© copyright by somdeep chatterjee may 2016

© Copyright by

Somdeep Chatterjee

May 2016

_______________

A Dissertation

Presented to

The Faculty of the Department

of Economics

University of Houston

_______________

In Partial Fulfillment

Of the Requirements for the Degree of

Doctor of Philosophy

_______________

By

Somdeep Chatterjee

May, 2016

THREE ESSAYS ON ISSUES IN DEVELOPING ECONOMIES: CREDIT

REFORMS, ELECTIONS AND WAGE DISPERSION



_________________________ Somdeep Chatterjee

APPROVED:

_________________________ Aimee Chin, Ph.D. Committee Co-Chair

_________________________ Gergely Ujhelyi, Ph.D.

Committee Co-Chair

_________________________ Chinhui Juhn, Ph.D.

_________________________ Francisco Cantú, Ph.D.

Department of Political Science University of Houston

_________________________ Steven G. Craig, Ph.D. Interim Dean, College of Liberal Arts and Social Sciences Department of Economics

ii

_______________

An Abstract of a Dissertation

Presented to

The Faculty of the Department

of Economics

University of Houston

_______________

In Partial Fulfillment

Of the Requirements for the Degree of

Doctor of Philosophy

_______________

By

Somdeep Chatterjee

May, 2016



iv

Abstract

This dissertation presents three essays on some important issues in two develop-

ing economies, India and Ghana. In the first essay I analyze a major agricultural credit

reform in India, known as the Kisan Credit Card (KCC) scheme, which intended to

simplify the process of credit delivery in the agricultural sector. I use plausibly exoge-

nous variation in the reach of the program to find the causal effects of the policy on

agricultural output and technology adoption using a district panel data set. I also use

a household dataset to analyze the effects of differential exposure to this policy on a

wide range of household outcomes. I find evidence of increases in agricultural output

of rice, which is the major crop of the country. I also find that on average the use of

high yielding variety (HYV) seeds increases at the district level providing evidence

of technology adoption. The increase in output at the district level is corroborated by

suggestive increase in sales revenue and output of rice farmers at the household level.

In addition, I find evidence that bank borrowing increased for households due to this

program and the estimated effects on income and production are higher for such a

sub-sample of borrowers.

In the second essay, co-authored with Gergely Ujhelyi and Andrea Szabó, we try

to answer the question as to why people vote? One possibility is that they derive con-

sumption utility from doing so, but isolating this has proven empirically challenging.

In this paper we study a recent natural experiment in India, where legislative elections

have to provide a "None Of The Above" (NOTA) option to voters. Using the fact that

NOTA cannot affect the electoral outcome we show that studying individual voters’

behavior with and without NOTA provides a way to identify various components of

the consumption utility of voting. To address the challenge that individual votes are

v

not observable, we borrow techniques from the Industrial Organization literature to

estimate a structural model of voter demand for candidates and perform counterfac-

tual simulations removing the NOTA option. We complement this with a reduced-

form analysis of NOTA in a difference-in-differences framework, exploiting variation

in the timing of the reform created by the electoral calendar. Using both methods,

we find that NOTA increased turnout. We find minimal substitution between candi-

dates and NOTA, indicating that NOTA votes are cast by new voters who turn out

to vote specifically for this option. This indicates the presence of an option-specific

consumption utility of voting.

In the final essay, I use a matched employer-employee dataset from the Ghanaian

manufacturing sector to analyze earnings dispersion in Ghana from 1992-2003, a pe-

riod post extensive economic reforms. I find that variance of earnings increased from

1992-1998 and decreased thereafter resembling an inverted u-shaped relationship. I

use ANOVA and variance decomposition approaches to understand the underlying

factors that led to such a pattern in earnings inequality. I find that between-firm

factors explain this pattern more than within-firm factors. I also find that the mean

earnings gap between workers above and below the 90th percentile of income distri-

bution can explain majority of the initial surge in inequality (61%) but only explains a

very small fraction of the eventual decline (9%). I run modified Mincerian regressions

and decompose the variance components to find that the decline in earnings inequal-

ity is consistent with decline in variance of firm level earnings whereas variance of

predicted wage from worker characteristics have increased. I also find suggestive evi-

dence of changing patterns of worker-firm sorting which contributes to the decline in

inequality.

vi

Acknowledgements

The role of the Chair of the committee without doubt deserves the most plaudits

for a successful dissertation of a PhD student. I consider myself extremely lucky to

have had not one, but two people serve as co-Chairs of my committee. This not only

doubled my impetus towards achieving this target but also made me doubly more

determined to put my best foot forward towards producing this research output. As a

result I cannot thank Aimee Chin and Gergely Ujhelyi enough for acting as co-chairs

of my committee and guiding me at every step throughout my graduate student life

at the University of Houston. Professor Chin has truly been an academic mentor to

me from whom I could seek help for almost anything related to my career, not just my

research. Prof Ujhelyi on the other hand was the one who first encouraged me to start

thinking about research, right in my first year as a graduate student. The fact that I

started attending the Empirical Microeconomics seminar series every Tuesday right

from then is owed to him.

I must acknowledge the very important role played by the third member of my

dissertation committee, Chinhui Juhn. It was with her that I did my first independent

study in Grad school and the way she walked me through the papers we read together

and discussed every Friday will remain one of the most fruitful experiences of my life.

I must also take a moment to specially thank Janet Kohlhase for being super support-

ive during the lowest point of my days here at U of H. I also thank Bent Sørensen for

encouraging and uplifting me particularly at that time and for being a great Grad Di-

rector to all of us in general. I thank Amber Pozo for her great administrative support

throughout my career at University of Houston.

vii

I would also like to thank Andrew Zuppann for his comments and suggestions

at various times and for giving the chance to present preliminary work in the Brown

Bag sessions and thank all the attendees. I have benefitted greatly from all of their

comments which I will perhaps miss the most, apart from the delicious food! I thank

all the other faculty members of the department of Economics for providing feedback

at various times during seminars. I would be failing in my duty if I do not specially

mention Andrea Szabó for being an amazing co-author. I am grateful to Francisco

Cantú from the Department of Political Science for being the external member of my

dissertation committee. Life at UH would not have been as great had it not been

for my fantastic classmates and colleagues. I thank Chon-Kit for the many useful

discussions. I thank Bocong for being such a great friend and confidante and Subash

for being extremely supportive. For the innumerable memories at UH, within and

outside the department, I must thank Gautham, Vinh, Max, Xuejing, Shoumen, Aritri,

Indrajit, Debashis and so on but I apologize to everyone whose names I cannot take

due to paucity of space.

While one is at grad school, he requires a solid support system from outside and

back home. I was lucky to have one. I cannot express in words the contribution of my

parents, Srabani and Somnath. Everything that I am today is due to them and there

are no other contenders befitting for me to dedicate this thesis to. I must thank my

professors from my alma mater in India, Ajitava Raychaudhuri, Saikat Sinha Ray and

Swapnendu Banerjee for believing in me, my greatest buddy and childhood friend

Sayantan for being there at all times, my closest friend from college Saumik for being

who he is.

Last but in no way the least, I was blessed to have unbelievable grandparents for

viii

super important emotional support and much needed pampering at times. The con-

tribution of Dadai, Chotdadu and Monididi cannot be quantified here. Even though

Mammam could not live to see this day, I am sure she has always been with me from

wherever she is now and would have been the happiest person on earth today.

ix

Contents

1 Effects of Agricultural Credit Reforms on Farming Outcomes: Evidence from

the Kisan Credit Card Program in India. 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.2.1 The Kisan Credit Card Program . . . . . . . . . . . . . . . . . . . 6

1.2.2 Conceptual Framework and Related Literature . . . . . . . . . . . 9

1.3 Empirical Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

1.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1.5.1 Results using District-Panel Dataset . . . . . . . . . . . . . . . . . 21

Effects on Crop Production . . . . . . . . . . . . . . . . . . . . . . 21

Technology Adoption . . . . . . . . . . . . . . . . . . . . . . . . . 22

Threats to Identification: Check for Pre-Trends . . . . . . . . . . . 23

1.5.2 Do Households Change their Borrowing Patterns? . . . . . . . . . 24

Total Borrowing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

Analysis of the Largest Loans . . . . . . . . . . . . . . . . . . . . . 26

1.5.3 Effects on Household Production, Income and Consumption . . . 27

Rice Production and Sales . . . . . . . . . . . . . . . . . . . . . . . 27

Income . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

x

Consumption Expenditures . . . . . . . . . . . . . . . . . . . . . . 29

1.5.4 Falsification Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . 30

1.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

1.7 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2 “None of the Above" Votes in India and the Consumption Utility of Voting

(with Gergely Ujhelyi and Andrea Szabó) 46

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

2.2 The consumption utility of voting . . . . . . . . . . . . . . . . . . . . . . . 51

2.3 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55

2.3.1 The Indian NOTA policy . . . . . . . . . . . . . . . . . . . . . . . . 55

2.3.2 NOTA-like options in other countries . . . . . . . . . . . . . . . . 56

2.3.3 Assembly elections in India . . . . . . . . . . . . . . . . . . . . . . 58

2.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

2.4.1 Samples used for analysis . . . . . . . . . . . . . . . . . . . . . . . 61

2.4.2 Election data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

2.4.3 Voter demographics . . . . . . . . . . . . . . . . . . . . . . . . . . 63

2.5 Patterns in the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

2.5.1 NOTA votes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

2.5.2 The effect of NOTA on turnout . . . . . . . . . . . . . . . . . . . . 67

2.6 Estimating the effect of NOTA from a demand system for candidates . . 70

2.6.1 Specification: demand . . . . . . . . . . . . . . . . . . . . . . . . . 71

2.6.2 Specification: supply . . . . . . . . . . . . . . . . . . . . . . . . . . 74

2.6.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

2.6.4 Practical issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

xi

2.6.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

Parameter estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 81

Counterfactual analysis: The impact of NOTA . . . . . . . . . . . 82

2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

2.8 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

2.9 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

2.9.1 The correlates of NOTA votes . . . . . . . . . . . . . . . . . . . . . 90

2.9.2 Robustness of the DD estimates . . . . . . . . . . . . . . . . . . . . 91

National elections . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91

Redistricting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

State-specific events . . . . . . . . . . . . . . . . . . . . . . . . . . 93

3 Firm Ownership and Wage Dispersion: Evidence from Ghana Using Matched

Employer Employee Data 110

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

3.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

3.2.1 The Ghanaian Manufacturing Sector . . . . . . . . . . . . . . . . . 113

3.2.2 Related Literature and Conceptual Framework . . . . . . . . . . . 115

3.3 Empirical Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

3.3.1 Analyis of Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

3.3.2 Percentile Analysis: Contribution of Earnings Gap to Variance . . 117

3.3.3 Worker Characteristics and Firm Fixed Effects . . . . . . . . . . . 117

3.4 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118

3.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

3.5.1 Within-Firm and Between-Firm Effects . . . . . . . . . . . . . . . 120

xii

3.5.2 Percentile Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

3.5.3 Returns to Schooling and Earnings Dispersion . . . . . . . . . . . 124

3.5.4 Does Predicted Wage Dispersion by Worker and Firm Charac-

teristics Vary by Ownership Structure? . . . . . . . . . . . . . . . 124

3.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126

3.7 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

xiii

List of Figures

1.1 Crop Production: Year-specific coefficients for aligneds ·morebanksd . . 38

1.2 HYV Use: Year-specific coefficients for aligneds ·morebanksd . . . . . . . 39

2.1 Constituencies in the merged dataset . . . . . . . . . . . . . . . . . . . . . 102

2.2 Distribution of NOTA vote shares across constituencies . . . . . . . . . . 107

2.3 Impact of NOTA on turnout . . . . . . . . . . . . . . . . . . . . . . . . . . 108

2.4 Impact of NOTA on candidates’ vote shares . . . . . . . . . . . . . . . . . 109

3.1 Share of Firms by Ownership Type . . . . . . . . . . . . . . . . . . . . . . 129

3.2 Time Series Plots of Firm Averages by Ownership Types: Firm Dataset

of 200 Firms over 12 Years . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

3.3 Variances of Log Real Hourly Earnings by Ownership Type . . . . . . . . 136

xiv

List of Tables

1.1 Comparing Means of statewise spread of KCC in 2000 by aligned . . . . 37

1.2 VDSA District Dataset: Effects on Crop Production and HYV Use . . . . 40

1.3 IHDS Dataset: Effects on Borrowing Composition . . . . . . . . . . . . . 41

1.4 IHDS Dataset: Effects on Rice Production and Sales . . . . . . . . . . . . 42

1.5 IHDS Dataset: Effects on Income . . . . . . . . . . . . . . . . . . . . . . . 43

1.6 IHDS Dataset: Effects on Consumption Expenditure . . . . . . . . . . . . 44

1.7 Falsification Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

2.1 Timeline of events in the study period . . . . . . . . . . . . . . . . . . . . 95

2.2 Summary statistics of the electoral data at the constituency level . . . . . 96

2.3 Candidate characteristics in the panel data . . . . . . . . . . . . . . . . . 97

2.4 Voter demographics at the state level (repeated cross-section) . . . . . . . 97

2.5 Demographic characteristics of the constituencies from the Indian Cen-

sus (panel dataset) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

2.6 The impact of NOTA on turnout, DD estimates . . . . . . . . . . . . . . . 99

2.7 Estimates of the linear parameters of the demand system . . . . . . . . . 100

2.8 Estimates of the nonlinear parameters of the full model . . . . . . . . . . 101

2.9 Impact of NOTA on vote shares by party . . . . . . . . . . . . . . . . . . 103

2.10 The correlates of NOTA votes . . . . . . . . . . . . . . . . . . . . . . . . . 104

xv

2.11 Effect of NOTA on turnout, excluding national election years . . . . . . . 105

2.12 Effect of NOTA on turnout, controlling for redistricting . . . . . . . . . . 105

2.13 Effect of NOTA on turnout, robustness to state-specific events . . . . . . 106

3.1 Descriptive Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

3.2 Analysis of Variances of Log Real Hourly Earnings of Workers . . . . . . 131

3.3 Variances of Log Real Hourly Earnings of Workers by Ownership Type . 131

3.4 Contribution of Earnings Gap to Variance in Earnings . . . . . . . . . . . 132

3.5 Returns to Schooling and Experience . . . . . . . . . . . . . . . . . . . . . 133

3.6 Variance Decomposition: Estimated Firm Effects and Predicted Wage

from Worker Characteristics . . . . . . . . . . . . . . . . . . . . . . . . . . 134

xvi

To Baba and Ma...for everything and more!!!

1

Chapter 1

Effects of Agricultural Credit Reforms

on Farming Outcomes: Evidence from

the Kisan Credit Card Program in India.

1.1 Introduction

Providing access to financial resources to the poor continues to be an important pol-

icy prescription in the literature even though the empirical evidence on impacts of

credit constraint relaxations and expansion of credit options for the unconstrained in

developing countries is mixed (Karlan and Murdoch 2009). To design effective poli-

cies that provide or expand access to credit, one would first need to understand the

mechanisms through which credit access helps the poor and also the impact of im-

plementing such a policy on the targeted beneficiaries. To estimate these impacts, the

ideal experiment would be to randomly provide credit products to households and

compare the outcomes of the ones getting access to the ones without access to this

product. The January 2015 issue of AEJ Applied has six papers on this subject using

randomized evaluations.1 These papers find little to no impact of providing access to

finance. Other papers like de Mel, Mckenzie and Woodruff (2008) using experimental

designs find positive impacts.

While randomized evaluations are ideal to identify the causal effects of credit con-

straint relaxations, by design these cater to a relatively small sample of the entire pop-

ulation. Whether a large scale national reform would replicate these findings is impor-

tant to understand. In this paper, I look at a major overhaul in the agricultural credit

delivery process in India in 1998, known as the Kisan Credit Card (KCC) program,

and evaluate the impacts of this policy. The targeted group for this credit reform was

rural agricultural households, generally involved in farming and other related occu-

pations. Ease of delivering agricultural credit, reasonable interest rates and relaxation

of monitoring norms were the key features of this program. Reports from the Planning

Comission of India (2002) suggest that by 2000-01, KCCs constituted almost 71% of the

total production credit disbursement by commercial banks. It was also the dominant

mode of production credit delivery for other banks. The report also suggests that in

the first two years, close to 4 million credit cards were issued with a total disbursal of

credit lines worth 50 bllion INR (1 billion USD approximately).

Although this was a major policy reform, to date there has been little convincing

evidence of the impacts of this program. Chanda (2012) uses post-policy state level

data from 2004-2009 to see if growth in KCC issues lines up with increases in agri-

cultural productivity. There are other government of India commissioned descriptive

reports like the Planning Commission report mentioned above and Samantara (2010).

In this paper, I use a country wide district panel dataset to evaluate the causal effects

1All the 6 papers are cited in the references.

2

of this program on agricultural output and technology adoption. I also use household

data to estimate the impacts of this program on a wide range of outcomes including

income, consumption and borrowing.

The reach of formal financial institutions is not universal in most developing coun-

tries. This is because banks would want to select into richer regions unless they are

administratively required to setup branches in unbanked locations. This makes for-

mal credit markets less accessible to the poor in these areas. The KCC reform therefore

provides an opportunity to add to this literature of how access to formal credit insti-

tutions can help the poor sections of the society in line with Burgess and Pande (2005).

The unique feature of the KCC program was that it catered exclusively to the agricul-

tural sector. Although in this paper, I am not able to distinguish whether the effects

of KCC operate through channels of new access to credit or expansion of credit to the

ones who already had some access. As a result most of my estimations should be

viewed as a bundle of reduced form effects.

This paper takes advantage of rules in implementation of the policy to generate

plausibly exogenous variation in access to this program to identify causal effects of

the reform. The identification strategy relies on variation across three main dimen-

sions. First is the time dimension. The policy was implemented in 1998 and I look at

the outcomes in years before and after the policy. Second, is the political alignment di-

mension, ie, whether the state government is ruled by a party aligned with the central

government in the federal structure of India. Political alignment has been widely re-

garded to be important for policy implementation and performance (see Chibber et al

2004, Iyer and Mani 2012 and Asher and Novosad 2015). The final source of variation

comes from how the rolling out of these credit cards was implemented. The KCCs

3

could only be issued through formal banks and not by any other agency. I use district

level variation in the number of bank branches already setup prior to the policy to

proxy for access to this program.

I propose to identify the causal effect of the policy by the interaction of these three

variables. The effect is identified by looking at the difference in outcomes after and

before the policy in districts with more bank branches over districts with fewer bank

branches in states that are ruled by political parties aligned with the central govern-

ment after controlling for these differences in districts in the states not aligned with

the center. I use pre-policy data to show that these regions were not already different

along the relevant dimensions to provide support to the identifying assumption that

any differences post-policy are attributable to the program.

I find that increased access leads to significantly higher production levels. Rice is

the major crop of India and I find an aggregate increase in production by 88 thousand

tonnes (metric ton) per year on average which is between 1/3 to 1/4 of an increase

compared to the mean. 2 Corresponding to this large change, I find that technology

adoption has also been significant. Crop production area under high yielding variety

(HYV) seeds increased by around 71 thousand hectares at an aggregate level which

is just under a 1/3 increase compared to the mean. This suggests that with increased

access to credit, districts exposed to the program fared significantly better in terms

of porduction and technology adoption. Using houseold data, I corroborate some of

these results. I find suggestive evidence of increases in rice production for farmers

even though estimated imprecisely. I am constrained by the fact that the household

2The Food and Agricultural Organization’s FAOSTAT indicates that in 2012, the value of rice pro-duced in India is over 40 billion US dollars which makes it the most valuable crop of India. Rice hasconsistently been the major crop of India in terms of overall value for years. See - faostat.fao.org

4

data comes only from a sample of farmers and not the universe, unlike the district

panel data described above which contains all rice production in the districts. I find

that revenue from sales of rice is higher for farmers potentially exposed to KCC.

The advantage of using household data is being able to observe borrowing pat-

terns. Using a cross-section of households, I find that households are more likely to

have fewer but larger loans with exposure to KCC. I also find that they are more likely

to have larger bank loan sizes if exposed to KCC. These effects seem to be larger for

those households which report cultivation as their main source of income and for rice

farmers. This is reassuring because most of the production effects observed using the

district data seem to suggest that rice farmers would be most affected by this policy.

