Word Count: 7,191. Do not cite or circulate without permission. Please find the latest version here.

Absence: Electoral Cycles and Teacher Absenteeism in India*

Emmerich DaviesHarvard Graduate School of Educationemmerich_davies_escobar@gse.harvard.edu

January 18, 2021

Front-line service worker absence has commonly been cited as a reason for the poor performance of developingcountry public services. Teachers and health care workers are often absent and, when present, not working. Thisabsenteeism is expensive: a nationally representative sample of villages across India finds that teacher absencecosts $1.5 billion a year. This paper argues that one explanation for variation in absenteeism is the differentialattention politicians pay to public services over the cycle of their tenure. Using a panel of all governmentschools across India between 2006 and 2017, I find that teacher absenteeism decreases substantially in electionyears. Placebo tests on private school absenteeism finds smaller and inconsistent effects of election years onabsenteeism in the private sector, lending support for a channel of political control of the bureaucracy duringelections. I argue that political control of the bureaucracy has a strong effect on bureaucratic performance andthat electoral accountability focuses political attention.

*I am grateful to Francisco Lagos, Sophie Litschwartz, and Fernanda Ramírez-Espinosa for fantastic research assistance on this project.Francesca Refsum-Jensenius was instrumental in acquiring electoral data and Sandip Sukhtankar provided the ultimate public good bymaking his electoral constituency maps publicly available. Comments from Rikhil Bhavnani, Jane Gingrich, Saad Gulzar, James Pickett,Rabia Malik, Gilles Vernier, Torsten Figueiredo Walter and conference participants at the 2017 Annual Conference on South Asia, the2018 Political Economy of Education Workshop at Nuffield College, the 2018 North-East Universities Development Consortium annualconference, and the 2019 annual meeting of the Midwest Political Science Association have helped to greatly improve the paper. Theresponsibility for all errors rests solely with me.

1

Public sector worker absence has commonly been cited as a reason for the poor performance of low- andmiddle-income country public services (Chaudhury et al., 2006; Alcázar et al., 2006; Callen et al., 2017). Teachersand health care workers are often absent and, when present, not working. This absenteeism is expensive. In anationally representative sample of villages across India, Muralidharan et al. (2017) find that teacher absenteeismcosts $1.5 billion a year. In contexts of weak state capacity, low levels of accountability, and large informationalasymmetries in the face of principal-agent problems, high absenteeism is not a mystery. There is, however, largevariation in the levels of absenteeism between and within regions (Chaudhury et al., 2006). What explains thisvariation?

This paper argues that one explanation for variation in absence is the differential attention that politiciansgive to public services over the course of their tenure. Using theory and evidence from education in India, I arguethat differential attention paid to education over an electoral cycle determines a large amount of front-line serviceworker absence. Although accountability for public sector workers is often low, elected officials wield a number oftools to encourage them to show-up including the power of hiring, firing, and transferring under-performing publicsector workers. If desired, politicians can encourage public sector workers to work.

Using a school-level dataset of all government schools in India from 2006 to 2018 matched to electoral dataover the same period, I find that teachers are less likely to be absent in the year immediately preceding an electionand in election years. Specifically, I find that in election years, within school absenteeism declines by 20 percent.I argue that this is a result of political pressure applied by elected officials in election years to show up for work.The effect is large and consistent across a number of specifications. The results are robust to a number of differentestimation strategies, including using the time to election, individual election year dummies, and modeling the fullelectoral cycle. I do not find a similar electoral cycle in the private sector.

As the school panel is self-reported, I re-run a similar analysis on an independently collected school survey – theIndian Human Development Survey (IHDS) – and find remarkably similar results. Using this data, absenteeism isagain lowest in the year before and election year, providing independent verification of the all India data. Together,these results suggests that increased political attention near election years is a strong explanation for bureaucraticperformance.

This joins a growing literature in political science on the workings of the bureaucracy and front-line serviceproviders in particular (Bertelli et al., 2020), and brings attention to the political control of front-line functionaries(Handler, 1996; Lipsky, 2010; Wade, 1992). Across low- and middle-income countries, front-line service workersare frequently involved in electoral politics as poll monitors, census enumerators, and more partisan politicalactivities as members of political parties or collective bargaining unions (Kingdon and Muzammil, 2009; Kingdonand Teal, 2010; Larreguy, Montiel Olea and Querubin, 2017). This paper adds to this literature on front-line serviceprovider absenteeism by exploring a previously unexplored channel for front-line service provider absenteeism: theelectoral cycle. Scholars have recently begun to turn to the bureaucracy as an area of study for its potentially largeeffects on economic and political outcomes. Pepinsky, Pierskalla and Sacks (2017) argue that there is a distinctionbetween literature that focuses on principal-agent problems and the developmental state literature. This papers sitsin between these two literatures by trying to understand the impact of principal-agent problems on developmentoutcomes. Additionally, the paper uses rich administrative data to answer an important political problem. UnlikeGulzar and Pasquale (2017) I cannot conclude that this is a positive result for development and service provision:increased monitoring is likely a result of the election cycle, not greater accountability.

2

theoretical expectations

There are two pressures teachers face to perform. First, embedded in a principal-agent relationship, there is thetop-down pressure from principals (in this case higher-level bureaucrats and elected politicians) to perform. Herethey are subject to sanctions, such as transfers to undesirable locations, for poor performance. Second, across theGlobal South, and in India specifically, teachers, either as members of political parties or other interests groups suchas unions, are frequently engaged in partisan political activities during elections. It is likely that the first pressure —punishment for poor performance — increases during elections and result in reduced absenteeism, while the secondpressure — partisan activities near elections — results in higher absenteeism.

Given these relationships, it is unclear whether elected politicians will exert more or less control in an electionyear. Theories of democratic accountability suggest that politicians would exert greater control over teachers,especially closer to an election year as they seek reelection (Lenz, 2012). On the other hand, teachers across thedeveloping world often engage in non-teaching activities that are closely related to, if not direct, partisan politicalactivities in support of elected politicians that would take them away from their schools in election years. In thiscase, politicians might reduce accountability pressures if they believe teachers can help them win elections.

Principal-Agent Problems in Education: Monitoring and Transferring Teachers

Elected politicians in India control teachers through two channels. First, politicians control postings either directlyor indirectly through putting pressure on mid-level bureaucrats who can re-assign teachers to favorable or lessdesirable teaching positions (Béteille, 2015).1 Second, politicians can monitor teachers directly by visiting schoolsand observing whether teachers are present.2

In a series of unannounced audit studies, Chaudhury and Hammer (2004), Kremer et al. (2005), and Alcázaret al. (2006) explored service provider absenteeism across the world. While rates of absenteeism varied from a lowof 11 percent for government school teachers in Peru (Alcázar et al., 2006), to a high of 74 percent of governmentdoctors in clinics in Bangladesh (Chaudhury and Hammer, 2004), public service delivery across these variedcontexts was characterized by high levels of absenteeism.

More recent work has attempted to move beyond this and understand why front-line service providers areabsent and address high levels of absenteeism. In an intervention that provided financial incentives as well asmonitoring through the use of cameras, Duflo, Hanna and Ryan (2012) found that providing financial incentives forattendance significantly reduced absenteeism while also raising learning outcomes. Callen et al. (2017) implement asmartphone monitoring system to facilitate inspections of health centers in Pakistan. They find that although theintervention increased inspections, it only decreased absenteeism in politically competitive constituencies. Takentogether, these suggest significant principal-agent problems in motivating front-line workers. When provided withextrinsic incentives or increased monitoring, agents, in this case teachers and health-care workers, are more likely toshow-up to work. Callen et al. (2017) also suggest a mechanism other than material incentives. Political pressures,in the form of increased electoral competition was the key driver behind decreasing absence.

