Do Informed Voters Make Better Choices? ExperimentalEvidence from Urban India

Abhijit V. Banerjee, Selvan Kumar, Rohini Pande and Felix Su∗

September 11, 2011

Abstract

In the run-up to elections in a large Indian city, we provided residents in a random sampleof slums with newspapers containing report cards. The report cards presented information,obtained under India’s disclosure laws, on performance of the incumbent legislator andqualifications of the incumbent and two main challengers. Relative to the control slums,treatment slums saw higher turnout, reduced vote buying and higher voteshare for betterperforming incumbents. Voters exhibited sophistication in their use of information – theyused heuristics on public good location to evaluate public good spending by the incumbentand used challenger qualifications as a yardstick to judge incumbents. These findings sug-gest a political environment in which (atleast some) voters seek to elect the most competentcandidate but face voting costs.

∗The authors are from MIT (Banerjee), Carnegie Mellon University (Kumar) and Harvard University (Pande).We thank our partners Satark Nagrik Sangathan, Delhi NGO Network and Hindustan Times and especiallyAnjali Bharadwaj, Amrita Johri and Mrinal Pande for enabling this study, Shobhini Mukherji for providingfield oversight and Jeff McManus and Swapna Reddy for exceptional research assistance. We thank HewlettFoundation and the Empowerment Lab at CID, Harvard for financial support and Tim Besley, Pascaline Dupas,Esther Duflo, Alan Gerber, Dominic Leggett, Tom Palfrey, David Stromberg and seminar participants for helpfulcomments.

1

1 Introduction

The poor numerically dominate the electorate in many low-income democracies, yet have largelyfailed to translate their political weight into effective service delivery and other economic gains(see, for instance, Mauro (1995); UNDP (2002)). Explanations for this failure abound. Bytargeting clientelistic policies along ethnic lines, politicians may cause poor voters to value acandidate’s group identity over his other qualifications (Horowitz, 1985; Chandra, 2004; Banerjeeand Pande, 2009). Weak electoral institutions may allow the political elite to subvert democracyby stuffing ballots, buying votes and intimidating voters (Acemoglu et al., 2010; Simpser, 2008).Voters may be unable to identify politicians who would serve them well, either because theylack information (Djankov et al., 2010) or because they are unable to interpret the availableinformation.

The empirical challenge in distinguishing between these possible culprits – weak institutions,clientelistic policies and poorly informed voter populations – is that they so often coexist in low-income settings. In this paper we use data from a field experiment conducted in urban Indiato isolate and evaluate one channel of influence – information about politician performance andqualifications. We ask whether providing such information via the media influences vote-buying,voter turnout and incumbent voteshare.

Our field experiment occurred in the run-up to the 2008 state legislature elections in Delhi,India’s capital city. The last decade has witnessed the adoption of stringent disclosure laws inIndia, prominent among them being India’s Right to Information Act and mandated disclosurerequirements for all candidates for elected office. We published information obtained under thesedisclosure laws as jurisdiction-specific report cards in a leading vernacular newspaper.

Each report card contained information about incumbent performance along three dimen-sions – legislative activity, committee attendance and spending of discretionary constituencydevelopment funds across eight public good categories. It also provided information on thewealth, education and criminal record of the incumbent and the two main challengers in thatjurisdiction. In a random sample of 200 slums, households received a pamphlet on legislator re-sponsibilities and a free copy of a newspaper that featured the report card for their jurisdiction.Households in the 575 control slums did not receive any informational material.

Relative to control slums, we observe several significant changes in voter behavior in treat-ment slums. First, average voter turnout increased by 3.5 percent, or two percentage points(from 57.5% to 59.5%). Second, cash-based vote-buying was 19 percent less likely to occur intreatment polling stations. Third, while the campaign did not influence the average incumbentvote share, worse performing incumbents and those facing better qualified challengers receivedsignificantly fewer votes. The increases in turnout were relatively higher in treatment slums

2

located in jurisdictions where the incumbent was a worse performer.1

The richness of the performance and qualification data allows us to examine how votersuse available information. Voters only react to performance information along dimensions thatdirectly affect their well-being. They reward legislators with a better attendance record inoversight committees (for fair price shops and police) but do not react to their attendancerecord in the legislature. Equally, they condition their vote not on overall spending for a publicgood category but on the extent of spending in slums. To do so, it appears that they combineheuristics on the extent of slum spending within a public good category (possibly derived fromwhat they observed in their slum) with report card information on spending (by public goodcategory). A single newspaper contained report cards for two neighboring jurisdictions. Alegislator’s attendance record in committees is likely comparable across jurisdictions and, in thiscase, voters engage in yardstick competition. In contrast, the extent of slum spending withina public good category is jurisdiction-specific, and consequently we do not observe yardstickcompetition for spending behavior.

Turning to qualifications, we find that voters condition their vote on incumbent qualificationsrelative to challengers in the jurisdiction. Incumbents who are richer or less educated thantheir challengers receive fewer votes. Moreover, consistent with a model of rational learning,voters ignore information about relative qualifications of irrelevant candidates (challengers inthe neighboring jurisdiction).

A simple model of voting where (some) citizens vote not to impact the outcome but toexpress their preferences provides a parsimonious explanation for these findings. We assumeexpressive voters favor better performing politicians. In this setting, more precise informationabout incumbent performance increases the expected benefits from voting for preferred candi-date and, therefore, increases turnout and reduces vote-buying. More precise performance andqualification signals also boost the vote share of better performing incumbents. However, if thequalification signal for the incumbent is less precise than the performance signal then voters willrationally choose to ignore the former. In contrast, they always use the qualification signal forthe challengers. Finally, whether turnout increases vary with incumbent performance dependson voters priors. If, for instance, voters priors led them to favor the challenger but the perfor-mance signal raises voters (relative) valuation of the incumbent then turnout may decline if theexpected benefit from voting falls below the cost.

Our report cards used data available under existing disclosure laws which provide broadinformation about the qualifications and performance of politicians. Arguably, the global movetoward disclosure is better accomplished by broad revelations about performance than by specific

1This finding resonates with several U.S. political economy papers which find that voters are often moreenergized by worse performers Bloom and Price (1975); Washington (2006); Hastings et al. (2007).

3

disclosures of corruption charges. The results support both the optimistic view of the power ofinformation disclosures suggested by Djankov et al. (2010), based on the negative cross-countrycorrelation between disclosure laws and corruption and the importance of an independent andcredible media source in enhancing the quality of government (Besley and Prat, 2006; Djankovet al., 2003).

A growing political economy literature demonstrates a net positive impact of informationon policy outcomes in democratic setting (Besley and Burgess, 2002; Stromberg, 2004). In thecase of politicians, Snyder and Stromberg (2010) have shown that increased newspaper coverageimproves politician behavior (from a voter perspective). For voters, a first set of papers show thataccess to more informative media sources increases turnout (Stromberg, 2004; Gentzkow, 2006).Turning specifically to electoral accountability, Ferraz and Finan (2008) use detailed Brazilianelectoral and audit data to show that new information about political corruption reduces theprobability of reelection for corrupt incumbents. These effects are stronger in areas with greaterradio coverage.

These papers suggest that informing voters about politician performance may be a powerfulaccountability mechanism but leave open two questions. First, can less educated voters processbroad performance information or is their ability to hold incumbents accountable limited torelatively clearcut cases of political corruption? Second, in settings with weak institutions, doelectoral malpractices limit the power of information? In this paper we provide evidence thatthe power of information remains high even in settings with weak institutions and relatively lesseducated populations. We also extend the ambit of political outcomes studied to include turnout,incumbent vote share and vote-buying outcomes. Reflecting this, our theoretical frameworkstudies voting outcomes when turnout is endogenous – a feature which is rarely present inexisting tests of political agency models.