I find that even though on average there is no effect on income but farm income

is higher by 129 Indian rupees (USD 2) per month for households whose main in-

come source was cultivation. Compared to the mean, the magnitude of this effect is

almost 25%. This is consistent with the findings on production and sales. With higher

sales, we might expect higher profits, ceteris paribus. I find no overall impacts on con-

sumption expenditure but composition of consumption changes. I find higher daily

expenditures but lower expenditure on consumption like tobacco and beetel leaves.

I do not find any effects on the margin of whether a household is likelier to have a

bank loan in response to the policy. Since I do find that households have a higher bank

loan size conditional on borrowing, this allows me to analyze a sub sample of house-

holds seperately. I look at all the outcomes for only those households who borrow

from banks and I find that all the effects described above are much more pronounced

for this group. Since KCC had to operate through banks, this gives us confidence that

our estimated effects are likely to be mediated through bank borrowing which should

5

include KCC borrowing.

The rest of the paper is organized as follows. Section 2 provides background infor-

mation. Section 3 describes the empirical strategy. Section 4 explains the Data. Section

5 presents results and Section 6 concludes.

1.2 Background

1.2.1 The Kisan Credit Card Program

Agriculture constitues roughly a fifth of the total GDP of India and employs two out

of three Indian workers. In the late nineties, agriculture started opening up to the mar-

ket rather than being limited to subsistence farming. Agricultural credit has played an

important role in developing the market for such produce and help improve the con-

dition of farmers in the country. However, the finance and credit institutions present

in the country prior to 1998 were deemed inefficient by several reports and experts and

as a result the Kisan Credit Card program was envisaged. This scheme was launched

in 1998 and was introduced for the first time in the budget speech of the Finance Min-

ister of India in the parliament. Within a year after its inception around 5 million cards

were issued to farmers. Prior to 1998, the system of agricultural credit delivery was

complicated. A multi agency approach was used where borrowers had to go through

several layers of bureaucracies depending on the purpose of their loans (Samantara

2010). KCC also brought about a revolving credit regime as opposed to the existing

demand loan system (Chanda 2012).

At its inception, the KCC was not a traditional credit card that is commonly used.

The card was a mere documentation for identifying the individual and his credit line

6

with a given bank. It did not have features that allowed payments at merchant outlets.

This also makes the presence of banks an important dimension for identifying the

intensity of reach of the program. The way to use a KCC was to visit the bank branch

in person and withdraw a certain amount of money which could then be used for

purchases. This also ruled out the possibility of banks monitoring the usage of the

loans.

The most important feature of this credit product was the ease of availability of

loan. Some banks laid down rules for eligibility like having title to an acre of irrigated

land. On fulfiling this criterion, the farmer would be eligible for a loan with a bank

without any collateral requirement for an amount upto 50,000 INR (around 1000 USD

back then). The KCC accounts were largely valid for 3 years and repayment time

frames spanned upto a year. On successful replayments and responsible credit use,

these accounts were renewable but the initial approval was given largely without any

bacground checks. As pointed out above, a big difference from existing crop loans

was that the usage of the KCC loans were not monitored whereas most agro-credit

was tied to agricultural use or purchase of inputs, fertilizers etc. So, a farmer could

get a KCC account and use the amount for personal consumption.

In a way KCC provided the best available source of personal credit to poor farmers.

The biggest advantage over microfinance institutions were that KCC was operational

through formal banks and charged a very reasonable interest rate of around 7% per

annum as opposed to as large as 36-40% rates charged by self help group microcredit

institutions. The approval process was also very simple and was a single window

exercise as the only criterion was ownership of an acre of irrigated land. Many banks

have recorded allowance of credit limits in excess of 50,000 INR but in such cases

7

they often asked for collaterals. Therefore larger scale farmers who are financially in

a better off situation were only likely to go for these loans. There was no clause to my

knowledge which restricted large farmers from opening a KCC account.

Samantara (2010) points out that a major reason why KCC was launched was to

integrate the various credit needs of farmers, from personal consumption to festival

expenditure, education, health and agricultural needs, into one comprehensive prod-

uct. Earlier a farmer had to weigh multiple options based on the purpose of his loan.

KCC made it a one stop procedure wherein he could withdraw the requisite amount

and use it for any purpose whatsoever. All the bank cared about was the timely re-

payment and not the usage. This was a major shift from the pre-existing agro-credit

policy in India which was called the Agricultural Credit Delivery System. Under that

system, a multi-product multi-agency approach was adopted. Policy makers in the

country had planned this in a way such that specific needs of farmers could be ad-

dressed by specific credit products. A farmer could go to a bank for purchase of a

particular input and get a loan against that purchase. The idea was more like financ-

ing purchases rather than giving out cash loans. From such a scheme KCC came as

a welcome change which sought to replace the multi-product approach in favor of

a cash credit approach in a single comprehensive product. As might be already evi-

dent from this discussion, KCC was intended to address the short term credit needs of

farmers and not the longer term needs. Since there was no monitoring, one could not

rule out the possibility of withdrawing cash from these accounts and using them for

consumption purposes. At present, Kisan Credit Cards are available as differentiated

products with various banks coming out with various varieties and features.

Overall the Kisan Credit Card program should be viewed as a bundle of reforms in

8

one. It not only aimed to relax credit constraints by making loans available to the ones

constrained prior to 1998 but also provided a source of flexible credit. KCCs could

potentially finance a lot of purchases, not just agricultural inputs and therefore have

wider social consequences. Since KCCs were a source of cheaper credit, one might

also view it as expansion of credit options for the ones already having access to other

forms of credit. Unconstrained farmers may now be attracted to borrow at cheaper

rates and finance their short term credit needs.

1.2.2 Conceptual Framework and Related Literature

To estimate the true causal effects of access to credit one would ideally want to gen-

erate random variation in access to financial institutions. There is a rich literature

comprising of experimental studies along these lines (Angelucci, Karlan and Zinman

2015, Attanasio et al 2015, de Mel, Mckenzie and Woodruff 2008, Augsburg et al 2015,

Banerjee et al 2015, Crepon et al 2015, Tarozzi, Desai and Johnson 2015). Apart from

this there is a quasi-experimental literature which looks at policy reforms in the formal

financial sector to answer a similar question (Burgess and Pande 2005, Banerjee and

Duflo 2014). Government policy reforms are usually not randomly assigned, there-

fore identifying the causal effects of such programs is challenging even though it is

important to understand the mechanisms behind such policies aimed at removal of

borrowing constraints.

Most recent studies on the role of credit access focus largely on this aspect of mech-

anisms of credit delivery (Karlan and Morduch 2009). This paper is the first to objec-

tively evaluate the Kisan Credit Card scheme using a district panel dataset and ex-

tends this literature by looking at this large scale national reform in credit delivery

9

mechanism. In the Indian context, Banerjee and Munshi (2004) and Banerjee and Du-

flo (2014) study the role of credit constraints on firms and businesses. However the

role of credit constraints in agricultural occupations has been little studied till date.

This paper also contributes to the literature by attemtping to fill this void.

An important question that arises here is whether this program should be viewed

as enhanced ‘access’ to agricultural credit or ‘expansion’ of credit to the ones who al-

ready had access to credit? The existence of credit constraints and impediments to

borrowing are major roadblocks in developing economies which is why governments

may want to innovate by reforming the system of credit delivery. If the main objec-

tive is to improve the condition of the poor, one would imagine that removing the

borrowing constraints would be important, or in other words a program like KCC

should have given ‘access’ to credit to the ones who never had the chance to borrow

before. The starting point of the analysis is to understand how we expect credit access

to affect the credit constrained? If KCC relaxed credit constraints and people unable

to borrow elsewhere could now borrow under this program, economic theory and

existing empirical evidence would lead us to expect multiple effects.

First, if households invest in productive assets or the borrowed funds are used to

finance improvements in technology of agricultural production, we expect their agri-

cultural income to be higher. Second, if we aggregate these effects, overall production

of crops should be higher and overall adoption of new technology should also be

higher. Third, composition of consumption may change. Banerjee et al (2015) find

such evidence in a microfinance experiment but the idea is applicable to a broader

10

country wide setting as well because in essence we are thinking of the impact of relax-

ation of credit constraints per se. Finally, since this was a national level formal lend-

ing program, one would expect that with enhanced access informal lending would go

down and be substituted by more formal sector loans.

The flip side however, is that from a lender’s perspective, such a policy may at-

tract poor quality borrowers. This leads to issues of adverse selection. Asubel (1991)

discusses credit card markets in the US and how lowering interest rates are far from

ideal from a bank’s perspective as bad borrowers may select into borrowing at lower

rates. KCC lending was usually at a much lower rate of interest than market rates or

informal lending rates prevalent among microfinance institutions. This would have

meant that the adverse selection issue was likely to be severe under this program.

Also since new borrowers are unlikely to have ever engaged in credit dealings, their

perception about their own future stream of income determining their repaying abil-

ity is likely to be myopic. Melzer (2011) and Bond, Musto and Yilmaz (2009) point

out these problems about ‘misinformed’ borrowers underestimating their future re-

payment commitments.

It is also important to think about potential general equilibrium effects of this pro-

gram. Are there any spillovers? For example, if some farmers get credit cards whereas

others do not, maybe they have a competitive advantage over the ones who did not get

this card and this might lead to perverse welfare implications. Similarly, if KCCs are

very attractive and result in high profits for farmers, this maybe an incentive for non-

farmers to take up agricultural occupations which in turn may affect non-agricultural

sectors in the rural areas.

11

1.3 Empirical Strategy

There are two parts in my empirical strategy. I have the twin objective of evaluating

the overall effects of access to credit on production outcomes on average and also

whether access to credit through such a reform is useful for intended beneficiaries.

To this end, I use two different datasets. The first is a district panel dataset and the

second is a cross-sectional household dataset.

Identifying the causal effects of having a KCC on agricultural outcomes using sur-

vey data is difficult because KCCs were not randomly assigned to households. Also,

using a cross sectional dataset, it is not possible to use time varying access to the

scheme either. 3 To overcome these issues, I propose an identification strategy that re-

lies on plausible exogenous variation in the reach of this program to find causal effects

of the program. Apart from the time dimension (program introduced in 1998) which

provides variation in the access to the program over the span of the data, there are two

different cross sectional dimensions that give us a sense of which regions might have

had more access to these cards after the policy. I use an interaction of these dimesions

to identify effect of the policy.

The KCC program was announced by the Finance Minister of India in his budget

speech in 1998 and the implementation began soon after. The government at the cen-

ter was ruled by the Bharatiya Janata Party (BJP) led National Democratic Alliance

(NDA) coalition. However, not all state governments were run by the NDA coalition.

Since the implementation of this policy required a lot of work at the grass roots in

terms of setting up infrastructure, spreading awareness, nudging banks to implement

3Even though there is no clear idea even in government documents in terms of how these cards wererolled out.

12

this policy and the like, one can understand that the role that state governments and

officials at the village and block levels who are employed by the state governments

would have had an important role to play in the penetration of this policy in those

states. This gives one potential source of variation in the policy. I use an indicator

variable aligned which takes the value 1 if the state in question was ruled by the BJP

or one of its NDA allies in 1998 and 0 otherwise. The idea is that aligned states would

probably have earlier or quicker access to this policy whereas the opposition parties

may choose to be slack in the policy implementation in the states where they are in

power, out of several motives including the fact that they would want the scheme to

be projected as a failure for the ruling coalition and take advantage of this in future

elections themselves.

The first real governmental study on the program outreach was done in 2002 by the

Planning Commission of India. They published a report with tables on the state wise

coverage of Kisan Credit Cards as of March 2000, which is 2 years into the program.

The coverage rates were basically the number of KCCs issued by various banks as a

percentage of total operational land holdings in the concerned state. So this gave an

idea as to how many farmers were potentially reached or covered under the policy

within the first two years of the policy at a state level. If we observe that aligned states

actually were implementing the policy faster than the other states, we might be more

confident about the use of this dimension to identify the effects of the program. Table

1.1 provides supportive evidence. I find that coverage in aligned states is almost 2.5

times the coverage in rest of the states and the difference is statistically significant at

the 99% level of confidence.

13

The second dimension that I bring to this analysis of variation in access is a techni-

cality that the policy had. These credit cards could only be given out through banks.

So it is understandable that areas with more banks are likely to be able to roll out these

cards faster than the ones which are unbanked or have fewer banks. However, there

may be concerns that banks opened up or positioned or repositioned themselves based

on the policy announcement in markets where KCC lending would flourish more. To

account for this issue I use bank data at the baseline year, ie, 1998 and not after the

policy. I use district level existing bank branches data from 1998 to enumerate the

number of branch offices of banks at the time of announcement of the policy. This

gives us another potential exogenous source of variation in the intensity of coverage

of the program. I create the variable bank98 to denote the number of bank branches

in a given district in 1998 and use the indicator variable morebanks which takes the

value 1 for districts with number of banks above the mean of bank98.

Finally, I use the indicator variable I(Y EAR > 1998) to capture the time of expo-

sure to the policy and controlling for pre-existing differences along the above cross

sectional dimensions over time. I run the following regression for district ‘d’ in state

‘s’ at time ‘t’:

Ydst = αs + δt + β1aligneds ·morebanksd · I(Y EAR > 1998) + β2morebanksd

+ β3aligneds ·morebanksd + β4morebanksd · I(Y EAR > 1998)

+ β5aligneds · I(Y EAR > 1998) + γXdst + εdst (1.1)

The coefficient of interest is β1 which captures the causal effect of the policy on

14

outcomes Y . I use state fixed effects captured by αs. Demographic controls at the

household level are included in X . I control for the number of persons in the fam-

ily, number of children, number of married men and women and also the age and

education levels of men and women.

The interpretation of β1 is that it gives us the difference over time (post- and pre-

policy) in Y for households in districts with more banks compared to households

in districts with lesser banks in aligned states after controlling for these same differ-

ences in non-aligned states. The identifying assumption is that the outcome Y would

not have been different for these groups of households had there been no KCC pol-

icy. There is no standard way to validate this assumption and identification always

assumes this, but the panel structure of the data provides an opportunity to check

whether these districts were historically different and already had differential trends

even before the policy. If we find that before the policy, differences in outcomes along

the above dimensions were not different, we gain confidence that the identifying as-

sumption is plausible. I describe a check for this at a later section and find that before

1998 there were indeed no differences in outcomes in these areas.

The fact that prior to the policy, the cross sectional dimensions seem to be similar,

leads us into the cross sectional analysis. The dataset that I use is from 2005 which is a

post-policy year. I still use the above cross sectional dimensions to generate exogenous

variation in access to the policy but do not have the time dimensions anymore. Since

there were no differences in these regions prior to the policy, any difference that I find

for 2005 can be attributed as a causal effect of the program.

Using the household dataset, I therefore propose to run the following regression

for household h in district d and state s:

15

Yhds = αs + θ1(aligned ·morebanks)ds + θ2(morebanks)d + ωXhds + uhds (1.2)

In this specification, θ1 is the causal impact of the policy on outcomes Y . The

identifying assumption here, similar to above, is that in the absence of the policy,

the differences in household outcomes between districts with more and less banks in

aligned states would not have been any different from the differences in household

outcomes in more and less bank districts in non-aligned states.

The main outcome that I look at is crop production. As mentioned earlier, rice is

the major crop of the country in terms of value. I focus primarily on rice production

but also look at the other important crops like wheat and maize. The idea is that

with access to credit, farmers may be able to invest more and increase output. Since

there is an element of investment behavior attached to credit access, I look at the use

of high-yielding variety (HYV) seeds. If farmers would adopt more HYV seeds to

increase their production, this would be evidence of technology adoption. I observe

all of these outcomes at the district level and use the panel dataset to find effects on

these. The cross sectional dataset however has a wide range of other outcomes that

are of interest. I briefly describe some of those below.

If access to credit leads to higher agricultural production, an immediate hypothesis

that follows is, access to credit leads to higher incomes for farmers. I use the household

survey data to test this hypothesis. I also hypothesize that since KCC is a formal

source of credit, this might lead to crowding out of informal lending sources like local

money lenders and employers. I do not observe usage of HYV seeds at the household

level but a feature of the agricultural sector is that most poor farmers are not able

16

to preserve and/or grow seeds for indigenous production. I hypothesize that with

access to credit, farmers become more efficient and will be able to use home grown

seeds as a result. I also look at various measures of consumption to see if household

consumption expenditure changed with exposure to the policy or if composition of

their expenditure on different types of consumption changed.

1.4 Data

District Production Data

The data for this study mainly comes from 2 sources. First, ICRISAT-VDSA database

provides a district panel data set for agricultural outcomes.4 For this analysis I am

only focussing on production of rice, wheat, maize and use of HYV seeds. The data

contains information on total production, total area under production, gross and net

cropped and irrigated areas„ number of markets in district, rainfall etc. Although

the dataset provides data from 1966-2011, I focus on the post-1985 period. This is

because of two reasons. Firstly, the empirical strategy would require that pre-trends

are accounted for among the geographic classifications used to identify the causal

effect of the program. One would be worried that in years long before the policy,

potential treatment and control groups would have had very different trends in out-

comes which would invalidate the analysis. Also, the period before 1986 marks a long

history of political turmoil including the emergency days and war with neighboring

countries. 1986 gives us a reasonable starting point for the analysis and it is at least 12

years before the KCC program began. Secondly, the dataset for the early 60s and 70s

4The ICRISAT has a rich database known as the Village Dynamics of South Asia (VDSA) and makesthis available for 19 major states of India

17

has lots of missing information, so analysis using those years would in any way lead

to lesser power.

Household Survey Data

The second dataset is the Indian Human Development Survey (IHDS)-2005. The

first official release of the survey was in 2008 for a survey they conducted in 1503 In-

dian villages and 971 urban neighborhoods in the year 2005. So, the data in this edition

of the survey is based on respondents interviewed in 2005. It was jointly conducted

by a team from the University of Maryland, USA and the National Council of Applied

Economic Research (NCAER), India. The 2005 survey covered 41,554 households and

compiled responses from two interviews each of which lasted for an approximate du-

ration of one hour. I have a wide range of outcome variables to look at including

income, consumption per capita, asset ownership, loan and debt details etc. I focus

only on the rural sample and exclude the urban households which yields a sample of

26734 households.

Household Crop Data

The IHDS-2005 also surveyed households to collect data at the crop level. There

are multiple households producing multiple crops. As will become clear later, most

of my main results appear to be driven by rice producers. So I merge the household

survey data with the crop files using only those farmers who produce any rice. For my

regressions using this dataset, I focus on the households below the 99th percentile to

exclude some large outliers. In the sample the mean of rice production for a household

is around 25 units measured in tenths of a quintal, the maximum is 2600 which is

18

unusually high. Therefore, I exclude the large outliers who produce above the 99th

percentile, which is 200 units in tenths of a quintal.

Household Data from 1993

To provide support to my identification strategy, described in the following sec-

tion, I do a falsification exercise using a cross section of households from the 1993

Human Development Profile of India (HDPI) which was a household survey and in-

terviewed several households who would later be reinterviewed in the IHDS.

Other Data

My identification strategy also relies on variations across three dimensions, cover-

age of KCCs, number of bank branches in 1998 and political alignment of state govern-

ments with the center as of 1998. I look up media reports and open source information

available online to match whether the political party ruling a state was part of the rul-

ing coalition at the center.5 I use data from the Reserve Bank of India website to list

the number of bank branches and branch offices in each district. I also use data from

the Planning Commission of India publication of 2002 for state level access to KCCs

by number of land holding covered under the scheme in 2000 to support the idea that

political alignment was important in terms of the reach of the program.

Do households own a Kisan Credit Card?5In particular I look up the name of the Chief Minister of the states in 1998 and note down his

political party. Then I check if that political party was part of the ruling coalition at the center, ie,National Democratic Alliance or NDA.

19

The IHDS-2005 includes a question for households on whether anybody in the

family owns a KCC or not. This is only a dummy variable. The ideal scenario to de-

scribe the true causal effect of access to credit on outcomes would be to do a 2SLS

regression by instrumenting for access to credit. So if the KCC program was an instru-

ment for access to credit, then ideally we would want to run a first stage regression

of access to credit on the identifying variables and divide the reduced form estimates

above by the first stage coefficient. However, regressions using this dummy variable

as the dependent variable should not be interpreted as the ‘first-stage’ because of two

main reasons explained as follows.

