Arjun Soin Final PDF
Arjun Soin Final PDF
Arjun Soin Final PDF
Abstract
Mohalla Clinics are community-based clinics in India’s capital city Delhi set up in
2015 by the Aam Aadmi Party, a newly founded political party (2012) currently at
the helm of Delhi’s government. Mohalla Clinics routinely provide essential health
services including medicines, diagnostics, and consultation free of cost to traditionally
underserved urban residents. From the outset, the initiative was touted as one that
would bring free health care to the poor as well as more public and national recognition
to the Aam Aadmi Party as a champion of the ordinary citizen’s cause. In this paper, I
study whether the establishment of Mohalla Clinics affected the share of votes received
by its main opposition the BJP in municipal elections. Through this, I test whether
programmatic policy which is expected to generate added public goods is effective for
political patronage. I find suggestive evidence from OLS and IV results that the clinics
did reduce BJP vote margins, but less so in places that were strongholds. In addition,
the clinics generated substantial public goods in the form of unprecedented health care
access even though they may have fallen short in vote-garnering.
*E-mail address: asoin@stanford.edu. I would like to thank my thesis advisor, Professor Marcella Alsan,
for her continuous support throughout this thesis’s conception and writing. Thank you also to the Delhi
government’s Ministry of Health for supporting me with the data procurement process. Finally, thank you to
Professor Marcelo Clarici-Arias for his patience over the past several months.
1
Arjun Soin - May 2019
1 Introduction
Political patronage is a situation in which a person is rewarded for supporting a particular
politician or political party, such as campaigning or voting for them. Distributing patronage
for electoral gains is a commonly observed practice in developing countries. Notable research
shows that while jobs and money, two highly individualistic goods, constitute the most-
preferred benefits for voters in less developed democracies, a universal health care or free
education programs are the most popular types of benefits for voters in more developed
democracies.
In this paper, I examine whether the introduction of community-based health centers that
cater to the urban poor, increased electoral gains for the incumbent party. The distribution
of patronage is shaped by the nature of electoral competition that local politicians face. I
focus on Delhi – India’s capital and the world’s second largest city – and examine whether
primary health care as programmatic policy can function as a viable method for consolidating
electoral gains.
In particular, I examine the effect of Mohalla (community/neighborhood in Hindi) Clinics,
established by the Aam Aadmi (translates to common man) Party (AAP) in order to help
citizens meet basic health necessities for free (2). With a population of over 18 million,
Delhi is among the world’s largest cities. Summary statistics reveal that the number of
recorded patient visitations at these clinics in the first two years of data availability is
approximately a third of the population of Delhi. Delhi has a two-tier elected governance
structure – a state legislature and a municipal corporation – each of which controls a different
set of public services. Different political entities are incumbent in Delhi’s state legislature
and at the local municipal level. The AAP heads the government at the state level of
Delhi’s Legislative Assembly, while the Bharatiya Janata (translates to Indian People) Party
(BJP) that runs India’s federal government, is in power for Delhi’s local municipal body.
The fact that the clinics were named after the implementing party (AAP), mentioned in
manifestos (see appendix ) and brought up multiple times during election campaigning suggests
a vote-garnering purpose as well.
To test the hypothesis on vote-garnering, I bring several datasets together. First, I obtain
Page 2
Arjun Soin - May 2019
proprietary data from the Delhi Government on clinic presence and effectiveness through
variables such as clinic coverage by assembly, patient visits and physician effort from recorded
tests and referrals. I combine this data with Delhi State Election Commission’s data (see
appendix) on municipal elections contested before and after the establishment of Mohalla
Clinics. This electoral data is aggregated at the assembly constituency (AC) level, with each
AC containing a fixed number of wards that were contested on 2012 and 2017. I then merge
these datasets by mapping clinics to ACs. A ward delimitation exercise that led to a different
set of wards being contested in successive municipal elections necessitates aggregation at the
AC-level. This merged dataset for 70 ACs within the boundaries of Delhi allows me to test
for political patronage by AAP through constituency-specific changes in vote share of the
BJP. I also add columns for various assembly-level controls and year-specific fixed effects.
In order to causally identify the effect of Mohalla Clinics on vote share for the main
competing, non-implementing party (BJP), I rely on the conditional independence assumption
(CIA) that the clinics are randomly assigned. The potential confounding factors include:
health care coverage from hospitals and smaller facilities in a given AC, physician effort at
each clinic, MCD zone (east, south or north) of an AC and the number of wards reserved
for members of disadvantaged groups. This CIA helps me show crucial associations such as
overall negative effect of clinics on assembly-aggregated vote share of the competing political
party (BJP). Because of endogenity concerns with the location of Mohalla Clinics, I also use
an instrumental variables strategy for which I leverage the delimitation exercise mentioned
above. In particular, I instrument for the number of wards added to or removed from each
AC from a delimitation exercise which changed the boundaries for the 2017 MCD election.
This variable can be thought of as indicative of expected primary health care demand (or
congestion). As shown by Iyer and Reddy (14) these boundaries were the result of technical
expertise and impervious to gerrymandering. While I am unable to control for the number of
contemporaneous wards to account for any direct effects of ward size/number on the outcome
of interest due to collinearity, I control for other electoral variables capturing complex voter
dynamics from the change in boundaries.
I consistently find a negative clinic effect on the opposing party’s assembly-aggregated
MCD vote share for most variations and parameterizations of clinic impact. Aggregated at
Page 3
Arjun Soin - May 2019
the assembly-level, when vote share for MCD elections in 2012 is less than 34%,which is close
h
to the average, the clinics are associated with a net loss of −0.051 + (0.158 ∗ V oteShare)
percentage points (where V oteShare < 0.34) in 2017 votes for the BJP. I find suggestive
evidence that the clinics did reduce BJP vote margins at the assembly level, but less so in
places that were BJP strongholds. The IV further confirms a negative clinic effect on the BJP’s
assembly-aggregated 2017 MCD vote share, with every clinic reducing assembly-aggregated
MCD vote share by 18 percentage points.
This paper informs the political economy of redistribution. It confirms that putting in
place individually-targeted policies for patronage might be more effective for electoral gains.
In the context of primary health care in one of the largest cities in the world specifically, I show
that programmatic policy that more generally targets the public to generate added public
goods (such as widespread access to health care) as an approach to vote-garnering comes with
many restrictions in the short-run such as the implementing party’s experience at different
levels of government and lack of political patronage prospects in areas that are traditional
strongholds of a competing, non-implementing party. As these clinics continue to roll out
they could include more and more traditionally underserved citizens and serve as electoral
planks for the implementing party. Whether or not these clinics meet long-term patronage
goals for the pertinent party, it seems worthwhile pooling more resources and facilitating
cooperation between different levels of government and their underlying political entities to
ensure all planned 1000 clinics are installed, even at the cost of patronage-related outcomes
that seem to work well only under certain conditions of previous electoral competition. More
notably, this paper offers insights to political actors and policymakers in developing countries,
especially India, about certain expectations of electoral outcomes from programmatic policies
and the drawbacks of attempting to gain patronage through large-scale and more diffuse
public goods such as primary health care.
This paper proceeds as follows: First I give a background on political patronage and
evidence of its effectiveness in political economy, followed by relevant background on Mohalla
Clinics and Delhi’s political landscape. The third section covers data sources and formulae.
From there, I outline my empirical approach with subsections for each of my strategies. The
final two sections discuss empirical results and pertinent conclusions respectively.