Callen et al. (2017) find their results are conditional on competitive elections, suggesting that the control ofthe bureaucracy by local politicians is an important channel through which accountability is ensured. Gulzar andPasquale (2017) find that in areas where politicians can fully internalize credit from successful development projects,

1These are most often District Education Officers (DEOs) who sit at the district-level, the third level of Indian bureaucratic adminis-tration, and under responsibilities for Indian education policy implementation, have control of teacher postings and transfers within theirdistricts.

2Interview with M. Somi Reddy, District Education Officer, Ranga Reddy District, Andhra Pradesh, September 2013.

3

development projects will be more successful. Better monitoring has also been found to reduce teacher absenteeism(Muralidharan et al., 2017).

Rogger and Rasul (2013) find that increased autonomy for Nigerian civil service workers led to improved projectcompletion rates and quality, while performance incentives (or extrinsic motivations) reduced completion ratesand quality. Taken together, these series of studies suggest a significant role for some combination of intrinsicand extrinsic motivations for bureaucrats to perform their job. Extrinsic incentives solve some basic performanceproblems by encouraging front-line workers to show-up to work, while they are less useful for encouraging betterperformance while at work.

Béteille (2009, 2015) finds that transfers are a powerful form of sanctioning teachers for poor performance inIndia, and the decision to transfer teachers lies with DEOs. Beteille’s work builds on a larger series of anthropologicalstudies by Robert Wade (1985) that looks at the power of bureaucrat transfer and assignment as a powerful sourceof patronage. Wade (1985) outlined the specific mechanisms through which the Indian state held its employeesaccountable. Through the use of transfers and premiums on more desired positions, various bureaucracies in theIndian state create an “internal labor market,” through which they can sanction non-performing workers as wellreward supporters. Béteille (2015) tests these mechanisms and finds that a para-statal organization has emerged tofacilitate the ability of the state to create an internal labor market and transfer teachers.

My own field work suggests that mid-level bureaucrats are subject to political pressure from elected officialsto ensure front-line service provider attendance. The responsibility to sanction teachers, including the hiring andfiring of teachers, rests at the district level, particularly with the DEO. DEOs across the state of Andhra Pradeshfrequently cited political pressure, particularly from elected officials, as a key driver in their decision to monitorcertain schools and sanction teachers. In a context of low capacity and high information asymmetries, DEOs reliedon information and pressure from elected representatives to decide which schools to monitor.

Politicians insert themselves in these relationships of accountability by putting pressure on mid-level bureaucratsto sanction underperforming teachers or reward supporters. Theories of democratic accountability suggest thatthey are most likely to exert these pressures in election years to ensure voters reward them for good performance(Lenz, 2012). As a result, we should see increased monitoring and sanctioning of teachers in election years anddecreased absenteeism as a result.

The paper closest to this one is Fagernäs and Pelkonen (2016) that uses the same data to look at teacher transfersover the electoral cycle. The key difference between this paper and theirs is that I am interested in top-downpressure from politicians on bureaucratic performance while theirs lies in bureaucratic sanctioning. This paper alsois interested in the effects of competition on bureaucratic performance by looking at the marginal effects of politicalcompetition and electoral alignment on bureaucratic performance.

Vested Interests in Education: Teachers as Partisan Political Actors

While embedded in principal-agent relationships, teachers are also vested interests in the political system (Moe,2015). Government teachers are engaged in a variety of political and administrative tasks unrelated to their officialrole as teachers. Government teachers are often the most educated members and the most constant representativeof the state in rural communities (Béteille, 2009), and serve as poll booth monitors and census enumerators wherethey work (Béteille, 2009; Neggers, 2018). They also engage in partisan political activities such as mobilizing voters,and acting as teachers union representatives. Teachers unions are particularly powerful in India (Kingdon andMuzammil, 2009; Moe and Wiborg, 2016), and have been credited with bringing down the Chief Minister ofAndhra Pradesh in the late 1990s (Rudolph and Rudolph, 2001).

4

In Mexico, Larreguy, Montiel Olea and Querubin (2017) find that the Mexican National Educational WorkersUnion (or by its Spanish acronym SNTE), the largest Mexican teachers union, serves as a partisan machine bydelivering votes to the parties they support in elections. They deliver votes by monitoring voters as this effect isonly present in polling stations located in schools, a monitoring function Mexican teachers share with teachers inIndia. During elections in India, teachers are called upon to man poll booths (Neggers, 2018), tally votes, and areoften affiliated with political parties. Neggers (2018) finds poll-booth monitors privilege co-ethnics, suggesting asecond mechanism of encouraging those similar to them to turn out.

This literature suggests two contradictory expectation. On the one hand, increased monitoring has reducedabsenteeism across a number of contexts. Politicians will be likely to increase monitoring in election years. Thissuggests we should see reduced absenteeism in election years as politicians look to win subsequent elections byimproving service provision near elections. On the other hand, the number of official and unofficial political activitiesthat teachers are engaged in India and other developing countries should increase in election years. Working intheir official capacity as poll booth monitors and in their unofficial capacity as part of teacher’s unions, teachers areengaged in a number of partisan activities around elections that suggest they may be more absent in election years.

The two mechanisms discussed above — increased monitoring of teachers by principals in election years andteachers own partisan political activities — provide contrary expectations for teacher absenteeism in elections years.On one hand, election years increase incentives for principals to increase monitoring of the agent. Looking to winre-election, politicians will use all the tools at their disposal such as transferring non-performing teachers, to ensurelower teacher absenteeism. On the other hand, teachers are also embedded in partisan networks of their own, eitheras members of teachers unions, poll booth monitors, co-ethnics, or members of partisan political machines. Thisshould increase election year absences as they are working to help politicians get re-elected. In the next section, Iattempt to tease out which mechanism should prevail and provide credibly causal evidence that increased monitoringby the principal is the mechanism that dominates.

data & methods

This paper draws on two sources of data to create a school-level panel of schools across India. I combine data fromthe District Information System for Education (DISE) School Report Cards with assembly constituency electiondata.

District Information System for Education School Report Cards

The primary data source used in this paper is the DISE School Report Cards. The data consists of self-reported dataon school-level infrastructure, enrollment, educational outcomes, resources, and labor for every year from 2006 to2017.3 School headmasters are responsible for reporting the data to the National University of Education Planningand Administration (NUEPA) in September of the beginning of the academic year for the previous academic year.4

All registered schools are mandated to report this data, meaning that all government schools in the country, as well3I refer to years here as the second year in the school year. For example, the 2005-2006 academic year is referred to as 2006. This is to

correspond with the electoral year each academic year would correspond to.4NUEPA is a federal public university tasked with training education administrators and researchers as well as collecting nationally

representative data on education at the primary and secondary level. Headmasters are responsible for filling out forms, which are thenchecked by cluster and district education officials. District officials compile the DISE data for all schools in a given district and send it to thestate office. Each state then collects the information and sends it NUEPA located in Delhi. There is a five percent back check to verifyinformation.

5

as private schools that meet government standards for registration are included in the dataset. NUEPA and DISEsend the data reporting sheet to unrecognized schools they are aware of, so the data represents an undercountof unrecognized schools as the Government often has poor records of unrecognized schools (Rangaraju, Tooleyand Dixon, 2012). Given that we are interested in absence in government run or aided schools, the missingness ofprivate unrecognized schools is less of a concern.

IHDS Data

I also use the second wave of the Indian Human Development Survey (IHDS) for data on teacher absenteeismand reasons for absenteeism. The IHDS survey is a nationally representative survey of 1,503 villages and 971 urbanneighborhoods across India. The 2011-2012 round was the second round of a panel survey that surveyed the samevillages in 2004-2005. The survey asked a small number of questions of the largest government and private schoolin each surveyed village, and the next two largest schools, irrespective of whether they were public or private. Forevery teacher in the school, the surveyors asked if the teacher was present on the day of the survey, and, if they wereabsent, absent on officially sanctioned government work. In this sense, the survey replicates the data collection inKremer et al. (2005), randomly auditing schools on staff absenteeism.