First, the ideal first stage we have in mind would be actual borrowings and usage

of the credit card and not the mere possession of this card. The only way that enhanced

access to credit through possession of this card would lead to increases in income is

if people actually borrowed using this card. Second, since we have just a single time

point, the year 2005, which is seven years after the policy was implemented, all the

coefficients reported using this dummy variable would be under-estimates of the first

stage coefficients. For example, if a household had the KCC for 7 years, and we believe

it was constrained prior to that, then the coefficients from the reduced form estimates

I report are relevant over a period of time while the household has benefitted from

access to credit. So if for this household we consider a change in some outcome Y , it is

not just an instantaneous rise but an overall change. If we divide this by the first stage

which just takes into account 1 period of time, the potential 2SLS estimate would be

hugely overestimated. So we would either need to multiply the so called first stage

coefficient by the number of years the household had the card for (the information for

which is unavailable) or deflate the reduced form by some factor.

20

Second, the dummy variable for having a KCC is not the perfect proxy for ‘access

to credit’ which would be the main dependent variable in our structural regression

model to do the 2SLS regressions. It is also quite possible that a single household had

multiple KCCs but this would show up as a 1 on the dummy, the same as a household

with just 1 KCC. To avoid these problems, I do not use this as an outcome variable

in my regressions. However, roughly comparing the means of this dummy variable

in areas potentially exposed more to the program to the areas exposed less, I seem to

find a positive difference, but this is merely suggestive and therefore I do not interpret

this as causal. The mean of this dummy variable for the entire sample is around 4%

which makes any estimation using this as a dependent variable less convincing.

1.5 Results

1.5.1 Results using District-Panel Dataset

Effects on Crop Production

Table 1.2 reports results on reduced form effects of credit access through more ex-

posure to KCC program on crop production outcomes. I run regressions using the

specification in equation 1.1 as above and report the coefficients β1 for each outcome.

Rice is by far the major crop of India in terms of value of output. I find from column

1 that annual district production of rice increases by about 88 thousand tonnes with

more exposure to KCC. This is quite a big effect compared to the mean of 285 thousand

tonnes which suggests that impediments to borrowing severely constrain the scale of

production. One possible interpretation of this is while farmers are credit constrained,

21

they can put a smaller area under crop production, use lesser inputs and have little or

no access to advanced production technology. With access to credit, these are less of

problems and as a result we expect to see a surge in production, to the extent that is

found in Table 1.2. 6

One possible concern could be that there are state specific or district specific time

trends that are driving these results. To address this concern, I allow for trends in

the identifying variables in columns (2) and (3) and I find that the point estimate is

robust. In columns 7 and 8, I look at two other crops and do not find any significant

effect of this policy. Again, a reason could be that these crops are much less important

in terms of value and not all states produce these whereas rice is a more universal

crop in a country like India. So, with access to credit, given rice is more profitable in

India, farmers are expected to invest more in rice production. However, it is reassuring

that even though not significant, the point estimates on these are still positive which

suggests an increase in overall production.

Technology Adoption

A possible mediating channel for an increase in rice production could be adoption

of technology. Existing studies have shown that credit constraints are important hin-

drances in adoption of technology (Croppenstedt, Demeke and Meschi 2003). Mukher-

jee (2012) uses Indian household data to show that access to banks leads to better

6India had a major drought in 2002 which affected several rice farmers. Rainfall was about 56%below normal in July and almost 22% less rain was recorded overall (see Bhat 2006). In general thisshould not impact my analysis. However, there maybe concerns that banked districts in aligned statesmight have responded differently in terms of providing support to the agricultural system and thereforeit confounds the estimate somewhat. I find that the point estimates are not very different if we exclude2002 which alleviates these concerns. These results are not reported but are available upon request.

22

adoption of High Yielding Variety (HYV) seeds in production. Since the KCC pro-

gram intended to provide more credit access, it is interesting to examine whether the

relaxing of credit constraints has a similar effect as Mukherjee (2012) on aggregate.

Column 4 in Table 1.2 suggests that overall crop area put under HYV seeds usage

is higher by 71 thousand hectares with exposure to KCC. This is suggestive evidence

that access to credit leads to some technology adoption. As with overall production,

the point estimate here is also robust to linear de-trending as reported in columns 5-6.

These reduced form effects can be viewed as mediating channels for an increase in rice

production.

Threats to Identification: Check for Pre-Trends

The identification strategy would be invalidated in the case of pre-existing differen-

tial trends in the areas plausibly exposed more to KCCs compared to the ones not

exposed as much. One example would be if some districts in aligned states are tra-

ditional strong holds of the political party in the center, those districts may in any

case get preferential treatment historically and the coefficient we are picking up is not

the true causal effect of the policy. To alleviate concerns such as these, in Figure 1.1 I

plot all the β1 coefficients for crop production outcomes by year instead of interacting

with I(Y EAR > 1998). The dotted lines represent 95% confidence intervals. In other

words, instead of using all the previous years as the omitted reference group, I exclude

the year 1986 and compare the year specific effects with respect to this excluded year.

Each point of the graph represents the following object for year t:

23

[(Yaligned,morebanks − Yaligned,lessbanks)− (Ynonaligned,morebanks − Ynonaligned,lessbanks)]t

− [(Yaligned,morebanks − Yaligned,lessbanks)− (Ynonaligned,morebanks − Ynonaligned,lessbanks)]1986

(1.3)

I find that these coefficients, for all the outcomes are not statistically different from

zero prior to the policy year (marked by a vertical line) and for rice production, they

become positive since 1998. These suggest that the areas identified as exposed more to

KCCs were not systematically different from the areas without as much KCC exposure

as per my identification strategy.

In figure 1.2, I perform the same exercise but for HYV area as an outcome. The

coefficient does not jump at 1998 as sharply as for rice production but at least prior to

1998 it is never significant, which supports the identifying assumption somewhat.

1.5.2 Do Households Change their Borrowing Patterns?

In this section I look at the impacts of these agricultural credit reforms on outcomes

related to borrowing and lending. I report reduced form regressions using the house-

hold data. Unfortunately the district panel dataset (VDSA) does not provide any in-

formation on credit and therefore it is not possible to compare these findings at the

district level. So all of the following analyses are based on the cross sectional dataset.

24

Total Borrowing

Panel A of Table 1.3 reports regression results for the outcomes I discuss here. Through-

out the table I report results for all available households and 2 sub categories. First,

columns titled ‘cultivator’ represent those households whose main income source is

cultivation. Second, columns titled ‘Rice’ are for those households who produce any

rice. I find from columns 1-3 that on average there is no effect on whether people ex-

posed to KCC are more likely to borrow. The dependent variable is based on answers

to the survey question of whether the household had any loans in the last 5 years. The

policy was implemented from 1998 and the survey is based on 2004-05, so it is hard to

make conclusive statements about the estimated coefficients, especially because of the

lack of precision. I also do not find any significant effect on total outstanding debt.

The more interesting results come from columns 7 to 12. I find that on average,

households have lesser number of loans in the last 5 years. For every 2 rice farmers,

I estimate 3 fewer loans with exposure to the KCC program. I do not find any evi-

dence of the policy impacting the margin of whether the main creditor is a bank for

the households. I define bank as the main creditor if the largest loan, conditional on

borrowing, comes from a bank. The fact that this margin is unaffected by the pol-

icy allows me to look at effects of the program on a sub sample of households who

borrow from banks. The KCC policy was expected to operate through banks, so the

households who actually borrow from banks are likely to be affected by this program

the most. I look at outcomes like production, consumption and income for this sub

sample of households in the following sections.

25

Analysis of the Largest Loans

In Panel B, I restrict attention only to the largest loans of households in the 5 years

before the survey. Columns 1-6 focus on the largest loan from any source. The rest

of the columns focus on the largest loans if the source is reported to be a bank. I do

not find any difference in interest rates across the board. Although for bank loans, the

negative coefficient (and the lower mean interest rates) are suggestive that the policy

led to availability of cheaper credit because one feature of the reform was to allow

borrowing at lower rates of interest. Again, these estimates are imprecise, so we have

to be cautious with interpreting these.

The effects on loan size are significant. Not only do I find that the average house-

hold increasingly exposed to KCC borrows almost 9 thousand INR more than the

average household less exposed to KCC, but this number is 16 thousand INR for the

average rice farmer. This is with respect to loans from any source. If I restrict the sam-

ple to largest loans coming from banks, these numbers are considerably higher. The

average household borrowing from banks and exposed more to KCC has a largest loan

that is 41 thousand INR bigger in size than the one with less exposure to KCC. These

numbers are very similar for the cultivator and the rice farmer samples. These results

are consistent with theories of expansion of credit as a result of KCC as well as access

to credit. Whether the higher borrowing is because in the counterfactual households

are constrained or due to the fact that loans are now cheaper cannot be seperated with

this exercise though the point estimates on the borrowing margins in panel A sug-

gests that most of the effect is driven by existing borrowers and not new borrowers.

Eitherway, this helps corroborate the findings on production. If borrowing increased,

irrespective of the channel, we would expect more investment and therefore higher

26

production.

1.5.3 Effects on Household Production, Income and Consumption

It is interesting to examine how the higher borrowing estimated above translates into

spending and income. The following sections are devoted to this exercise. I first check

if production and sales increased for rice at the household level, which was the crop

that appeared to have been most affected by the policy in the district analysis. Then I

estimate effects of the program on household income and finally look at consumption

expenditure.

Rice Production and Sales

I use the IHDS crop level data, as described above, to estimate the reduced form ef-

fects of the KCC program on production outcomes. Results are reported in Table 1.4.

Most estimates are imprecise with large standard errors clustered by dsitrict. I restrict

attention to only those households that produce some positive amount of rice. Col-

umn 1 suggests minimal effects on overall household production levels but if I restrict

the sample to only those farmers who sell their output, as in columns 3 and 5, I find

suggestive evidence of large increases in production levels and revenue. The increase

in revenue is almost 40 thousand INR per year.

In columns 2, 4 and 6 I look at these outcomes for the subsample of bank borrowers

only. I find significant increases in production and revenue from sales of rice. This is

consistent with earlier findings of increase in production at the district level and bigger

bank loans. In the coutnerfactual, if households did not have access to larger loans

prior to the introduction of KCC, they may have faced difficutlies in financing their

27

production technology. With KCC they can secure larger formal sector loans which

allows productive investments and that transpires into higher output and revenue.

Income

Another way to corroborate the idea of higher agricultural output with increased ac-

cess to credit is to see if this translates into effects on household outcomes. Increased

agricultural output is only expected to have welfare effects if there is an observable

increase in income of the farmers. Table 1.5 reports results that look at this dimension.

When I restrict the sample to households whose main income source is cultivation

and look at the reduced form policy effects on incomes from their farms, I find in-

comes higher by 129 INR (USD 2) which is about 25% compared to the mean. This is

approximately a 24 INR monthly increase per capita for rice farmers. 7

I also check for non-farm income and find no effects. If the reduced form effects are

operating through enhanced credit access, especially for households with previously

no access to credit, then we would onlyt expect farm incomes to be higher because the

policy was directed towards farming households.

When I restrict the sample to only bank borrowers, which we have now identified

as the group of people most likely to be affected by the policy, I largely find significant

effects on income. Both per-capita income and per-capita farm income is likely to be

higher for these households if exposed to the KCC program more.

7Effects are imprecise as before but if we compare this to the estimated effects on revenue of ricefarmers we can do some rough calculations. A 24 INR per capita increase in income (profits) of ricefarmers would imply a yearly per capita increase in income of 288 INR. The average household hasfive or six members, so this translates to a total annual profit of around 1600 INR. With estimated salesrevenue increases of 40 thousand INR annually, this implies that costs and investments would havebeen higher by around 38 thousand for rice farmers.

28

Consumption Expenditures

I do not find any effect of enhanced credit access on overall consumption expendi-

tures but as reported in Table 1.6, composition of consumption expenditure changes.

Banerjee et al (2015) in their microfinance experiment find that spending categories are

sensitive to credit access and my results are consistent with their findings in a larger

nationwide setting. Similar to their experimental results, I find a decrease in expen-

diture on what is coined as ‘temptation goods’. These are expenses on tobacco, beetel

leaves etc and credit access has been believed to be a ‘disciplining device’ of sorts and

therefore exposure to credit reduces expenditure on these items. I also find an increase

in expenditure on recurring purchases of day to day household items.

The vast health economics literature also predicts that with increases in income,

stress levels decline and as a result consumption of goods like tobacco and alcohol

would go down (Cotti, Dunn and Tefft 2015). If access to credit led to higher produc-

tion and higher income, it is not surprising that consumption expenditures decrease

on temptation goods.

Expenditure on temptation goods is lower by 29 INR per month whereas day to

day spendng is higher by about 11 INR. For cultivators, this effect is 36 INR and and 16

INR respectively and is estimated with greater precision. The effects are even higher

for the sample of households who borrow from banks. One possible explanation con-

sistent with the findings would suggest that with credit constraints being relaxed,

households now can plan out their future spending stream better and spend money

on more productive uses that would be welfare enhancing in the long run whereas

they cut back on less productive consumption like tobacco etc. Even if the effects

are not through relaxation of credit constraints, expansion of credit could also have

29

similar effects.

1.5.4 Falsification Exercise

In this section, I perform a robustness check for my identification strategy and de-

scribe a falsification test. The identifying assumption for my analysis was that any

differences in districts with more banks compared to districts with less banks in 2005

in aligned states is attributable to the KCC program after controlling for trend differ-

ences in these districts using the non-aligned states. However, one maybe worried

that prior to the policy, these areas were already different and what we are picking

up is an existing trend. Figures 1.1 and 1.2 using the district panel data alleviates this

concern but here I present an alternate test using household data.

The cross sectional data is from the 2005 Indian Human Development Survey

(IHDS). A portion of the households interviewed in 2005 were drawn from an ear-

lier survey known as the Human Development Profile of India (HDPI) conducted in

1993. Since HDPI was in a year before the policy, I use the above identification strat-

egy for the households that can be traced back and run the regression equation (1) for

some comparable outcome variables but for the year 1993. If θ1 is the potential effect

of the KCC policy with the 2005 data, then for the same regression with the 1993 data,

we would not expect it to be significantly different from zero. Columns (1) and (2) of

Table 8 report the θ1 coefficients for the above regression for both 1993 and 2005 data

using the comparable Xs and for the comparable Y s. I report the regression results

for per capita income in Table 1.7.

I find that coefficients are systematically higher in 2005 whereas they are never sig-

nificantly different from zero in 1993. This suggests that the regions being compared

30

in my estimation were not different prior to the policy and any difference arising post-

policy may therefore be attributable as a reduced form impact of the program.

In column (3), I repeat the regressions from column (2) using the same sample but

adding other controls as used in the main analysis above. These additional controls

like number of persons in the family, number of children and married persons were

not available in 1993. I find that most of the effects are still pretty much the same as in

column (2) though the point estimates are marginally bigger.

1.6 Conclusions

In this paper I looked at a major agricultural credit reform in India known as the Kisan

Credit Card policy which simplified the functioning of the agricultural credit market.

A stated goal of the policy was to relax credit constraints on the poor. I used plausibly

exogenous geographic variation in the outreach of the program to identify the causal

impact of the policy. I find evidence that the reform led to large scale increases in ag-

gregate agricultural output. Rice, the major crop of India seems to have been the most

affected with a surge in production post-policy. There seems to have been significant

adoption of technology by putting more area under cultivation to use of HYV seeds.

Using a household dataset, I estimated effects of this policy on borrowing compo-

sitions. I find households are likely to have fewer but bigger loans with exposure to

the program. Also, size of the largest loan coming from banks is bigger for households

in areas exposed to KCC. No significant effects are estimated on interest rates.

I further looked at the impacts of this policy on outcomes like consumption, pro-

duction and income. I find suggestive evidence of increase in rice production and sales

31

revenue. I estimate an increase in farm income of around 129 INR (USD 2) monthly

per capita with enhanced credit access. The reduced form effects of the policy further

suggest that credit access acts a potential disciplining device where people spend less

on unproductive consumption and spend more on productive or investment goods.

There is no effect however, on overall consumption expenditure.

I identify households with a bank loan as the most affected category and find that

all estimated effects are much more pronounced for this sub sample of households

suggesting that a mediating channel for the reduced form estimates are bank loans.

Since KCC by design had to operate through banks, this provides confidence about

our identification strategy picking up the effect of the KCC policy.

32

1.7 References

1. Angelucci, M., Karlan, D. and Zinman, J (2015). ‘Microcredit Impacts: Evidence

from a Randomized Microcredit Program Placement Experiment by Comparta-

mos Banco’ American Economic Journal - Applied Economics

2. Asher, S and Novosad, P. (2015). ‘Politics and Local Economic Growth’ Working

Paper - Dartmouth - http://www.dartmouth.edu/ novosad/asher-novosad-politicians.pdf

3. Asubel, L. (1991). ‘The Failure of Competition in the Credit Card Market’ Amer-

ican Economic Review

4. Attanasio, O., Augsburg, B., De Haas, R., Fitzsimons, E. and Harmgart, H. (2015).

‘The Impacts of Microfinance: Evidence from Joint Liability Lending in Mongo-

lia’ American Economic Journal - Applied Economics

5. Augsburg, B., De Haas, R., Harmgart, H., and Meghir, C. (2015). ‘The Impacts of

Microcredit: Evidence from Bosnia and Herzegovina’ American Economic Journal

- Applied Economics

6. Banerjee, A. V. and Duflo, E. (2014). ‘Do firms want to Borrow More? Testing

Credit Constraints using a Directed Lending Program.’ Review of Economic Stud-

ies

7. Banerjee, A. V., Duflo, E., Glennerster, R. and Kinnan, C. (2015) ‘The Miracle

of Microfinance. Evidence from a Randomized Evaluation’. American Economic

Journal - Applied Economics

33

8. Banerjee, A. V. and Munshi, K. (2004). ‘How Efficiently is Capital Allocated?

Evidence from the Knitted Garment Industry in Tirupur.’ Review of Economic

Studies

9. Bhat, G. S (2006). ‘The Indian Drought of 2002 - a sub seasonal Phenomenon?’

Quarterly Journal of the Royal Meteorological Society

10. Bond, P., Musto, D and Yilmaz, B. (2009). ‘Predatory Mortgage Lending’ Journal

of Financial Economics

11. Burgess, R. and Pande, R. (2005) ‘Do Rural Banks matter? Evidence from the

Indian Social Banking Experiment.’ American Economic Review

12. Chanda, A. (2012). ‘Evaluating the Kisan Credit Card Scheme’ International Growth

Center Working Paper 12/0345

13. Chibber, P., Shastri, S and Sisson, R. (2004). ‘Federal Arrangements and the Pro-

vision of Public Goods in India’ Asian Survey

14. Chin, A., Karkoviata, L and Wilcox, N. (2011) ‘Impact of Bank Accounts on Mi-

grant Savings and Remittances: Evidence from a Field Experiment.’ Working

Paper, University of Houston

15. Cotti, C., Dunn, R and Tefft, N. (2015). ‘The Dow is Killing Me: Risky Health

Behaviors and the Stock Market’ Health Economics

16. Crepon, B., Devoto, F., Duflo, E and Pariente, W. (2015) ‘Estimating the Impact of

Microcredit on Those Who Take It Up: Evidence from a Randomized Experiment

in Morocco’ American Economic Journal - Applied Economics

34

17. Croppenstedt, A., Demeke, M. and Meschi, M. (2003) Technology Adoption in

the Presence of Constraints: the Case of Fertilizer Demand in Ethiopia’ Review of

Development Economics

18. De Mel, S., McKenzie, D. and Woodruff, C. (2008). ‘Returns to Capital in Mi-

croenterprises. Evidence from a Field Experiment.’ Quarterly Journal of Economics

19. Iyer, L. and Mani, M (2012). ‘Traveling Agents: Political Change and Bureau-

cratic Turnover in India’ Review of Economics and Statistics

20. Karlan, D. and Murdoch, J (2009). ‘Access to Finance’. Handbook of Development

Economics - Chapter 2

21. Melzer, B. (2011). ‘The Real Costs of Credit Access: Evidence from the Payday

Lending Market’ Quarterly Journal of Economics

22. Mukherjee, S (2012). ‘Access to Formal Banks and Technology Adoption: Evi-

dence from Indian Household Panel Data’ University of Houston - Working Paper

23. O’Donoghue, T. and Rabin, M. (1999). ‘Doing it Now or Later’ American Economic

Review

24. Planning Commission of India (2002). ‘Support from the Banking System: A

case Study of the Kisan Credit Card’ Study Report 146, Socio Economic Research

Division

25. Samantara, Samir (2010). ‘Kisan Credit Card - A Study” Occasional Paper 52,

National Bank of Agriculture and Rural Development, Mumbai

35

26. Tarozzi, A. Desai, J and Johnson, K (2015). ‘The Impacts of Microcredit: Evidence

from Ethiopia’ American Economic Journal - Applied Economics

36

TABLE 1.1: Comparing Means of statewise spread of KCC in 2000 byaligned

Aligned State Not Aligned State ∆(1) (2) (1)-(2)

KCC Coverage (in percentages) 30.596 12.710 17.886

Standard Deviation 17.223 12.617 H0 : ∆ = 0p-value < 0.001

Notes: KCC coverage is obtained from Planning Commission of India reports. It is calculated as thenumber of KCCs issued as percentage of total operational holdings in a given state in the year 2000. Iuse the definition of aligned as described for states aligned in 1998 and use the coverage figures 2 yearson. This table suggests that aligned states had much higher initial growth of the policy which providessupport to the use of aligned as a dimension of identification.

