V., Green, D., McManus, J., & Pande, R. (2014). Are Poor Voters indifferent to whether elected

leaders are criminal or corrupt? Political Communication(April 2015). doi: 10.1080/10584609.2014

.914615

Besley, T., & Burgess, R. (2002). THE POLITICAL ECONOMY OF GOVERNMENT RESPONSIVENESS:

THEORY AND EVIDENCE FROM INDIA. The Quarterly Journal of Economics. Retrieved from

http://econ.duke.edu/~psarcidi/lunchf08/besburgess.pdf

Besley, T., Persson, T., & Sturm, D. M. (2010, oct). Political Competition, Policy and Growth: Theory and

Evidence from the US. Review of Economic Studies, 77(4), 1329–1352. Retrieved from http://restud

.oxfordjournals.org/lookup/doi/10.1111/j.1467-937X.2010.00606.x doi: 10.1111/j.1467-937X

.2010.00606.x

Bursztyn, L. (2016, oct). Poverty and the Political Economy of Public Education Spending: Evidence

From Brazil. Journal of the European Economic Association, 14(5), 1101–1128. Retrieved from http://

doi.wiley.com/10.1111/jeea.12174 doi: 10.1111/jeea.12174

Butler, D. M., & Nickerson, D. W. (2011, aug). Can Learning Constituency Opinion Affect How Legislators

Vote? Results from a Field Experiment. Quarterly Journal of Political Science, 6(1), 55–83. Retrieved

from http://www.nowpublishers.com/article/Details/QJPS-11019 doi: 10.1561/100.00011019

De, M. F. P., Hidalgo, F. D., Kasahara, Y., Arguelhes, D. W., Brady, H., Cohon, A., . . .

Zucco, C. (2011). When Do Voters Punish Corrupt Politicians? Experimental Evi-

dence from Brazil *. Retrieved from https://law.utexas.edu/wp-content/uploads/sites/25/

figueiredo when do voters punish.pdf

de Kadt, D., & Lieberman, E. S. (2017, dec). Nuanced Accountability: Voter Responses to Service De-

livery in Southern Africa. British Journal of Political Science, 1–31. Retrieved from https://www

.cambridge.org/core/product/identifier/S0007123417000345/type/journal article doi:

10.1017/S0007123417000345

Ferraz, C., & Finan, F. (2008, may). Exposing Corrupt Politicians: The Effects of Brazil’s Publicly Re-

leased Audits on Electoral Outcomes. Quarterly Journal of Economics, 123(2), 703–745. Retrieved

from https://academic.oup.com/qje/article-lookup/doi/10.1162/qjec.2008.123.2.703 doi:

10.1162/qjec.2008.123.2.703

Grossman, G., & Michelitch, K. (2016). Information Dissemination, Competitive Pressure, and Politician

29

Performance between Elections: A Field Experiment in Uganda. Retrieved from file:///C:/Users/

W/Downloads/GM 2016.pdf

Harding, R. (2015). Attribution and accountability: Voting for roads in ghana. World Politics, 67(4),

656–689.

Kramon, E., Posner, D. N., et al. (2016). Ethnic favoritism in education in kenya. Quarterly Journal of

Political Science, 11(1), 1–58.

Larreguy, H., Marshall, J., & Trucco, L. (2015). Breaking Clientelism or Rewarding Incumbents? Evidence

from an Urban Titling Program in Mexico. (September).

Lieberman, E. S., Posner, D. N., & Tsai, L. L. (2014). Does Information Lead to More Active Citizenship?

Evidence from an Education Intervention in Rural Kenya. World Development, 60, 69–83. doi: 10.1016/

j.worlddev.2014.03.014

Maffioli, E. M. (2017). The political economy of slow-onset disasters: Evidence from

the ebola outbreak. Retrieved from https://sites.duke.edu/elisamaffioli/files/2017/11/

maffioli jobmarket nov6.pdf

Manacorda, M., Miguel, E., & Vigorito, A. (2011, jul). Government Transfers and Political Support.