Together, the two sources of data allow me to triangulate between two otherwise imperfect data sources(Herrera and Kapur, 2007), and draw broader conclusions on the drivers of absence in Indian public service. Whilethe IHDS data effectively provides a randomized audit of a select number of government and private schools inthe country, designed to be representative of the country as a whole, the DISE data allows me to generalize theresults to all schools in India. The consistency in the results between the two sources provides support that theelectoral cycles I discover are real and not the product of data quality, manipulation by teachers or school officials,or a result of self-reporting.

Electoral Data

I then match the School Report Card Data with electoral data at the assembly constituency level from 2004 to2018.5 Assembly constituencies are India’s state-level legislative assemblies and are equivalent to state houses inthe United States and other two-tiered federal systems. Each assembly constituency elects one Member of theLegislative Assembly (MLA) in first-past-the-post single-member district.

I match the school report cards data to assembly constituencies using the postal pincodes provided for schoolsin the school report cards data. Each school observation in the school report cards data reports the postal pincodein which the school is located. I geo-locate these pincodes to spatial points, and then merge these points to assemblyconstituencies.6 Postal pincodes are small units of and are nested within electoral constituencies, allowing for aclean identification of the assembly constituency in which a school is located.

Responsibilities for government schools lies at the district level, which is the third tier of decentralization belowthe Centre and States in India’s administrative system. Assembly constituencies are nested within districts andthere are between four and ten assembly constituencies in each district, so one DEO will respond to various MLAs.Ultimately, DEOs are appointed at the state-level by the Chief Minister (CM) of each state.

5Data was downloaded from the Trivedi Centre for Political Data at Ashoka University and more details of the data collection processcan be found in Jensenius (2016) and Jensenius and Verniers (2017).

6Sandip Sukhtankar has provided the ultimate public good by making his assembly constituency maps publicly available here. Theschools in the state of Madhya Pradesh do not provide any geo-located information, so all analyses are conducted without data from thestate of Madhya Pradesh and accounts for much of the unmatched data.

6

I run two analysis. First, I run an OLS of the form:

Yi,t = β1Election Yearc,t + β2yi,t−1 + Zi,t + γi + ζt + µi,t,d, (1)

Where Yi,t is the either a binary indicator for whether any teachers were absent in school i in year t, or thelogged total number of days teachers were absent in school i in year t. Election Year is the main variable of interestthat takes the value of 1 if there is an election in constituency c in year t, and Zi,t is a vector of controls, includingwhether the school is urban or rural, private or government run, and the number of teachers in the school, γi areschool-level fixed effects, and ζt are year fixed effects. µi,t,d is the error term clustered at the district level. Next, Iuse an event study model to understand timing over the electoral cycle, with the staggered state-level electionsacross India as the events. Specifically, I estimate the following equation using school-level yearly panel data:

Yit =∑j 6=0

αj · 1{j = t− ei}+ β1yi,t−1 + Zi,t + γi + ζt + εit (2)

where i represents schools, t represents years, and Y stands for either an indicator for any absence in theschool-year, or the number of missed days (in log terms) in a given school-year. αj · 1{j = t− ei} is an indicatorvariable that equals one when school i is j years away from the state election. The model also includes schooland time fixed effects, γi + ζt, where there is an indicator for each school and year. These fixed effects parameterscontrol for unobserved national-level trends, as well as any unobserved school-specific characteristics.

The primary variable of interest is αj · 1{j = t− ei} over the electoral cycle from two years to an election(j = −2) to 2 year after an election (j = 2).

Finally, I include a lag of the dependent variable, yi,t−1 to explicitly model the temporal dependence of thedata as absenteeism in one year is likely influenced by earlier absenteeism. I am also concerned about the presenceof serial correlation in the data, so including a lag makes sense from a modeling perspective.

I run the analyses on two sets of outcomes: whether there is any absence in a school in a year, and the totalnumber of absences in a school. We can think of these two sets of results as the extensive and intensive marginsrespectively.

results

I provide summary statistics for the two sources of data in this paper, the 2011-2012 round of the Indian HumanDevelopment Survey and a school-level panel from the District Information System for Education (DISE) from2006-2014, in Table 1. There are several differences of note from the summary statistics. First, form the IHDS data,we see that private schools in the data are in general larger, with about 65 more students and 3 more teachers ineach school. This is consistent with the fieldwork strategy for the IDHS where surveys are conducted of the largestgovernment and private school in each village and private schools, in particular larger private schools, are likely tohire more teachers in line with parental demand (Muralidharan and Sundararaman, 2015). There are few teacherlevel differences between government and private schools other than that private school teachers are likely to be ofhigher caste status than their government school counterparts, with private school teachers more likely to be uppercaste and from religions other than Hinduism and Islam, suggesting schools that both cater to economic elites andreligious minorities.

Although I use fewer variables from the DISE data (Panel A in Table 1), there are significant differences acrossall variables between private and public schools. Teachers in private schools are 12 percent less likely to be absent,

7

private schools are more likely to be located in cities, and private schools are also likely to be larger. Of note is thatthe size of schools, government schools, and private schools in the IHDS and DISE datasets are remarkably similar,suggesting that IHDS does not sample from a particular type of school when looking at private or governmentschools. Most schools in the sample are also rural and government schools, consistent with the distribution ofschools in India.

Table 1: Summary Statistics

Panel A: DISE Summary StatisticsGovernment Schools Private Schools Difference All Schools

Mean SD N Mean SD N Mean SD N

Absent (%) 0.14 0.35 11,888,443 0.03 0.16 2,672,423 -0.115*** 0.12 0.32 15,378,545Rural (%) 0.93 0.26 12,016,513 0.63 0.48 2,692,784 -0.298*** 0.86 0.35 15,531,546Teachers 4.52 4.30 11,646,051 8.56 7.36 2,579,633 4.04*** 5.49 5.61 15,032,804

Panel B: IHDS School Level Summary StatisticsGovernment Schools Private Schools Difference All Schools

Mean SD N Mean SD N Mean SD N

Number of Students 169.36 158.83 2,039 234.75 205.34 1,846 65.4*** 200.43 185.29 3,885Number of Teachers 5.19 3.47 2,042 8.51 5.45 1,847 3.33*** 6.77 4.81 3,889

Panel C: IHDS Teacher Level Summary StatisticsGovernment Schools Private Schools Difference All Schools

Mean SD N Mean SD N Mean SD N

Absent from school (%) 0.16 0.37 14,826 0.12 0.33 16,129 -0.04 0.14 0.35 30,955Absent on official duty (%) 0.03 0.18 14,826 0.02 0.12 16,129 -0.016 0.02 0.15 30,955Male (%) 0.36 0.48 14,855 0.36 0.48 16,175 0 0.36 0.48 31,030Age 40.60 9.63 14,804 33.21 9.88 16,075 -7.389 36.75 10.44 30,879Hindu (%) 0.85 0.36 14,848 0.83 0.38 16,138 -0.024 0.84 0.37 30,986Muslim (%) 0.08 0.27 14,848 0.07 0.26 16,138 -0.007 0.07 0.26 30,986Other Religion (%) 0.07 0.26 14,848 0.10 0.30 16,138 0.031*** 0.09 0.28 30,986Upper Caste (%) 0.35 0.48 14,677 0.44 0.50 15,850 0.084*** 0.40 0.49 30,527OBC (%) 0.40 0.49 14,677 0.40 0.49 15,850 0.008 0.40 0.49 30,527SC/ST (%) 0.24 0.43 14,677 0.15 0.35 15,850 -0.092 0.19 0.39 30,527Other Caste (%) 0.01 0.11 14,677 0.01 0.10 15,850 -0.001 0.01 0.10 30,527Distance from school (km) 5.65 9.60 14,773 3.06 5.68 15,977 -2.589 4.31 7.92 30,750

Notes: For all panels, Columns 1-3 present summary statistics for government schools, Columns 4-6 present summary statistics for private schools, Column 7 presents a t-test ofdifferences between government and private schools, and Columns 8-10 present summary statistics for all schools together. For column 7, * p < 0.1, ** p < 0.05, *** p < 0.01. Panel Apresents summary statistics for the school-level panel of District Information Systems for Education data from 2006-2014, Panel B presents summary statistics from the 2011-2012wave of the Indian Human Development Survey (IHDS) school survey at the school level, Panel C presents summary statistics for the IHDS data at the teacher level.