Finally, our paper directly contributes to a growing experimental literature on voter behaviorin low income countries (Pande, 2011). This literature builds on the insights of the Get Outthe Vote literature for the U.S. (Gerber and Green, 2000), and has found significant turnouteffects for nonpartisan motivational campaigns (Gine and ÊMansuri, 2010; Banerjee et al., 2010)and campaigns that exhort voters to use their electoral influence to protest against malpractices(Collier and Vicente, 2008). Evidence on whether information disclosures improve electoralaccountability is more limited but tends to support the view that citizens seek to base theirvoting choices on incumbent performance (Chong et al., 2010; Humphreys and Weinstein, 2010).2

2Chong et al. (2010) evaluates an information campaign in Mexico where program information was provided tovoters. Since incumbent politicians cannot stand for re-election the voters faced the more complicated metric ofusing this information to hold parties (not candidates) accountable. One interpretation of their main finding (thatvoter turnout is lower when incumbents perform worse) is that party affiliations are relatively strong in Mexico.In ongoing work, Humphreys and Weinstein (2010) examine the incentive effects of providing information.

4

The rest of this paper is structured as follows: Section 2 provides a conceptual frameworkto help interpret what we find. Section 3 describes the context, the experimental interventionand empirical design. Section 4 provides the results and Section 5 concludes.

2 Conceptual Framework

In large elections, no single voter is pivotal. Much of the literature modelling voter behaviorin large elections has, therefore, focussed on benefits to voting that relate to voters expressingown preferences. In keeping with this, we use a simple model of expressive but costly voting toexamine how the release of report cards influences a potential voter’s decisions of whether tovote and who to vote for.

2.1 Basics

In our data, the unit of observation is a neighborhood within a jurisdiction. Voters in differentneighborhoods in a single jurisdiction face the same set of candidates and elect a single legislator(via plurality rule). We, therefore, develop a two period voting model for a single jurisdictionwhich consists of a continuum of neighborhoods. In each period the incumbent legislator withcharacteristic θ produces output y using a technology y = θ for all neighborhoods in the juris-diction.3 Each neighborhood is populated by a continuum of voters with measure one. Thereare two types of voters. A fraction μ are partisan and always vote for the candidate from theirpreferred party. Among them, the partisan vote share of the incumbent is ξ. The rest areexpressive voters. An expressive voter is a voter who cares about θ and is risk neutral. Weassume expressive voters face a cost of voting c with a distribution uniform on [0, c].

In terms of information structure, the set of expressive voters observe y = y + ε, whereε ∼ N(0, σ2

ε ), where y is first period output. They also receive a signal on both incumbentand challenger characteristics: θX = θX + ηX , X = I, C where ηX ∼ N(0, σ2

ηX). ηI , ηC , ε are

distributed independent of each other. Everyone in the same neighborhood observes identicalsignals.4 There is common knowledge that candidate X ′s type (i.e θX) is drawn from a knowndistribution: θX ∼ N(θX , σ

2X). This gives voters their priors. A key feature of this signal

structure is an information asymmetry between the incumbent and the challenger – the votersobserve both a performance signal (based on period 1 performance) and a characteristic signalfor the incumbent whereas for the challenger there is only a characteristic signal.

3This is equivalent to assuming that the actual output is random and varies across neighborhoods in thejurisdiction and y is its mean.

4The idea is that citizens in the same neighborhood hear the same rumors about candidates and observe thesame projects that they initiate (or not) and are exposed to the same campaign speeches etc.

5

We are interested in the expressive voter’s decision problem at the beginning of period 2,when the incumbent is running for re-election against a challenger. An expressive voter’s period2 utility is the gain in the expected period 2 output if the candidate she believes is the highestquality wins. She will vote if the expectations about the candidates warrant paying the votingcost, i.e. if | E[θI |θI , y]− E[θC |θC ] |≥ c.

2.2 Results

We model the release of report cards as improving the precision of signals available to the voters.We first show that report card release improves the effectiveness information signal available tovoters and then examine how voter behavior is influenced.

2.2.1 Report Cards and Information Signal

It is standard that if νX is the random variable θX − θX (with mean 0 and variance σ2X for

X = I, C) then

E[θI |θI , y] = θI +hηI (ηI + νI) + hε(ε+ νI)

hI + hηI + hε

where hI , hη and hε are the precisions (i.e. inverses of the variance) of θI , η and ε respectively.Likewise

E[θC |θC ] = θC +hηC (ηC + νC)

hC + hηC

where, once again, hC and hηC are defined in a corresponding way.Finally z ≡ hηI

(ηI+νI)+hε(ε+νI)

hI+hηI+hε

− hηC(ηC+νC)

hC+hηCis the random variable defining the effective infor-

mation signal–it is the extent to which the information signals have moved the posterior of thegap between the two candidates’ types away from the prior. By our assumptions, z ∼ N(0, σ2)

whereσ2 = [

hηI + hε

hI + hηI + hε

]h−1I + [

hηC

hC + hηC

]h−1C .

We model the release of report cards as raising signal precision, such that all three ofhηI , hηC , hε increase. It is easy to see that increasing each of these increases σ2.

Result 1: Release of report cards will cause the gap between the posterior and the prior valuesof the difference in the two candidates’ types to be larger on average in the sense of increasingthe variance of effective information signal z.This result captures the idea that more information will on average make the candidates lookmore different from each other.

Given this result, we now examine the impact of report cards on three outcomes of interested– expected turnout, incidence of vote-buying and realized candidate vote-share.

6

2.2.2 Report Cards and Expected Turnout

In this economic environment, expected turnout is given by:

μ+ (1− μ)Ω

where

Ω ≡ Proby,θI ,θC[D ≤ c] · Ey,θI ,θC

[D | D ≤ c]

c+ Proby,θI ,θC

[D > c]

E[] represents the expectations operator and D ≡∣∣∣E [

θI | θI , y]− E

[θC | θC

]∣∣∣.Result 2: As long as there is not too large a bias in favor of one candidate (Δ ≡ θI − θC isnot too large) and the precision of the signals (hI , hη and hε ) are not all too close to zero, thenturnout goes up when the variance of the effective information signal goes up.

This result is intuitive. A more informative signal should typically increase expected turnoutas long as the two candidates are not too different ex ante because information, on average,will spread them apart. However if candidates are already very different (based on the priorinformation) then further information may not boost turnout very much when it further increasesthe gap between candidates (because turnout is already very high) but may reduce turnoutsubstantially when it reduces the gap between them.

Combining Result 1 and Result 2, it seems reasonable to expect turnout to, on average,increase with the information campaign.

2.2.3 Incumbent vote share

The impact of more precise information on incumbent’s realized vote share V depends on whetherthe information is good or bad news for the incumbent. Here, unlike the case of turnout, weare interested not in the average effect but in how the effect of more precise information variesbased on the realized values of y, θI , θC .

For a given value of y, θI , θC , V varies with c and is defined in the Appendix. Let κ be aparameter which shifts hε and hηX ,X = I, C. For fixed values of y, θI , θC , we evaluate dV

dκ.

Result 3: As long as μ or c are not too small, a more precise challenger’s characteristic signalimproves incumbent vote share when challenger’s characteristic is worse, while an increase inthe precision of performance and characteristic signals for the incumbent in the same proportion( 1hηI

dhηI

dκ= 1

hε

dhε

dκ) increases the incumbent’s vote share more when his measured performance

and characteristic signal are better and his opponent’s characteristic signal is worse.This result captures the fact that expressive voters wish to vote for the high quality candidate,

and an increase in the precision of signals enables this.However, the result requires two conditions. To see what can go wrong when improvement

7

in precision is not proportional, suppose that the precision of the incumbent’s performancesignal increases much faster than his characteristic signal. This reduces the weight given to hischaracteristic signal in the posterior and as a result, having a more positive characteristic signalwill be less rewarded when overall signal quality goes up. Our second condition (on μ and c)places restrictions on the extent of turnout change in response to more precise signals. Supposethe information about the incumbent is such that the posterior belief about him is that he isworse than the challenger and moreover, a more precise signal causes the posterior about hisquality to go down (in other words there is a lot of bad news about him). Now suppose thenews about his performance became slightly less negative. This would bring him closer to thechallenger in terms of quality and reduce turnout. The reduction in turnout means that thenegative effect on his vote share coming from the improvement in signal quality is stronger,because it is on a smaller base. In other words, if turnout changes are unbounded then animprovement in signal can hurt him more in situations where his performance is better.