American Economic Journal: Applied Economics, 3(3), 1–28. Retrieved from http://pubs.aeaweb.org/

doi/10.1257/app.3.3.1 doi: 10.1257/app.3.3.1

Pop-Eleches, C., & Pop-Eleches, G. (n.d.). Government Spending and Pocketbook Voting: Quasi-

Experimental Evidence from Romania. princeton.edu. Retrieved from https://www.princeton.edu/

~gpop/Pop-ElechesRDpocketbookvotingJune09.pdf

Romero, M., Sandefur, J., & Sandholtz, W. (2017). Outsourcing Service Delivery in a Fragile State: Experi-

mental Evidence from Liberia. Retrieved from file:///C:/Users/W/Downloads/Romero JMP.pdf

Stigler, G. J. (1972). Economic competition and political competition. Public Choice, 13(1), 91–106.

Wantchekon, L. (2003, apr). Clientelism and Voting Behavior: Evidence from a Field Experiment

in Benin. World Politics, 55(03), 399–422. Retrieved from http://www.journals.cambridge.org/

abstract S0043887100003798 doi: 10.1353/wp.2003.0018

A Additional policy experiment results

The midline survey for the PSL evaluation included a number of questions on political attitudes. This

section shows the policy effects on most of them.

30

Table A.1: Household survey outcomes (policy midline)

Treatment Control Difference Difference(F.E.)

Household midline survey (N = 1271)

In the past 7 days HH head worked 0.518 0.522 -0.004 -0.006(0.500) (0.500) (0.030) (0.022)

Can read: what is the capital of Liberia 0.302 0.296 0.006 -0.011(0.459) (0.457) (0.032) (0.022)

Can name ≥ 1 senator of own cty 0.574 0.571 0.003 0.004(0.495) (0.495) (0.034) (0.024)

Sees own child schools as gov school 0.917 0.937 -0.020 -0.011(0.276) (0.244) (0.020) (0.015)

Has heard of PSL 0.182 0.146 0.036∗ 0.042∗∗∗

(0.386) (0.354) (0.021) (0.014)Has heard of operator 0.584 0.236 0.349∗∗∗ 0.350∗∗∗

(0.493) (0.425) (0.039) (0.025)Prefers private school to gov school 0.500 0.531 -0.031 -0.042

(0.500) (0.499) (0.033) (0.026)Foreign NGOs have too much control 0.632 0.650 -0.018 -0.023

(0.483) (0.477) (0.034) (0.024)Satisfied w/ Pres. Sirleaf 0.651 0.632 0.019 0.020

(0.477) (0.483) (0.032) (0.023)Satisfied w/ Representative 0.545 0.535 0.009 0.004

(0.498) (0.499) (0.036) (0.024)Satisfied w/ local leaders 0.730 0.690 0.040 0.040∗

(0.444) (0.463) (0.031) (0.021)Roads are highest priority 0.086 0.121 -0.035∗ -0.033∗∗

(0.281) (0.327) (0.021) (0.014)Hospitals are highest priority 0.071 0.091 -0.020 -0.020

(0.257) (0.288) (0.016) (0.012)Gov performance on health is good 0.453 0.480 -0.027 -0.022

(0.498) (0.500) (0.034) (0.023)Gov performance on roads is good 0.360 0.391 -0.032 -0.036

(0.480) (0.488) (0.037) (0.023)Gov is best at public goods 0.474 0.519 -0.046 -0.042∗

(0.500) (0.500) (0.034) (0.025)Own tribe never treated unfairly by gov 0.600 0.603 -0.003 0.001

(0.490) (0.490) (0.032) (0.023)Registered to vote 0.982 0.978 0.004 0.003

(0.133) (0.146) (0.009) (0.006)Plans to vote in October 2017 0.969 0.960 0.008 0.010

(0.174) (0.195) (0.012) (0.009)Feels close to any party 0.213 0.200 0.013 0.010

(0.410) (0.400) (0.027) (0.019)Expressed a voting preference 0.369 0.410 -0.041 -0.044∗

(0.483) (0.492) (0.034) (0.023)Would vote for UP if election tomorrow 0.178 0.198 -0.020 -0.021

(0.383) (0.399) (0.028) (0.017)

This table presents the mean and standard error of the mean (in parentheses) for the control (Column 1)and treatment (Column 2) groups, as well as the difference between treatment and control (Column 3), andthe difference taking into account the randomization design (i.e., including “pair” fixed effects) in Column4. Standard errors are clustered at the school level. The sample is the original (intent-to-treat) treatment andcontrol allocation.∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

31

Table A.2: Teacher survey outcomes (policy midline)

Treatment Control Difference Difference(F.E.)