Page 4
Arjun Soin - May 2019
2 Background
With regards to patronage, two distinctive practices carried out by political actors have
been described: Clientelistic Politics and Programmatic Politics (6). Clientelism refers to
the discretionary provision of private or local public goods or privileges such as jobs, water
or sanitation by government officials and political parties to particular groups of citizens,
in exchange for their votes. On the other hand, programmatic politics involves policies
such as social security with clearly defined eligibility rules based on publicly observable
characteristics such as location, age, gender, occupation, asset ownership, or citizenship. This
entitlement is not subject to discretion exercised by any political agent, elected official or their
representatives. According to Bardhan and Mookerjee (2017), platforms in programmatic
politics may be designed by political contestants to influence (future) political support from
specific constituencies, whereas vote buying is highly representative of a clientelist system
(6). Clientelistic practices tend to be more direct and transparent than programmatic ones
which are more diffuse and less targeted. It is for this reason that programmatic practices
are believed to be less efficient towards garnering votes. By nature, Mohalla Clinics fall more
into the programmatic category though this paper finds they serve (limited) clientilestic
ends as well. The aforementioned distinction is important since I interpret outcomes from
my regressions later in the paper in terms of what programmatic policy represents for
vote-garnering.
Previous work has explored whether programmatic policies affect electoral outcomes, both
by assessing voter responses and outcomes as well as macro-level trends of such policies. A
body of work that I talk about in this section also looks at clientilistc practices and their
associated political implications. I also discuss observational and experimental work that
addresses the social and electoral importance of public goods to citizens in India. Each unique
strand of aforementioned research on redistribution goes on to inform my story of Mohalla
Clinics being associated with electoral outcomes.
Frey (2015) shows that the expansion of Bolsa Familia, an arguably programmatic platform,
in Brazil reduced incumbency advantages of local mayors, increased political competition,
Page 5
Arjun Soin - May 2019
and increased health care and education spending shares (5). Bardhan et al (2017) estimate
the effects of political competition on voter responses to receipt of different kinds of private
benefits and local public goods for local governments in West Bengal, India. (7). They find
that voters were more responsive to benefits received from an incumbent that was considered
more likely to win the next election. It is also worth noting that most relevant literature takes
as given that broad public policies such as universal education or health care may be good
for growth but may be electorally ineffective (Vicente and Wantchekon, 2016) (8). The idea
behind such a claim is that such policies tend to be more diffuse and are less immediately
recognizable. With heightened electoral competition, implementing futuristic policies that
are likely to expand welfare in the long-run can harm political prospects.
Another salient consideration is whether the Mohalla Clinics inititiave would even change
voter behavior. In a discussion on India’s government-run health centers, Banerjee and Duflo
(2009) claim that they are unlikely to be highly valued by the poor (9). Their reasoning stems
from an observational approach towards outlining the shortcomings of India’s traditional
public health system. However, they also claim that affecting demand for health services
may not be too difficult as feedback mechanisms can provide more insight once demand is
stimulated. While their results could be extrapolated to conclude that the median voter
would not value free health care provision with high consideration, they examine provision
practices through randomized experiments. Extracting preferences from elections as I do in
this paper, however, can function as a more natural way to attach value to citizens’ perception
of healthcare centers.
As further evidence on primary health care as a public good in India, Banerjee, Pande
and Walton (2012) collect detailed survey data on the quality of social services available to
Delhi slum-dwellers to highlight governance constraints for a working paper published by
the International Growth Center (13). They find that health is only mentioned by some 7%
of survey respondents as a common problem for the household. Yet 70% choose a private
facility to treat minor health problems, and 43% for major health problems. Based on survey
results, it is claimed that people choose facilities primarily based on convenience/location and
perceived quality. Two-thirds noticed problems in government hospitals, especially time taken,
busy staff, no medicine, and rude staff. Though the authors do not present direct measures
Page 6
Arjun Soin - May 2019
on health status, 93% of households reported visiting a clinic or other health facility for a
minor problem in the six months before the survey was administered and 18% of households
report visiting a health facility visit for a major problem in the last year. These percentages
are roughly stable across wealth quintile. The authors claim that the use of government
facilities decreases with wealth for minor ailments, but there is no clear pattern of use for
major ailments among slum-dwellers/urban poor. This is consistent with the fact that people
are found to have a somewhat negative view of government health facilities (Das and Sanchez
2003), which was generalized from an observation of 1,621 individuals for 16 weeks through
detailed weekly interviews on morbidity and health-seeking behavior (12).
Tariq Thachil, a political scientist at Vanderbilt University, examines nonstate service
provision (provided by political parties as opposed to public goods by government) as an
electoral strategy in India (11). Taking a purely theoretical approach, he claims that service
provision is a “socially embedded tactic” well suited to helping elite parties with organizational
resources, but without pro-poor policy agendas, win over underprivileged electorates. While
inframarginal segments of the population will use them, the argument dictates that a pro-poor
policy is not the ostensible purpose as much as gaining political patronage.
Vicente and Wantchekon (2016) shed light on conditions under which clientelism works
particularly well in a context similar to India. This work helps me build a framework within
which to contrast the programmatic policy of this paper and possible clientilistic outcomes
(vote-buying) expected from it. They provide an overview of, discuss implementation, and
interpret the main results of recent field experiments conducted by West African countries
(8). According to them, the big picture is that while clientelism works particularly well for
incumbents, vote buying seems to be more effective for challengers. This pattern may be due
to the fact that clientelism is a safer way to buy voters – since voters then have the incentive
to vote accordingly (the counterpart for the vote comes only after the election and only if
the buying candidate is elected). On the contrary, cash-for votes before elections does not
embed any obvious enforcement (of transactions) mechanism. Cash payments and promises
of work are in any case more directly related to ensuring voter registration than carrying out
clientelistic politics.
Findings from this paper are sure to complement related work on redistribution in political
Page 7
Arjun Soin - May 2019
economy, some of which is discussed above. While much of the aforementioned research seems
to assume de facto that there is little reason for ordinary citizens to repose faith in political
entities for their provision of various public goods, the case for testing patronage through a
programmatic policy with a variation in level of government for political competition is what
makes this paper unique.
Delhi’s Mohalla Clinics were established in 2015 by the Delhi Government to routinely provide
essential health services including medicines, diagnostics, and consultation free of cost to
traditionally underserved urban residents. A new patient can walk into a Mohalla Clinic —
no appointment necessary — and receive a full physical examination from the doctor on staff
(2). The clinic offers all consultations, screenings, lab tests, and medications free of charge,
and asks the patient to register on a hand-held tablet for follow-up visits. Until July of 2018,
these clinics had recorded over 6 million patient visits (see data section). There are 200
clinics currently functioning and another 800 slated for opening by 2020. While former UN
Secretary-General Ban Ki-Moon and Norwegian Premier Gro Harlem Brundtland recently
visited and lauded Mohalla Clinics, the Stanford Social Innovation Review has also covered
them in a special article feature (1).