The assumption that the precisions of output and quality signals increase in the same pro-portion is, of course, relatively strong. Arguably, the report cards were more effective in captur-ing important dimensions of performance than in providing information about characteristics.Hence, it is reasonable to assume that hε is more elastic with respect to κ than hηI :

κ

hε

dhε

dκ>

κ

hη

dhηI

dκ.

Under this assumption an increase in κ raises the slope of incumbent vote share with respect toincumbent performance by even more than in the proportional case, but the slope with respectto incumbent qualifications will go down and in extreme cases may become negative for reasonssuggested as a part of the discussion of why Result 3 might fail.Result 4: As long as μ or c are not too small, if the precision of the incumbent’s performancesignal increases proportionally faster than the precision of the characteristic signal with respectto the treatment (κ), then an improvement in signal quality will increase V for better performingincumbents by more. However in this scenario improved signal quality may not always lead to agreater increase in V for incumbent with better characteristics.

Finally, since the incumbent is likely to be better known than the challengers we anticipatehI > hC and dhηI

dκ<

dhηC

dκ, i.e. the priors about the incumbent’s qualifications are likely to be

more precise and less easy to affect. If this is the case we would expect the effect of an increasein κ on the slope with respect to θC will be stronger than the effect on the slope with respectto θI even under the proportional signals assumption made in Result 3.

It is also interesting to ask how the incumbent’s performance affects turnout. It should beevident that while the effects on incumbent vote share reported above are entirely driven by

8

changes in turnout: more precise positive information about the incumbent can either reduceturnout of those who would have voted against him or increased turnout of those who wouldhave voted in favor. Positive information (z > 0) from the point of view of the incumbent willtend to reduce turnout when the bias in the expressive voter population is against the incumbentand increase it when it is in his favor.

2.2.4 Vote buying

Finally, we consider a simple extension of the model that allows us to examine vote-buying.We assume that while parties cannot contract on how citizens vote, they can identify whichcitizens are favorably inclined towards their candidate and can encourage them to vote bypaying them (in cash or kind) to vote or by providing them transportation to the voting booth.They may also pay these citizens to give up their voter identity card to a party worker whowould use it. Specifically we assume that each party observes the realization of Δ + z in eachneighborhood (neighborhoods vary in �, say) and then decide whether or not to buy votesin that neighborhood. We assume that parties behave non-strategically, in the sense of notconditioning their vote purchase on what they expect will happen in the election as a whole.This is, arguably, a reasonable approximation if the result is sufficiently uncertain (since allneighborhoods belong to the same jurisdiction, one vote from anywhere counts as one vote andthe decision to buy votes should not depend on Δ+ z in one particular neighborhood). Buyingvotes involves paying the voter a fixed amount s conditional on voting.5 If parties buy votes,they do so in the most cost-effective manner – identify voters who are closest to being indifferentbetween voting and not voting (on the side of not voting of course) and offer to pay them ifthey vote. As a result those with c ∈ [|z +�|, |z +�|+ s] will end up selling their votes to theincumbent as long as z +� > 0 (if z +� < 0, the incumbent’s party would never make theman offer) and to the challenger if z +� < 0. If z +� = 0, the voter can be bought by eitherside but since this is zero probability event, we can ignore it. Clearly if |z +�| > c, there is noreason to buy votes in that neighborhood.

It follows that the expected fraction of vote buying in a neighborhood with bias � is

c−Δσ∫

−c−Δσ

s

cφ(z)dz =

s

c[Φ(

c−Δ

σ)− Φ(

−c−Δ

σ)]

under the assumption that there is vote buying even in the case where c lies in the interval[|z +�|, |z +�|+ s] (as long as s is small this should not make much of a difference). Clearly

5Potential voters are typically paid before they vote. This is either enforced by explicitly taking the voteridentity card or it is coordinated through locally influential people who have the power to enforce this contract.

9

this difference is decreasing in σ if c > |Δ|.Result 5: Vote buying goes down when the variance of the effective information signal goes upas long as the bias in the neighborhood is not too large.

This result relies heavily on the fact that once the signal is strong enough in either direction,no one is indifferent and hence there is no reason to buy votes. While the sharpness of thisconclusion makes use of the fact that voting cost is uniformly distributed, a natural general-ization would be to assume that the cost of voting has a unimodal distribution, possibly withan unbounded support. Raising the precision of the information would then mean that theindifferent voters will tend to be more in the tails of the distribution and hence less numerous.

2.3 Testable Predictions

To summarize, the predictions we will take to the data are:

1. Under reasonable assumptions, neighborhoods that receive report cards (and, therefore,more precise information) should witness higher turnout and lower vote-buying.

2. In neighborhoods that receive the report cards, the vote share of better performing incum-bents should go up while that of worse performing incumbents should go down. The samepattern is also expected for the effect of more precise information about the qualificationof the incumbent but those effects are likely to be weaker.

3. More precise information about the qualification of the challenger should help the incum-bent when the challenger is poorly qualified and hurt him the reverse case. The effect ofinformation about the challenger’s qualifications is likely to be stronger than that of theincumbent.

4. Whether the increase in turnout varies with performance and qualifications of candidateswill depend on voter priors. If information serves to lower the expected benefit from voting(because, for instance, voters downgraded their beliefs about a candidate’s quality) thenthe release of report cards may lower turnout.

3 Experimental Design and Data

This section describes the context of our intervention, the design of report cards and the data-sets we use.

10

3.1 Setting

Delhi is India’s national capital and second-largest metropolis. It is home to over 15 millionpeople, close to 20% of whom live in slums (2001 Indian census). Widely dispersed across Delhi,these slums generally consist of poorly built congested tenements with inadequate infrastructure.Most slum dwellers lack legal property rights and are relatively poor; they are, however, a polit-ically active group that candidates target during elections. They are also the target populationof our intervention.

Delhi has an independent state legislature composed of seventy legislators. Elections occurevery five years with each legislator elected via plurality rule from a single member jurisdiction.Each legislator represents over 100, 000 citizens. There has been a steady accretion in legislatorresponsibilities over the last two decades, as a part of an overall push towards decentralizationand devolution of powers away from the bureaucracy. Voters have limited contact with theirlegislators and know relatively little about what they should expect from them. A large fractionof slum residents do not regularly read newspapers, which remain the main source of relativelyunbiased information about politics and politicians.6

Our field experiment was timed to coincide with the official two-week campaign period (end-ing 48 hours before polling started) for the November 2008 Delhi state election. Three majorparties were contesting: the incumbent party, Congress, Bhartiya Janata Party (BJP) and Bahu-jan Samaj Party (BSP). While all three made issues relating to the urban poor central to theircampaigns, each took a different angle: Congress campaigned on a platform of local develop-ment, and emphasized slum regularization undertaken since 2007, BJP emphasized controllingprice rise and combatting terrorism, and BSP targetted lower-caste poor households.7

During the two week campaign period, party campaigning in slums was extensive and in-cluded rallies and door-to-door canvassing. Such campaigning is often accompanied by gift-giving in the form of liquor, food, milk and clothing.8 These gifts are widely considered asinducements to vote for a specific party but the diffuse nature of gift-giving and unclear mecha-nism for enforcement suggests that their impact may be limited. Parties also often make directcash payment to subsets of voters. In many cases, the party worker explicitly exchanged cashfor the voter’s election identity card (to be subsequently used by a party worker who would

6In a household survey among slum dwellers in our sample 40% of the men and 66% of the women statedthat they do not read newspapers.