I present results from two similar analyses. Columns 1-4 in Table 2 presents a regression of the form in Equation1 on assembly constituencies using a dummy variable for whether a school reports any absence in a school that yearwith school and year fixed effects. This can be considered as the extensive margin on absenteeism and represents awithin-school-year variation on absenteeism. In election years, schools report two percent lower levels of absenteeismrelative to non-election years in the same school. In Columns 5-8 in Table 2, I re-run the same specification inEquation 1, using the log number of days all teachers in the school were absent instead of the probability of anyabsence. Again, there is significantly less absenteeism in an election year.

Next, I turn to the specification in Equation 2, using the same dependent variables in Table 2. Figure 1 reportswhether there is any reported absenteeism in a school over the entire five year electoral cycle. The timing of absencebecomes more apparent in this specification. Again, Figure 1 includes school and year fixed effects, so I am estimatingwithin school variation from year to year with the election cycle.

The results in Figure 1 make the nature of the electoral cycle clear. Absenteeism is highest the further we move

8

Table 2: Any Absence in an Election Year

Absent Log Absence

(1) (2) (3) (4) (5) (6) (7) (8)

Election Year −0.024∗∗∗ −0.026∗∗∗ −0.022∗∗∗ −0.026∗∗∗ −0.049∗∗∗ −0.058∗∗∗ −0.048∗∗∗ −0.061∗∗∗

(0.001) (0.001) (0.001) (0.001) (0.004) (0.003) (0.004) (0.003)

Year FE No Yes No Yes No Yes No YesSchool FE No No Yes Yes No No Yes YesMean Absence in Non-Election Year 0.137 0.078 0.089 0.105 0.327 0.206 0.239 0.275Number of Schools 1,222,939 1,222,939 1,222,939 1,222,939 1,220,024 1,220,024 1,220,024 1,220,024Observations 10,760,056 10,760,056 10,760,056 10,760,056 10,592,084 10,592,084 10,592,084 10,592,084

Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. Robust standard errors clustered at the school level in parentheses. The dependent variable in columns 1-4 is a dummy variablethat takes the value of one if the school reports any teacher absenteeism in that year and the log number of absences per teacher in columns 5-8. Each specification includescontrols for the number of teachers in each school, a dummy for whether the school is in a rural area, and a lagged dependent variable. “Election Year” is a dummy variableequal to 1 when the school is in a constituency in an election year.

from an election, with absenteeism two or more years before and after an election three percent higher than it is inan election year, from an election year mean of about 8 percent. This represents a 20 percent decline from yearsfurther away from an election. All estimates are also significantly different from each other (Gelman and Stern,2006). In short, teacher absenteeism is significantly higher the further we move from election years, decreasing byabout 20 percent to an election year low of about 8 percent.

Teacher absenteeism is lower in government schools in election years across all specifications. In the nextsection, I turn to teacher absenteeism in the private sector. If these results are indicative of increased politicalattention to the bureaucracy in election years, we should not see similar results for private schools as politiciansdo not exert the same level of control on private schools as they do government schools. Otherwise, if results aresimilar, this would be suggestive of other effects specific to election years I am unable to pick-up with this data.

Absence in the Private Sector

As a placebo check on these results, I turn to absence in the private sector. In independent audits, Kremer et al.(2005) found that private school teachers are also likely to be absent, although the levels of absenteeism are muchlower than government schools in the same village. With that, although absenteeism is also likely to be high inprivate schools, private schools are not subject to the same political interference as government schools. Similarlevels of absence in the private sector would raise concerns either over the quality of the data or whether we arepicking up real effects in government schools. These concerns should be partially alleviated through looking atabsence in the private sector. Politicians should have far less control over teacher absenteeism in the private sectoras private schools do not report to elected representatives. To test this, I repeat the analysis in Equations 1, and 2 inprivate schools. Table 3 reports the same results as Table 2 for private schools. While the results are similar to thosein Table 2, the size of the point estimates are between five to eight times smaller.

Next, I turn to estimating Equation 2 for private schools, repeating the analysis from Figure 1 for the subset ofprivate schools in the data. This presents absenteeism over the entire electoral cycle with school and year fixedeffects. While absenteeism is lower in election years similar to Figure 1, there are two important differences. First,there is no clear pattern of absenteeism over the election cycle, with the highest levels of absenteeism one year afterthe election. Second, similar to Table 3, absenteeism is significantly lower in all specifications, with an election yearmean of two percent.

Taken together, the results from Table 3 and Figure 2 suggest that there is not as clear an election cycle toabsenteeism in private schools as there is in government schools. Expanding the specifications to include and exclude

9

Figure 1: Any Absence in a School Year over the Electoral Cycle

0.00

0.01

0.02

0.03

−2 or more −1 Election Year 1 2 or moreYears to Election

Pro

babi

lity

Abs

ent

A

0.000

0.025

0.050

0.075

−2 or more −1 Election Year 1 2 or moreYears to Election

Ave

rage

Num

ber

of A

bsen

ces

B

Notes: In Panel A, the dependent variable is a dummy variable that takes the value of one if the schoolreports any teacher absenteeism in that year. In Panel B, the dependent variable is the log number Theregression includes controls for the number of teachers in a school, a dummy for whether the school isin a rural area, a lagged dependent variable, and year and school fixed effects. The line represents 95%confidence intervals. There are 8,796,007 school-year observations and 1,317,498 total schools in the sample.This figure corresponds to Column 4 in Table A1.

10

Figure 2: Any Absence in a School Year over the Electoral Cycle in Private Schools

0.000

0.002

0.004

0.006

0.008

−2 or more −1 Election Year 1 2 or moreYears to Election

Pro

babi

lity

Abs

ent

A

0.000

0.005

0.010

0.015

−2 or more −1 Election Year 1 2 or moreYears to Election

Ave

rage

Num

ber

of A

bsen

ces

B

Notes: In Panel A, the dependent variable is a dummy variable that takes the value of one if the school reports any teacher absenteeismin that year. In Panel B, the dependent variable is the log number The regression includes controls for the number of teachers in a school,a dummy for whether the school is in a rural area, a lagged dependent variable, and year and school fixed effects. The line represents 95%confidence intervals. There are 1,664,926 school-year observations and 401,401 total schools in the sample. This figure corresponds toColumn 4 in Table A2.

11

Table 3: Any Absence in an Election Year in Private Schools

Absent Log Absence

(1) (2) (3) (4) (5) (6) (7) (8)

Election Year −0.006∗∗∗ −0.007∗∗∗ −0.003∗∗∗ −0.005∗∗∗ −0.008∗∗∗ −0.011∗∗∗ −0.005∗∗∗ −0.009∗∗∗

(0.0002) (0.0002) (0.0002) (0.0002) (0.001) (0.001) (0.001) (0.001)

Year FE No Yes No Yes No Yes No YesSchool FE No No Yes Yes No No Yes YesMean Absence 0.04 0.013 0.028 0.026 0.09 0.031 0.07 0.066Number of Schools 604,113 604,113 604,113 604,113 595,957 595,957 595,957 595,957Observations 3,325,522 3,325,522 3,325,522 3,325,522 3,272,096 3,272,096 3,272,096 3,272,096

Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. Robust standard errors clustered at the school level in parentheses. The dependent variable in columns 1-4 isa dummy variable that takes the value of one if the school reports any teacher absenteeism in that year and the log number of absences per teacher incolumns 5-8. Each specification includes controls for the number of teachers in each school, a dummy for whether the school is in a rural area, and a laggeddependent variable. “Election Year” is a dummy variable equal to 1 when the school is in a constituency in an election year.

school and year fixed effects in Table A2 shows that these specifications are also sensitive to modeling choices. Allmodels that do no include year fixed effects show no difference in absence over the electoral cycle or when electionsare modeled as a dummy indicating the election year. Next, I turn to considering the alternative explanation thatschool leaders might be “cooking the books” given that the data is self-reported by school leaders.