According to the World Health Organization data for the year 2015, more than 65% of
the population in India paid for health from their own pockets (3). In a country with one of
the largest out-of-pocket health expenditures in the world, the free health services offered by
Mohalla Clinics help reduce the financial burden on low-income households by saving travel
costs and lost wages, as they are spread across various neighborhoods in the city. While
Mohalla Clinics are fast becoming viable points of entry into the formal health system for
the urban poor, there have been no quantitative studies that measure their health/societal
outcomes. Considerations on voter outcomes and related incumbency effects of local municipal
elections post their establishment can begin to tell a story backed by the notion of political
patronage and clientelism. The idea is that since constituencies contested in local elections
Page 8
Arjun Soin - May 2019
spatially correspond to areas covered by various clinics, there is much insight to be obtained
from mapping clinic presence to voter outcomes. The timing of this initiative is of special
interest in this paper. The project was conceptualized in July 2015 by the Minister of Health
and Family Welfare, Government of NCT of Delhi, shortly after an assembly (state-level)
election. The pilot program with 105 Mohalla Clinics was launched in March 2016 (16).
With a population of over 18 million, Delhi is regularly ranked among the world’s largest
cities. Delhi has a two-tier elected governance structure – a state legislature and a municipal
corporation – each of which controls a different set of public services. Local elections in
Delhi are contested for the Municipal Corporation of Delhi (MCD). The MCD is among the
largest municipal bodies in the world; providing civic services to more than 11 million citizens
in the capital city. It has three sub-regions: North, South and East. This political body
was constituted as part of India’s local body decentralization (under the 74th constitutional
amendment), with the first set of elections held in 1997. MCD Elections occur at the
ward-level via plurality rule every five years. MCD elections tend to be party-based. In
2007, India’s two main national parties Bharatiya Janata Party (BJP) and Indian National
Congress (INC) had the largest share of councillors at 65% and 25% respectively. To take
this into account, I also control for INC vote share in my regression specifications. Turning
to the role of parties, India’s two main national parties – BJP and INC – were also the most
important parties in Delhi elections in the previous decade. The party candidate choice for
each ward is directly determined by party leaders and there are no primaries (13).
Today, a notable discrepancy arises from the fact that different political entities are
incumbent in Delhi’s state legislature and the MCD. The Aam Aadmi Party (AAP), a
party born out of an anti-corruption crusade in 2012 that set up Mohalla Clinics in 2015, is
incumbent at the state level from winning the 2015 elections to Delhi’s Legislative Assembly.
However, the Bharatiya Janata Party (BJP) that runs India’s federal government triumphed
in both, the 2012 and 2017 MCD elections and is currently at the helm of the MCD. As the
AAP did not exist as a full-fledged political entity in 2012, it did not contest MCD elections
in that year. This also means there is no baseline electoral performance for the AAP in the
Page 9
Arjun Soin - May 2019
2012 MCD elections. A detailed timeline of electoral events is given on the following page.
Delhi contains 272 wards (councils) making up the Municipal Corporation of Delhi, with
each of these wards roughly being situated in 70 assembly constituencies (ACs) composing the
state legislature. A political party must cross the simple majority figure, which is 136 wards
for MCD and 35 assemblies for Legislative Assembly, and have the highest number of votes
within each captured ward or AC, to win the election. Delhi is unusual in many respects.
The state government, referred to as the Government of the National Capital Territory (here
referred to just as the Delhi state government), covers an area that spans the boundaries
of the city of Delhi. In addition, there are two municipalities for the city. The MCD covers
most of Delhi’s population, but does not cover an important area in the center, that includes
the major national and state official buildings, parliament, the prime minister’s home, nearby
residential areas and some commercial areas. This area falls under the New Delhi Municipal
Corporation of Delhi (NDMC). Finally, some areas associated with the military fall under
the Delhi Cantonment area. In terms of spatial coverage, this paper includes only areas that
fall under the MCD. Given that two constituencies (New Delhi Municipal Corporation and
Delhi Cantonment) do not fall within the jurisdiction of the MCD and are missing Ward-AC
combinations, I am left with 68 ACs for analysis.
Due to a delimitation exercise post the 2011 Census of India (15), 160 of the 272 ward
boundaries were redrawn. While precise differences in area or composition of new wards
contested post the delimitation exercise are not available publicly, I pursue analysis at the
AC-level, given that a detailed ward, Mohalla Clinic and electoral composition of each AC
can be constructed for both pre and post delimitation periods (see appendix ). I aggregate
all variables across each ward contained in each of the 68 assemblies. The data matrix then
consists of 68 rows for each AC, instead of 272 for each ward. In 2012, each AC contained 4
wards while in 2017 after delimitation took effect, each AC contained 3-7 wards. To be clear,
elections for both years were contested and polled at the 272 MCD wards but are aggregated
at the AC-level in this paper for consistent follow up over time.
Page 10
Arjun Soin - May 2019
TIMELINE OF EVENTS
April 2012 • BJP wins MCD, winning 138 of the total of 272 seats (local municipal election)
December 2013 • AAP forms minority government having won 28 of 70 Legislative Assembly seats (state-level)
May 2014 • BJP wins National Election, including all 7 parliamentary seats in Delhi (federal election)
Febuary 2015 • AAP wins 67/70 seats to Delhi Legislative Assembly to form majority government (state-level)
September 2015 • Mohalla Clinics set up by AAP govt., announced as a part of health budget (state-level)
March 2017 • Over 100 Mohalla Clinics functioning and recording daily patient numbers
25th April 2017 • BJP wins MCD, winning 181 wards with AAP second at 48 wards (local municipal election)
Mohalla Clinics were repeatedly brought up during election campaigns. Universalizing primary
healthcare through Mohalla Clinics was at the center of AAP’s vision for the health sector
through each election it contested (manifestos in appendix). While the AAP repreatedly
emphasized its flagship initiative to appeal to the MCD voter, the BJP deemed it a colossal
failure backed by claims of doctor absenteeism and lack of medicine stockpiles. Moreover, the
BJP made an appeal to the Election Commission, the autonomous constitutional authority
responsible for administering election processes in India, crying foul over Mohalla Clinics
being officially named AAMCs or Aam Aadmi Mohalla Clinics as exemplifying “political
nepotism” (4). The appeal succeeded; the Election Commission instructed the AAP to
remove any identification of the party’s name from the government-funded clinics. I use this
fact to support an assumption that the attachment of AAP’s name with the clinics in some
way influenced and continually reinforced an association with the political party for patients
visiting. This in turn could have served as a mechanism by which health care delivery by a
party incumbent at one level (state) could guide voter actions towards it at another level
(MCD).
Page 11
Arjun Soin - May 2019
This defining event of the opposition party’s official request to the Election Commission
to have the AAP remove its name from Mohalla Clinics further establishes the idea that
public goods influence voter perception especially in the presence of identification that ties
an initiative to a political entity. The spike in the number of news articles containing the
keyword Mohalla Clinics closer to the 2017 MCD elections shows that primary healthcare
and electoral settings are not isolated from one another. To help visualize this claim, I explore
the NOW (News on the Web) Corpus (17) which contains over 7.4 billion words of data
from web-based newspapers and magazines to construct a time-series graph on mentions
of Mohalla Clinics (normalized by the “MCD elections” key phrase) in the news. Figure 2
shown below depicts the frequency of the phrase “Mohalla Clinics”, normalized by the phrase
“MCD Elections”.
The hypothesis I test is whether the Mohalla Clinics introduced by AAP affected electoral
outcomes. The analysis will enable me to assess how effective primary health care is as a
method of vote-buying in a developing country context.