7The Congress government initiated slum regularization in 2007, whereby slum dwellers could purchase prop-erty rights to their dwellings. It also included a government drive to provide basic infrastructure in slums. Incontrast, BJP’s main campaign slogan was Mehengi Padi Congress (“Congress is Expensive”); during 2008, Delhisaw sharp price inflation of food, fuel, and other consumer goods. Coincidentally, Delhi elections occurred threedays after the 26/11 Mumbai terrorist attacks, which many predicted would bolster the BJP in elections. Thiselection marked BSP’s first entry in Delhi elections.

8The Delhi excise department registered over 1,500 bootlegging cases during the election month (IANS, 2008).

11

claim to be the voter). According to most reports, maximum vote-buying occurs during the 24hours prior to the election.

3.2 Report Cards

The central plank of our intervention was a door-to-door distribution of newspapers in a randomsample of slums. Each newspaper contained report cards on legislator characteristics and per-formance for two jurisdictions (typically, geographic neighbors). Below, we describe the sourceand nature of these data and Table 1 reports summary statistics.

A. Public Disclosure Laws

In October 2005 the Indian Right to Information (RTI) Act was implemented, giving citizensaccess to non-classified government records. Under the Act, a citizen may request informationfrom a public authority and be legally entitled to an expeditious reply (typically within 30 days).Roughly a million RTI petitions have been filed annually since 2005 (PricewaterhouseCoopers,2009). Our partner NGO filed over 70 RTIs in 2008, through which it obtained information onlegislator responsibilities and incumbent performance along several dimensions. This paper isone of the first evaluations that measures the impact of any form of RTI activism.9

Our second source of disclosures was furnished by a 2003 Supreme Court ruling that requirescandidates contesting national and state elections to submit affidavits at the time of filing theirnomination revealing their educational qualifications, assets and liabilities and any past criminalcharges.10 We use this affidavit information to identify qualifications of candidates of the threemajor parties.

B. Performance and Qualification measures

Performance Using data from the official response to RTI petitions we constructed report cardstatistics on three dimensions of legislator performance – activism in the legislature, attendanceof oversight committees and spending of discretionary funds.

On legislative activism, we reported legislator attendance and number of questions asked inthe legislature during 2007. Table 1 shows that the legislature met infrequently but attendancerates were high. The average legislator attended 16.9 out of 18 sessions but asked no questions.

On participation in oversight committees, we reported legislator attendance in the mostrecent meeting for each of three oversight committees: the Ration Vigilance Committee, the

9To the best of our knowledge, the only other experimental evaluation of RTI activism is Peisakhin and Pinto10This judgment was implemented by the Indian Election Commission which made filing an affidavit disclosing

this information a precondition for appearing in the ballot.

12

Police Vigilance Committee, and the District Development Committee. The first of these istasked with ensuring that local ration shops are effective in providing subsidized food to below-poverty-line residents, the second, with ensuring that police stations operate well and the policedo not harass locals. Finally, the District Development Committee provides oversight of devel-opment projects. Across Delhi 70% of legislators attended the most recent Ration Committeemeeting, 46% attended the Police Vigilance committee meeting and only 29% attended the Dis-trict Development Committee meetings. No legislators in our sample attended the DevelopmentCommittee Meeting.

Finally, each legislator receives 20 million Rupees a year (roughly $ 45,000) known as theMLA Local Area Development Scheme Fund (MLALADS) to spend on development in his juris-diction, along with five million Rupees annually to be spent exclusively on water development.The legislator is responsible for fund allocation, with implementation undertaken by the involvedmunicipal corporation.11

We reported MLALADS spending between 2004 to 2007 under the following categories;roads (including sidewalks), water (referring to water supply infrastructure such as borewells,pumps, and tanks), parks and statues, sewage (sewage pipes and public toilets), drains, lights,community halls, and boundary walls and others. (Unspent money can be rolled over into thenext year, but is forfeited at the end of the legislative term). The average legislator spent morethan 80% of the funds available to him. In contrast to MLALADS spending, all legislators spentthe full water fund.

Qualifications Our report cards also made use of affidavits filed by candidates to present in-formation on assets, criminal charges and education qualifications for the three major partycandidates in each jurisdiction, always including the incumbent.

First, we reported whether the candidate had any pending criminal charges. Overall, 91candidates had pending criminal charges. These candidates featured prominently on the rollsof the major parties (a quarter of the major party candidates faced criminal charges). Half theincumbents in our sample faced criminal charges.

Second, we reported the value of assets owned by the candidate and his/her spouse. In ouranalysis we use a widely used summary measure of wealth – being a crorepati, i.e. have assetsin excess of Rs. 10 million. In the 2008 election close to 20% of the candidates (153 candidates)were crorepatis. 12

Finally, we reported candidate education. Overall, only 3% (18 candidates) were illiterate.11After deciding on a particular development project, the legislator allocates funds to the municipal corpora-

tion, which carries out the work.12Delhi Election Watch, a consortium of NGOs that independently monitors elections, analyzed the change

in personal assets of the 45 incumbents who were recontesting. The average increase in assets per MLA over asingle five-year term was 211%, amounting to an average of almost 1.8 crore.

13

18% had up to 10 years of schooling, and 19% had up to twelve years of schooling. 19% held acollege degree and 15% a post graduate or professional degree.

3.3 Experimental Design

A. Sample

To select our sample we first identified jurisdictions where the incumbent was likely to stand forre-election. Then with our NGO partner we identified a subset of ten jurisdictions with highslum density and where the NGO had sufficient manpower to organize the doorstep newspaperdelivery.

For each jurisdiction the sample frame consisted of all polling stations that served slum areas.A polling station serves roughly 400 households (1,000 adult voters) who live in the same oradjacent neighborhoods.13 In each jurisdiction, we randomly selected twenty polling stationsfor treatment. This yielded a sample of 200 treatment and 575 control polling stations withstratification at the jurisdiction-level.

B. Intervention

In each treatment polling station our intervention targeted all households with at least one adultvoter who featured on the voter list.

Roughly three days before the newspaper distribution, two NGO workers visited each house-hold in a treatment polling station and gave the respondent a pamphlet. This pamphlet informedthe household that they will get a free newspaper with a report card on politician performanceon a specific date and primed the voter about the relevance of the report cards in several ways(see Figure 1).

First, it encouraged citizens to obtain information about politicians by improving their knowl-edge of the voting process. To do so, the pamphlet laid out the actual mechanics of voting, suchas how to determine eligibility to vote, accepted forms of identity proof, and complaint proce-dures. Second, it motivated citizens about the importance of performance by listing legislatorresponsibilities. Voters were informed about politician disclosure laws and encouraged to readour partner newspaper to learn about candidates’ backgrounds. No candidate-specific informa-tion was provided in the pamphlet.

On average, a two-member NGO team covered the households associated with a pollingstation in one and half days. Monitoring reports show that, on average, two-thirds of thehouseholds in a polling station were reached with the NGO spending 15 minutes per household.

13Households are assigned to polling stations on the basis of a door-to-door survey conducted by the IndianElection commission.

14

Roughly ten days before the election (between November 20 and 25, 2008) our partner news-paper Dainik Hindustan published report cards (An example is provided in Figure 2). On themorning of publication, NGO workers placed a free newspaper on the doorstep of each house-hold in the treatment slums. Four hundred newspapers were disseminated per slum, yielding adelivery of 80, 000 newspapers. After newspaper distribution, independently hired monitors vis-ited 20 households in 172 of the 200 treatment polling stations to check for newspaper presence.Newspapers were observed in 80% of households.