“Cooking the Books” or Accountability?

A concern with the self-reported nature of the DISE school report cards is that head teachers and district-levelofficers might be “cooking the books”, or falsely reporting lower rates of absenteeism, around elections to makebureaucratic effort look greater with no underlying change in behavior. The empirical evidence of “cooking thebooks” would be substantively similar – lower absenteeism in election years – although the mechanisms wouldbe different – perceived pressure to modify government data rather than hold providers accountable. Reportedabsenteeism of 12 percent in the DISE data and 14 percent in the IHDS data in Table 1 is lower than absenteeism inindependent audits conducted in 2005 (Chaudhury et al., 2006; Muralidharan et al., 2017).

To separate whether schools are “cooking the books” or there is electoral pressure on teachers, I use the IHDSschool surveys, an independent survey of schools conducted in 2011-2012 and contemporaneously to the DISEschool report cards data collection to test whether I find similar patterns in this data. In Table 4, I look at theprobability that a teacher is absent from the school on the day of the survey, absent from the school on the day of thesurvey to conduct official work, or present at the school on the day of the interview and present for the interviewas a function of whether their school is a government school or private school. As expected, teachers are fourpercent more likely to be absent from a government school (columns 1-2), and about two percent more likely to beabsent for officially sanctioned duty (columns 3-4). In results that are suggestive of a certain level of performance infront of IHDS enumerators, conditional of being at the school on the day of the interview, teachers in governmentschools are between four to six percent more likely to be present at the interview (columns 5-6). These resultsconfirm common sense expectations of what we think differences in teacher absenteeism should look like betweengovernment and private schools, with government schools showing consistently higher levels of absenteeism.

Next, I replicate Equation 1 on the IHDS data, to see if absenteeism is lower in election years. Similar tothe results in Tables 2 and 3, absenteeism is lower in government and private schools in election years, althoughabsenteeism decreases by more in government schools in election years. There is no effect of elections on a teacherbeing absent on official duty (Columns 3 and 4), and this is consistent across government and private schools.

12

Table 4: Absence by School Type

Absent On Official Duty Absent from Interview(1) (2) (3) (4) (5) (6)

Government School 0.042∗∗∗ 0.046∗∗∗ 0.014∗∗∗ 0.016∗∗∗ −0.035∗∗∗ −0.038∗∗∗

(0.004) (0.004) (0.002) (0.002) (0.007) (0.006)

Village Fixed Effects N Y N Y N YObservations 30,105 30,105 30,105 30,105 25,772 25,772Adjusted R2 0.007 0.329 0.005 0.162 0.041 0.478

Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. Robust standard errors in parentheses. Linearmodels of the effects of working in a government school on either absence from school onthe day of the interview, absent from the school for official government sanctioned work,or absent from the interview conditional on being present at the school on the day of theinterview. All models control for gender, age, religion, and caste, while even numberedcolumns also include village fixed effects.

Table 5: Any Absence in an Election Year

Absent On Official Duty

Government Private Government Private(1) (2) (3) (4)

Election Year −0.076∗∗∗ −0.068∗∗∗ −0.009 −0.007(0.009) (0.009) (0.009) (0.009)

Observations 12,875 14,184 12,875 14,184Adjusted R2 0.008 0.010 0.002 0.005

Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. Robust standarderrors in parentheses. Linear models of the effects of electionson Absenteeism (Columns 1 and 2) or whether a teacher isabsent from the school for official government sanctioned work(Columns 3 and 4). All models control for gender, age, religion,and caste.

Finally, I replicate the election cycle analysis from Equation 2 using IHDS data. In Figure 3, I present resultsfor the event study model using IHDS data. As the IHDS data contains information on whether a teacher wasabsent from the school on the day of the survey and whether they were on official duty, I plot both results together.The results using self-reported data from the DISE panel data and the IHDS independent audit data are remarkablyconsistent on absence. In both data sources, there are strong effects of the electoral cycle, with teachers far morelikely to be absent from the school the further we move from the election. Teachers have a 10 percent probability ofbeing absent from the school on the day of the IHDS survey, and this increases to greater than 20 percent morethan two years from the election on either side.

Turning to absence for official work, teachers are far more likely to be absent on official work in an electionyear, conditional on being absent from the school. Of all election year absences, 12 percent of those are of teacherson official duty. The electoral cycle of teachers on official duty is the inverse of teacher absence more generally.In the year before and after the election, teachers are more likely to be absent on official duty than in an electionyear, but the further we move from the election year, the less likely they are to be absent on official duty, with theprobability of being absent on official duty approaching zero.

13

Figure 3: Absence Over Electoral Cycle in Government Schools

0.0

0.1

0.2

−2 or more −1 Election Year 1 2 or moreYears to Election

Pro

babi

lity

Abs

ent

Notes: This figure presents results of a linear model of the probability of a teacher being absent in government schools the day IHDSsurveyors surveyed the school running the model in Equation 2. The dependent variable is a dummy variable that takes the value of one ifthe teacher was absent from the school on the day of the survey. The lines represent 95% confidence intervals with robust standard errors.There are 12,875 teacher observations for any absence, and the election year mean level of absence is 13.24 %. This figure corresponds tocolumns 1 & 2 in Table A3. Both models control for gender, age, religion, caste, and the distance the teacher lives from the school.

Private School Electoral Cycles

Next, I explore the same electoral cycles of absence in private schools using the DISE and IHDS data. Theoretically,we should expect that politicians should be able to exert lower levels of control on private schools as they cannotsanction private school teachers through transfers and other mechanisms of control. Empirically, as we can seefrom Table 1, there are significant differences between government and private schools. It is unclear, however, howthese differences should matter for teacher absenteeism. While absenteeism is lower in general, teachers in privateschools are also more likely to be from non-dominant religions and upper caste. Private schools are also more likelyto be in urban areas and employ a greater number of teachers. Each of these variables result in differing predictionsfor absenteeism: while upper caste teachers are more likely to be able to resist pressures from politicians, the greaterurban location of schools should allow for easier monitoring from politicians who do not have to travel as far tomonitor teachers. I test whether patterns in the data are suggestive of greater or lower levels of monitoring in theprivate sector below.

Once again, the results from the IHDS data are remarkably similar to those of the DISE data. I present resultsfrom Equation 2 in Figure 4 for the subset of private school teachers. While we still observe differential absenceacross the electoral cycle, once again the rate of absenteeism in private schools is much lower than in governmentschools. As we start from a lower baseline level of absenteeism, the differences over the electoral cycle are muchsmaller.

The second set of results explores absence on official duty conditional on being absent from the school on theday of the survey. Here, results are indeterminate and no effects of the electoral cycle on absenteeism. Teachers inprivate schools show no consistent pattern of absenteeism over the electoral cycle.

14

Figure 4: Absence Over Electoral Cycle in Private Schools

−0.05

0.00

0.05

0.10

0.15

−2 or more −1 Election Year 1 2 or moreYears to Election

Pro

babi

lity

Abs

ent

Notes: This figure presents results of a linear model of the probability of a teacher being absent in private schools the day IHDS surveyorssurveyed the school, running the model in Equation 2. The dependent variable is a dummy variable that takes the value of one if theteacher was absent from the school on the day of the survey. The lines represent 95% confidence intervals with robust standard errors.There are 14,184 teacher observations for any absence, and the election year mean is 8.67 %. This figure corresponds to column 1 in TableA4. The model also controls for gender, age, religion, caste, and the distance the teacher lives from the school.