3 Data
I analyze 2012 and 2017 MCD election data with an AC-level aggregation to maintain
consistency. It has already been mentioned that the 2011 delimitation exercise makes it
Page 12
Arjun Soin - May 2019
unfeasible to test at the ward level, given the 2012 and 2017 elections were contested in a
(relatively) different set of wards. While both elections were won by the BJP, the latter was
contested against a new opponent (AAP) after its flagship initiative Mohalla Clinics had
begun. The number of clinics within each AC (summed up over each ward within that AC)
as well as related parameterizations of clinic presence serve as proxies for health care coverage
through Mohalla clinics within a given area.
Data on vote share for the 2012 and 2017 MCD elections is available on the Delhi Government
online portal for each MCD ward contested (see appendix). For both elections, I obtain
ward-level data on electoral outcomes, which provides information on turnout, candidate list
(with party) and candidate-wise vote-share. The relevant files contain a tabular listing of
each ward with each candidate’s name, political affiliation and votes accumulated. I import
the file for each ward as a dataframe and manipulate the data on python to arrive at total
BJP votes for each ward within an AC as the numerator for the equation below. Summing
the total votes polled for each ward within an AC leads to the denominator for the Vote
Share equation. For a given assembly constituency,
Page 13
Arjun Soin - May 2019
Votes to BJPyear,AC
Vote.Shareyear,AC = (1)
Votes Polledyear,AC
Note: Votes to BJP is the sum of votes from each ward i located within that AC
I generate BJP vote shares as variables because the BJP is consistently the main com-
peting political party at all levels of Delhi’s government to the AAP, the party responsible for
setting up and maintaining Mohalla Clinics as a public service. The empirical strategy to be
outlined in this paper is based on using BJP’s vote share in each AC as a proxy for electoral
competition that a programmatic initiative like Mohalla Clinics could dampen, hence giving
quantitative insight into the distribution of political patronage by the AAP.
After drafting a project proposal to the Delhi Government, I was able to obtain data on
Mohalla Clinics since the early months of 2016, which is roughly when the clinics opened and
began collecting variables like OPD numbers, tests prescribed and number of referrals on
data tablets daily. I capture the first two years of data availability in the given summary
tables. Numbers on patient attendance and tests prescribed show a relatively quick uptake
of these clinics:
Table 1: Patient visits and tests recorded over two year period of April 2016 - March 2018
The recorded patient visitation numbers correspond well to the order of magnitude of the total
votes polled across the 2012 and 2017 MCD elections. Though not directly indicative of an
Page 14
Arjun Soin - May 2019
overlap, it is safe to assume that many MCD voters repeatedly interfaced with Mohalla Clinics.
Hence it is feasible to investigate whether the political patronage mechanism stemming from
giving these clinics an AAP name or the recognition of quality healthcare free of cost by
voters is at play. The dataset obtained from the government also contains an address/location
for each functioning clinic. Using a shapefile containing geospatial information on Delhi’s AC
boundaries, I create a mapping between each functioning clinic and AC at the time of the
2017 election. I also use this shapefile to extract the area (in sq. km) of each constituency
that is later used as a control variable in my regressions.
I obtained multiple monthly datasets containing patient OPD visits, tests and referral
numbers for each Mohalla Clinic. Aggregating these, I arrive at yearly numbers for each AC
in the dataset. The year in consideration is April 2016 - April 2017, in sync with the election
year for which vote share is calculated.
Summary statistics for relevant variables used in the analysis are reported in Table 3.
On average, each AC in Delhi contained close to 2 Mohalla Clinics at the time of analysis.
Furthermore, 77% of the assemblies contain at least one Mohalla Clinic. The average ratio of
patient visits on AC population is approximately 0.162. Electorally, it is interesting to note
that BJP’s average vote share does not really change that much, only falling 0.4 percentage
points from 2012 to 2017. Based on this, I expect the variation between constituency vote
shares to tell the story of political patronage more than the average vote share numbers.
Panel (b) of Table 3 reports descriptive statistics for select control, fixed-effect and other
variables used later in my analysis. Outside of healthcare being provided by Mohalla Clinics,
each AC has on average almost 2 hospitals and 10 smaller health facilities providing fewer
services at a smaller scale. The average AC area in sq. km (extracted in relative terms from
Page 15
Arjun Soin - May 2019
the shapefile) is 25.7 sq.km. At the time of the 2017 MCD elections, 64 of the 68 ACs were
represented by a non-BJP Member of Legislative Assembly.
Page 16
Arjun Soin - May 2019
4 Empirical Approach
To estimate the impact of the health care provision through Mohalla Clinics on electoral
outcomes, my initial identifying assumption is that the presence of clinics is more or less
random, conditional on covariates described before. Meetings with government officials
confirmed that locations of clinics are not strategically determined (other than keeping them
localized enough to cater to the urban poor). Before the clinics opened up, the government
gave out 3 advertisements asking for people to rent out their properties which the government
was willing to rent at the market price. Once the places were finalized, they were modified as
per a common model developed with direct involvement of the Minister of Health. They were
set up and began functioning as and when private property/government land owners were
willing to rent out space, a process whose outcome is challenging to predict and so is likely
to be randomly distributed. Given that only 200 of 1000 planned clinics are functioning at
various locations, I begin by treating clinics as good as randomly assigned.
As there could still be some doubt over the random assignment of clinics, I run multivariate
ordinary least squares (OLS) regressions under the conditional independence assumption
(CIA) where I include potential confounding factors as control variables in the regression
specifications. These control variables account for complex health care and electoral dynamics
that could independently explain vote share of the BJP and any understanding of a clinic effect
on the BJP’s vote share only seems complete when these dynamics are taken into account.
Within the multivariate regressions, equations with necessary controls and fixed-effects
represent my preferred specifications.
The argument on random assignment underlying the above multivariate regressions could
be countered on the grounds of the inconsistency between the decision to start Mohalla
Clinics being strategic/political and the location being random. It is precisely due to such an
endogeneity concern that I incorporate an instrumental variables strategy as the third and final
strand of my empirical section, where I instrument for Mohalla Clinics with the ward-number
difference for each AC after the MCD delimitation exercise of 2011 was conducted.
Page 17
Arjun Soin - May 2019
I first use multivariate OLS regressions. Under the CIA, dependence between treatment
assignment and treatment-specific outcomes can be mitigated by conditioning on potential
confounding factors. In this model, some salient confounding factors are: other hospitals and
health facilities in the area, structure of reserved wards for certain population categories,
MCD zone within which AC falls and whether or not the AC state legislator belongs to the
non-implementing party (BJP). Variables representing these factors are included as control
variables in my regression equations.
I generalize a framework that allows for the treatment variable to take more than 2 values.
Suppose we focus on a given parametrization of the clinic variable, define the potential
outcome associated clinic impact c as:
Yci = fi (C)
where we put an i -subscript on the f (.) function to show that the potential vote share
obtained is AC specific. The CIA in this general setup then becomes:
Page 18
Arjun Soin - May 2019
BJPVote.Sharei,2017 = α + β1 Mohallai + , (2)
BJPVote.Sharei,2017 = α + β1 Mohallai + β2 (BJPVote.Sharei,2012 ) + β3 Mohallai ∗ BJPVote.Sharei,2012
+X T Ω + , (3)
BJPVote.Sharei,2017 = α + β1 lnMohallai + β2 BJPVote.Sharei,2012 + β3 lnMohallai ∗ BJPVote.Sharei,2012
+X T Ω + , (4)
where equation (2) is a simple specification to test β1 for clinic variable effect (without
interaction), equation (3) is a multiple framework with an interaction between the clinic
variable and previous (2012) vote share and equation (4) replicates equation (3) with a linear-
log model. Furthermore, BJPVote.Share2017 and BJPVote.Share2012 are vote shares of the
BJP in the 2017 and 2012 MCD elections (aggregated across all wards in AC i) respectively,
M ohallai is a variable that takes on various parameterizations of Mohalla Clinic health care
coverage variables outlined in the previous section (Clinics, ClinicDummy, Physian Effort
and Patient Visits), and vector X includes confounding factors as control variables.