37

FIGURE 1.1: Crop Production: Year-specific coefficients for aligneds ·morebanksd

-100

-50

0

50

100

150

200

250

1986 1991 1996 2001 2006

Maize Production

95% ConfidenceIntervals

-100

-50

0

50

100

150

200

250

1986 1991 1996 2001 2006

Wheat Production

95% ConfidenceIntervals

Notes: The vertical line marks the policy year whereas the dotted line represent 95% confidence intervals. Regressions use the VDSA district panel dataset and control forcrop specific area under production, rainfall, crop specific iirigated area, markets nearby, gross cropped and irrigated areas and state fixed effects.

38

FIGURE 1.2: HYV Use: Year-specific coefficients for aligneds ·morebanksd

Notes: The vertical line marks the policy year whereas the dotted line represent 95% confidence intervals. Regressions use the VDSA district panel dataset and control forrainfall, markets nearby, gross cropped and irrigated areas and state fixed effects.

39

TAB

LE

1.2:

VD

SAD

istr

ictD

atas

et:E

ffec

tson

Cro

pPr

oduc

tion

and

HY

VU

se

Ric

ePr

oduc

tion

Are

aun

der

HY

VSe

eds

Whe

atM

aize

(1)

(2)

(3)

(4)

(5)

(6)

(7)

(8)

aligneds·m

orebanks d·I

(YEAR>

1998

)88

.7**

*89

.2**

*89

.8**

*71

.5**

*72

.7**

*72

.6**

*9.

996.

51(2

1.4)

(21.

5)(2

1.4)

(20.

8)(2

1.2)

(21.

6)(1

6.1)

(6.7

)

Line

arTr

end

inaligned

Yes

Yes

Yes

Yes

Line

arTr

end

inmorebanks

Yes

Yes

R2

0.94

0.94

0.94

0.75

0.75

0.75

0.97

0.83

Obs

erva

tion

s49

9249

9249

9241

7241

7241

7249

7849

92

Mea

nof

Dep

Var

284.

321

3.7

181.

738

.1

Not

es:A

naly

ses

inth

ista

ble

are

base

don

the

dist

rict

pane

ldat

aset

.Eac

hco

lum

npr

esen

tsa

diff

eren

treg

ress

ion.

Prod

ucti

onfig

ures

are

annu

al,u

nits

1000

tonn

esan

dH

YV

area

isin

term

sof

1000

hect

ares

.A

llre

gres

sion

sin

clud

est

ate

and

year

fixed

effe

cts

and

cont

rolf

oral

lthe

doub

lein

tera

ctio

nte

rms

and

base

line

vari

able

saligned

,after

andmorebanks.

Ico

ntro

lfor

the

area

put

unde

rri

cecu

ltiv

atio

nan

dal

soir

riga

tion

spec

ific

tori

ce.A

ddit

iona

lcon

trol

sin

clud

era

infa

ll,gr

oss

crop

ped

and

irri

gate

dar

ea,p

rese

nce

ofm

arke

ts.C

lust

ered

Erro

rsar

eat

the

dist

rict

leve

inpa

rent

hese

s.**

*p<

0.01

**p<

0.05

*p<

0.1

40

TAB

LE

1.3:

IHD

SD

atas

et:E

ffec

tson

Borr

owin

gC

ompo

siti

on

PAN

ELA

IfBo

rrow

sTo

talO

utst

andi

ngD

ebt

Num

ber

ofLo

ans

Mai

nC

redi

tor

isBa

nk

All

Cul

tiva

tor

Ric

eA

llC

ulti

vato

rR

ice

All

Cul

tiva

tor

Ric

eA

llC

ulti

vato

rR

ice

(1)

(2)

(3)

(4)

(5)

6)(7

)(8

)(9

)(1

0)(1

1)(1

2)

aligned·m

orebanks

-0.0

72-0

.059

-0.1

38*

-0.5

14-7

.984

-0.0

27-1

.351

**-0

.998

-1.5

90**

-.039

-.067

-.068

(0.0

49)

(0.0

60)

(0.0

75)

(6.5

89)

(9.4

41)

(9.6

70)

(0.6

51)

(0.6

67)

(0.7

43)

(0.0

23)

(0.0

47)

(0.0

62)

Mea

n0.

458

0.50

80.

509

38.6

3849

.252

34.8

273.

170

3.23

23.

385

0.12

10.

174

0.15

1

Obs

erva

tion

s21

450

7907

5700

1025

040

2126

7810

111

4147

2827

2145

079

0757

00

PAN

ELB

Onl

yLa

rges

tLoa

nsO

nly

Larg

estL

oans

from

Bank

s

Inte

rest

Rat

esLo

anSi

zeIn

tere

stR

ates

Loan

Size

All

Cul

tiva

tor

Ric

eA

llC

ulti

vato

rR

ice

All

Cul

tiva

tor

Ric

eA

llC

ulti

vato

rR

ice

(1)

(2)

(3)

(4)

(5)

6)(7

)(8

)(9

)(1

0)(1

1)(1

2)

aligned·m

orebanks

0.08

00.

014

0.19

29.

414*

**12

.068

16.7

25*

-0.0

37-0

.085

-0.1

1441

.364

***

37.8

17**

44.1

81**

(0.2

39)

(0.2

07)

(0.3

10)

(4.7

59)

(7.6

72)

(10.

197)

(0.0

72)

(0.0

82)

(0.0

82)

(2.9

73)

(14.

576)

(19.

470)

Mea

n2.

105

1.88

62.

157

32.7

1940

.828

32.9

281.

059

1.07

91.

045

62.9

6165

.314

58.9

79

Obs

erva

tion

s10

114

4145

2827

1011

741

5128

7826

9614

3788

026

9614

3788

0

Not

es:E

ach

colu

mn

repr

esen

tsa

diff

eren

treg

ress

ion.

The

sam

ple

inPa

nelB

incl

udes

answ

ers

toqu

esti

ons

abou

tthe

larg

estl

oan

inla

st5

year

sfo

rth

eho

useh

olds

.M

onet

ary

Val

ues

(for

loan

size

and

outs

tand

ing

loan

s)ar

ein

INR

1000

unit

s.C

olum

ns1-

3in

Pane

lAre

port

regr

essi

ons

whe

reth

ede

pend

entv

aria

ble

isa

dum

my

for

whe

ther

the

hous

ehol

dha

san

ybo

rrow

ing

inth

epa

st5

year

s.To

talo

utst

andi

ngde

btis

the

vari

able

for

how

muc

hth

eho

useh

old

curr

entl

yow

esot

hers

cond

itio

nal

onno

n-ze

roou

tsta

ndin

gde

bt.T

henu

mbe

rof

loan

sva

riab

leis

also

wit

hre

spec

tto

num

ber

oflo

ans

inpa

st5

year

s.Th

ede

pend

entv

aria

ble,

Mai

nC

redi

tor

isBa

nk,

isa

dum

my

indi

cati

ngif

the

larg

estl

oan

ofa

borr

ower

com

esfr

oma

bank

and

take

sth

eva

lue

zero

for

borr

ower

sfr

omot

her

sour

ces

asw

ella

sno

nbo

rrow

ers.

The

coef

ficie

nts

repo

rted

are

foraligned·m

orebanks.

All

regr

essi

ons

incl

ude

stat

efix

edef

fect

san

dco

ntro

lfor

base

linemorebanks

vari

able

.A

ddit

iona

ldem

ogra

phic

cont

rols

incl

ude

num

ber

ofpe

rson

sin

each

fam

ily,n

umbe

rof

child

ren

inea

chho

useh

old,

num

ber

ofm

arri

edm

enan

dm

arri

edw

omen

,age

and

educ

atio

nle

vels

ofm

enan

dw

omen

.C

ulti

vato

rre

pres

ents

the

sam

ple

ofho

useh

olds

who

sem

ain

inco

me

sour

ceis

repo

rted

tobe

cult

ivat

ion

and

allie

dag

ricu

ltur

e.R

ice

farm

ers

are

hous

ehol

dsw

hopr

oduc

ea

posi

tive

amou

ntof

rice

.Clu

ster

edSt

anda

rdEr

rors

are

atth

edi

stri

ctle

vel.

***

p<0.

01**

p<0.

05*p<

0.1

41

TAB

LE

1.4:

IHD

SD

atas

et:E

ffec

tson

Ric

ePr

oduc

tion

and

Sale

s

All

Prod

ucer

sIf

Sells

Out

put

Qua

ntit

yQ

uant

ity

Pric

eX

Qua

ntit

y

Full

Sam

ple

Bank

Borr

ower

sFu

llSa

mpl

eBa

nkBo

rrow

ers

Full

Sam

ple

Bank

Borr

ower

s(1

)(2

)(3

)(4

)(5

)(6

)

aligned·m

orebanks

0.91

710

.809

**7.

719

18.9

82**

40.8

810

7.63

**(4

.033

)(5

.299

)(6

.355

)(7

.636

)(3

8.62

)(4

3.12

)

Obs

erva

tion

s71

1899

725

8346

425

8346

4

Mea

n21

.067

25.2

4440

.736

40.3

6523

8.54

228.

67

Not

es:

The

sam

ple

isre

stri

cted

toon

lyri

cefa

rmer

sin

the

IHD

Sda

tase

t.Ea

chco

lum

nre

pres

ents

adi

ffer

entr

egre

ssio

n.A

llre

gres

-si

ons

incl

ude

stat

efix

edef

fect

san

dco

ntro

lfor

base

linemorebanks

vari

able

.A

ddit

iona

ldem

ogra

phic

cont

rols

incl

ude

num

ber

ofpe

rson

sin

each

fam

ily,

num

ber

ofch

ildre

nin

each

hous

ehol

d,nu

mbe

rof

mar

ried

men

and

mar

ried

wom

en,a

gean

ded

ucat

ion

leve

lsof

men

and

wom

enan

dar

eaun

der

rice

prod

ucti

on.T

hede

pend

entv

aria

ble

inco

lum

ns(1

)to

(4)i

sri

cepr

oduc

tion

inte

nths

ofa

quin

tal.

The

depe

nden

tvar

iabl

ein

colu

mn

5is

the

reve

nue

from

sale

ofri

ceco

ndit

iona

lon

selli

ngri

ce.T

heun

its

are

INR

1000

.Ba

nkBo

rrow

ers

repr

esen

tth

esa

mpl

eof

hous

ehol

dsw

hoha

vebo

rrow

edin

the

past

5ye

ars

and

thei

rla

rges

tlo

anco

mes

from

aba

nk.C

lust

ered

stan

dard

erro

rsat

the

dist

rict

leve

lin

pare

nthe

ses.

Num

ber

ofcl

uste

rsis

282.

***

p<0.

01**

p<0.

05*p<

0.1

42

TAB

LE

1.5:

IHD

SD

atas

et:E

ffec

tson

Inco

me

Per

Cap

ita

Inco

me

Per

Cap

ita

Farm

Inco

me

Per

Cap

ita

Non

Farm

Inco

me

All

Cul

tiva

tor

Bank

Borr

ower

All

Cul

tiva

tor

Bank

Borr

ower

sA

llBu

sine

ssPe

rson

(1)

(2)

(3)

(4)

(5)

6)(7

)(8

)

aligned·m

orebanks

-0.0

160.

053

0.32

3*0.

043

0.12

90.

424*

0.03

0.01

2(0

.069

)(0

.141

)(0

.197

)(0

.063

)(0

.140

)(0

.223

)(0

.060

)(0

.159

)

Mea

n0.

715

0.73

10.

951

0.20

70.

476

0.41

10.

463

0.80

8

Obs

erva

tion

s21

117

7634

2634

2029

172

6225

0137

6967

7

Not

es:E

ach

colu

mn

repr

esen

tsa

diff

eren

treg

ress

ion.

The

coef

ficie

nts

repo

rted

are

foraligned·m

orebanks.

All

regr

essi

ons

incl

ude

stat

efix

edef

fect

san

dco

ntro

lfor

base

linemorebanks

vari

able

.Add

itio

nald

emog

raph

icco

ntro

lsin

clud

enu

mbe

rof

pers

ons

inea

chfa

mily

,num

ber

ofch

ildre

nin

each

hous

ehol

d,nu

mbe

rof

mar

ried

men

and

mar

ried

wom

en,a

gean

ded

ucat

ion

leve

lsof

men

and

wom

en.

Inth

ista

ble,

mon

etar

yfig

ures

are

inm

onth

lyIN

R10

00w

hich

isap

prox

mon

thly

USD

20.

The

sam

ple

size

isva

riab

lein

this

tabl

eba

sed

onth

egr

adat

ions

‘ifcu

ltiv

ator

’and

‘ifbu

sine

sspe

rson

’.A

hous

ehol

dis

deno

ted

asa

cult

ivat

orif

the

resp

onde

ntre

port

edth

atth

eir

mai

nin

com

eso

urce

isfr

omcu

ltiv

atio

nor

allie

dac

tivi

ties

.Si

mila

rly

ifth

ere

port

edm

ain

sour

ceof

inco

me

isbu

sine

ss,I

clas

sify

thes

eho

useh

olds

as‘b

usin

ess

pers

ons’

.Ban

kBo

rrow

ers

repr

esen

tthe

sam

ple

ofho

useh

olds

who

have

borr

owed

inth

epa

st5

year

san

dth

eir

larg

estl

oan

com

esfr

oma

bank

.Clu

ster

edSt

anda

rdEr

rors

are

atth

edi

stri

ctle

vel.

***

p<0.

01**

p<0.

05*p<

0.1

43

TAB

LE

1.6:

IHD

SD

atas

et:E

ffec

tson

Con

sum

ptio

nEx

pend

itur

e

Per

Cap

ita

Mon

thly

Con

sH

ouse

hold

Item

sTe

mpt

atio

nG

oods

All

Cul

tiva

tor

Bank

Borr

ower

sA

llC

ulti

vato

rBa

nkBo

rrow

ers

All

Cul

tiva

tor

Bank

Borr

ower

s(1

)(2

)(3

)(4

)(5

)6)

(7)

(8)

(9)

aligned·m

orebanks

-0.0

019

0.00

020.

0084

0.01

130.

0161

*0.

0215

**-0

.028

3***

-0.0

360*

**-0

.011

4(0

.003

7)(0

.006

4)(0

.007

4)(0

.007

3)(0

.009

3)(0

.010

8)(0

.010

3)(0

.012

5)(0

.017

8)

Mea

n0.

0644

0.06

580.

0854

0.07

290.

0791

0.08

770.

0817

0.08

290.

0928

Obs

erva

tion

s21

437

7899

2695

2143

778

9926

9521

437

7899

2695

Not

es:E

ach

colu

mn

repr

esen

tsa

diff

eren

treg

ress

ion.

The

coef

ficie

nts

repo

rted

are

foraligned·m

orebanks.

All

regr

essi

ons

incl

ude

stat

efix

edef

fect

san

dco

ntro

lfor

base

linemorebanks

vari

able

.Add

itio

nald

emog

raph

icco

ntro

lsin

clud

enu

mbe

rof

pers

ons

inea

chfa

mily

,num

ber

ofch

ildre

nin

each

hous

ehol

d,nu

mbe

rof

mar

ried

men

and

mar

ried

wom

en,a

gean

ded

ucat

ion

leve

lsof

men

and

wom

en.M

onet

ary

figur

esar

ein

INR

1000

unit

she

re.T

hete

mpt

atio

ngo

ods

incl

ude

beet

elle

aves

/tob

acco

etc.

Bane

rjee

etal

(201

5)fin

dth

atm

icro

finan

ceas

adi

scip

linin

gde

vice

redu

ces

per

capi

taco

nsum

ptio

nof

thes

ego

ods

byIN

R9.

The

per

capi

tale

quiv

alen

tof

my

findi

ngof

INR

28re

duct

ion

isIN

R4.

The

HH

item

sin

clud

ere

gula

rpu

rcha

ses

like

soap

s,bu

lbs,

buck

ets,

inse

ctic

ides

etc.

Iuse

the

IHD

S20

05sa

mpl

eof

rura

lhou

seho

lds

whi

chha

ve26

734

obse

rvat

ions

.C

ulti

vato

rre

pres

ents

the

sam

ple

ofho

useh

olds

who

sem

ain

inco

me

sour

ceis

repo

rted

tobe

cult

ivat

ion

and

allie

dag

ricu

ltur

e.Ba

nkBo

rrow

ers

repr

esen

tth

esa

mpl

eof

hous

ehol

dsw

hoha

vebo

rrow

edin

the

past

5ye

ars

and

thei

rla

rges

tloa

nco

mes

from

aba

nk.

Clu

ster

edSt

anda

rdEr

rors

are

atth

edi

stri

ctle

vel.

***

p<0.

01**

p<0.

05*p<

0.1

44

TAB

LE

1.7:

Fals

ifica

tion

Exer

cise

1993

2005

2005

Com

para

ble

Con

trol

sA

llC

ontr

ols

(1)

(2)

(3)

Per

Cap

ita

Mon

thly

Inco

me

0.01

90.

094

0.12

7(0

.045

)(0

.125

)(0

.120

)

Per

Cap

ita

Mon

thly

Inco

me

(ifp

rinc

ipal

occu

pati

onis

cult

ivat

ion)

0.03

70.

292

0.32

9(0

.064

)(0

.198

)(0

.191

)

Not

es:I

run

redu

ced

form

regr

essi

ons

follo

win

gth

esa

me

iden

tific

atio

nst

rate

gyas

desc

ribe

din

the

text

for

2005

and

usin

gth

eH

DPI

data

from

1993

.Ius

eth

esa

me

hous

ehol

ds(8

947)

for

both

year

s.Fo

rth

esa

mpl

elo

okin

gon

lyat

cult

ivat

ors,

Ihav

e37

31ho

useh

olds

.In

colu

mns

1an

d2,

Irun

regr

essi

ons

usin

gon

lyth

eco

mpa

rabl

eco

ntro

lsw

hich

are

age,

educ

atio

nof

mal

e,ed

ucat

ion

offe

mal

ean

dst

ate

fixed

effe

cts.

Inco

lum

n3,

Ire-

run

regr

essi

ons

ofco

lum

n2

usin

gth

esa

me

sam

ple

buta

ddin

gth

eco

ntro

lsus

edin

the

orig

inal

anal

ysis

whi

char

eav

aial

able

for

2005

.The

addi

tion

alco

ntro

lsin

colu

mn

3ar

enu

mbe

rof

pers

ons

infa

mily

,num

ber

ofch

ildre

nan

dnu

mbe

rof

mar

ried

mal

esan

dfe

mal

es.

Ion

lyre

port

the

com

para

ble

outc

ome

vari

able

s.C

lust

ered

stan

dard

erro

rsat

the

dist

rict

leve

lin

pare

nthe

ses.

***

p<0.

01**

p<0.