Election Work

Next, I look at whether teachers are more likely to be absent for government sanctioned purposes in an electionyear. Teachers often lament that their time is taken up by official work, comparing themselves to postal workerswho move paper from one destination to another (Aiyar and Bhattacharya, 2016), and that this work increases inelection years as they are required to engage in preparing election booths that are often located in schools (Béteille,2009; Neggers, 2018).

Exploring absence for official work gives us purchase on a number of remaining question. First, it allows usto explore if schools are “cooking the books” for absenteeism and are more likely to mark teachers as absent forofficial work in particular years. Second, it allows us to see whether teachers do engage in a greater amount ofwork and sanctioned absence during election years. Finally, it allows us to separate bureaucratic pressure – whichshould lead to a greater level of absenteeism in an election year as the electoral bureaucracy, independent of politicalactors, would pressure teachers to be more absent in election years – from political pressure, which should reduceabsenteeism in election years.

I run the model in Equation 2 on a subset of all teachers reported as absent on the day of the survey using adummy for whether they were absent on official government duty such as census collection or election relatedadministration. I present the results for Government schools in Figure 5 and for private schools in Figure 6.

In Figure 5, we observe an opposite electoral cycle than we do in 1 and 3: teachers are more likely to be absenton official Government duty the year before, in an election year, and the year after an election than they are furtheraway from elections. Teachers are between 10 and 20 percent less likely to be absent for official work the furtherwe move from an election. While this is consistent with official policy where teachers are commissioned to helpwith formal government duties, it also provides evidence that the reduced absenteeism we see in 1 and 3 is in spite oftheir increased absenteeism for official work. We see increased absenteeism outside of the electoral period even

15

Figure 5: Absence for Government Work Over Electoral Cycle in Government Schools

−0.4

−0.3

−0.2

−0.1

0.0

0.1

−2 or more −1 Election Year 1 2 or moreYears to Election

Pro

babi

lity

Abs

ent f

or G

over

nmen

t Wor

k

Notes: This figure presents results of a linear model of the probability of a teacher being absent on officialduty conditional on being absent from the school in government schools the day IHDS surveyors surveyedthe school, running the model in Equation 2. The dependent variable is a dummy variable that takes thevalue of one if the teacher was absent from the school on official duty on the day of the survey. The linesrepresent 95% confidence intervals with robust standard errors. There are 2,187 teacher observations forany absence, and the election year mean is 12.17 %. This figure corresponds to column 2 in Table A3. Themodel also controls for gender, age, religion, caste, and the distance the teacher lives from the school.

while formal demands on a teacher’s time outside of school are increasing around elections.Finally, Figure 6 presents the results of absence for official work among private school teachers. Lending support

to political attention turning solely to Government schools, results in private schools are noisy and inconsistent.Teachers in private schools are no more or less likely to be absent for official work the further we move from anelection.

This section suggests that teachers and headmasters are unlikely to be “cooking the books” and systematicallyforging data around elections. Using data from an independent school-level survey, I find that remarkably similarresults from those from a national census of Government and registered private schools. Teachers in governmentschools are between 10 and 15 percent more likely to be absent from schools on the day of the unannounced school-level survey the further we move from an election, despite being more likely to be absent for official governmentduty around elections. Again, similar to the national-level census of schools, I do not observe these patterns inprivate schools, with noisy estimates for teacher absenteeism over the electoral cycle.

discussion

Combining individual school-level data from the DISE and data on election timing, I find that there is a strong andpersistent electoral cycle to absenteeism in government schools. While reported rates of absenteeism are lowerin this data than independent audits, approximately 13 percent of schools report some absenteeism in any givenschool-year, and an average of 5.5 teaching days are lost to absenteeism in each school. These numbers declinesignificantly in election years. The probability of any absence and the total number of days lost to absenteeism

16

Figure 6: Absence for Government Work Over Electoral Cycle in Private Schools

−0.4

−0.2

0.0

0.2

−2 or more −1 Election Year 1 2 or moreYears to Election

Pro

babi

lity

Abs

ent f

or G

oven

rmen

t Wor

k

Notes: This figure presents results of a linear model of the probability of a teacher being absent on official duty conditional on being absentfrom the school in private schools the day IHDS surveyors surveyed the school, running the model in Equation 2. The dependent variableis a dummy variable that takes the value of one if the teacher was absent from the school on official duty on the day of the survey. The linesrepresent 95% confidence intervals with robust standard errors. There are 1,810 teacher observations for any absence, and the electionyear mean is -0.09 %. This figure corresponds to column 2 in Table A4. The model also controls for gender, age, religion, caste, and thedistance the teacher lives from the school.

declines by 20 percent in a government school in a constituency election year. Furthermore, these effects are notas strong, if non-existent, in private schools and are not robust to alternative modeling decisions. The results areconsistent across models that only take account of election years and models that model the entire electoral cycle,as well as the inclusion of school and year fixed effects.

These results suggests that there is a strong link between bureaucratic performance, as measured by absenteeism,and democratic accountability. Teachers in government schools are more likely to show-up to work in electionyears. These results mirror those in Muralidharan et al. (2017) who find that there is lower absenteeism wheretop-down monitoring is greater. The question for policy, however, is how to extend monitoring beyond certaingeographic areas or election years.

Limitations

A key question surrounding data quality is how self-reported data provided by organizations like DISE compare toindependent evaluations of absence from random audits such as in Banerjee and Duflo (2006); Béteille (2009);Chaudhury and Hammer (2004); Chaudhury et al. (2006). The levels of absence found in this paper are muchlower than absence found by independent evaluations of service worker absenteeism from other papers in India.Average levels of absence self-reported in the DISE dataset reach 13 percent for the year, far shorter than the levelsof absence recorded on random spot checks in Chaudhury et al. (2006) of 25 percent on any given day.

It is important to note that this data should not be taken as a census of absence of in government schools in India.As the data is self-reported, there are strong incentives to misrepresent absence and furthermore, as DISE only asksabout officially sanctioned absences, the level of unofficial absence is likely higher as we find in independent auditssuch as Kremer et al. (2005); Muralidharan et al. (2017). Additionally, the variable used for absence is whether

17

there are teachers on non-teaching assignments, a specific question on whether teachers are working on officialdesignation. While teachers are can be requisitioned for official duties, much of the absence from teachers, muchthat is undocumented, is not for official duties.

With this in mind, the DISE data serves as the only independent and broadly comparable source of dataavailable to the government and broader public, and is used by the former to assess the state of schools. Whilethe data is almost certainly biased downwards, it does have important implications for decision making as this isthe dataset used by policy makers. Finding similar results in the IHDS data adds confidence that the self-reporteddata is not being systematically manipulated in election years and that we are seeing real decreases in absenteeism.Across all specifications, results are comparable in direction between both sources of data.

conclusions

Building on previous studies of public sector worker absence in developing countries, I provide theory and evidencefrom a large administrative dataset on the sources of absence: the political-electoral cycle. Using an administrativedataset of over one million school over an eleven year period and an average of four elections per school, I find thatteachers are far less likely to be absent in the year of a state-level election. These results are robust to a series ofspecifications, including year and school-level fixed effects that compare variation within schools across the entiretime period. Finally, we do not see the same effect in private schools, adding support for the channel of politicalcontrol of the bureaucracy.

Like other studies on the political interference of the bureaucracy (Asher and Novosad, 2017; Béteille, 2015;Gulzar and Pasquale, 2017; Kapur and Vaishnav, 2011), I find that empirical evidence of a clear channel of local levelpoliticians interfering in service provision. Unlike Gulzar and Pasquale (2017), however, the results do not suggestthe benefits of political interference, but raises the question of how to sustain political pressure in non-electionyears. This mirrors results in the United States where off-cycle elections result in vested interests winning power,suggesting differential attention by voters instead of politicians outside of high profile elections (Anzia, 2013).Teachers are present more in election years, with high levels of absenteeism in non-election years.