For the regression coefficient for BJPVote.Share2012 , while I expect to observe a positive
sign, it is to be kept in mind that there are still multiple unobserved factors influencing this
coefficient, such as measures of performance, accountability and public goods’ provision that
ultimately determine the fate of incumbents. The interaction between BJPVote.Share2012
and Mohallai allows me to test whether the Mohalla Clinics effect differed in BJP assembly-
aggregated MCD strongholds. The effect could be positive or negative. i.e. there could have
been a bigger effect of clinics on vote share in places that were heavily BJP or a muted effect
Page 19
Arjun Soin - May 2019
(the latter is what I expect to find). The test is whether the places with a higher BJP vote
share were more or less affected by the introduction of Mohalla clinics.
Based on the interactions, I compute inflection points in assembly-aggregated 2012 MCD
vote share below which clinics lead to a net loss in 2017 votes for the competing political
party (BJP). The inflection points and related inferences are presented in the Results section.
The inflection point calculation serves as a step towards identifying a political patronage
distribution through the Mohalla Clinics’ programmatic approach.
The multiple regressions outlined above are still susceptible to endogeneity issues. A
specific endogeneity problem stems from the fact that even with controls, clinics might
not be considered completely random. As they were set up and began functioning as and
when private property/government land owners were willing to rent out space, there is an
inconsistency between the decision to start Mohalla Clinics being strategic/political and the
location being random. The advertisements for renting out space could have been targeted
in assemblies known to be especially weak for assembly-aggregated municipal vote share of
the BJP. The public’s willingness to rent out space for an initiative by a state government
formed by a political party born out of a populist anti-corruption crusade (see timeline)
could have been at an all-time high in the initial few months of the advertisements. Such
theoretical possibilities create pathways that necessitate a more sophisticated statistical
approach towards isolating any clinic effect as causal. Despite the concerns, the directionality
of association between clinic presence and vote share of opposing party that are presented in
the CIA results acts as suggestive evidence of a trend in voter behavior that points towards a
rising interaction with primary health care and increasing salience of programmatic public
services like primary health in electoral considerations.
I now employ a widely used technique for dealing with endogeneity problems – namely, when
one or more of the regressors (X) could be correlated with the error term (u). The possibility
of inconsistent parameter estimation due to endogenous regressors in multiple regressions from
the previous section cannot be discounted. In this case, the regression estimates measure only
the magnitude of association, rather than the magnitude and direction of causation, which
Page 20
Arjun Soin - May 2019
is still encouraging as a first for Mohalla Clinics. For instance, AAP might have targeted
Mohalla Clinics (and related advertisements) in areas where BJP was weak; thus suggesting
a negative bias. Yet another mechanism for negative bias could be the fact that clinics in
areas with weaker BJP vote share have been generating higher volumes of patients and higher
associated physician effort, due to lack of BJP political influence on clinics’ functioning from
political rivalry or vendetta. I use IV to overcome such biases. The main advantage of using
IV is that I am explicit about the source of variation used to evaluate the impact of clinics
on vote share.
In this paper, I use the difference in wards from delimitation in 2011 as instrument for
the Mohalla Clinics. Because this variation is arguably orthogonal to vote share, it mitigates
concerns about the causal interpretation of the results. For this context, the orthogonality
comes from a few notable facts. First, 33 out of 68 ACs in my data have had relatively
static ward compositions, while for the others delimitation primarily lead to geographically
coherent areas with a similarly sized electorate. Second, since 2017 was the implementing
party AAP’s first MCD election, its MCD voter dynamics are defined by the 2017 MCD
election alone. This means that no 2017 voter would have previously voted for the AAP
in an MCD election. Despite delimitation, the first-time votes for the AAP would almost
certainly be redirected from previous BJP or Congress votes, with Congress vote share
controlled for and BJP vote share being the dependent variable in my equations. Most
importantly, the official delimitation directive claims that wards were redistricted within the
relevant AC, not outside of it (15). Since I aggregate at the assembly-level, the composition
of voters within that assembly remains the same, with just the populations of wards being
adjusted in accordance with the 2011 Census. This context is unique since even if delimitation
introduces uncertainties in electoral outcomes, the cumulative outcome at the assembly-level
still represents distinctive voter preferences from each ward.
The IV addition to my empirical approach allows me to better estimate the effect of
Mohalla Clinics on BJP vote share while bringing me closer to establishing causality, even
with observational data. Inferences and conclusions on political patronage can also still be
drawn from IV regressions. In essence, the IV approach in this paper is adopted to confirm
the observed negative direction of clinic presence on BJP vote share.
Page 21
Arjun Soin - May 2019
Page 22
Arjun Soin - May 2019
could bring competing political influence from previous elections into the fray. However,
another finding from the paper above is that the process of delimitation appears to place
few restrictions on the ability of political incumbents to contest for re-election. Given this
was observed at a higher-level of government than the MCD with higher numbers of citizens
involved, I expect the same to be the case for MCD wards. To clarify, I look into the
official Delhi State Election Commission ‘Principles of Delimitation’ document available on its
website and find pertinent guidelines (appendix) that strengthen the argument that redrawn
boundaries remain sacrosanct and are outcomes of a well-defined delimitation process.
Another real concern for the IV is whether changing the boundaries for the MCD directly
affected electoral outcomes. Even in the absence of gerrymandering, this could well have
happened. Since altering ward boundaries alters voter composition, traditional stronghold
areas are subject to change. The delimitation led to critical changes in the demography
and character of the wards. However, any delimitation of wards only occurred within the
boundaries of a relevant assembly and not outside it. Aggregating ward outcomes at the
assembly level ensures that inter-ward compositional changes are accounted for. I realize that
no instrument can be perfect and that I will have already shown suggestive evidence using
CIA from the previous regression specifications. Ideally, I would like to control for the current
number of wards to capture the independent effect that the 2017 ward structure had on the
BJP vote share. However, my structural model would then be subject to collinearity since
each AC contained the same number of wards before delimitation. Nevertheless, I control for
other electoral variables capturing complex voter dynamics from the change in boundaries.
The concern on accounting for 2017 electoral dynamics is also mitigated by the fact that half
of the ACs in my sample remained static in terms of ward structure.
BJPVote.Sharei,2017 = π0 + π1 Clinicsi + X T Ω + ,
(5)
where vector X includes confounding factors as control variables for ACs. Now we
have wardChangesi (our instrument), a variable not included in the structural model which
Page 23
Arjun Soin - May 2019
becomes the explanatory variable for the following first stage estimating equation. The
reduced form equation in the “first stage” is given by:
Clinicsi = π0 + π1 wardChangesi + X T Ω + ,
(6)
which allows me to test more formally the relevance through corr(Clinics, wardChanges) 6=
0. I focus on Clinicsi as the main endogenous variable and wardChangesi as the instrument
to keep the model ‘identified’. While my CIA preferred specification normalized the Clinicsi
variable by Areai , I now control for area of each AC in the structural model to justify having
the Clinicsi variable as the preferred approach for the IV. Using estimates from the above
equation, I now generate a new set of values for the variable:
I present results for the above model in the next section. Overall, I am able to better
isolate the negative clinic effect using wardChanges as an instrument given of the arguments
outlined earlier. The instrumental variable regression and associated results are generated
from linearmodels package in python, from which I use the IV-2SLS function. The package
also lets me output a first-stage result for each IV-2SLS model along with framing a comparison
with OLS models.