05*p<

0.1

45

46

Chapter 2

“None of the Above" Votes in India and

the Consumption Utility of Voting(with Gergely Ujhelyi and Andrea Szabó)

2.1 Introduction

One possible solution to the “paradox” of why people bother to vote in large elections

is that voting yields consumption utility. Such consumption utility could be derived

from performing one’s civic duty, expressing one’s political views, or participating

in a democracy (Downs, 1957; Riker and Ordeshook, 1968; Brennan and Lomasky,

1993). Empirically distinguishing these consumption motives from each-other and

from other possible goals, such as a desire to affect the electoral outcome, is notori-

ously difficult. In this paper we propose to do this by using data from a natural ex-

periment in Indian elections and estimating a structural model of voter turnout using

techniques from the consumer demand literature.

We propose to identify different components of the consumption utility of voting

by exploiting a natural experiment in the world’s largest democracy. Following a

decision by the Indian Supreme Court, since September 2013 all state and national

elections in the country must offer a “None Of The Above” (NOTA) option to voters.

Votes cast for NOTA are counted (rather than simply discarded as invalid) but do not

affect the outcome of the election (the winner is still the candidate with a plurality of

votes among votes cast for candidates). In the five states that held local elections after

the Supreme Court ruling 1.7 million voters chose NOTA, and in the 2014 national

election 6 million voters voted for this option (representing 1.1% of all votes cast).

While elections with a NOTA-type option have been used elsewhere, none of them

came close to the scale of the Indian experiment.1

Because in the Indian system NOTA votes cannot affect the outcome of the elec-

tion, voters who choose this option must be motivated by a consumption utility to

vote. Such consumption utility can arise from two broad sources. It can be a general

utility obtained from showing up at the polls (such as complying with a social norm

to participate in the election), or it can be a utility specific to the option chosen by the

voter (such as utility derived from expressing one’s views).

Intuitively, we can distinguish between these two types of consumption utility by

asking how a NOTA-voter would have behaved in the absence of the NOTA option.

If without the NOTA option this voter would have voted for one of the candidates,

this is consistent with both a general and an option-specific utility of voting. By con-

trast if without NOTA this voter would have abstained, then the NOTA vote cannot

be explained by a general utility derived from showing up at the polls. Instead, the

voter must be voting for NOTA in order to obtain a utility specific to this option. Thus,

1In most elections the only way for a voter to participate without voting for a candidate is to cast aninvalid vote and these are difficult to distinguish from voting mistakes. In systems where a NOTA-typeoption is explicitly available to voters, it typically has electoral consequences, affecting who gets electedor whether the election has to be repeated. We review these different systems in section 2.3 below.

47

studying the counterfactual behavior of NOTA voters can test apart these two compo-

nents of the consumption utility from voting.

Our empirical work seeks to test whether, following the introduction of NOTA,

new voters showed up at the polls in order to vote for this option. This question

is challenging because it requires making statements about individual voter behavior

in the counterfactual no-NOTA scenario. Because ballots are secret, individual voter

behavior is observed neither with nor without the NOTA option. Instead, it must be

inferred from aggregate data.

To begin, we first ignore individual behavior and study the impact of NOTA on

aggregate turnout in a reduced form framework. This exercise exploits variation in

the effective timing of the NOTA reform created by the Indian electoral calendar: elec-

tions to the states’ legislative assemblies occur at different times in different states.

This allows us to study the impact of NOTA in a difference-in-differences framework

by comparing the change in voter turnout in states not yet affected by the policy to

changes in those that were already affected. From this analysis we estimate that, in the

average electoral district, the introduction of the NOTA policy significantly increased

turnout. This finding survives a variety of robustness checks and the magnitude of

the effect (2-3 percentage points) is similar to the vote share of NOTA observed in the

data.

While suggestive of new voters turning out to vote for NOTA, these aggregate pat-

terns do not provide conclusive evidence because they mask the substitution between

abstention, candidates, and NOTA at the individual level. To study this, we relate

the aggregate voting returns to individual voter behavior using a structural model

of voter demand for candidates. We adapt the BLP model of Berry, Levinsohn and

48

Pakes (1995) from the consumer demand literature, where consumers (voters) choose

between the products (candidates) of firms (parties) in various markets (electoral dis-

tricts). Voters have preferences over observed and unobserved candidate characteris-

tics (including NOTA) and abstention. The model explicitly allows for heterogeneity

in these preferences and links them to the aggregate vote shares we observe in the

data. Estimating the model allows us to recover the parameters of individual voters’

utility functions from this aggregate data. Using the estimates, we study how voters

substitute between choosing NOTA, one of the candidates, and abstention in counter-

factual simulations where the NOTA option is removed.

The results of this analysis indicate that NOTA increased turnout, which is in line

with the aggregate patterns observed in the reduced form exercise. Furthermore we

find that the magnitude of this increase explains virtually all the NOTA votes ob-

served in the data. We find negligible substitution towards NOTA away from the

candidates running for election. These results indicate that most voters who voted for

NOTA would normally abstain. In turn, this provides evidence for the existence of

consumption utility specifically from voting for NOTA. In this context, models that

do not include an option-specific utility of voting would have a hard time explaining

the data.

To the extent that participation in a democracy is valuable, our finding that having

a NOTA option on the ballot can increase voter turnout is relevant in its own right,

and provides support for the arguments of the Indian Supreme Court in introducing

this policy.

Our paper is related to the vast literature on voter turnout, some of which we

review in section 2.2 below. Instead of proposing a new model of turnout, we focus on

49

testing models apart by asking whether the presence of consumption utility specific to

the NOTA option (as opposed to consumption utility from simply showing up at the

polls) is necessary to explain our data. This approach is similar in spirit to Coate and

Conlin (2004) and Coate et al. (2008) who estimate and compare competing structural

models of turnout on data from Texas liquor referenda.

Methodologically, our paper offers a novel way to estimate vote returns in multi-

party elections. Some earlier approaches to this problem (e.g., Glasgow and Alvarez,

2005) have used discrete choice models with individual-level survey data, but such

data is subject to well-known biases in voters’ self-reported behavior (see, e.g., Selb

and Munzert (2013) and the literature cited therein). Other studies use aggregate ad-

ministrative data and purely statistical models to deal with the problem of conducting

“ecological inference” regarding voter preferences (see Cho and Manski (2008) for a

review). By contrast our BLP-based approach combines the advantages of a micro-

founded discrete choice model with those of aggregate administrative data. It allows

for rich heterogeneity in voter tastes for candidate characteristics and, because it is

micro-founded, offers the possibility of conducting counterfactual experiments. In

a different context, Rekkas (2007) also exploits some of these advantages of the BLP

model in her study of campaign expenditures in the 1997 Canadian election. Our pa-

per goes further by using panel data, allowing for heterogeneity in voters’ preferences

driven by demographics as in Nevo (2001), allowing for endogenous candidate choice

by the competing parties, and by conducting counterfactual experiments using the

estimated model.

Finally, our paper relates to previous studies of NOTA-type votes in the political

science literature (reviewed in section 2.3 below). We differ from this literature by

50

using NOTA votes to isolate the consumption utility from voting and by estimating a

structural model that can be used to answer normative questions about the desirability

of having this option on the ballot.

In the rest of the paper, section 2.2 explains how we propose to use NOTA votes

to identify various components of the consumption utility from voting. Section 2.3

describes the NOTA policy, explains how it differs from similar options available to

voters in other countries, and describes the Indian electoral setting we analyze. Sec-

tion 2.4 describes our data and section 2.5 documents the pattern of NOTA votes and

presents a difference-in-differences analysis of the effect of NOTA on turnout. Section

2.6 estimates the structural model and presents the counterfactual results. Section 2.7

concludes.

2.2 The consumption utility of voting

Why people vote is one of the classical questions of economics and political science.

In the “calculus of voting” model (Downs, 1957; Riker and Ordeshook, 1968; Fiorina,

1976), voters consider both instrumental and consumption benefits. They vote for

candidate j if

PjBj − c+ (Uj + U0) > 0 (2.1)

and abstain otherwise (where j = arg maxj′

(Pj′Bj′ + Uj′) is the voter’s preferred can-

didate). The first term in (2.1) is the expected instrumental benefit, where Pj is an

individual’s probability of being pivotal in the election of candidate j and Bj is the

benefit of the candidate winning (without loss of generality, all terms in (2.1) are as-

sumed to be non-negative). The second term represents any direct or opportunity

51

costs from voting. The final term is the consumption utility of voting, which captures

a wide range of factors sometimes referred to as “expressive utility” or “civic duty”:

“1. the satisfaction from compliance with the ethic of voting [...] 2. the satisfaction

from affirming allegiance to the political system [...] 3. the satisfaction from affirm-

ing a partisan preference [...] 4. the satisfaction of deciding, going to the polls, etc.

[...] 5. the satisfaction of affirming one’s efficacy in the political system” (Riker and

Ordeshook, 1968, p28). We separate this consumption utility into two components to

highlight that part of the utility (Uj) may depend on voting for a specific candidate

j (e.g., the satisfaction from expressing partisan support), while part of it (U0) only

depends on showing up at the polls regardless of who one votes for (e.g., satisfaction

from compliance with an ethical norm to vote).

Observing that in large elections the probability Pj of being pivotal is close to 0,

the recent literature seeking to explain turnout within the framework of the calculus

of voting equation (2.1) has followed various routes.2 First, voters could overestimate

Pj . Lab experiments show that, indeed, voters often overestimate the probability that

their vote will matter and suggest that this can explain turnout decisions (Duffy and

Tavits, 2008; Dittman et al., 2014). Relatedly, Ortoleva and Snowberg (2015) show

that turnout is higher in populations with more overconfident voters. Under these

conditions, turnout can be explained even if (Uj + U0) = 0.

A second set of studies present models that create an option-specific utility Uj . In

2Missing from equation (2.1) are instrumental motivations other than those related to winning. Forexample, it is possible that a voter votes in order to signal his preferences to affect the policies chosenafter the election. Or he may vote in order to encourage a candidate to run again in the future. Itis possible to treat such motivations in a strategic setting but the likelihood that a voter’s vote willbe pivotal in affecting policy or encouraging a candidate is likely to be small (see, e.g., Razin (2003)).Here we follow most of the literature in assuming that if such motivations exist, they are sources ofconsumption utility and hence part of Uj .

52

Coate and Conlin (2004) and Feddersen and Sandorini (2006), this utility represents

ethical considerations regarding what would be best for everyone in one’s group. In

other models, such as Shachar and Nalebuff (1999), Uj is created by the mobilization

efforts of political leaders. In Degan and Merlo (2011), Uj includes a psychological

disutility from the possibility of voting for the “wrong” candidate.

Another set of papers focus on the general utility U0 from showing up to vote.

For example, members of a group may observe turnout and draw inferences about

whether an individual is an “ethical type” (Bénabou and Tirole, 2006; Ali and Lin,

2013). Similarly, a voter may vote to avoid a feeling of shame from not having done

his duty, especially if others will ask whether one has voted (Harbaugh, 1996; Blais,

2000; DellaVigna et al., 2015).

While these studies convincingly demonstrate that the proposed models have ex-

planatory power, it is not always clear to what extent these models are necessary to

explain the data.3 In particular, can we rule out that one or both components of the

consumption utility Uj + U0 is 0?

As we explain below, the Indian NOTA policy allows us to study this question by

creating a “None Of The Above” option that voters can vote for but that, by design,

cannot affect the electoral outcome. First, we document that voters actually choose

NOTA. Because PNOTABNOTA = 0, from equation (2.1) a voter who chooses NOTA

must have

UNOTA + U0 > c (2.2)

i.e., there has to be a positive consumption utility of voting.

3Examples of papers that highlight the value of testing models apart include the pair of studies byCoate and Conlin (2004) and Coate, Conlin, and Moro (2008) which explicitly compare several differentmodels of turnout on the same dataset.

53

Second, we ask how a voter who chose NOTA would have voted in the absence of

the NOTA option. If the voter would have abstained, then PjBj − c + (Uj + U0) ≤ 0.

Combining with (2.2), we have

UNOTA > PjBj + Uj,

i.e., there has to be a positive option-specifc utility from voting for NOTA. For exam-

ple, a voter may derive utility from expressing his disapproval of all the candidates.

Conversely, if in the absence of NOTA the voter would have voted for one of the candi-

dates, then it is possible that there is no option-specific utility but U0 > 0 (for example,

a voter may vote to satisfy social pressure while deriving no specific utility from vot-

ing for any of the options on the ballot.). Thus, studying voters’ behavior with and

without NOTA offers a test for the existence of an option-specific utility of voting.

In this way, although not an actual candidate, studying NOTA votes gives us an

opportunity to test both for the existence of a consumption utility from voting and for

the existence of a consumption utility from voting for this particular option.4 The sec-

ond exercise is empirically challenging because it requires making statements about

individual behavior both with and without NOTA while, due to the secret ballot, such

behavior is never observed.4It is difficult to imagine a similar experiment where an actual candidate’s probability of wining

is administratively set to 0. (Having small-party candidates on the ballot who have little chance ofwinning is not the same experiment since voters’ believing that Pj > 0 cannot be ruled out.) Anotherpossibility is to run a lab experiment where Pj is set to 0, but creating an artificial environment wherepeople derive utility from voting may be difficult. In an interesting experiment Shayo and Harel (2012)create variation in P and find evidence that the moral superiority of an alternative affects voters’ be-havior when P is small.

54

2.3 Background

2.3.1 The Indian NOTA policy

In elections where a paper ballot is used, voters can participate without voting for

any of the candidates: they can hand in an empty ballot or otherwise intentionally

invalidate their vote. With the advent of electronic voting machines Indian voters

lost this possibility. In 2004, the citizen’s group People’s Union for Civil Liberties

(PUCL) filed a petition with the Supreme Court to rectify this and give voters the

ability to have their participation recorded without forcing them to vote on any of the

candidates.5 In its 2013 decision, the Supreme Court agreed:

“For democracy to survive, it is essential that the best available men

should be chosen as people’s representatives for proper governance of the

country. This can be best achieved through men of high moral and ethical

values, who win the elections on a positive vote. Thus in a vibrant democ-

racy, the voter must be given an opportunity to choose none of the above

[...] Democracy is all about choice. This choice can be better expressed by

giving the voters an opportunity to verbalize themselves unreservedly and

by imposing least restrictions on their ability to make such a choice. By pro-

viding NOTA button in the Electronic Voting Machines, it will accelerate

5Under the electronic voting machines, the only way for a voter to have his non-vote recorded wasto inform the clerk at the voting booth of his desire to do so. The clerk would then record this on thevoter ledger together with the voter’s thumbprint for identification. The PUCL argued that this wasunconstitutional, violating the secret ballot.

55

the effective political participation in the present state of democratic sys-

tem and the voters in fact will be empowered.” (PUCL vs. Union of India,

2013, p43-44).

Following the Supreme Court’s decision, since September 2013, all state and na-

tional elections in India give voters the option of recording a “None Of The Above”

vote on the voting machine. These votes are counted and reported separately but have

no role in the outcome of the election. In particular, votes cast on NOTA affect neither

the validity nor the winner of an election. Even if NOTA were to receive a majority

of the votes, the winner of the election would be the candidate who received the most

votes among the non-NOTA votes.

The NOTA policy received wide news coverage in both national and local me-

dia. In its decision the Supreme Court directed the Election Commission to undertake

awareness programs to inform the electorate of the new policy, and voter education

programs explicitly focused on explaining this new option to voters. As a result we

expect that most voters would be well-informed about the NOTA policy, including the

fact that NOTA votes would not affect the electoral outcome.6

2.3.2 NOTA-like options in other countries

In most countries voters can effectively cast a “none of the above” vote by intentionally

returning an invalid vote (e.g., leaving the ballot blank, writing on the ballot, or mark-

ing more than one candidate). Because it is typically impossible to know whether such

votes occur intentionally or by mistake, it is difficult to use them to draw conclusions

regarding voters’ intentional behavior (see, e.g., McAllister and Makkai, 1993; Herron6As we discuss in section 2.5, several patterns in the data also support this assumption.

56

and Sekhon, 2005; Power and Garand, 2007; Uggla, 2008; Driscoll and Nelson, 2014).

For some applications, the fact that invalid votes also include voting mistakes will

simply add measurement error to the “true” measure intended to capture negative

votes. In other cases, however, this will have an important impact on the interpreta-

tion of the results. For example, more invalid votes among the less educated can mean

either that these voters are more likely to make mistakes when filling out the ballot, or

they are particularly dissatisfied and intentionally cast invalid votes to express this.7

In some countries, while there is no NOTA option on the ballot, blank votes are

counted separately from invalid votes and are believed to represent a negative vote.

In principle, this system could be equivalent to the Indian NOTA, but in practice the

equivalence is unlikely to be perfect. First, blank votes could still represent voting

mistakes, especially if there is a judgement call to be made about whether a vote is

truly blank when it is being counted (for example, there could be markings on the

side of the ballot, a small dot inside the checkbox, etc.). Fujiwara (2015) finds that

the introduction of voting machines in Brazil reduced both blank and invalid votes

among the less educated, which is consistent with both of these containing voting

mistakes. Second, using the blank vote as an expression of dissatisfaction requires

a shared understanding among voters regarding what the vote represents. Whether

this social norm is operative in a given election is difficult to test. This is illustrated

by the findings of Superti (2015) who studies a set of municipal elections in Spain - a

country where the blank vote is generally understood to mean “none of the above.”

She shows that despite this common understanding, voter dissatisfaction following

7These two interpretations also have different welfare implications regarding the desirability of hav-ing a NOTA option on the ballot. In the first case, NOTA only serves to confuse the less educated; inthe second case, it gives disadvantaged segments of the population a voice.

57

a ban which prevented the Basque nationalist party from contesting an election was

likely expressed through an increase in invalid rather than blank votes.

Another feature that makes India a cleaner case study than other systems for the

analysis of voters’ motivations is the electoral impact of the NOTA vote. Recall that in

India the NOTA vote can never “win,” and due to the first-past-the-post system it has

no impact on the allocation of legislative seats. By contrast in Colombia if the “blank

vote” wins, new elections must be called with the rejected candidates prohibited from

running again. In Spain, while the blank vote can never win, seats are allocated in

a proportional system and a minimum 3% threshold must be reached for a party to

enter parliament. In both of these systems choosing the blank vote as opposed to

choosing one of the parties has immediate electoral consequences, affecting the mix

of candidates eventually elected for office. In the Indian case, NOTA votes cannot be

driven by electoral motivations in the current election.8

2.3.3 Assembly elections in India

We study voters’ behavior under NOTA in the context of Indian state elections. In

the Indian federal system, state governments are responsible for most areas of local

significance, including health care, education, public works, police and security, and

disaster management. State legislative assemblies are elected in single-member elec-

toral districts (called “constituencies”) in a first-past-the-post system. The party or

8A voter’s motivation (with any vote under any system) can always include a desire to affect long-run outcomes, e.g., by signaling his political preferences to the eventual winner in order to affect policy,or by encouraging a candidate to run in future elections. Because a single vote is just as unlikely to bepivotal in affecting these outcomes as it is in affecting who wins, we think that these motivations arebest viewed as alternative sources of the consumption utility derived from voting.

58

coalition that wins the most number of seats in an assembly forms the state govern-

ment headed by a Chief Minister and his council of ministers.9

Table 2.1 shows the timing of state assembly elections in our study period. Elec-

tions are held every 5 years but the electoral calendar varies widely across states. For

example some states held assembly elections in 2007 and 2012 while others in 2008

and 2013; some states always go to the polls in March while others always do so in

November. This variation in the timing of elections creates an important source of

identification for the analysis below.

In most states assembly elections are conducted separately from other elections.

Four states, Andhra Pradesh, Arunachal Pradesh, Odisha and Sikkim, hold elections

simultaneously with national elections. We will exclude these states from the analysis

below.

All state and national elections in India are conducted by the Election Commission

of India under the supervision of the chief election commissioner. Since independence,

the Commission has emerged as a highly regarded institution with a large degree

of autonomy (McMillan, 2010). Election dates are set well in advance and declared

as local holidays to reduce the cost of participation. Polling stations (“booths”) are

spread out throughout each constituency and enlisted voters are assigned to specific

booths. Voters go to their designated booth to cast their vote with their Elector’s Photo

Identification Card.10 Generally these booths are set up in neighboring schools or

9In states that have a bicameral legislature, the system just described applies to the lower house.Members of the upper house are either elected by the lower house or appointed by the Chief Ministeror the Governor (the representative of the federal government in the states).