Teacher midline survey (N = 765)

Member of teacher union 0.328 0.214 -0.114∗∗∗ -0.099∗∗∗

(0.026) (0.017) (0.037) (0.027)On gov payroll 0.706 0.659 -0.047 -0.008

(0.025) (0.020) (0.049) (0.036)Has heard of PSL 0.229 0.621 0.392∗∗∗ 0.371∗∗∗

(0.023) (0.020) (0.048) (0.036)Has heard of operator 0.393 0.935 0.542∗∗∗ 0.557∗∗∗

(0.027) (0.010) (0.049) (0.036)Op. should expand next year | has heard 0.822 0.882 0.060 0.031

(0.041) (0.014) (0.053) (0.042)Rather work in op. school | has heard 0.510 0.731 0.221∗∗∗ 0.159∗∗∗

(0.051) (0.019) (0.065) (0.047)Foreign NGOs have too much control 0.625 0.564 -0.061 -0.052

(0.030) (0.023) (0.042) (0.036)Gov should work w/ private school companies 0.823 0.849 0.026 0.008

(0.023) (0.016) (0.031) (0.025)Satisfied w/ Pres. Sirleaf 0.768 0.755 -0.013 -0.034

(0.026) (0.020) (0.038) (0.032)Satisfied w/ Representative 0.460 0.412 -0.048 -0.022

(0.031) (0.022) (0.053) (0.037)Satisfied w/ local leaders 0.644 0.673 0.029 0.016

(0.030) (0.021) (0.040) (0.033)Registered to vote 0.985 0.992 0.007 0.005

(0.007) (0.004) (0.008) (0.008)Plans to vote in October 2017 0.966 0.976 0.010 0.010

(0.011) (0.007) (0.013) (0.011)Schools are highest priority 0.796 0.810 0.013 0.021

(0.025) (0.018) (0.033) (0.031)Roads are highest priority 0.083 0.077 -0.006 -0.011

(0.017) (0.012) (0.024) (0.022)Hospitals are highest priority 0.057 0.057 0.000 -0.010

(0.014) (0.010) (0.019) (0.018)Gov performance on edu. is good 0.447 0.397 -0.051 -0.052

(0.031) (0.022) (0.040) (0.035)Gov performance on health is good 0.562 0.506 -0.056 -0.065∗∗

(0.030) (0.023) (0.042) (0.032)Gov performance on roads is good 0.361 0.333 -0.028 -0.016

(0.030) (0.021) (0.045) (0.033)Feels close to any party 0.257 0.215 -0.041 -0.044

(0.027) (0.019) (0.035) (0.030)Would vote for UP if election tomorrow | vote pref 0.633 0.573 -0.060 -0.077∗

(0.037) (0.029) (0.059) (0.045)Expressed a voting preference 0.566 0.498 -0.068 -0.056

(0.030) (0.022) (0.046) (0.039)

This table presents the mean and standard error of the mean (in parentheses) for the control (Column 1) and treatment(Column 2) groups, as well as the difference between treatment and control (Column 3), and the difference taking intoaccount the randomization design (i.e., including “pair” fixed effects) in Column 4. Standard errors are clustered at theschool level. The sample is the original (intent-to-treat) treatment and control allocation.∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01

32

B Which information matters?

We’re underpowered to detect differences between the impact of our two types of information, but there

is some suggestive evidence that the popularity information matters more than the impact information.

Table B.1: Children in primary school

ExpandPSL

Fund PSLw/ taxes

Gov shouldwork w/pvt. edu.