Page 24
Arjun Soin - May 2019
5 Results
First, I report results from a simple specification without the interaction between a Mohalla
Clinic variable and BJP’s 2012 vote share. From column (1) in Table 4, we see that each
additional Mohalla Clinic leads to a 1.4 % decline in BJP vote share, relative to the average.
The coefficient average generates a negative effect across the board - if 100% of the people
in a given AC visited Mohalla Clinics, AC-aggregated BJP vote share would drop by 2.3
percentage points.
ClinicDummy −0.002
(0.019)
PhysEffort 0.019
(0.135)
ClinicVisitsByPop −0.023
(0.050)
Observations 68 68 68 68
∗ ∗∗ ∗∗∗
Note: p<0.1; p<0.05; p<0.01
Page 25
Arjun Soin - May 2019
In tables 5 and 6, I include interactions between clinic presence variables and 2012 BJP
AC-aggregated vote share. I report results with standard (Table 5) and log parameterizations
(Table 6) of clinic variables from the multiple framework specification.
ClinicDummy −0.023
(0.065)
PhysEffort −0.445
(0.454)
ClinicVisits −0.310
(0.231)
Observations 68 68 68 68
∗ ∗∗ ∗∗∗
Note: p<0.1; p<0.05; p<0.01
Page 26
Arjun Soin - May 2019
logPhysEffortByArea −4.194
(2.749)
logClinicVisitsByArea −0.004
(0.007)
logClinicVisitsByPop −0.354
(0.270)
Observations 68 68 68 68
R2 0.572 0.545 0.543 0.542
Adjusted R2 0.448 0.414 0.411 0.421
Residual SE 0.049 (df = 52) 0.050 (df = 52) 0.050 (df = 52) 0.050 (df = 53)
∗
Note 1: p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01
Note 2: :(N) represents (V oteShare2012 ∗ ClinicV ariable) from Column (N)
Page 27
Arjun Soin - May 2019
Table 6, specifically Column (1), contains my preferred specification. Results from Table 6
contain clinic variables normalized by area and population to account for these historical
demographic differences of ACs. Log transformations of the variables make the distributions
less skewed and results more interpretable. From tables 5 and 6, we consistently see a negative
clinic effect for every variation and parameterization of clinic impact. However, the clinic
effect is less negative where the BJP had a higher vote share in the 2012 MCD elections.
This nuance is what makes β3 (the coefficient associated with the interaction) positive.
To look at specific inflection points, I take columns (1) - (4) from Table 6 and determine
an upper bound (detailed steps in appendix) for x on the inequality:
Based on the interactions and the given inequality, when vote share for 2012 is less than
34%, which is close to the average, the clinics lead to a net loss in 2017 votes for the BJP.
But for places above the mean, the presence of the clinic actually bolstered the BJP. While
on one hand, having a previous assembly-aggregated MCD vote share below the mean for the
BJP is associated with a negative effect in 2017 from AAP’s Mohalla Clinics, clinics do not
dampen vote share where the BJP was already above the mean for a given AC. For assemblies
where the BJP was already above mean vote share (assembly-aggregated), Mohalla Clinics
seem purely programmatic - generating added public goods but no political patronage gains.
That BJP vote share is actually bolstered above the mean is also interesting. The BJP has
been running the MCD since 2007. Mohalla Clinics, though catering to a significant fraction
of voting citizens, were unable to meet clientilistic outcomes in areas where previous levels of
electoral confidence in the BJP was relatively high. That the Election Commission’s decision
to remove the AAP’s name from the clinics deterred an electorate that had already supported
the BJP previously from succumbing to a new party’s clientilism is one likely explanation.
However, the diffusion argument from Banerjee and Duflo about immediate recognizability of
public goods can also account for the association - it is highly likely that in areas with high
previous vote share, the BJP had already been focusing on providing ‘electorally effective’
and less programmatic goods and services through the MCD.
With log parameterizations, results from Table 6 show a probabilistic picture with
Page 28
Arjun Soin - May 2019
Similar calculations for other variables confirm the 34% vote share cutoff for Mohalla
Clinics’ negative effect on the competing party. Interpreting this in log terms: for an
AC i where BJP secured less than 34% cumulative vote share across its wards in 2012; a
Page 29
Arjun Soin - May 2019
100% change in (or doubling of) the number of clinics within its boundaries, with other
electoral and health-related factors controlled for, would have dampened 2017 vote share
by −0.051 + (0.158 ∗ V S), where V S (< 0.34) is the exact vote share received. Beyond the
34% mark, vote share is bolstered by −0.051 + (0.158 ∗ V S), where V S > 0.34 (significant
at the 5% level). This suggests a distribution on political patronage that is based on a
cutoff from previous electoral competition. A cutoff-based distribution further suggests that
programmatic policy is an effective tool for vote-garnering only under certain conditions.
While the AAP could leverage its incumbent position at the Assembly level to provide Mohalla
Clinics, the clinics are effective for electoral gains on one side of the inflection point and
detrimental on the other. The suggestive evidence from my regression specification shows
that an association between programmatic policy and electoral favor is neither absolute nor
uniform, but that such policy provides true public services (universal access to health care)
nevertheless – which can also be seen in a significant and negative correlation between 2012
MCD vote share and clinic presence in 2017, considering AAP did not even contest in 2012.
The fact that there is significant suggestive evidence of Mohalla Clinics, a programmatic
policy, being effective towards electoral favor in itself is salient in supplementing existing
literature on the political economy of redistribution.
5.2 IV Results
I first present the first-stage regression where Clinics is the outcome and wardChanges is the
independent variable. The first-stage result is presented with and without year-AC controls.
As mentioned earlier, in order to capture confounding factors existing within the boundaries
of an AC, the model with control variables included is the preferred one. While I realize there
exist fluctuations in voter dynamics from changes in assembly-ward composition, electoral
controls such as reservation structure, zone of MCD and national parliamentary constituency
of an AC are included with an aim to mitigate concerns on these changing dynamics. I then
present model comparisons between OLS and IV to confirm the more robust estimator in
this context.
Table 7 presents the first-stage relationship between the wards added or removed from an
assembly and the clinics in an assembly. As the table shows, more wards added to an assembly
Page 30
Arjun Soin - May 2019
from delimitation is strongly positively correlated with higher clinic presence. Assemblies
where one ward was added from delimitation saw a 0.63 increase in chance of having at least
one functioning Mohalla Clinic within its boundaries, a value significant at the 1 percent
level. Moreover, the F-statistic of 7.87 suggests that wardChanges is unlikely to be a weak
instrument. On running the Wu-Hausman Test for Endogenity on the given model for the
null hypothesis that all endogenous variables are exogenous, I get a statistic of 33.8642 with
a p-value of 0.0001. The test evaluates the consistency of an estimator when compared to
an alternative, less efficient estimator which is already known to be consistent (OLS and
IV). One interpretation of the small p-value is that OLS estimates are not consistent and IV
should be applied.