10Voter Registration is a one time procedure. Except in special cases (such as for convicted criminals),once registered as a voter, a person can vote in all subsequent elections without having to go throughany further registration process. Once registered the voter’s name is on the voters’ list and he or shegets the identification card which needs to be produced at the polling station before being allowed tovote.

59

public buildings within a very small radius of one’s residence. Participation rates in

Indian elections tend to be high. In our state election data, average turnout is 71% and

only 7% of the constituencies had turnout lower than 50%. (By comparison turnout in

US midterm elections is typically around 40%.) The voting age is 18, and the average

constituency has approximately 180 thousand eligible voters.

Since 2004 all voting in India has taken place using electronic voting machines

(EVMs).11 Each candidate running in an election has a separate button assigned to

him on the machine. Next to the button is the symbol identifying the candidate (to

accommodate illiterate voters) and the voter pushes the button to record his vote. A

light illuminates confirming that the vote was successful.12 Under the NOTA policy,

one of the buttons on the machine is assigned to the NOTA option.

In the system of political reservation, some constituencies are designated Sched-

uled Caste (SC) and some Scheduled Tribe (ST). In these, only candidates from the

given caste can run (to win, they must still obtain a plurality of all votes regardless of

voters’ caste). The reserved status of SC and ST constituencies is set at the same time

as the electoral boundaries are drawn. In contrast to local (village) governments, state

elections have no political reservation for women.

The current electoral constituencies were set in April 2008 by a commission work-

ing under the Election Commission (see Table 2.1). This was the first time in over 30

years that electoral redistricting (“delimitation”) took place in India. All constituency

boundaries as well as the reservation status of the constituencies was fixed by the de-

limitation commission in order to reflect population figures of the 2001 census. As

11Electronic voting machines in India were introduced gradually beginning in 1999. Since 2004 allgeneral and state elections are conducted using these machines.

12These machines are simpler to operate than some of the EVMs used in other countries that some-times require a voter to follow written instructions, enter a candidate’s number on a keypad, etc.

60

described below, this redistricting poses challenges for the construction of our dataset

and our empirical strategy.

2.4 Data

2.4.1 Samples used for analysis

Our analysis uses two samples of constituencies: a panel serves as our primary dataset,

and we use a repeated cross-section as a secondary sample.

The instrumental variables used in the structural analysis require panel data, and

our main sample is a constituency level panel dataset of the 6 states that conducted

assembly elections in both 2008 and 2013 under the new electoral boundaries: Kar-

nataka, Chhattisgarh, Rajasthan, Madhya Pradesh, Delhi, and Mizoram (see Table 2.1).

One of these states, Karnataka held elections in both years without a NOTA option,

while the remaining 5 states had a NOTA option in 2013 but not in 2008.

The main obstacle to extending the panel data to more constituencies is the delim-

itation (electoral redistricting). This makes it impossible to include elections before

April 2008 in the panel as there is too little overlap between the old and new con-

stituencies to make constituency-level matching meaningful.13 For example, although

3 other states also held elections in both 2008 and 2013, they did so in February-March

and had their constituency boundaries redrawn between the two elections in April

13Using GIS software we have computed the maximum overlap of each current constituency’s areawith an old constituency. For example, a maximum overlap of 80% indicates that 80% of the currentconstituency’s area came from one constituency, while 20% came from one or more other constituencies.We find that half of the current constituencies have a maximum overlap of 62% or less and a quarter ofthe constituencies have a maximum overlap of 50% or less. This makes it impossible to match electoraldata across constituencies in a meaningful way.

61

2008 so we cannot include these states in the panel. Other states with consistent elec-

toral boundaries in our study period are those holding elections in 2014. However,

2014 was a national election year that made headlines around the world for its un-

usual outcome (the BJP led by Narendra Modi won by a landslide, the first time in 30

years that a single party won a majority of the legislative seats). Because 2014 state

assembly elections took place either simultaneously with or after the national election

(and in the latter case more than a year after the NOTA policy was introduced), the

national election could confound the impact of NOTA in these states. We therefore

decided to exclude these states from the panel analysis.

To obtain more power for a reduced form analysis, we use as a secondary dataset a

repeated cross section of constituencies in 25 states that conducted elections between

2006 and 2014. Like the panel, this dataset excludes the states that held assembly elec-

tions simultaneously with national elections (Andhra Pradesh, Arunachal Pradesh,

Odisha and Sikkim) since turnout considerations in these states are likely to be very

different.14 It also excludes the state of Bihar because its unique election calendar (2005

and 2010) would require earlier data on voter demographics than we have access to.

We next describe the information available in the primary (panel) and secondary

(repeated cross-section) datasets.

2.4.2 Election data

The electoral data comes from the Election Commission of India, which provides in-

formation on assembly elections at the candidate level. For each constituency we

14Since our main goal with the repeated cross section is to increase power, we include the states thatheld elections in 2014 but not simultaneously with the national election. Excluding these states makeslittle difference for the results.

62

know the list of candidates running, their party, age, gender, and caste (General, ST or

SC). We also know the number of votes received by each candidate (including NOTA

for the relevant elections) as well as the number of eligible voters in each constituency.

Table 2.2 shows summary statistics of the electoral data at the constituency level

for the panel and the repeated cross section. In the 6-state panel, for each year there

are 854 constituencies. In 2013, 630 of these constituencies were affected by NOTA

because state elections were held after the Supreme Court decision. In the repeated

cross section, we have a total of 6685 constituency-year observations from 25 states.

1176 of these observations were affected by NOTA in either 2013 or 2014. The average

constituency has approximately 180 thousand eligible voters and 11 candidates com-

peting. The overwhelming majority of candidates are male: the average constituency

has less than one female candidate. The median age of candidates in a constituency

is typically between 38 and 53. Approximately 13% of the constituencies are reserved

for SC and 15% for ST. The average non-reserved constituency has 1.3 SC candidates

and less than 0.5 ST candidates. Average turnout is 71% and the average vote share of

the winning candidate is 45%. Summary statistics of the electoral data in the 6-state

panel and the 25-state repeated cross section tend to be similar.

Table 2.3 shows a more detailed distribution of candidate characteristics in the

panel dataset. Each year we have approximately 10,000 candidates in this dataset.

40% of these candidates run as independents not affiliated with any party.

2.4.3 Voter demographics

Our first source of demographic information is various waves of the National Sample

Survey, conducted by the Indian Ministry of Statistics and Program Implementation

63

since 1950. Each wave contains close to half a million individual surveys covering all

Indian states, and is designed to be representative of the population at the subdistrict

level. We obtained the individual level data and use it to create characteristics of the

voting age population at the state-year or the district-year level for the reduced-form

analysis. Table 2.4 summarizes these variables for the 25 states in the repeated cross

section. We complement this with data on the growth rate of per capita state domestic

product from the Reserve Bank of India.

For the structural exercise, demographic characteristics are needed at the con-

stituency level. We are not aware of any existing dataset with appropriate coverage.

We create the necessary dataset using the 2011 Indian Census by aggregating village-

level information and matching it to constituencies using GIS coordinates. Specifically,

we obtained GIS boundary files for the 2013 electoral constituencies and the 2001 cen-

sus. To use data from the 2011 census, we proceed in two steps. First, we match

sub-districts (“tehsil”) in the 2001 census to the 2011 census using village names.15

Administrative boundaries in India change over time, with tehsils, districts, and even

states splitting up into new units. This step of our matching procedure is based on

the smallest administrative unit available in the census, the village. Second, we match

the 2011 census data to each 2013 electoral constituency using the 2001 sub-district

boundaries. We use area-weighted averages to compute values for constituencies that

overlap several sub-districts.

Of the 854 constituencies, we have constituency boundary files for 850. We were

able to match 723 of these to the sub-district data from the census. The location of

these constituencies is shown on Figure 2.1. Most of the constituencies we lose during

15Sub-districts, called tehsils in most states, are administrative units above the villages and below thedistricts and the states.

64

the matching (70) are in NCT Delhi. We lose this entire state because the census data

is not sufficiently disaggregated. Of the matched constituencies, 520 are affected by

NOTA in 2013 and 203 are not (in the full panel, these numbers are respectively 630

and 224).

Table 2.5 shows the summary statistics of the census data at the constituency level.

The variables include basic demographic characteristics such as gender, caste, literacy,

and employment as well as economic characteristics of the households (infrastructure

and asset ownership).

2.5 Patterns in the data

2.5.1 NOTA votes

The first noteworthy feature of the data is that a positive number of voters voted for

NOTA. Despite the fact that voting for NOTA could not affect the results of the elec-

tion, in the 9/25 states in our data affected by the policy a total of 2.51 million voters

chose this option.16 The distribution of the NOTA vote share is shown on Figure 2.2.

NOTA was chosen by a positive number of voters in every constituency, receiving an

average vote share of 1.5% with a range of 0.1-11%. As a fraction of all eligible vot-

ers (including abstainers) 1% voted for NOTA.17 In the average constituency, NOTA

16While the fact that people voted on an option that could not affect the election might seem sur-prising to some readers, this behavior is not qualitatively different from votes cast on small extra-parliamentary parties, or from voting in an election where voters have no trust in the integrity of theelection and that their vote will actually be counted. For example, in Cantú and García-Ponce (2015),despite having just voted, around 5% of Mexican voters exiting the election booth say that they haveno confidence that “the vote you cast for president will be respected and counted for the final result,”and another 15% state that they have little confidence.

17In the 5 states affected by NOTA in our panel data, the average vote share of NOTA among totalvotes cast (eligible voters) was 1.9% (1.4%).

65

received more votes than 7 of the candidates running for election. In 97 constituen-

cies out of 1176, the vote share of NOTA was larger than the winning margin (the

difference between the vote share of the winner and the runner up).

One consequence of the introduction of NOTA is simply the appearance of another

option on the ballot. A potential concern is that this new option confused some vot-

ers who chose it by mistake. Our findings below on increased turnout are difficult to

reconcile with this interpretation. If NOTA had simply confused voters at the voting

booth, we would not expect to find a positive impact on voter turnout. An alternative

way that voters might be confused is if they mistakenly thought that voting for NOTA

would somehow affect the electoral result (for example, that the election would be

invalid if NOTA obtained a majority). We find this interpretation implausible for two

reasons. First, given the 1.5% actual vote share on NOTA, voting for NOTA to in-

validate the election would have required not just confusion about electoral rules but

also extremely unrealistic expectations about the number of voters planning to vote

for NOTA. Second, if voters mistakenly thought that NOTA would affect elections, we

would expect them to be less likely to vote for NOTA as they gain more experience.

To check for this, we looked at the 2014 general elections, held at the same time in

all states. Some of these states already had experience with NOTA at the assembly

elections in 2013, while others did not. If the use of NOTA in 2013 was due to voter

confusion, we would expect the experienced states to vote for NOTA less than the

inexperienced states. In fact, the opposite is true: in this general election the average

NOTA vote share among the experienced states was 1.28%, compared to 1.09% among

non-experienced states. Voters in states that had more experience with NOTA were

significantly more likely to use it (p = 0.027).

66

Figure 2.2 reveals some heterogeneity in NOTA votes across constituencies. Our

structural analysis below will relate this heterogeneity to voter demographics and

model how different groups of voters choose between the different options on the bal-

lot. As a precursor to this analysis, in the Appendix we run cross-sectional regressions

of the NOTA vote share on a variety of constituency characteristics. We find evidence

of systematic heterogeneity: the NOTA vote share is significantly higher in reserved

constituencies and in constituencies with more illiterate voters, more women, more

ST, and a lower share of rural workers. This may suggest that economically disadvan-

taged and / or politically disenfranchized voters obtain more utility from expressing

themselves by voting for NOTA.

2.5.2 The effect of NOTA on turnout

As discussed in section 2.2, testing for a consumption utility of voting requires infer-

ring the changes in individual voters’ behavior following the introduction of NOTA.

Specifically, did the NOTA policy lead to some voters choosing to vote for NOTA in-

stead of abstaining? While this question is difficult to answer using reduced form

methods, in the Indian case we can use a simple difference-in-differences approach to

answer a related question: did the NOTA policy lead to some voters choosing to vote

instead of abstaining?

While identifying the impact of NOTA on turnout is challenging because the intro-

duction of the policy took place at the same time across India, we can exploit variation

in the Indian electoral calendar for a difference-in-differences analysis. Specifically, we

use the fact that elections to the state assemblies are held at different times in different

67

states (see Table 2.1). For the panel dataset, our specification is the following:

Ycst = α0 + α1NOTAst + α2Xcst + γc + ηt + εcst, (2.3)

where Ycst is turnout in constituency c of state s in year t, NOTAst equals 1 if the

NOTA policy is in place and 0 otherwise, Xcst are control variables, and γc and ηt are

constituency and year fixed effects, respectively. Using states that held elections in

2008 and 2013, the parameter of interest, α1 is identified by comparing the change in

turnout in the states that held elections in both years without NOTA to the change

in turnout in the states that were affected by NOTA in 2013 (but not in 2008). For

the repeated cross-section sample, the specification is identical to (2.3) except that we

replace the constituency fixed effects γc with state fixed effects γs.18

Table 2.6 shows the results from estimating equation (2.3). In column (1), we use

the panel dataset and control for the number of eligible voters in a constituency, state

labor force participation, weekly household earnings, and education (as well as con-

stituency and year fixed effects). The coefficient estimate on NOTA indicates a positive

turnout effect but with the small number of states and state-level variation in the pol-

icy, the estimate is highly imprecise.19 In column (2) we repeat the same specification

for the repeated cross-section, replacing the constituency fixed effects with state fixed

18Recall that the panel includes constituencies from 6 states, 5 of which were affected by NOTA. Therepeated cross section contains constituencies from 25 states, 9 of which were affected by NOTA (5 in2013 and 4 in 2014).

19Because the NOTA policy varies at the state level, inference needs to account for clustering. Giventhe small number of clusters, we obtain the p-value for the effect of NOTA by using a wild bootstrapprocedure as recommended by Cameron and Miller (2015) with the 6-point weight distribution of Webb(2013).

68

effects. The point estimate on NOTA remains similar but the precision improves dras-

tically, indicating a statistically significant turnout effect of 3 percentage points.20 In

column (3) we add as additional controls a dummy for reserved constituencies as well

as the following state-level variables: unemployment, sex ratio, urbanization, and the

growth rate of state per capita net domestic product. The estimated effect of NOTA

remains robust to these additional controls.

The main threat to identification in the regressions presented in Table 2.6 is other

events or policies that may affect changes in turnout between assembly elections held

before and after the introduction of NOTA. In the Appendix, we present a number

of robustness checks: we exclude national election years, control for the extent to

which constituencies were affected by electoral redistricting, and we drop specific

states where political events during our period of studies might potentially confound

the effect of NOTA. We find that our estimates are robust to these alternative samples

and specifications.

Overall, these results indicate that the presence of the NOTA option on the election

ballot increased turnout in the average constituency. It is interesting to note that in

each case the 95% confidence interval around the estimates includes the fraction of

eligible voters who voted for NOTA in the data (1% in the repeated cross-section).

This may indicate that NOTA voters turn out to vote specifically for this option and

would abstain if this option was not present on the ballot. At the same time, our

ability to infer NOTA voters’ counterfactual behavior using this aggregate analysis is

fundamentally limited. Did the NOTA policy lead to some voters choosing to vote for

NOTA instead of abstaining? Our structural analysis below will make such inference

20We obtain similar inference using either standard errors clustered by state or the wild bootstrapprocedure.

69

regarding individual voters’ voting patterns possible.

2.6 Estimating the effect of NOTA from a demand sys-

tem for candidates

In this section we estimate a model of voter choice among candidates, NOTA and

abstention using techniques from the discrete consumer choice literature. Two factors

make this approach particularly useful in the voting context. First, the rules governing

elections imply that several assumptions of the model naturally hold. Second, the

inference problem addressed by the method - inference regarding individual behavior

from aggregate data - is central to any voting application.

While the typical Industrial Organization applications view the static discrete choice

framework as an approximation, the rules governing elections actually make this

model quite realistic in the electoral context. By the nature of elections, voters are

restricted to a discrete choice between voting for a candidate, voting for NOTA, or

abstaining.21 Choices are made simultaneously by all voters in a given race - unlike

consumers, voters cannot adjust the timing of their choice. When making a choice,

the voter has before him a complete list of all available options on the ballot, in con-

trast to a consumer who may not be aware of all available brands of the product he is

considering buying. Electoral competition takes place in markets (electoral districts)

21By contrast, when choosing which product to buy, consumers may purchase a mixture of productseven if they only consume one at any given time.

70

that are administratively defined and where vote shares are fully observed by the re-

searcher.22 Finally, abstention (the “outside good” chosen by somebody who does not

choose any of the alternatives on the ballot) is well defined, and administrative data

on its prevalence is readily available.23

The attractiveness of the estimation framework used here arises from its ability to

deliver inference on individual behavior from aggregate data. In IO, this is an attrac-

tive feature for researchers who have access to market-level data. While for the con-

sumer demand application one could in principle obtain individual data (e.g., from

loyalty programs), in the voting context this is virtually impossible. Due to the secrecy

of the ballot, in most cases administrative data on individual choices simply does not

exist. Because of this, most existing research either relies exclusively on aggregate

analysis, or uses voter survey data to analyze individual behavior. Because voter sur-

vey data is subject to well-known biases, being able to infer individual behavior from

aggregate data is of major importance in this literature.

2.6.1 Specification: demand

Our specification of voter preferences adapts the consumer demand model of Berry,

Levinsohn and Pakes (1995) (BLP).24 Consider a constituency t ∈ {1, ..., T}where vot-

ers can vote for candidates j ∈ {1, ..., J}. If available, we include the NOTA option in

the list of candidates, and we let j = 0 indicate abstention (the outside option). Each

22By contrast, studies of consumer choice have to rely on proxying the true market in which a setof products compete, e.g., based on geographic areas. It is also common in these studies to rely on asample of products and stores, while in the electoral context complete data on the vote shares of allcandidates (including abstention) is readily available.

23By contrast, defining the relevant outside good for, e.g., the market for new cars, requires makingassumptions (is it a used car, public transportation, nothing?).

24See also Nevo (2001) and Petrin (2002).

71

candidate is described by a set of characteristics observed by the researcher and a set

of unobserved characteristics. Besides the candidate’s party, characteristics observed

in our data include gender, age, and caste. Unobserved characteristics include, for

example, the candidate’s experience. Assume that the utility that voter i derives from

voting for candidate j ∈ {1, ..., J} can be specified as

Uijt = βixjt + ξj + ξjt + εijt, (2.4)

where xjt = (x1jt, ..., x

Kjt)′ is a vector of the observed characteristics of party j’s candi-

date, ξj is the average popularity of the party, ξjt captures voters’ common valuation

of unobserved candidate characteristics, and εijt is a stochastic term with mean zero

drawn from a Type-I extreme value distribution (the role of this assumption will be

made clear below). Unless stated otherwise we treat the NOTA option as another can-

didate, including a NOTA indicator to identify characteristics (e.g., gender) which are

only defined for actual candidates.25

Voter preferences for the various candidate characteristics are represented by the

coefficients βi = (β1i , ..., β

Ki ). These vary across individuals based on demographic

variables and unobserved characteristics:

β′i = β + Πdi + Σvi, (2.5)

where di = (d1i , .., d

Di )′ is a vector of “observed” demographic variables, vi = (v1

i , ..., vKi )′

are “unobserved” voter characteristics, and the parameters are in the (K × 1) vector

25For example, if the only candidate characteristic is gender (gj), then xj = [(1− nj)gj , nj ] where njis equal to 1 for NOTA and 0 otherwise.

72

β, the (K ×D) matrix Π, and the (K ×K) scaling matrix Σ. We assume that the vi are

drawn from independent Normal distributions with mean 0. As in most consumer de-

mand applications, “observed” variables are individual characteristics whose empir-

ical distribution is known (from census data), while the distribution of “unobserved”

characteristics has to be assumed. While individual level consumption data is some-

times available, in the voting context, given the secrecy of the ballot, it is generally

impossible to directly match individual characteristics to votes.