Finally, within 48 hours of newspaper distribution NGO workers organized a public readingof newspaper report cards. Monitor reports (for 130 meetings) reveal that the average readinglasted 1.5 hours but attendance was low. (Average attendance was 20 women and 14 men, outof a target population of 1000 adults). In 90% of the cases attendees audience had copies of thereport card during the campaign and 95% of the monitoring reports state that the readout wasnon-partisan.

In augmenting newspaper delivery with a pamphlet campaign and public readings, we soughtto maximize the likelihood that slum-dwellers received the report card information. However,the use of a multi-pronged strategy may raise the concern that, independent of newspaperdistribution, the door-to-door campaign or citizen meeting may have influenced voter behavior.As we describe in detail below, several pieces of evidence suggests that the report cards directlyinfluenced voter behavior. First, while the pamphlet provided no incumbent-specific informationwe show that voter behavior was sensitive to incumbent performance. This suggests that theimpacts relate to actual information. Moreover, the magnitude of the impacts we observedsuggests an influence on the voting behavior of a significantly larger population than thoseattending the public readings. Finally, we have checked that the treatment effects do not varywith the quality of NGO staff conducting the public reading.(Results available from the authors).

3.4 Data

Our empirical analysis utilizes several datasets. The first is official polling-station electoralreturns. Here, the two outcomes of interest are voter turnout and incumbent votes (as a fractionof total votes cast). Average voter turnout in the control polling stations was 57%, and theaverage incumbent vote share was 46%. Nine of the ten incumbents were from the ruling party(Congress). 90% of the incumbents won. Some victories were narrow – the margin of victoryvaried from 0.53% to 30%.

Our second dataset is an observational survey: in 29 treatment and 32 control polling sta-tions, a surveyor spent approximately four hours on the eve of the election noting any visibleevidence of party campaigning and whether this was accompanied by the distribution of liquor,

15

food, clothes or milk/refreshments as enticement. The observer also noted whether any cashbased vote buying occurred. Over 95% of the polling stations witnessed door-to-door campaign-ing, and public rallies occurred in over 70%. Gift-giving was prevalent in roughly 80% of theslums.

Finally, we use data from a household survey that was conducted in the six-day intervalbetween election day and when results were announced. The survey was conducted in the 200treatment polling stations and a randomly selected 200 control polling station localities. In eachpolling station ten randomly selected individuals were administered a brief pop-quiz on politicianperformance and perceptions of politician spending behavior. The survey was administered byan independently hired survey company that had no links with the intervention.

We use the report card data to ask whether voter responsiveness varied with incumbent per-formance and qualifications. The report card provided category-wise information on spending,but it is likely that categories differ in their relevance for slum dwellers. For example, roadspending may be less useful for slum dwellers with unpaved roads. Therefore, after the electionswe provided our NGO partners with a list of all projects in their jurisdiction that had beenallocated funding by the incumbent. The NGO dispatched fieldworkers to visit the site of eachspending item and identified whether spending occurred in a slum. This allows us to identifydevelopment spending in slums (overall and category-wise). Figure 3a shows the average distri-bution of slum and non-slum spending for each public good category. In Figure 3b we show thisbreakdown for roads for the ten jurisdictions – we observe significant cross-jurisdiction variationin slum spending per category.

In Table 2 we report a randomization check. Panel A uses electoral roll data and Panel Bhousehold survey data. The average polling station had a thousand electors, and Panel B showsthat these electors are relatively poor – the average per capita household income is a dollar aday. In column (3) we observe balance on electoral roll covariates across treatment and controlpolling stations for the electoral data sample (775 polling stations). In column (6) we observesimilar balance for electoral and survey data for the household sample (3,896 respondents across388 jurisdictions). In column (9) we consider the smaller sample of 61 polling stations for whichwe have observational data. Once again, the joint F-test suggests overall balance with a p-valueof 0.25.14

14While we focus on the pure experimental estimates, we have checked that our results are robust to controllingfor average household covariates.

16

4 Did Information Influence Voter Behavior?

We first report the average impacts of performance and qualification information on voter knowl-edge and electoral outcomes. We then show that, consistent with a model in which voters careabout politician quality, changes in voter behavior are correlated with incumbent performanceand qualifications. Voter sophistication in the use of information provides further evidence thatreport card statistics (and not just priming) influenced voter behavior.

4.1 Average Impacts

To examine the average treatment impacts we estimate regressions of the form

Ysj = αj + βTsj + εsj (1)

Tsj is a dummy indicating whether polling station s in jurisdiction j was assigned to treat-ment, and β is the unbiased ITT effect. We include jurisdiction fixed effects αj to account forstratification.

Table 3 reports the results. In columns (1)-(5) we start by using the survey data to examinewhether political knowledge among slum-dwellers varied by treatment status of their slum. Inthis case, as the unit of observation is an individual, we cluster standard errors by polling-stationand allow treatment impacts to vary by education. We define a respondent as educated if shehas at least 5 years of formal education.15

In columns (1)-(3) we consider respondent knowledge of legislator responsibilities, spendingperformance and candidate qualifications. We asked each respondent a series of eleven factualquestions and code the response for each question as correct or incorrect. 16 In column (1)we consider the total score across the eleven questions. Respondent knowledge levels are lowand the average respondent has a score of 2.7 out of 11. We see that the campaign improvedknowledge levels by a very significant 10% among educated respondents. In columns (2) and(3) we divide the questions into information on legislator responsibilities that the respondentsreceived during the pamphlet phase and incumbent and challenger specific information provided

15In our sample, roughly a third of the respondents state that they are illiterate or have basic literacy, slightlyover a quarter state that they have less than 5 years of education and the remaining 37% have 5 years or moreof education. We observe qualitatively identical but slightly noisier estimates if we include more educationcategories.

16We code the answer to how much money was spent by the legislator as correct if the answer was within1 standard deviation of actual spending. Similarly, we code the answer to how much did you legislator spendrelative to the average legislator as correct if the answer is "more" and the legislator spent more than average, ifanswer is "less" and the legislator spent less than average, or if answer is "same" and legislator the spent within1 standard deviation of average.

17

by the report cards. For both, the campaign significantly improved information levels amongeducated respondents.

While the gains in political knowledge are significant, they are relatively small in size andlimited to educated respondents. One possible reason is that remembering precise numbers (forsay spending) is hard and the ability to do so is correlated with being able to read newspapers.Arguably, in terms of voting behavior what matters is changes in respondent perceptions oflegislator performance. The survey asked the respondent to rank the amount of work doneby the incumbent across multiple categories. We consider responses regarding the two largestspending categories which featured in the report card and for which we observe spending inevery jurisdiction (and know the extent of slum spending).17 For both categories we create anindicator variable for whether the respondent believed that the incumbent did a lot of work.

32% of the respondents state that the incumbent had done a lot of work on roads and 12%believe this to be the case for drains. In column (4) we see that, conditional on treatment, theperception that the incumbent did a lot of work is increasing in the fraction of road spendingthat occurred in slums. In column (5) we observe the same pattern for drains. These resultsare striking because the report cards provided category-specific spending but not broken downby whether it occurred in slums. Thus, voters are able to use other available information tocorrectly evaluate the incidence of slum spending within a category. Here, we do not observe anydifference in responsiveness by education status suggesting that information spread to illiterateresidents as well.

Given these results on voter knowledge, we now turn to the evidence on average campaignimpacts. Here, we start by considering electoral outcomes - turnout and incumbent vote share.To flexibly estimate turnout effects we consider the log number of voters as the outcome variableand include log number of registered voters as a control variable.18 In column (6) we see thattreatment slums saw a 3.6% increase in turnout. Next, in column (7) we consider incumbentvoteshare and observe no change in incumbent voteshare in treatment slums (relative to controlslums).