(1) (2)
Observations 68 68
F-Statistic 13.15∗∗∗ 7.87∗∗∗
R-Squared 0.1740 0.5692
∗ ∗∗ ∗∗∗
Note: p<0.1; p<0.05; p<0.01
Standard Errors Reported in Parentheses for Table 7
Page 31
Arjun Soin - May 2019
Table 8 shows comparative results for OLS and Instrumental Variables Models. The table
shows variations of each model with and without pertinent controls. Since I test for political
patronage of primary health care clinics, a model that does not account for confounding
electoral and health care factors of ACs is almost certainly incomplete. Adding relevant
controls to remove the effect of these factors is likely to make my analysis more robust,
and leads to a model including confounding variables being the preferred one (Column (4),
Table 8). I find a relatively large percentage point gap in IV and OLS estimates, with a
significant result at the 1 percent level only for the IV model (Column (4), Table 8) that
includes year-fixed effects and AC controls and is the preferred specification as mentioned
earlier. This model is also found to have an R2 value of 0.7391, suggesting a high explanatory
power. There is a possibility that OLS estimates might be underestimating the effect of
AAP’s clinics on BJP’s vote share due to unobserved variations in demand for primary health
care, which are brought into the IV model from looking at the difference in ward-structure
after delimitation. Specifically focusing on the IV model from column (4) in Table 8, the
presence of an additionally functioning Mohalla Clinic in an assembly is associated with
an 18.2 percentage point decrease in the assembly-aggregated vote share for BJP in the
2017 MCD elections. OLS estimates suggest the same negative clinic effect though at the
magnitude of of 0.005 percentage points and insignificantly. For the AAP, the IV result points
towards an estimated 0.182 higher probability of gaining political favor from the functioning
of programmatic policy through Mohalla Clinics. Note that Table 8, Column (1) repeats the
findings from Table 4, Column (1) covering the OLS-specification of clinic variable effect.
Page 32
Arjun Soin - May 2019
The results significantly confirm a negative Mohalla Clinic effect on BJP vote share. While
a cutoff that reverses electoral favor can still exist even within the IV framework, moving
from an IV model without controls to one with controls further breaks down the story of
political patronage through primary health care. Since a significant number of controls
indicate whether clientelistic or programmatic practices can occur, more robust results from
the controlled-IV model indeed supplement findings from regressions in the previous empirical
section. A caveat here is that the IV model could well be capturing complex voter dynamics
from changes in ward compositions that I claimed cannot be ruled out in the exclusion-
restriction. The −0.182 coefficient value is in the absence of a cutoff on previous (2012) vote
share as seen in the CIA section and so can seem quite large. From the model of patronage I
have been following, this would mean there must still be a clear welfare-driven, programmatic
goal behind Mohalla Clinics. Again, considering that primary health care tends to be diffuse,
a consistent negative clinic effect is still remarkable. Granted that in the IV model this
Page 33
Arjun Soin - May 2019
negative effect could also be reflective of delimitation-specific outcomes, I believe that adding
an independent variable that controls for the current (2017) number of elections could go
a long way. However, since each AC contained 4 wards pre-delimitation, adding a control
for the current number of wards would have resulted in multicollinearity issues. That said,
this finding of a negative clinic effect on BJP vote share in sync with my preferred CIA
specification is invaluable in confirming that programmatic policy can dampen electoral
prospects of the non-implementing party.
Page 34
Arjun Soin - May 2019
or places that had higher than average vote share for municipal wards aggregated by AC
containing them. Since the AAP was ruling the state government at the time of municipal
elections yet aiming to take over the MCD, dampening BJP’s vote share at the municipal
level can be viewed as gaining electoral favor through patronage. The numeric on a cutoff
vote share showing strongholds resisting patronage informs me that programmatic politics is
a crucial element of public service offerings like Mohalla Clinics and that they generate added
public goods in the absence of absolute political favor. The instrumental variable strategy
more conclusively confirms a negative effect of Mohalla Clinics on BJP’s vote share when
treating expected demand from wards added/removed as orthogonal to assembly-aggregated
vote share for the 2017 elections.
This paper informs the political economy of redistribution. It confirms that putting in
place individually-targeted policies for patronage might be more effective for electoral gains
as is concluded in relevant literature covering less diffuse and quicker public services. This
paper steps in to show, in the context of primary health care in one of the largest cities in
the world, that programmatic policy that more generally targets the public might generate
added public goods (such as widespread access to health care) but such an approach to
vote-garnering comes with many restrictions in the short-run such as the implementing party’s
experience at different levels of government and lack of political patronage prospects in areas
that are traditional strongholds of the non-implementing party. For instance, a significant
and negative correlation between the BJP’s 2012 MCD vote share before Mohalla Clinics
were setup and clinic presence in 2017 points towards an earnest provision of primary health
care by the AAP Delhi government especially considering the AAP did not contest in 2012.
The paper complements existing work in confirming that programmatic policy will always
have limitations when it comes to vote-garnering, and that clinetelistic policy may be more
effective despite being unable to generate long-term, tangible public goods. In other words,
with a negative clinic effect dwindling in assemblies where aggregated municipal ward vote
share for the main opposing party (BJP) was above average (34%), and the pre-electoral
removal of the AAP’s name from the scheme — it is a formidable possibility that Mohalla
Clinics generated widespread primary health care access and were electorally ineffective. I
expect that as clinics increase coverage across Delhi, strongholds become less competitive and
Page 35
Arjun Soin - May 2019
AAP’s sweeping initiatives like Mohalla Clinics at the state level will have to be balanced by
increased accountability of BJP at the MCD.
Considering preconceived notions on people’s negative beliefs of primary health care
mentioned earlier in the paper, one would not expect to see much faith reposed in Mohalla
Clinics by ordinary citizens. However, the fact that, even after controlling for pertinent
variables at the assembly level, the trends in electoral outcomes differentially favor AAP
based on Mohalla Clinic outcomes shows that the initiative has been put in place despite
non-uniform political favor/electoral gains at avail. It is well possible that in the future
there are larger clientilistic implications of Mohalla Clinics. However, only if a standard
of care at these clinics is maintained will primary health care continue to function as (a
limited) means of political patronage (due to its programmatic nature). As claims of doctor
absenteeism, unavailability of medicines, lack of promptness in completing the 1000-clinic
project or political nepotism claims by opposition parties ramp up, Mohalla Clinics could
also end up as most other primary health experiments in developing nations - present but
merely in the negative beliefs of voters.
Page 36
Arjun Soin - May 2019
References
[1] Noël Duan, Health Care in the Mohallas. Stanford Social Innovation Review, Stanford,
2017.
[2] C Lahariya, Mohalla Clinics of Delhi, India: Could these become platform to strengthen
primary healthcare?. J Family Med Prim Care, 2017;6:1–10, [PMC free article]
[3] World Health Organization Global Health Expenditure database, Out-of-pocket expen-
diture (% of current health expenditure). apps.who.int/nha/database.
[4] Naming clinics as Aam Aadmi political nepotism: BJP, Economic Times, July, 2015.
[5] Anderson Frey, The local political economy of conditional cash transfers in Brazil (T).
UBC Theses and Dissertations, University of British Columbia, 2016.
[6] Pranab Bardhan and Dilip Mookherjee, A Theory of Clientelistic Politics versus
Programmatic Politics, 2018.