To complete the choice set, the utility of the “outside good” (abstention) must be

specified. In consumer demand applications constructing the outside choice typically

involves two sets of assumptions: assumptions about what consumers do when they

don’t purchase a specific product, and assumptions about what constitutes a market.

In the voting context neither of these is necessary, since electoral constituencies are

exogenously given and voters who do not vote necessarily abstain. We let

Ui0t = π0di + σ0v0i + εi0t, (2.6)

which allows for the utility of abstention (hence the cost of voting) to vary by observed

demographics and unobserved voter characteristics. As discussed below, we also in-

clude in (2.4) state and year fixed effects and indicators for whether the constituency

is reserved for SC or ST candidates. Since voter choices will be determined by the dif-

ferences in utilities, including these variables in (2.4) is equivalent to including them

in the specification of the utility of abstention in (2.6). Thus, we are also allowing for

further heterogeneity in voting costs as captured by these variables.

73

Let ξ = (ξ1, ..., ξJ), θ1 = (β, ξ), and θ2 = (Π,Σ), and let θ = (θ1, θ2) represent the

parameters of the model. Substituting (2.5) into (2.4), we can write

Uijt = δjt + µijt + εijt,

where δjt ≡ βxjt+ξj + ξjt and µijt ≡ (Πdi+Σvi)xjt. Voters choose to vote for one of

the candidates (including NOTA) or abstain. This implicitly defines the set of demo-

graphics and unobserved variables for which voter i will choose candidate j:

Ajt(x, δt(θ1), θ2) = {(di,vi, εit)|Uijt ≥ Uilt for l = 0, 1, ..., J} ,

where x are all the candidate characteristics, δt = (δ1t, ..., δJt), and εit = (εi1t, ..., εiJt).

Given the distribution of (di,vi, εit), we can integrate overAjt to obtain the vote shares

sjt(x, δt(θ1), θ2) predicted by the model. Under the assumed Type-I extreme value

distribution for εijt, these are given by

sjt(x, δt(θ1), θ2) =

∫exp [δjt + µijt − µi0t]

1 +∑q≥1

exp [δqt + µiqt − µi0t]dF (di,vi) , (2.7)

where µi0t ≡ π0di + σ0v0i and F (di,vi) denotes the distribution of the voter character-

istics. These predicted vote shares are a function of the data (x), the parameters (θ),

and the unobserved candidate characteristics ξ.

2.6.2 Specification: supply

While some political economy models treat candidates as exogenously given, others,

notably the citizen-candidate literature, emphasize that politician characteristics may

74

emerge endogenously in the political process (Osborne and Slivinski, 1996; Besley

and Coate, 1997). To allow for this possibility while keeping the problem tractable,

we adopt a simple simultaneous-moves specification of the supply of candidates. We

will use this framework to justify the instrumental variables we use in the estimation

below.26

As in the citizen-candidate literature suppose that implemented policies depend

on elected politician’s characteristics and that candidate characteristics emerge en-

dogenously in the political process. In particular, suppose that candidates are cho-

sen by a political party that cares about winning as well as the policy implemented

by the winner. In constituency t, party j’s payoff is given by vjt(xt, st), where xt =

(x1t, ...,xJt) are the characteristics of all candidates running in the election and st =

(s1t, ..., sJt) are the vote shares that determine the winner. Vote shares are determined

by candidates’ characteristics as well as the voter valuations ξjt, as in equation (2.7).

Thus, sjt = sjt(xt, ξt) where ξt = (ξ1t, ..., ξJt) (to simplify the notation, we set ξ = 0).

Given a party’s membership, fielding candidates with some characteristics may be

easier than others. For example, a lower caste party may find it difficult to field general

caste candidates. A simple way to capture this is by supposing that party j faces a

budget constraint m =∑k

qktjxktj ≡ qtjxjt, where m is the budget available to spend

on candidates (assumed constant for simplicity) and qkjt is the “price” of increasing a

candidate’s characteristic k in constituency t. For example, if xk = 1 denotes a general

caste candidate, qkjt may be the extra cost of finding such a candidate and convincing

him to run. Prices will generally depend on such factors as a party’s membership, the

26In the consumer demand literature, it is common to model firms that compete on prices but takeall other product characteristics as exogenously given. In our context, there is no natural separationbetween endogenous and exogenous candidate characteristics so we will treat all characteristics exceptNOTA as potentially endogenous.

75

economic and demographic characteristics of a constituency, the prestige associated

with a political career in the local population, etc. We assume that parties take these

prices as given.

Suppose that parties choose the characteristics of their candidates simultaneously,

after voter valuations ξt have been realized. In a Nash equilibrium, the characteristics

of party j’s candidate will satisfy

x∗jt ∈ arg maxxjt

(vjt(xt, st(xt, ξt))|m = qjtxjt)

or

x∗jt = x∗jt(xt, ξt,qjt). (2.8)

In words, candidate j’s characteristics depend on the characteristics of all candidates

running, voters’ valuation of the unobserved characteristics, and party j’s cost of in-

creasing the various characteristics in the given constituency. This has two implica-

tions. First, the dependence of observed characteristics xjt on voter valuations ξt cre-

ates an endogeneity problem for the estimation of the utility functions (2.4). Second, it

is plausible that the prices qjt for a given party are correlated across constituencies t.

For example, a lower caste party is likely to face a higher price to field a general caste

candidate in all constituencies within a state. This implies that the characteristics of a

given party’s candidates will be correlated across constituencies. As explained below,

this opens the possibility of using candidate characteristics in neighboring constituen-

cies as instrumental variables in the estimation.

76

2.6.3 Estimation

Estimation follows the algorithm proposed by BLP. The idea is to treat the unobserved

characteristics ξ as the econometric error and derive moment conditions that can be

used to estimate the parameters using Generalized Method of Moments (GMM). De-

tailed treatments of the procedure can be found in BLP and Nevo (2000, 2001) so we

only provide a brief summary below.

Consider a dataset with information on candidate characteristics x and actual vote

shares Sjt. BLP show that, for given θ2, it is possible to numerically solve for δt from

the equations sjt(x, δt, θ2) = Sjt, i.e., equating the model-predicted vote shares to those

observed in the data. Using the resulting values of δjt(θ2), we express the unobserved

candidate characteristics as ξjt(θ) = δjt(θ2) − ξj − βxjt. Given our data and with δjt

computed, this is a standard econometric error, which depends nonlinearly on the pa-

rameters of the model. While we do not expect ξjt(θ) to be independent of xjt, we can

find a suitable set of instruments Zjt and use the moment conditions E[ξjt(θ)|Zjt] = 0

to estimate the parameters using GMM. Thus, we find

θ = arg minθ

ξ(θ)′ZW−1Z′ξ(θ),

where ξ(θ) is the vector of errors, Z is the matrix of instruments, and W is the weight-

ing matrix.

To compute the estimate, we use the standard two-step GMM procedure (Greene,

2003, p206). We first set W = Z′Z and compute an initial estimate of the parame-

ters, θ1. We then use this initial estimate to recompute a robust weight matrix W =

1n

n∑j,t

[ξjt(θ1)]2Z′jtZjt, and use this updated weight matrix to compute the final parameter

77

estimates.

In this framework, the need for instrumental variables arises for two reasons. First,

instruments are needed to generate enough moment conditions to identify the nonlin-

ear parameters in voters’ utility functions. Thus, instruments are necessary even if ξjt

and xjt are uncorrelated. Second, instruments are needed because some of the candi-

date characteristics are likely to be endogenous, as suggested by equation (2.8).27 In

the context of consumer demand estimation, where “voters” are the consumers and

“candidates” are the products, it is common to use instruments based on the character-

istics of other products produced by the same firm and the characteristics of products

produced by other firms (e.g., BLP; Nevo, 2001). A natural counterpart in our setting

is to think of firms as the parties that field the candidates. Using this analogy, we

use as instruments the average characteristics of a given party’s candidates in other

constituencies within the state. For example, we create a candidate gender IV for a

particular candidate by taking the fraction of female candidates of the given party

in other constituencies in the state for both 2008 and 2013.28 What makes these in-

struments possible in our case is the variation in the characteristics of a given party’s

candidates across constituencies as each election is contested by a different set of in-

dividuals. This avoids the difficulties that sometimes arise in the consumer demand

literature from insufficient variation in product characteristics across markets (e.g.,

Nevo, 2001).27The endogeneity problem is likely to be present even if candidate characteristics are assumed to be

fixed at the time of the election. We only observe a short list of characteristics xjt, and unobserved char-acteristics that influence ξjt (including experience, voting record, qualifications, physical appearance,etc.) are likely to be correlated with these.

28As usual, variables that enter the utility function and are treated as exogenous serve as their owninstruments. In our case, this includes state, year, and party fixed effects, as well as the NOTA indicatorand its interaction with demographics.

78

Beyond the analogy to product characteristics, a rationale for these instruments

in our case may be given based on the supply of candidates available to each party,

as in section 2.6.2. For a given party, the “price” of increasing a particular candidate

characteristic is likely to be correlated across constituencies due to the characteristics

of the party’s membership, demographic characteristics of the constituency, etc. This

implies that the characteristics of a particular party’s candidates are likely to be corre-

lated across constituencies.

The identifying assumption, expressed in the moment conditions, is that unob-

served voter valuations for a particular candidate are conditionally independent of

these instruments. One case in which this assumption will hold is if, controlling for

party-specific means and demographics, constituency-specific voter valuations ξjt are

independent across constituencies (but may be correlated for a given constituency

over time). This rules out a popularity shock to some of a party’s candidates as would

be caused, e.g., by a regionally coordinated advertising campaign (a campaign rais-

ing the popularity of all candidates would be captured by the party dummies). See

Hausman (1996) and Nevo (2001) for analogous assumptions in the consumer demand

literature. A second case in which the identifying assumption will hold is if parties do

not condition their choice of candidates on the popularity shocks ξjt in (2.8). For ex-

ample a party with an SC base may find it impossible to respond to a popularity shock

by finding a candidate from a different caste in time for the election. In this case, the

mix of candidate characteristics offered by a party would reflect the supply of charac-

teristics in the relevant population, rather than respond to popularity shocks among

voters.

Under either interpretation, (2.8) implies that characteristics of candidates fielded

79

by a given party in different constituencies will only be correlated due to the common

prices qjt. Thus, for a given party, the characteristics of its candidates xjt′ is con-

stituencies t′ 6= t are valid instruments for the characteristics of the candidate running

in constituency t. Since the introduction of the NOTA option took place at the national

level by the Supreme Court, the NOTA characteristic is treated as exogenous.

2.6.4 Practical issues

As described above, parties play an important role in our specification: we estimate

party fixed effects and we define our IVs based on parties. One difficulty arises be-

cause of the presence of many small parties. There are a total of 202 parties in the data,

but half of them field candidates in only 1 of every 40 constituency within a state. A

second, related difficulty is the presence of independent candidates (candidates not

affiliated with any party). There are 6751 of these candidates in the data, but 70% of

them receive less than 1% of the votes in a constituency and only 3% receive more

than 10%. Each of these parties and candidates adds a new parameter that is difficult

to identify due to the small number of constituencies where the party is represented

(in the extreme case of an independent candidate running in only one year, identify-

ing the fixed effect is not possible). To circumvent these difficulties, we create a “Small

Party” category comprising parties fielding candidates in less than a third of the con-

stituencies in any given state and we average all small party candidates’ characteristics

within a constituency (we do this after constructing the instruments so that the indi-

vidual IVs are aggregated also). We also create an “Independent Party” containing all

independent candidates within a state, and aggregate them within constituencies in

the same way. After this aggregation, we are left with a total of 22 parties.

80

We include in the analysis the full list of candidate characteristics available in the

data: gender, caste, age and party. We select the constituency characteristics to be

included based on the variables that indicated significant heterogeneity in voter pref-

erences for NOTA in the regressions in section 2.5. We include percent male, literacy,

percent SC, percent ST, and the share of rural workers.

The BLP algorithm requires numerically solving the integral in (2.7) to obtain the

predicted market shares. We do this in the standard way by drawing individual voters

from the distribution of demographics in each constituency, computing the predicted

individual probabilities of voting for each candidate, and averaging across simula-

tions to obtain the simulator for the integral.

2.6.5 Results

Parameter estimates

We present parameter estimates for different specifications of the above model in Ta-

bles 2.7 and 2.8. First, we set Π and Σ equal to 0 so that voter heterogeneity only enters

through the εijt terms in equation (2.4). This is the conditional Logit model, and we

estimate it with and without instruments in columns (1) and (2) of Table 2.7. We report

coefficient estimates on the candidate characteristics as well as a subset of the control

variables (reserved constituency indicators and dummies for 3 major parties).

In column (3), we keep Π = 0 but allow for random coefficients on all the candi-

date characteristics as well as the outside option (through Σvi in equation (2.5) and σ0

in equation (2.6)). The estimates of Σ are in the first column of Table 2.8. Finally, we

present the full model which allows for both observed and unobserved heterogeneity

in voters’ evaluation of the various candidate characteristics. These estimates are in

81

column (4) of Table 2.7 (β) and the remainder of Table 2.8 (Σ and Π). Moving from col-

umn (1) to column (2) of Table 2.8, we kept those elements of Σ that were statistically

significant.

In Table 2.7, coefficients change substantially between the OLS and IV specifica-

tions, suggesting that instrumenting is important. Tables 2.7 and 2.8 reveal that voters

value candidate characteristics other than party affiliation. For example, older male

candidates tend to receive more votes. Allowing for taste heterogeneity among voters

regarding these characteristics in the full model substantially reduces the estimated

impact of party affiliation on votes (column (4) of Table 2.7). In Table 2.8, areas with

a higher share of ST voters yield more votes for ST candidates (this effect is identified

from variation across non-reserved constituencies). NOTA is a less popular option in

constituencies with more literate voters and in rural areas.

Counterfactual analysis: The impact of NOTA

In this section we use the estimated model to evaluate the impact of introducing

NOTA. We restrict attention to those constituencies in our data that had the NOTA

option available in 2013 and perform a counterfactual experiment where the NOTA

option is removed. We compute new vote shares and turnout rates under this coun-

terfactual scenario, and calculate the impact of NOTA as the difference between the

actual and the counterfactual outcomes.

The estimated impact of NOTA on turnout is shown in Figure 2.3. The average

increase in turnout is above 0.5 percentage points in 74% of the constituencies, with an

average of 1.2 percentage points. This is somewhat lower than the effect we obtained

82

from the reduced form exercise in section 2.5.2, but very close to the 1.4 percent of

eligible voters who voted for NOTA in the data.

The estimated impact of NOTA on individual candidates’ vote shares is shown

in Figure 2.4. The reduction in vote shares is smaller than 0.5 percentage points in

absolute value for 96% of the candidates, with a mean of 0.06 percentage points. This

indicates that substitution from voting for a candidate to voting for NOTA is minimal.

To gauge the impact of NOTA on parties, Table 2.9 aggregates voter choices by

party. For each party in the data, the first column gives the number of candidates and

the second column shows the fraction of the 101.2 million eligible voters who voted

for that party. The third column shows the difference relative to the counterfactual

without NOTA. For the two largest parties, BJP and INC, we estimate a loss in total

votes of around 0.15 percentage points. The change in the votes cast on other parties is

even smaller. By contrast the change in the overall abstention rate (and hence turnout)

is a magnitude larger at 1.2 percentage points. This is similar to the estimated effect

across all constituencies.

It should be emphasized that allowing for the flexible random coefficients spec-

ification above was crucial to obtain these results. The more restrictive Logit model

could not possibly have resulted in these effects. As is well known, the Logit specifica-

tion implies that substitution patterns only depend on observed choice shares. In our

case, this would imply that adding the NOTA option would, by construction, cause

the biggest change in the most popular candidate’s vote share. This is illustrated by

the last column of Table 2.9 which shows the counterfactual implications that would

be obtained from the Logit specification. As can be seen, this would imply that substi-

tution towards NOTA is similar for voters of the two largest parties and for abstaining

83

voters.

The patterns emerging from the demand model are similar to those observed in

the difference-in-differences analysis. The results indicate that NOTA votes are mostly

cast by voters who would have chosen to abstain in the counterfactual without NOTA

and turn out to vote specifically for NOTA. This provides strong evidence that voters

derive positive consumption utility from voting for this option.

2.7 Conclusion

This paper analyzed India’s NOTA policy which gives people the option to partici-

pate in elections and cast a valid “None of the Above” vote without the possibility of

affecting the electoral outcome. Individuals who choose to vote for NOTA but would

abstain otherwise must derive a consumption utility from voting that is specific to this

option. Thus, the NOTA policy makes it possible to test apart various components of

the consumption utility of voting. To address the challenge that, due to the secret bal-

lot, individual choices are not observable, we estimate counterfactual voter behavior

using a structural model and techniques borrowed from the consumer demand liter-

ature. The model allows for rich heterogeneity in voter preferences and relates these

parameters to the aggregate vote returns. In counterfactual simulations, we find that

the NOTA policy resulted in increased turnout. Based on the estimated model, vir-

tually all the NOTA votes observed in the data represent new voters who showed up

specifically to vote for NOTA and who would have abstained in the absence of this op-

tion. These patterns are also supported by the findings from a reduced-form analysis.

Our results show that, in this context, the presence of an option-specific consumption

84

utility from voting is necessary to explain the data. For example, voters may derive

utility from expressing their protest against one or more of the candidates running

for election. In this case, models that do not incorporate an option-specific utility of

voting would have a hard time explaining the data.

To the extent that voter participation is valuable in a democracy, our results suggest

that having a NOTA option on the ballot may be a desirable policy. It creates both

political participation and individual utility.

2.8 References

1. Ali, N., and C. Lin (2013): “Why People Vote: Ethical Motives and Social Incen-

tives,” American Economic Journal: Microeconomics

2. Bénabou, R. J., and J. Tirole. (2006). “Incentives and Prosocial Behavior,” Ameri-

can Economic Review

3. Berry, S., J. Levinsohn, and A. Pakes (1995): “Automobile Prices in Market Equi-

librium,” Econometrica, 63(4)

4. Besley, T., and S. Coate (1997): “An Economic Model of Representative Democ-

racy,” Quarterly Journal of Economics

5. Blais, A. (2000): To vote or not to vote, University of Pittsburgh Press, Pittsburgh,

PA.

6. Brennan, G., and L. Lomasky (1993): Democracy and Decision: The Pure Theory of

Electoral Preference. Cambridge University Press, Cambridge, UK.

85

7. Cameron, A.C., and D.L. Miller (2015): “A Practitioner’s Guide to Cluster-Robust

Inference,” Journal of Human Resources

8. Cantú, F., and O. García-Ponce (2015): “Partisan losers’ effects: Perceptions of

electoral integrity in Mexico,” Electoral Studies

9. Cho, W.K.T., and C.F. Manski (2008): “Cross-Level/Ecological Inference,” in:

J.M. Box-Steffensmeier, H.E. Brady, and D. Collier (eds.) The Oxford Handbook

of Political Methodology, Oxford University Press, Oxford, UK.

10. Coate, S. and M. Conlin (2004): “A Group Rule: Utilitarian Approach to Voter

Turnout: Theory and Evidence,” American Economic Review

11. Coate, S., M. Conlin, and A. Moro (2008): “The performance of pivotal-voter

models in small-scale elections: Evidence from Texas liquor referenda,” Journal

of Public Economics

12. Degan, A. and A. Merlo (2011): “A Structural Model of Turnout and Voting in

Multiple Elections,” Journal of the European Economic Association

13. DellaVigna, S., J.A. List, U. Malmendier, and G. Rao (2015): “Voting to Tell Oth-

ers,” working paper, UC Berkeley.

14. Dittmann, I., D. Kubler, E. Maug, and L. Mechtenberg (2014): “Why Votes Have

Value: Instrumental Voting with Overconfidence and Overestimation of Others’

Errors,” Games and Economic Behavior

15. Downs, A. (1957): An economic theory of democracy, Harper and Row, New York,

NY.