Finally, in columns (8)-(11) we consider three outcomes from our observational sample for62 polling stations (collected the night before elections). Observers report a very high incidenceof party campaigning – Door-to-door campaigning by political parties was observed in 96% ofthe polling stations and a public meeting or rally was ongoing in over 70%. In columns (8) and

17A printing error in the questionairre meant that the third such category of parks mistakenly featured twice,once in combination with community hall and once with lights. Thus, we are unable to identify perceptionson spending on parks (or the other two categories). The other categories of sewage, schools, and crime do notcorrespond to categories on the report card and cannot be matched. The last category we can match to aggregatespending is water. However, as we lack slum spending measures for this category we do not use it.

18This allows the turnout response to vary as a function of number registered voters. Our results are verysimilar if we instead use fraction of registered voters who voted as the dependent variable.

18

(9) we observe no difference in party campaigning - via rallies or door to door campaigning -across treatment and control slums. In column (10) we examine the impact on gift-giving duringcampaigning by the three main parties. Gift-giving included distribution of liquor, milk, clothesand food and occurred (as part of the party campaign) in over 30% of the slums. However,we observe no differences across treatment and control slums. In contrast, we observe a 19percentage point decline in cash vote-buying in treatment slums relative to the control slums(column 11). Reflecting the illegal (and targeted) nature of such transactions, observers typicallyreported these cash based transactions as occurring between specific households and undercoverparty workers.

These turnout and vote-buying results are consistent with the idea that in low-informationenvironments the introduction of performance information increases the expected benefit ofvoting for some subset of voters. These voters shift from not voting to voting. Since the groupof (close to) indifferent voters is now smaller vote-buying via cash goes down. In contrast, theless targetted (and cheaper) forms of gift-giving are unaffected by the campaign. Under thisview, the lack of an average impact on incumbent vote share may reflect the fact that treatmentimpacts vary with incumbent performance.

However, it is also possible that the campaign enthused citizens about voting and this di-rectly translated into a higher turnout and reduced vote-buying. While we cannot rule outthat the campaign may have had such direct effects we can provide evidence on the value ofinformation provision by examining whether treatment effects vary by politician performanceand qualifications.

4.2 Do Incumbent Performance and Relative Qualifications Matter?

We start by examining whether treatment effects varied with incumbent performance and qual-ifications and then examine voter sophistication in the use of information.

4.2.1 Basic Results

A report card provided information on several dimensions of performance for the incumbent andon three qualification measures for the incumbent and two challengers. As a summary statisticof incumbent quality, we use the highest eigenvalue obtained from a principal component anal-ysis on the entire set of performance and qualification measures provided in the report card.The performance measures include legislature attendance and questions asked, committee at-tendance, total spending and fraction spent in slums (we normalize the spending measures). Forqualifications, we use three dummy variables – whether, relative to the challengers, the incum-bent was the most educated, owned the least assets and had the fewest criminal charges. We also

19

construct separate summary measures for incumbent performance and candidate qualifications.For performance we use the variables listed above. For qualifications, we separately constructsummary measures for the incumbent and challengers using three indicator variables – whetherthe candidate has less than one crore in assets, no criminal charges and a college education.We use the average for the two challengers as our measure of average challenger qualificationsin the jurisdiction. In the qualification regressions we include both incumbent and challengerqualifications (interacted with treatment) as explanatory variables.

The basic form of our estimating equation is:

Ysj = αj + βTsj + γTsj ×Xj + εsj (2)

where Xj is the relevant summary statistic on performance and/or qualifications. As before,we include jurisdiction fixed effects. While treatment assignment coincides with the unit ofobservation (polling-station), incumbent quality varies at a more aggregate (jurisdiction) level.However, the small number of jurisdictions makes clustering standard errors by jurisdictioninappropriate. We instead report the results from a randomization inference which tests thesharp null of no treatment effect (details of the procedure are in the Appendix).

Columns (1)-(3) of Table 4 considers turnout as the outcome of interest. In column (1) wesee that turnout was lower in treatment slums (relative to control slums in the same jurisdiction)when the incumbent performs better. We calibrate the estimated effects for two values of thequality index. For the median legislator in our sample (PCA value of −0.31) the turnout intreatment slums was 5.3% higher relative to control slums and there was no impact on voteshare. For the best performing legislator in our sample (PCA value of 3.62) the turnout was3.8% lower in treatment slums than control slums. In terms of our model these turnout resultsare consistent with the set of expressive voters starting with a bias in favor of the challenger. Incolumn (2) we consider the performance PCA and observe a very similar effect. In column (3)we include the incumbent qualification measure and the average qualification measure for thechallengers. Qualification information has no independent effect on turnout.

Next we consider incumbent vote share. Relative to control slums, incumbent vote share intreatment slums is increasing in incumbent performance and qualifications (column 4). For themedian legislator in the sample treatment had no impact on vote share. For the best perform-ing legislator the incumbent vote share was 6.9% higher in treatment slums relative to control.Better performing incumbents receive more votes in treatment slums (column 5). In contrast,in the case of qualifications the average challenger qualifications but not the incumbent’s ownqualifications influences incumbent voteshare. The latter finding is consistent with the ideathat incumbent performance is more informative than qualifications and hence voters ratio-nally ignore the latter. Finally, as a robustness check we also included the average challenger

20

qualifications for the neighboring jurisdiction (the report card for which featured in the samenewspaper) interacted with treatment as an additional explanatory variable. Reassuringly, wefind that voters ignore information on irrelevant challengers, i.e. those who contest elections inthe neighboring jurisdiction.

Finally, we consider vote-buying as the outcome. In columns (7)-(9) the point estimatessuggest that the decline in cash bribes was concentrated in jurisdictions where the incumbentperformed better or the challengers were relatively better qualified. However, in no case arethese impacts significant. This potentially reflects the small sample size.

We now discuss a couple of identification concerns. Incumbents and challengers were notrandomly assigned to jurisdictions and it is likely that (for instance) jurisdictions with moreeducated voters also elected better incumbents. This difference in jurisdiction and politiciancharacteristics per se is not a problem for our analysis. Since our regressions include jurisdictionfixed effects we always identify the impact from a comparison of treatment and control pollingstations within a jurisdiction.

To see our identification strategy more clearly, we also estimate jurisdiction-wise regressionsof the form given by equation (1) for turnout and incumbent voteshare and present the treatmentestimates visually (our randomization of polling-stations into treatment or control was stratifiedby jurisdiction). Figures 4a and 4b present these for two performance measures PerfPCA andslum spending (a lower score reflects better performance). In both figures we observe a similarpattern. In treatment slums, relative to control slums, turnout is lower and incumbent voteshare is higher when an incumbent performs well. The converse is true if the incumbent was apoor performer.

A related but distinct concern is that jurisdiction-specific characteristics may directly influ-ence the responsiveness of voters to information. For instance, if more educated voters bothelect better incumbents and respond to receiving newspapers by becoming more pro-incumbentthen we may observe the effects seen in Figures 4a and 4b even when voters do not care aboutperformance per se. To address this concern in Appendix Table 1 we report regressions wherewe include additional jurisdiction-level covariates (by treatment). We first consider two mea-sures of political preference in the jurisdiction – whether the incumbent belonged to the rulingparty (Congress) and turnout in the jurisdiction in the previous election. In columns (1) and(3) we see that the treatment effects are robust to controlling for these two variables. Next, weexamine whether our impacts are robust to interacting treatment with demographic variables.Unfortunately, census data is not available at the political jurisdiction level. We, therefore, usevariables constructed from our post-poll data. We have roughly 400 respondents per jurisdiction.In columns (2) and (4) we see that our basic results remain robust to adding these controls

21

4.2.2 Which Aspects of Performance and Qualification Matter?

Our simple model of voter behavior suggested that faced by multiple sources of informationvoters will discount relatively noisy and uninformative signals on incumbent quality. Motivatedby this observation, we now examine the impact of different aspects of performance and qualifi-cation information. In doing so, we are also able to provide evidence of significant sophisticationin how voters process information.