[8] Pedro C. Vicente, Leonard Wantchekon, Clientelism and vote buying: lessons from field
experiments in African elections. Oxford Review of Economic Policy, Volume 25, Issue
2, Summer 2009, Pages 292–305.
[9] Abhijit Banerjee, Esther Duflo, Improving health care delivery in India. MIT, 2009.
[10] Simeon Nichter, Electoral Clientelism or Relational Clientelism? Healthcare and Steril-
ization in Brazil. SSRN Electronic Journal, 10.2139/ssrn.1919567, 2011.
[11] Tariq Thachil, Embedded Mobilization: Nonstate Service Provision as Electoral Strategy
in India. World Politics. 63. 434 - 469. 10.1017/S0043887111000116, 2011.
Page 37
Arjun Soin - May 2019
[12] Jishnu Das and Carolina Sanchez-Paramo, Short but not sweet - new evidence on short
duration morbidities from India (English). Policy, Research working paper ; no. WPS
2971. Washington, DC: World Bank, 2003.
[13] Abhijit Banerjee, Rohini Pande and Michael Walton, Delhi’s Slum-Dwellers: Depri-
vation, Preferences and Political Engagement among the Urban Poor. International
Growth Centre, London School of Economic and Political Science, 2012.
[14] Lakshmi Iyer and Maya Reddy, Redrawing the Lines: Did Political Incumbents Influence
Electoral Redistricting in the World’s Largest Democracy? Harvard University Working
Paper, 2013.
[16] Directorate General of Health Services, Aam Aadmi Mohalla Clinics, Write Up. Health
and Family Welfare, www.delhi.gov.in, 2015.
[17] Mark Davies, The new 4.3 billion word NOW corpus, with 4-5 million words of data
added every day. The 9th International Corpus Linguistics Conference, Birmingham,
2017, Available at: https://www.english-corpora.org/now/.
Page 38
Arjun Soin - May 2019
APPENDIX
Variable Definitions
Electoral Variables
• Assemblyi : A list of size N = 68, with each entry representing an assembly constituency
in Delhi that falls under the MCD jurisdiction. There are two missing values for the
two assembly constituencies outside of the MCD purview.
• BJP V ote.Shareassembly,year : Vote Share obtained by the BJP aggregated for each
ward located inside Assemblyi during an MCD election of given year, where year =
{2012, 2017}.
• reservedW ardsSC: The number of wards in Assemblyi that are reserved for a member
of the Scheduled Caste. Candidates of General category are not eligible to contest from
these wards. Reservation in a ward is an outcome of the share of population comprising
the specific community.
• Clinicsi : The number of active Mohalla Clinics in assembly i at the time of the 2017
MCD elections. An “active” clinic is one from which regular patient numbers were being
recorded during the assembly year before the elections. The image on the next page
shows a sample portion of the assembly - clinic mapping available from the government.
• ClinicV isitsi : The number of recorded patient visits in assembly i at Mohalla Clinics
for the assembly year before 2017 MCD elections. This variable is aggregated from
monthly datasets obtained from the government.
• ClinicV isitsByP opi : The number of recorded patient visits in assembly i normalized
(divided) by assembly population for the assembly year before 2017 MCD elections.
Page 39
Arjun Soin - May 2019
This variable is aggregated from monthly clinic-specific datasets obtained from the
government. In other words:
ClinicV isitsi
ClinicV isitsByP opi = (A1)
P opulationi
• ClinicT estsi : The number of recorded medical tests in assembly i at Mohalla Clinics
for the assembly year before 2017 MCD elections. These are tests conducted from the
official list of free 212 diagnostic tests available at Mohalla Clinics.
• ClinicRef erralsi : The number of recorded patient referrals to more specialized facilities
in assembly i at Mohalla Clinics for the assembly year before 2017 MCD elections.
These are tests conducted from the official list of free 212 diagnostic tests available at
Mohalla Clinics.
Page 40
Arjun Soin - May 2019
• hospitalsP resenti : The number of hospitals catering to assembly i, with daily patient
volumes exceeding 1000.
• otherF acilitiesi : The number of other, smaller health facilities in assembly i. These
facilities include Dispensaries, IPP VIII Centers, Maternity and Child Welfare Centers,
Tuberculosis Clinics, Primary Health Centers (PHCs) and Railway Clinics.
• Areai : The area of assembly i, a variable drawn from a shapefile obtained at Stanford
University’s digital library archives. I use GeoPandas, an open source python project,
to calculate the area from the geospatial shapefile obtained from the digital archives.
The code creates a new columnn in my dataset that converts the embedded information
in the shapefile to area of each assembly constituency.
• BJP M LAi : A dummy for whether or not assembly i is represented by a BJP Member
of Legislative Assembly. This variable controls for discrepancies observed between MCD
and Legislative Assembly of Delhi electoral composition.
• otherP artiesi : Number of competing political parties other than AAP and BJP (spe-
Page 41
Arjun Soin - May 2019
cially Congress) contesting in the wards of assembly i for the 2017 MCD election.
• wardChangesi : Variable shows the difference between wards after and wards before
the 2011 delimitation exercise. It is calculated as:
• ZoneEasti : Dummy for whether assembly i falls within East MCD jurisdiction.
• ZoneN orthi : Dummy for whether assembly i falls within North MCD jurisdiction.
• ZoneSouthi : Dummy for whether assembly i falls within South MCD jurisdiction.
Page 42
Arjun Soin - May 2019
Figure A4: Mohalla Clinic with image of Delhi Chief Minister (AAP President)
Page 43
Arjun Soin - May 2019
Page 44
Arjun Soin - May 2019
The idea of calculating an inflection point for the given multiple regression specifications is
to isolate a cutoff point of previous electoral confidence (proxied as BJP V ote.Sharei,2012 )
below/above which the effect of clinics seems to show opposing trends. In the paper, it
represents the point of previous vote share where between a clientilistic and programmatic
outcome. (from a theoretical standpoint). Suppose we estimating equation (3) from the paper:
BJPVote.Sharei,2017 = α + β1 Mohallai + β2 (BJPVote.Sharei,2012 ) + β3 Mohallai ∗ BJPVote.Sharei,2012
X X
+ γi Ai + ηi 2017i + , (A4)
i=1 i=1
Mathematically, the inflection point will be calculated by solving the following inequality:
where we use the coefficients of the Clinic variable and that of its interaction with 2012
(previous) vote share of the opposing party. Now using inequality solution rules:
β1,M ohallaClinics
where β3,M ohallaClinics
is the quantity known as the ‘inflection (or cutoff) point’ referred to in
the paper, above which clinic effect is no longer negative.
Page 45
Arjun Soin - May 2019
Page 46
Arjun Soin - May 2019
• For the instrumental variable, I propose using an outcome from the ward delimitation
exercise based on the 2011 Cenus of India, represented by wardsChanges.
• This variable calculates the difference between wards today (for 2017) and wards before
the 2011 delimitation exercise and can function as a proxy for expected demand/appetite
for primary health care treatment.
• Boundaries of almost 150 wards out of 272 wards were redrawn. The number of wards
in each MCD sub-region - North, South and East - remained constant at 104, 104 and
64 respectively. But the number of municipal wards within each AC seat changed - in
2012, each AC contained 4 wards while in 2017 after delimitation took effect, each AC
contained 3-7 wards.
Page 47