Too muchforeigncontrol

PSL⇒↑learning

Voterssupport PSL

Impact info 0.136 -0.044 -0.017 -0.093 -0.057 -0.041(0.089) (0.086) (0.056) (0.092) (0.061) (0.091)

Kids in primary 0.243∗∗∗ -0.002 0.047 -0.167∗∗ 0.020 -0.092(0.079) (0.076) (0.049) (0.082) (0.053) (0.080)

Impact info ×Kids in primary -0.139 0.152 0.050 0.093 0.049 0.058

(0.113) (0.109) (0.071) (0.117) (0.076) (0.115)

Popularity info 0.282∗∗∗ 0.046 0.035 -0.111 -0.043 -0.096(0.089) (0.086) (0.056) (0.092) (0.059) (0.091)

Popularity info ×Kids in primary -0.325∗∗∗ -0.020 -0.045 0.133 -0.071 0.034

(0.115) (0.110) (0.072) (0.118) (0.076) (0.117)

N 471 457 460 454 428 448DV Mean 0.531 0.678 0.896 0.548 0.890 0.596Controls N N N N N Np: Impact = Pop 0.112 0.123 0.191 0.732 0.123 0.842Robust standard errors in parentheses. ’Expand PSL’ means the candidate asked us to tell their voters they supportexpanding the school policy. Other columns are candidate survey outcomes.* p<0.10, ** p<0.05, *** p<0.01

33

Table B.2: Tight races

ExpandPSL

Fund PSLw/ taxes

Gov shouldwork w/pvt. edu.

Too muchforeigncontrol

PSL⇒↑learning

Voterssupport PSL

Impact info -0.000 0.061 -0.017 -0.022 -0.019 0.029(0.068) (0.066) (0.043) (0.070) (0.045) (0.069)

Margin < 5 pp -0.142∗ 0.058 -0.081 0.094 -0.058 0.055(0.083) (0.080) (0.052) (0.086) (0.055) (0.084)

Impact info ×Margin < 5 pp 0.136 -0.055 0.073 -0.052 0.037 -0.076

(0.116) (0.113) (0.074) (0.119) (0.078) (0.118)

Popularity info -0.039 -0.003 -0.029 0.026 -0.125∗∗∗ -0.016(0.069) (0.067) (0.044) (0.072) (0.046) (0.071)

Popularity info ×Margin < 5 pp 0.317∗∗∗ 0.069 0.106 -0.179 0.108 -0.195

(0.117) (0.113) (0.074) (0.120) (0.078) (0.119)

N 484 466 472 464 437 459DV Mean 0.514 0.663 0.892 0.547 0.888 0.584Controls N N N N N Np: Impact = Pop 0.119 0.270 0.649 0.280 0.363 0.315Robust standard errors in parentheses. ’Expand PSL’ means the candidate asked us to tell their voters they supportexpanding the school policy. Other columns are candidate survey outcomes.* p<0.10, ** p<0.05, *** p<0.01

34

Table B.3: Viable candidates

ExpandPSL

Fund PSLw/ taxes

Gov shouldwork w/pvt. edu.

Too muchforeigncontrol

PSL⇒↑learning

Voterssupport PSL

Impact info 0.032 0.029 0.018 -0.034 0.018 0.020(0.068) (0.066) (0.043) (0.070) (0.044) (0.069)

Candidate > 5 pct -0.158∗ -0.017 0.028 0.029 -0.073 -0.008(0.083) (0.081) (0.053) (0.087) (0.056) (0.085)

Impact info ×Candidate > 5 pct 0.038 0.040 -0.029 -0.048 -0.061 -0.020

(0.116) (0.113) (0.074) (0.121) (0.078) (0.119)

Popularity info -0.018 -0.063 -0.009 -0.061 -0.157∗∗∗ -0.124∗

(0.069) (0.066) (0.044) (0.071) (0.045) (0.071)

Popularity info ×Candidate > 5 pct 0.259∗∗ 0.246∗∗ 0.048 0.035 0.209∗∗∗ 0.109

(0.118) (0.114) (0.075) (0.121) (0.078) (0.120)

N 480 463 468 460 433 455DV Mean 0.512 0.661 0.891 0.548 0.887 0.582Controls N N N N N Np: Impact = Pop 0.059 0.067 0.298 0.486 0.000 0.279Robust standard errors in parentheses. ’Expand PSL’ means the candidate asked us to tell their voters they supportexpanding the school policy. Other columns are candidate survey outcomes.* p<0.10, ** p<0.05, *** p<0.01

35