It is this question that the paper leaves unaddressed: how can policy makers ensure that either politicians exertthe same pressure in non-election years, or teachers respond to this pressure in non-election years. The paradox isthe form this pressure takes is also problematic as it is often coercive and detrimental to the provision of high qualityeducation (Béteille, 2015; Wade, 1985). Politicians influence teacher performance through the threat of transfers,hiring, and firing, and this market is often run through middle men who do not sit in the education bureaucracy(Béteille, 2015). These networks are also embedded in larger networks of patronage that run from the local-level upto the state bureaucracy (Wade, 1985).

18

Bibliography

Aiyar, Yamini and Shrayana Bhattacharya. 2016. “A Case Study of the Block Level Education Bureaucracy: ThePost Office Paradox.” Economic And Political Weekly 51(11).

Alcázar, Lorena, F. Halsey Rogers, Nazmul Chaudhury, Jeffrey Hammer, Michael Kremer and Karthik Muralidha-ran. 2006. “Why Are Teachers Absent? Probing Service Delivery in Peruvian Primary Schools.” InternationalJournal of Educational Research 45(3):117–136.

Anzia, Sarah. 2013. Timing and Turnout: How Off-Cyle Elections Favor Organized Groups. Chicago, IL: Universityof Chicago Press.

Asher, Sam and Paul Novosad. 2017. “Politics and Local Economic Growth: Evidence from India.” AmericanEconomic Journal: Applied Economics 9(1):229–273.

Banerjee, Abhijit V. and Esther Duflo. 2006. “Addressing Absence.” Journal of Economic Perspectives 20(1):117–132.Bertelli, Anthony M., Mai Hassan, Dan Honig, Daniel Rogger and Martin J. Williams. 2020. “An Agenda for the

Study of Public Administration in Developing Countries.” Governance 33(4):735–748.Béteille, Tara. 2009. Absenteeism, Transfers and Patronage: The Political Economy of Teacher Labor Markets in

India PhD thesis.Béteille, Tara. 2015. “Fixers in Indias Teacher Labor Markets: Behind the Scenes.” Asian Survey 55(5):942–968.Callen, Michael, Saad Gulzar, Ali Hasanain, Yasir Khan and Arman Rezaee. 2017. The Political Economy of Public

Employee Absence: Experimental Evidence from Pakistan. Technical report Cambridge, MA: .Chaudhury, Nazmul and Jeffrey Hammer. 2004. “Ghost Doctors: Absenteeism in Rural Bangladeshi Health

Facilities.” The World Bank Economic Review 18(3):423–441.Chaudhury, Nazmul, Jeffrey Hammer, Michael Kremer, Karthik Muralidharan and F. Halsey Rogers. 2006. “Missing

in Action: Teacher and Health Worker Absence in Developing Countries.” Journal of Economic Perspectives20(1):91–116.

Duflo, Esther, Rema Hanna and Stephen P. Ryan. 2012. “Incentives Work: Getting Teachers to Come to School.”American Economic Review 102(4):1241–1278.

Fagernäs, Sonja and Panu Pelkonen. 2016. “Teachers, Electoral Cycles and Learning in India.”.Gelman, Andrew and Hal Stern. 2006. “The Difference Between “ Significant ” and “ Not Significant ” Is Not

Itself Statistically Significant.” The American Statistician 60(4):328–331.Gulzar, Saad and Benjamin J. Pasquale. 2017. “Politicians, Bureaucrats, and Development: Evidence from India.”

American Political Science Review 111(01):162–183.Handler, Joel F. 1996. Down from Bureaucracy: The Ambiguity of Privatization and Empowerement. Princeton:

Princeton University Press.Herrera, Yoshiko M. and Devesh Kapur. 2007. “Improving Data Quality: Actors, Incentives, and Capabilities.”

Political Analysis 15(4):365–386.

19

Jensenius, Francesca R. 2016. “Competing Inequalities? On the Intersection of Gender and Ethnicity in CandidateNominations in Indian Elections.” Government and Opposition 51(3):440–463.

Jensenius, Francesca R. and Gilles Verniers. 2017. “Studying Indian Politics with Large-Scale Data: Indian ElectionData 1961–Today.” Studies in Indian Politics 5(2):269–275.

Kapur, Devesh and Milan Vaishnav. 2011. “Quid Pro Quo: Builders, Politicians, and Election Finance in India.”.Kingdon, Geeta and Francis Teal. 2010. “Teacher Unions, Teacher Pay and Student Performance in India: A

Pupil Fixed Effects Approach.” Journal of Development Economics 91(2):278–288.Kingdon, Geeta G. and Mohd. Muzammil. 2009. “A Political Economy of Education in India: The Case of Uttar

Pradesh.” Oxford Development Studies 37(2):123–144.Kremer, Michael, Karthik Muralidharan, Nazmul Chaudhury, Jeffrey Hammer and F Halsey Rogers. 2005. “Teacher

Absence in India: A Snapshot.” Journal of the European Economic Association 3(2-3):658–667.Larreguy, Horacio, Cesar E. Montiel Olea and Pablo Querubin. 2017. “Political Brokers: Partisans or Agents?

Evidence from the Mexican Teachers’ Union.” American Journal of Political Science pp. n/a–n/a.Lenz, Gabriel S. 2012. Follow the Leader?: How Voters Respond to Politicians’ Policies and Performance. Chicago ;

London: University of Chicago Press.Lipsky, Michael. 2010. Street-Level Bureaucracy: Dilemmas of the Individual in Public Services. 30th anniv ed. New

York: Russell Sage Foundation.Moe, Terry M. 2015. “Vested Interests and Political Institutions.” Political Science Quarterly 130(2):277–318.Moe, Terry M. and Susanne Wiborg, eds. 2016. The Comparative Politics of Education: Teachers Unions and

Education Systems around the World. Cambridge, United Kingdom ; New York: Cambridge University Press.Muralidharan, Karthik, Jishnu Das, Alaka Holla and Aakash Mohpal. 2017. “The Fiscal Cost of Weak Governance:

Evidence from Teacher Absence in India.” Journal of Public Economics 145:116–135.Muralidharan, Karthik and Venkatesh Sundararaman. 2015. “The Aggregate Effect of School Choice: Evidence

from a Two-Stage Experiment in India.” Quarterly Journal of Economics 130(3):1011–1066.Neggers, Yusuf. 2018. “Enfranchising Your Own? Experimental Evidence on Bureaucrat Diversity and Election

Bias in India.” American Economic Review 108(6):1288–1321.Pepinsky, Thomas, Jan H. Pierskalla and Audrey Sacks. 2017. “Bureaucracy and Service Delivery.” Annual Review

of Political Science 20:249–268.Rangaraju, Baladevan, James Tooley and Pauline Dixon. 2012. The Private School Revolution in Bihar: Findings

from a Survey in Patna Urban. Policy Paper India Institute New Delhi: .Rogger, Daniel and Imran Rasul. 2013. “Management of Bureaucrats and Public Service Delivery: Evidence from

the Nigerian Civil Service.”.Rudolph, Lloyd I. and Susan H. Rudolph. 2001. “Iconisation of Chandrababu: Sharing Sovereignty in India’s

Federal Market Economy.” Economic and Political Weekly 36(18):1541–1552.Wade, Robert. 1985. “The Market for Public Office: Why the Indian State Is Not Better at Development.” World

Development 13(4):467–497.Wade, Robert. 1992. “How to Make ‘Street Level’Bureaucracies Work Better: India and Korea.” ids bulletin

23(4):51–54.

A1

a1 appendix

Full Results Tables

This section provides the full results tables for the figures presented in Figures 1 to 6. Those figures are representedin column 4 in each of the tables below. Columns 1 to 3 also provide robustness checks with and without year andschool fixed effects. All specifications include controls for the number of teachers in the school, a dummy for ruralschools, and a lagged dependent variable.