Here, we restrict attention to incumbent vote share as the outcome of interest. This is bothfor expositional ease and because the impact of information on turnout depends on voters priorbeliefs (on which, unfortunately, we lack data.

For each performance measure we estimate two specifications. First, we estimate basicregressions of the form (2). Second, we exploit the fact that a newspaper contained report-cardson two neighboring jurisdictions to ask whether voters engage in yardstick competition, i.e. useinformation on the performance of the incumbent in the neighboring jurisdiction to benchmarktheir incumbent’s performance. In this case our estimating equation is of the form

Ysj = αj + β1Tsj + β2Pj × Tsj + β3Pk × Tsj + εsj (3)

where Pk is the performance measure for the incumbent in the neighboring jurisdiction.The results are in Table 5. Columns (1)-(2) show that voting outcomes are unaffected

by information about an incumbent’s attendance in the legislature and his record of askingquestions in the legislative assembly. This is consistent with our survey data that suggests thatslum dwellers see the main responsibilities of their legislator as relating to local developmentand grievance redressal, not the enacting of bills.

Columns (3)-(4) consider committee attendance. The report cards provided informationon whether last meeting of the committee was held according to schedule and whether theincumbent attended the meeting. We construct an aggregate committee attendance index basedon attendance record in the Ration Committee and Police Committee meetings. Going fromattending neither committee to attending both increases the incumbent’s vote share by over 7percentage points (column 3). We also find evidence of yardstick competition (column 4). Thevote share of the incumbent is lower when the neighbor’s committee attendance is higher.

In columns (5)-(9) we examine whether incumbent vote share is sensitive to the extent of dis-cretionary fund spending and the fraction of spending in slums. A failure to spend all availablefunds may measure an unwillingness on the part of the incumbent to exert effort. Alternatively,the widespread belief that discretionary spending is subject to significant corruption may leadrespondents to associate higher spending with greater corruption. Possibly reflecting this ambi-guity, we fail to observe voter responsiveness to total spending by the incumbent. In contrast,

22

incumbent vote share is increasing in the extent of slum spending. However, in column (6) weobserve no evidence of yardstick competition. This is consistent with the observation that votersuse jurisdiction-specific information about how spending in a public good category translatesinto slum spending in evaluating the incumbent. To the extent that such jurisdiction-specificinformation relies on personal exposure to slum spending we would not expect voters to be ableto evaluate spending in other jurisdictions.

In columns (7)-(9) we separately consider the three largest spending categories – roads, parksand drains. These are also the three categories for which we see spending in every jurisdiction. Inevery case, voters respond not to the overall level of spending but only to slum-specific spendingin the category. Incumbent vote share is increasing in slum spending in each category. This isfurther evidence that voters use information about slum spending within public good categories.

The spending results suggest significant sophistication on the part of voters. They have a(correct) heuristic about the fraction of spending within each category that is in slums and usethis to evaluate the relevance of total spending by the incumbent in a category. The most likelyexplanation for how voters form the heuristic is by observing spending in that category in ownor nearby slum. The report card information then helps them translate this information into anestimate of how much of the spending within a category was relevant for them (i.e. occurred inslums). Voters rewarded incumbents who spent more in slums.

Next we consider the three qualification measures. In the case of assets, we see in column(10) that the incumbent gets fewer votes if a higher fraction of challengers are not crorepatis(i.e. have less than Rs. 10 million in assets). Turning to education, we again see that theincumbent does worse if a higher fraction of challengers have a college education (column 11).In contrast, voters appear to not pay heed to challengers’ criminal charges. Finally, in no casedo voters place weight on the incumbent’s own qualifications. This is consistent with the viewthat performance indicators provide much more precise information about incumbent qualitythan does qualification data.

5 Conclusion

The idea that voters in an otherwise well-functioning democracy might be severely constrained byinformation about the candidates’ qualifications and past record is both striking and important.We see that voters when given the information move quite substantially and if this informationhad reached the entire jurisdiction, outcomes may have been quite different. We also see evidencethat voters are somewhat sophisticated in how they use the information, allaying fears thatinformation would simply confuse them.

In particular, voters appear to have good knowledge of public goods available in their neigh-

23

borhood. What they lack is knowledge about cost of public goods. (Olken (2009) finds evidenceof similar information gaps in the context of local roads in Indonesia). The last decade has seena very significant rise in the the extent of constituency development funds available to legisla-tors, even though it has been associated with significant malfeasance (Tshangana, 2010).19 Ourresults suggest that improving knowledge about how these funds are spent may significantlyimprove voters ability to hold legislators accountable for how they are spent.

References

Acemoglu, D., J. Robinson, and R. Santos-Villagran (2010). The monopoly of violence: Evidencefrom colombia.

Banerjee, A., D. Green, J. Green, and R. Pande (2010). Can voters be primed to choose betterlegislators? Experimental evidence from rural India.

Banerjee, A. V. and R. Pande (2009). Parochial Politics Ethnic Preferences and PoliticianCorruption. mimeo, Harvard.

Besley, T. and R. Burgess (2002, November). The political economy of government responsive-ness: Theory and evidence from india*. Quarterly Journal of Economics 117 (4), 1415–1451.

Besley, T. and A. Prat (2006). Handcuffs for the grabbing hand? media capture and governmentaccountability. American Economic Review 96 (3), 720–736.

Bloom, H. and D. Price (1975). Voter response to short run economic conditions: The asym-metric effect of prosperity and recession. American Political Science Review 69, 1254–60.

Chandra, K. (2004). Why Ethnic Parties Succeed: Patronage and Ethnic Headcounts in India.Cambridge: Cambridge University Press.

Chong, A., A. L.DeLao, D. Karlan, and L. Wantchekeon (2010). Information dissemination andlocal governments’ electoral returns, evidence from a field experiment in mexico.

Collier, P. and P. Vicente (2008). Votes and violence: Evidence from a field experiment inNigeria. Center for The Study of African Economies Working Paper Series.19India was one of the first countries to introduce CDFs in 1993; since then, the following countries have

adopted CDFs: Southern Sudan, Philippines, Honduras, Nepal, Pakistan, Jamaica, Solomon Islands, Tanzania,Malawi, Namibia, Zambia, Uganda, Ghana, and Malaysia. The last six countries distribute a fixed amount offunds annually per constituency/legislator, while Kenya and Tanzania partially adjust the formula according toto population, poverty levels, and geographical size of each constituency (Tshangana, 2010).

24

Djankov, S., R. LaPorta, F. L. de Silanes, and A. Shleifer (2010). Disclosure by politicians.American Economic Journal: Applied Economics, forthcoming.

Djankov, S., C. McLiesh, T. Nenova, and A. Shlefier (2003). Who owns the media? Journal ofLaw and Economics 46 (2), 341–381.

Ferraz, C. and F. Finan (2008). Exposing corrupt politicians: The effects of brazil’s publiclyreleased audits on electoral outcomes. Quarterly Journal of Economics 123 (2), 703–745.

Gentzkow, M. (2006). Television and voter turnout. The Quarterly Journal of Economics 121 (3),931–972.

Gerber, A. and D. Green (2000). The effects of canvassing, telephone calls, and direct mail onvoter turnout: A field experiment. The American Political Science Review 94 (3), 653–663.

Gine, X. and G. ÊMansuri (2010). Together we will:ÊEvidence fromÊa field experiment onfemale voter turnout in Êpakistan.