Table A1: Any Absence in a Government School over the Electoral Cycle

Absent Log Absence

(1) (2) (3) (4) (5) (6) (7) (8)

-2 Years from Election 0.031∗∗∗ 0.030∗∗∗ 0.033∗∗∗ 0.037∗∗∗ 0.077∗∗∗ 0.074∗∗∗ 0.084∗∗∗ 0.089∗∗∗

(0.0003) (0.0003) (0.0003) (0.0003) (0.001) (0.001) (0.001) (0.001)

-1 Year from Election 0.004∗∗∗ 0.004∗∗∗ 0.010∗∗∗ 0.012∗∗∗ 0.007∗∗∗ 0.011∗∗∗ 0.027∗∗∗ 0.032∗∗∗

(0.0003) (0.0003) (0.0003) (0.0003) (0.001) (0.001) (0.001) (0.001)

1 Year from Election 0.027∗∗∗ 0.029∗∗∗ 0.021∗∗∗ 0.023∗∗∗ 0.043∗∗∗ 0.053∗∗∗ 0.034∗∗∗ 0.043∗∗∗

(0.0003) (0.0003) (0.0003) (0.0003) (0.001) (0.001) (0.001) (0.001)

2 Years from Election 0.033∗∗∗ 0.042∗∗∗ 0.024∗∗∗ 0.037∗∗∗ 0.071∗∗∗ 0.100∗∗∗ 0.055∗∗∗ 0.091∗∗∗

(0.0003) (0.0003) (0.0003) (0.0003) (0.001) (0.001) (0.001) (0.001)

Year FE No Yes No Yes No Yes No YesSchool FE No No Yes Yes No No Yes YesMean Absence in Election Year 0.117 0.052 0.068 0.078 0.327 0.147 0.19 0.213Number of Schools 1,222,930 1,222,930 1,222,930 1,222,930 1,220,015 1,220,015 1,220,015 1,220,015Observations 10,759,357 10,759,357 10,759,357 10,759,357 10,591,394 10,591,394 10,591,394 10,591,394

Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. Robust standard errors clustered at the school level in parentheses. The dependent variable in columns 1-4is a dummy variable that takes the value of one if the school reports any teacher absenteeism in that year, and the log number of absences in columns5-8. Each specification includes controls for the number of teachers in each school, a dummy for whether the school is in a rural area, and a laggeddependent variable. Column 4 corresponds to Panel A of Figure 1, and Column 8 corresponds to Panel B of Figure 1.

A2

Table A2: Any Absence over the Electoral Cycle in Private Schools

Absent Log Absence

(1) (2) (3) (4) (5) (6) (7) (8)

-2 Years from Election 0.003∗∗∗ 0.003∗∗∗ 0.002∗∗∗ 0.003∗∗∗ 0.005∗∗∗ 0.005∗∗∗ 0.003∗∗∗ 0.004∗∗∗

(0.0003) (0.0003) (0.0003) (0.0003) (0.001) (0.001) (0.001) (0.001)

-1 Year from Election 0.004∗∗∗ 0.005∗∗∗ 0.003∗∗∗ 0.004∗∗∗ 0.001 0.003∗∗∗ 0.004∗∗∗ 0.007∗∗∗

(0.0003) (0.0003) (0.0003) (0.0003) (0.001) (0.001) (0.001) (0.001)

1 Year from Election 0.011∗∗∗ 0.011∗∗∗ 0.007∗∗∗ 0.007∗∗∗ 0.020∗∗∗ 0.021∗∗∗ 0.013∗∗∗ 0.015∗∗∗

(0.0003) (0.0003) (0.0003) (0.0003) (0.001) (0.001) (0.001) (0.001)

2 Years from Election 0.006∗∗∗ 0.010∗∗∗ 0.002∗∗∗ 0.007∗∗∗ 0.006∗∗∗ 0.014∗∗∗ −0.002∗∗∗ 0.009∗∗∗

(0.0003) (0.0003) (0.0003) (0.0003) (0.001) (0.001) (0.001) (0.001)

Year FE No Yes No Yes No Yes No YesSchool FE No No Yes Yes No No Yes YesMean Absence in Election Year 0.041 0.007 0.024 0.021 0.089 0.021 0.064 0.056Number of Schools 527,097 527,097 527,097 527,097 521,152 521,152 521,152 521,152Observations 3,026,562 3,026,562 3,026,562 3,026,562 2,980,601 2,980,601 2,980,601 2,980,601

Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. Robust standard errors clustered at the school level in parentheses. The dependent variable in columns 1-4 isa dummy variable that takes the value of one if the school reports any teacher absenteeism in that year, and the log number of absences in columns 5-8.Each specification includes controls for the number of teachers in each school, a dummy for whether the school is in a rural area, and a lagged dependentvariable. Column 4 corresponds to Panel A of Figure 2, and Column 8 corresponds to Panel B of Figure 2.

A3

Table A3: Absence Over Electoral Cycle in Government Schools

Absent On Official Duty(1) (2)

2 or More Years Before Election 0.164∗∗ −0.243∗∗∗

(0.067) (0.033)

1 Year Before Election 0.080∗∗∗ −0.006(0.012) (0.038)

1 Year After Election 0.014 0.012(0.009) (0.037)

2 or More Years After Election 0.116∗∗∗ −0.086∗∗∗

(0.010) (0.032)

Male −0.001 0.060∗∗∗

(0.007) (0.018)

Age −0.001∗∗∗ 0.002∗∗

(0.0003) (0.001)

Muslim −0.026∗∗ 0.029(0.012) (0.036)

Other Religion 0.041∗∗∗ 0.114∗∗∗

(0.014) (0.038)

Upper Caste 0.003 0.037∗

(0.008) (0.020)

SC/ST 0.001 0.008(0.009) (0.022)

Other Caste −0.055∗∗ −0.178∗∗∗

(0.023) (0.027)

Distance from School 0.002∗∗∗ −0.001(0.0004) (0.001)

Constant 0.132∗∗∗ 0.122∗∗

(0.017) (0.048)

Observations 12,875 2,187Adjusted R2 0.021 0.025

Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. Robust standard errors in paren-theses. Linear models of the probability of a teacher being absent from agovernment school the day IHDS surveyors surveyed the school, as wellas the probability a teacher was absent on official duty as outlined in Equa-tion 2. The dependent variable is either a dummy variable that takes thevalue of one if the teacher was absent from the school on the day of the sur-vey, or, conditional on being absent, whether they were absent on officialduty on the day of the survey.

A4

Table A4: Absence Over Electoral Cycle in Private Schools

Absent On Official Duty(1) (2)

2 or More Years Before Election 0.051 −0.104∗∗∗

(0.056) (0.034)

1 Year Before Election 0.035∗∗∗ 0.069∗

(0.009) (0.039)

1 Year After Election 0.011 −0.032(0.007) (0.032)

2 or More Years After Election 0.114∗∗∗ −0.052∗

(0.008) (0.028)

Male 0.006 0.025(0.006) (0.015)

Age −0.0005∗ 0.005∗∗∗

(0.0003) (0.001)

Muslim 0.009 0.001(0.011) (0.031)

Other Religion 0.039∗∗∗ 0.065∗∗

(0.010) (0.030)

Upper Caste −0.033∗∗∗ −0.029∗

(0.006) (0.017)

SC/ST 0.012 −0.057∗∗∗

(0.009) (0.019)

Other Caste 0.024 −0.133∗∗∗

(0.032) (0.048)

Distance from School 0.002∗∗∗ −0.001(0.001) (0.001)

Constant 0.087∗∗∗ −0.001(0.012) (0.038)

Observations 14,184 1,810Adjusted R2 0.027 0.049

Notes: * p < 0.1, ** p < 0.05, *** p < 0.01. Robust standard errors in paren-theses. Linear models of the probability of a teacher being absent from aprivate school the day IHDS surveyors surveyed the school, as well as theprobability a teacher was absent on official duty as outlined in Equation2. The dependent variable is either a dummy variable that takes the valueof one if the teacher was absent from the school on the day of the survey,or, conditional on being absent, whether they were absent on official dutyon the day of the survey.