Hastings, J., T. Kane, D. Staiger, and J. Weinstein (2007). The effect of randomized schooladmissions on voter participation. Journal of Public Economics.

Horowitz, D. L. (1985). Ethnic Groups in Conflict. Berkeley: University of California Press.

Humphreys, M. and J. Weinstein (2010). Policing politicians: Citizen empowerment and politicalaccountability in uganda.

IANS (2008, November 25). Indo-Asian News Service. Bootlegging in Delhi up 400 pc duringelection week.

Mauro, P. (1995). Corruption and growth. Quarterly Journal of Economics 110 (3), 681–712.

Pande, R. (2011). Can informed voters enforce better governance? experiments in low incomedemocracies. Annual Review of Economics 3, forthcoming.

PricewaterhouseCoopers (2009). Final report: Understanding the “key issues and constraints”in implementing the RTI act. http://rti.gov.in/rticorner/studybypwc/index-study.htm.

Simpser, A. (2008). Cheating big: On the logic of electoral corruption in developing countries.

Snyder, J. and D. Stromberg (2010). Press coverage and electoral accountability. Journal ofPolitical Economy 118 (1), 355–408.

25

Stromberg, D. (2004). Radio’s impact on public spending. The Quarterly Journal of Eco-nomics 119 (1), 189–221.

Tshangana, A. H. (2010, January 18). Consituency Development Funds: Scoping Paper.

UNDP (2002). Human development report 2002: Deepening democracy in a fragmented world.http://hdr.undp.org.

Washington, E. (2006). How Black Candidates Affect Voter Turnout. Quarterly Journal ofEconomics 121 (3), 973–998.

6 Theory Appendix

Proof of Result 2 The expression for turnout (Ω) can be rewritten as

Ω = Proby,˜θI ,˜θC [Δ + z ≤ c]Ey,˜θI ,˜θC

[Δ + z | Δ+ z ≤ c]

c+ Proby,˜θI ,˜θC [Δ + z ≥ c]

where Δ ≡ θI − θC . Using the notation developed above we can write expected turnout as

μ+ (1− μ){1cP rob[c−Δ ≥ z ≥ −Δ]E[z +Δ|c−Δ ≥ z ≥ −Δ]−

1

cP rob[−c−Δ ≤ z ≤ −Δ]E[z+Δ| − c−Δ ≤ z ≤ −Δ]+Prob[z ≥ c−�] +Prob[z ≤ −c−Δ]}

Using the expression for the expectation of a truncated normal variable this can be rewritten as

μ+ (1− μ){(Φ( c−Δ

σ)(Δ− c

c) + Φ(

−c−Δ

σ)(Δ + c

c)− Δ

c2Φ(−Δ

σ)}+

σ(1− μ)

c{2φ(−Δ

σ)− φ(

c−Δ

σ)− φ(

−c−Δ

σ)}+ (1− μ).

where Φ and φ are, respectively, the distribution function and density function corresponding tothe standard normal distribution. Taking derivatives with respect to σ and using the fact thatfor the standard normal distribution,

φ′(x) + xφ(x) = 0,

26

we get the radically simplified expression

1− μ

c[2φ(−Δ

σ)− φ(

c−Δ

σ)− φ(

−c−Δ

σ)]

for the derivative of turnout with respect to the variance of the effective information signal, σ2.

For Δ = 0 this expression is clearly positive since φ(0) > φ( cσ) and φ(0) > φ(−c

σ). Turnout

goes up on average when the variance of the effective signal goes up. More generally note that

−Δ

σ=

1

2[c−Δ

σ+

−c−Δ

σ]

and hence as long as φis concave over the range from −c−Δσ

to c−Δσ

, 2φ(−Δσ)−φ( c−Δ

σ)−φ(−c−Δ

σ) >

0. However we know that φ eventually becomes convex and therefore this property does notnecessarily hold globally. The point of inflection for the standard normal distribution is 1, so asufficient condition for turnout to go up with σ is that

c−Δ

σ< 1

or, symmetrically,−c−Δ

σ> −1.

Two things have to be true to make this condition hold: Δ should not be too large and σ shouldnot be too small.

Proof for Result 3: Let D = |E[θI |θI , y]−E[θC |θC ]| = |Δ+z|. For a given value of y, θI , θC ,V

1. V = μξ

μ+(1−μ)Dc

if − c ≤ E[θI |θI , y]− E[θC |θC ] < 0

2. V = μξif − c > E[θI |θI , y]− E[θC |θC ]

Let κ be a parameter which shifts hε and hηX ,X = I, C. For fixed values of y, θI , θC .Takingderivatives of V with respect to κ gives us that

dV

dκ=

1

c

μ(1− ξ)

[μ+ (1−μ)c

D]2

d(Δ + z)

dκ(4)

when ¯c ≥Δ+ z ≥ 0,dV

dκ=

1

c

μξ

[μ+ (1−μ)c

D]2

d(Δ + z)

dκ(5)

27

whenc ≤ Δ+ z < 0

anddV

dκ= 0

otherwise.We are interested in how dV

dκchanges when y, θI go up or when θC goes down. We can ignore

the cases when c < |Δ+ z| because then V does not depend on κ. Assume that c > |Δ+ z|.Therelevant expressions for dV

dκfor this scenario are given by equations 4 and 5.

Let us start with the d(Δ+z)dκ

term. It is easy to check that the expression for d(Δ+z)dκ

can berewritten as

[(hI

dhηI

dκ+ hε

dhηI

dκ− hηI

dhε

dκ)θI + (hI

dhε

dκ+ hηI

dhεdκ

− hεdhηI

dκ)y − hI(

dhε

dκ+

dhηI

dκ)θI

(hI + hηI + hε)2−hC(θC − θC)

dhηC

dκ

(hC + hηC )2

].

However using the fact that 1hηI

dhηI

dκ= 1

hε

dhε

dκ,this expression reduces to

hIdhηI

dκ(θI − θI) + hI

dhε

dκ(y − θI)

(hI + hηI + hε)2− hC(θC − θC)

dhηC

dκ

(hC + hηC )2

. (6)

This expression is clearly increasing in both y and θIand decreasing in θC .

Next turn to the first term in the expressions for dVdκ. While there are two cases, the expressions

are identical but for a constant and therefore we can treat them as one. An increase in y or θI ora reduction in θC will always increase z but this can either increase or decrease D, depending onwhether Δ+ z < 0 to start with or not. Moreover the effect of an increase in D on dV

dκdepends

on the sign of d(Δ+z)dκ

. Therefore this effect is potentially ambiguous. However we can make thisproblematic effect as small as we want by assuming that μis close enough to one or cis largeenough. This what this proposition assumes. Hence the result.

7 Data Appendix

Our randomization inference p-values are based on 10,000 replications within each sample frame.We explain the procedure for regressions using polling station electoral data. We took as thesample frame polling stations and then simulated treatment assignment by randomly selecting20 polling stations in each jurisdiction. We repeated this simulation 10,000 times so that we had10,000 different treatment-control assignments in our 10 jurisdictions. Thus the basic idea is toreassign treatment while leaving outcomes unaffected to represent the hypothesis of no effect.

28

Next we run regressions with simulated treatment and compare with actual treatment effects.After creating each simulated treatment variable we re-ran the regressions. These simulationsyield a distribution for each coefficient, centered around 0, with standard deviations that couldbe compared to the standard errors on the coefficients from regressions using actual treatmentassignment. We then perform t-tests on whether the actual coefficients are comparable to thesimulated coefficients. Since the simulated coefficients are approximately 0, this is equivalentto measuring the significance levels of the actual coefficients but replacing the actual standarderrors with the simulated standard errors.

29

Figure 2: Report Cards in The Hindustan Times on November 24, 2008