86

16. Driscoll, A., and M.J. Nelson (2014): “Ignorance or Opposition? Blank and

Spoiled Votes in Low-Information, Highly Politicized Environments,” Political

Research Quarterly

17. Duffy, J. and M. Tavits (2008): “Beliefs and Voting Decisions: A Test of the Pivotal

Voter Model,” American Journal of Political Science

18. Feddersen, T. and A. Sandorini (2006): “A Theory of Participation in Elections,”

American Economic Review

19. Fiorina, M. P. (1976): “The Voting Decision: Instrumental and Expressive As-

pects,” Journal of Politics

20. Glasgow, G., and R.M. Alvarez (2005): “Voting behavior and the electoral context

of government formation,” Electoral Studies

21. Greene, W.H. (2002): Econometric Analysis, Fifth Ed., Prentice Hall, Upper Saddle

River, NJ.

22. Harbaugh, W. T. (1996): “If People Vote Because They like to, then Why do so

Many of Them Lie?,” Public Choice

23. Hausman, J. (1996): “Valuation of New Goods Under Perfect and Imperfect

Competition,” in: T. Bresnahan and R. Gordon (eds.): The Economics of New Goods,

Studies in Income and Wealth Vol. 58, National Bureau of Economic Research,

Chicago, IL.

87

24. Herron, M. C. and J. S. Sekhon (2005): “Black Candidates and Black Voters: As-

sessing the Impact of Candidate Race on Uncounted Vote Rates,” Journal of Poli-

tics

25. McAllister, I. and T. Makkai (1993): “Institutions, Society or Protest? Explaining

Invalid Votes in Australian Elections,” Electoral Studies

26. McMillan, A. (2010): “The Election Commission,” in: N.G. Jayal and P.B. Mehta

(eds): The Oxford Companion to Politics in India, Oxford University Press, Oxford,

UK.

27. Nevo, A. (2000): “A Practitioner’s Guide to Estimation of Random Coefficients

Logit Models of Demand,” Journal of Economics and Management Strategy

28. Nevo, A. (2001): “Measuring Market Power in the Ready-to-Eat Cereal Indus-

try,” Econometrica

29. Ortoleva, P. and E. Snowberg (2015): “Overconfidence in Political Behavior,”

American Economic Review

30. Osborne, M., and A. Slivinski (1996): “A Model of Political Competition with

Citizen-Candidates,” Quarterly Journal of Economics

31. Petrin A. (2002): “Quantifying the Benefits of New Products: The Case of the

Minivan,” Journal of Political Economy

32. Power, T. J. and J. C. Garand (2007): “Determinants of invalid voting in Latin

America,” Electoral Studies

88

33. Razin, R. (2003): “Signaling and Election Motivations in a Voting Model with

Common Values and Responsive Candidates,” Econometrica

34. Rekkas, M. (2007): “The Impact of Campaign Spending on Votes in Multiparty

Elections,” Review of Economics and Statistics

35. Riker, H. W. and P. C. Ordeshook (1968): “A Theory of the Calculus of Voting,”

American Political Science Review

36. Selb, P., and S. Munzert (2013): “Voter overrepresentation, vote misreporting,

and turnout bias in postelection surveys,” Electoral Studies

37. Shachar, R., and B. Nalebuff (1999): “Follow the Leader: Theory and Evidence

on Political Participation,” American Economic Review

38. Shayo, M. and A. Harel (2012): “Non-consequentialist voting,” Journal of Eco-

nomic Behavior and Organization

39. Superti, C. (2015): “Vanguard of the Discontents: Blank and Null Voting as So-

phisticated Protest,” working paper, Harvard University.

40. Uggla, F. (2008): “Incompetence, Alienation, or Calculation? Explaining Levels

of Invalid Ballots and Extra-Parliamentary Votes,” Comparative Political Studies

41. Webb, M. (2013): “Reworking Wild Bootstrap Based Inference for Clustered Er-

rors,” working paper, University of Calgary.

89

2.9 Appendix

2.9.1 The correlates of NOTA votes

In this section we investigate the correlation between NOTA votes and constituency

characteristics. We use the same sample as in our structural exercise (the panel dataset

matched to the census data) and run simple cross-sectional regressions on the 520 con-

stituencies that are affected by the NOTA policy in 2013. We include state fixed effects

and, to avoid confounding our estimates by differential turnout across constituencies,

we measure NOTA vote shares as a fraction of total votes cast.29

The results are in Table 2.10. We find substantial heterogeneity in NOTA votes

across constituencies. For example, the NOTA vote share is significantly higher in re-

served constituencies and in constituencies with more illiterate voters, more women,

more ST, and a lower share of rural workers. Each of these patterns is consistent with

a variety of possible explanations. One possible interpretation is that NOTA votes are

higher in more economically disadvantaged populations, reflecting a general dissatis-

faction with elected leaders in these constituencies. Note however that the coefficients

remain unchanged if we add controls for various indicators of infrastructure and eco-

nomic activity in column (2). Another possible interpretation is that NOTA votes come

from politically underrepresented voters, such as women and non-SC or ST voters in

reserved constituencies.

In columns (3) and (4) we add candidate characteristics to the regression. These

estimates are merely suggestive because the list of candidates running will be affected

by their expected vote share, and this could be correlated with the fraction of voters

29Using NOTA votes as a share of eligible voters yields very similar results.

90

choosing NOTA. We find that constituencies with more candidates running result in

lower NOTA vote shares, which is consistent with NOTA reflecting dissatisfaction

with the menu of candidates being offered. We do not find evidence that the presence

of female, SC or ST candidates affects NOTA votes.

2.9.2 Robustness of the DD estimates

This section explores the robustness of the difference-in-difference estimates presented

in section 2.5.2 of the paper to various political events affecting one or more states.

National elections

In our study period, Indian national elections took place in Spring 2009 and 2014.

Recall that we do not include in the analysis the four states that hold their assembly

elections simultaneously with the national election. In the remaining states, because

split-ticket voting (constituencies voting for different parties at the state and national

levels) is common in India, it is ex ante not obvious that events affecting national

turnout, like the wave of support for the BJP in the 2014 national elections, would

affect assembly elections.30 If national elections did have an impact, we would expect

this to be the strongest for states that held assembly elections in October and December

of the national election years. If national elections impacted turnout in the assembly

elections in these states, this has the potential to confound our estimates of the NOTA

policy introduced between the two national election years in September 2013.

In Table 2.11 we exclude the national election years from the sample. Columns

(1) and (2) exclude 2014 and columns (3) and (4) exclude both 2009 and 2014. Odd30Note also that increased support for the BJP would lead to more BJP votes rather than NOTA votes.

91

numbered columns correspond to the specification in column (2) of Table 2.6 with the

basic controls and even numbered columns to column (3) with the extended controls.

We find that all point estimates are similar to, and if anything slightly larger than the

3 percentage points effect we found in Table 2.6. National elections do not appear to

confound the estimates reported in the main text.

Redistricting

Another potential confound is the electoral redistricting that took place in April 2008.

Because elections are held every 5 years and NOTA was introduced in September

2013, none of the states that were affected by NOTA in our period of study were re-

districted, while most states that were not affected by NOTA were redistricted. Thus,

redistricting has the potential to confound our estimates of NOTA.31

To control for this, we create a constituency-level measure of redistricting by us-

ing GIS boundary files to compare constituencies before and after the delimitation.

Our first measure calculates for each current constituency that was redistricted in

our study period the largest area that was part of a single constituency before the

redistricting. For example, a value of 0.8 for this “maximum overlap” measure indi-

cates that 80% of the current constituency’s area was part of a single constituency

pre-delimitation (while the remaining 20% was part of one or more different con-

stituencies). The higher the maximum overlap, the less a constituency was affected

by redistricting. Our second measure, rather than focus on the largest area of over-

lap, uses each overlapping area to create an index of “territorial fractionalization.” If

a constituency overlaps with n pre-delimitation constituencies with s1, ..., sn denoting

31For example, if redistricting lowered turnout, our estimate of NOTA’s effect of turnout would likelybe biased upward.

92

the share of its area falling in each of these, then the fractionalization index is 1−n∑i=1

s2i .

The larger this value, the more the current constituency was affected by redistricting.

Both of these measures are available for 22 states (constituency boundary files are not

available for the states of Assam, Manipur, and Nagaland).

Table 2.12 presents regressions corresponding to specifications (2) and (3) in Ta-

ble 2.6 controlling for these measures of redistricting. The first two columns repeat

columns (2) and (3) in Table 2.6 on the 22 states with available redistricting measures.

Columns (3) and (4) then add the maximum overlap measure and columns (5) and (6)

the territorial fractionalization index. As can be seen, adding either measure of redis-

tricting to the regressions causes very little change in the estimated effect of NOTA.

The estimates also retain their significance, except for column (6) where the standard

error increases just enough to yield a p-value of 0.101.32

State-specific events

Turning to state-specific events that may confound our estimates, we identified four

states where various events may plausibly affect 2013 or 2014 turnout relative to the

previous election (that is, turnout in the with-NOTA election relative to turnout in

the without-NOTA election). In Chhattisgarh, Maoist insurgents conducted terrorist

attacks in 2010 and May 2013, between the 2008 and 2013 elections in this state. In

Jammu & Kashmir, various incidents occurred between its 2008 and 2014 elections,

including a border skirmish in January 2013 between India and Pakistan described by

observers as one of the worst in 10 years. In Delhi, a new anti-corruption party, Aam

Aadmi entered politics in 2012, energized voters, and emerged as the second-largest

32The coefficients on the redistricting measures are never statistically significant.

93

party in the 2013 assembly election. Finally, Maharashtra held its 2009 election a year

after the 2008 terrorist attacks in Mumbai on several hotels and public buildings, and

security concerns may have depressed voter turnout there.

In Table 2.13, we repeat specifications (3) and (4) from Table 2.6 excluding each

of these states one at a time and then all four of them. The results corresponding

to the first specification are in column (1) and column (2) corresponds to the second

specification with the extended set of controls. All these coefficients are close to the

3 percentage point effect found in Table 2.6. The events in these four states do not

appear to drive the estimated effect of NOTA on turnout reported in the main text.

94

TABLE 2.1: Timeline of events in the study period

Year Month State assembly elections Other events2006 4 Assam

5 Kerala, Puducherry, Tamil Nadu,West Bengal

2007 2 Manipur, Punjab, Uttarakhand5 Uttar Pradesh6 Goa12 Gujarat, Himachal Pradesh

2008 2 Tripura3 Meghalaya, Nagaland4 Delimitation5 Karnataka11 Madhya Pradesh, NCT of Delhi12 Chhattisgarh, Jammu & Kashmir,

Mizoram, Rajasthan2009 4 Andhra Pradesh*, Arunachal Pradesh*, National elections

Odisha*, Sikkim*10 Haryana, Maharashtra12 Jharkhand

2010 10 Bihar*2011 4 Assam, Kerala, Puducherry, Tamil Nadu

5 West Bengal2012 1 Manipur, Punjab, Uttarakhand

3 Goa, Uttar Pradesh11 Himachal Pradesh12 Gujarat

2013 2 Meghalaya, Nagaland, Tripura5 Karnataka9 NOTA policy introduced11 Chhattisgarh, Madhya Pradesh12 Mizoram, NCT of Delhi, Rajasthan

2014 4 Andhra Pradesh*, Arunachal Pradesh*, National electionsOdisha*, Sikkim*

10 Haryana, Maharashtra12 Jammu & Kashmir, Jharkhand

Notes: * excluded from the dataset.

95

TABLE 2.2: Summary statistics of the electoral data at the constituencylevel

Variable Obs Mean Std. Dev. 10% 90%A. PanelNumber of candidates 1708 11.32 4.96 6 18Female candidates 1708 0.80 1.00 0 2Median candidate age 1708 44.07 5.45 38 51Eligible voters (1000) 1708 175.54 47.15 139.76 218.02Turnout 1708 0.71 0.09 0.58 0.82Winning vote share 1708 0.44 0.09 0.33 0.55NOTA votes / total votes 630 0.019 0.013 0.006 0.035NOTA votes / eligible voters 630 0.014 0.009 0.004 0.026Non-reserved constituencies:Number of SC candidates 1144 1.31 1.43 0 3Number of ST candidates 1144 0.48 1.04 0 2

B. Repeated cross sectionNumber of candidates 6685 10.54 5.35 5 17Female candidates 6685 0.73 0.97 0 2Median candidate age 6685 44.99 5.78 38 53Eligible voters (1000) 6685 180.75 88.23 41.20 292.90Turnout 6685 0.71 0.13 0.53 0.87Winning vote share 6685 0.45 0.10 0.32 0.56NOTA votes / total votes 1176 0.015 0.116 0.004 0.030NOTA votes / eligible voters 1176 0.010 0.009 0.003 0.022Non-reserved constituencies:Number of SC candidates 4842 1.24 1.56 0 3Number of ST candidates 4842 0.26 0.79 0 1Notes: The panel dataset contains the 2008 and 2013 state assembly elections in the states of Karnataka, NCTDelhi, Mizoram, Rajasthan, Madhya Pradesh, and Chhattisgarh. The repeated cross-section contains all assemblyelections between 2006 and 2014 in 25 states. Turnout is total votes divided by the number of eligible voters.Winning vote share is the winner’s share of all non-NOTA votes. Source: Election Commission of India.

96

TABLE 2.3: Candidate characteristics in the panel data

Variable All 2008 2013Age 44.42 43.98 44.87Female 6.87 6.87 6.88General caste 61.47 62.35 60.63SC 20.79 21.3 20.3ST 14.58 16.35 12.89Nationally recognized party 27.12 27.58 26.68State recognized party 3.89 4.9 2.93Other state-recognized party 8.05 6.9 9.15Unrecognized party 18.77 18.21 19.29Independent 39.01 42.41 35.76NOTA 3.16 - 6.18N 19957 9762 10195Notes: 2008 and 2013 state assembly elections in the states of Karnataka, NCTDelhi, Mizoram, Rajasthan, Madhya Pradesh, and Chhattisgarh. All numbersexcept Mean age are percentages. Source: Election Commission of India.

TABLE 2.4: Voter demographics at the state level (repeated cross-section)

Obs Mean Std. Dev. Min MaxLabor force participation 50 0.58 0.07 0.46 0.75Unemployment rate 50 0.02 0.03 0.00 0.15Real household earnings (Rp per week) 50 1708 665 864 4335Fraction illiterate 50 0.25 0.13 0.04 0.52Fraction primary school or less 50 0.23 0.09 0.13 0.53Female per 1000 male 50 987 72 790 1172Fraction urban 50 0.32 0.18 0.09 0.97State NDP growth rate 50 6.11 4.57 -5.38 24.31Notes: Source for all variables except NDP growth rate: National Sample Survey, rounds 62, 64, 66,68, 71. Individual surveys for respondents above 18 were aggregated to the state level. Householdearnings deflated to 2001 prices using the CPI from the Reserve Bank of India. Source for NDPgrowth rate: Reserve Bank of India.

97

TABLE 2.5: Demographic characteristics of the constituencies from theIndian Census (panel dataset)

Variable Mean Std. Dev. 10% 90%Literate 0.58 0.09 0.47 0.69Percent male 0.51 0.01 0.50 0.53Percent SC 0.18 0.07 0.08 0.26Percent ST 0.29 0.18 0.16 0.58Percent employed 0.46 0.05 0.40 0.52Percent rural workers 0.66 0.17 0.44 0.84Fraction of households with infrastructure:No latrine 0.71 0.22 0.37 0.91Water near premises 0.47 0.10 0.35 0.59Water on premises 0.26 0.16 0.10 0.47Fraction of households owning asset:Car 0.03 0.02 0.01 0.06Computer 0.01 0.01 0.002 0.02Phone 0.55 0.20 0.25 0.78TV 0.37 0.17 0.15 0.61Notes: Source: Indian Census, 2011. Village level census data was merged to assemblyconstituency GIS boundary files as described in the text. N = 723.

98

TABLE 2.6: The impact of NOTA on turnout, DD estimates

(1) (2) (3)NOTA 0.035 0.029** 0.030*s.e. (0.013) (0.016)bootstrap p-value [0.587] [0.036] [0.060]Eligible voters, labor force participation, x x xhh earnings, educationPolitical reservation, unemployment, sex xratio, urbanization, NDP growth rateConstituency FE xState FE x xR2 0.78 0.18 0.19N 1708 6685 6685States 6 25 25Notes: Estimates of the effect of the NOTA policy on turnout from Eqn. (2). All regressions control for time fixedeffects, the log number of eligible voters in a constituency and its square, and the following state-level variables:labor force participation, real weekly household earnings, fraction of illiterates, fraction with primary schoolor less as highest education. Column (3) also controls for reserved constituencies and the following state levelvariables: unemployment, sex ratio, fraction urban, and the growth rate of net domestic state product. Standarderrors clustered by state in parentheses. The bootstrap p-value was computed using a wild bootstrap procedurewith a 6-point weight distribution. ***, **, and * indicate significance at 1, 5, and 10 percent, respectively.

99

TABLE 2.7: Estimates of the linear parameters of the demand system

InstrumentalVariable Random

OLS Logit Logit Coefficients Full modelVariable (1) (2) (3) (4)Female -0.090** -0.153 -1.547* -3.725

(0.044) (0.268) (0.928) (2.992)Age 1.029*** 8.925*** 8.013*** 10.231*

(0.116) (1.423) (2.097) (5.767)SC candidate -0.370*** 0.273 0.377 0.120

(0.047) (0.223) (0.385) (1.266)ST candidate -0.043 0.082 0.143 2.002

(0.067) (0.148) (0.386) (2.878)NOTA 0.269*** -3.439*** -4.263*** -61.830***

(0.074) (0.076) (0.890) (11.052)Reserved SC 0.302*** -0.202 -0.258 -0.140

(0.048) (0.189) (0.360) (0.954)Reserved ST 0.290*** 0.351*** 0.734*** -1.923

(0.064) (0.115) (0.194) (2.688)Party: BJP 2.664*** -5.135*** -4.677*** -1.994

(0.042) (0.728) (0.990) (3.104)Party: INC 2.738*** -5.157*** -4.724*** -2.089

(0.041) (0.745) (1.017) (3.177)Party: BSP 0.325*** -7.039*** -6.802*** -4.348

(0.045) (0.644) (0.895) (2.933)Notes: The table reports estimates of the linear parameters of the model (β). All specifications include afull set of party dummies (three of which are reported in the table) as well as dummies for states, years,and constituency reservation status. Columns (2)-(4) include instrumental variables as described in thetext. Robust standard errors in parentheses. ***, **, and * indicates significance at 1, 5, and 10 percent,respectively.

100

TAB

LE

2.8:

Esti

mat

esof

the

nonl

inea

rpa

ram

eter

sof

the

full

mod

el

Stan

dard

Stan

dard

Dev

iati

ons

Dev

iati

ons

Inte

ract

ions

wit

hD

emog

raph

icV

aria

bles

Var

iabl

e(1

)(2

)Pe

rcen

tSC

Perc

entS

TPe

rcen

tmal

eLi

tera

cyR

ural

Fem

ale

-3.8

73**

-7.0

81*

--

6.21

4-

-(1

.513

)(3

.813

)(6

.498

)A

ge-0

.915

--

--

--

(2.4

53)

SCca

ndid

ate

0.00

3-

0.06

0-

--

-(1

.799

)(2

.188

)ST

cand

idat

e1.

951*

*4.

971*

**-

6.45

2*-

--

(0.8

59)

(1.7

08)

(3.5

56)

NO

TA1.

343

--

-43.

704

-42.

581

-45.

274*

-24.

270*

*(0

.839

)(4

3.32

2)(3

4.23

1)(2

4.27

6)(1

1.30

5)C

onst

ant

-0.0

60-

--

-13

.066

***

6.08

2***

(1.6

50)

(1.8

81)

(1.9

27)

Not

es:

The

tabl

ere

port

ses

tim

ates

ofth

eno

nlin

ear

para

met

ers

ofth

em

odel

(Πan

dΣ

).’S

tand

ard

devi

atio

ns’r

efer

toth

eΣ

para

met

ers

ofth

era

ndom

coef

ficie

nts.

The

spec

ifica

tion

inco

lum

n(1

)co

rres

pond

sto

colu

mn

(3)

inTa

ble

7an

dre

stri

cts

Π=

0.

The

rem

aini

ngco

lum

nsof

the

tabl

eco

rres

pond

toth

esp

ecifi

cati

onin

colu

mn

(4)i

nTa

ble

7.R

obus

tsta

ndar

der

rors

inpa

rent

hese

s.**

*,**

,and

*in

dica

tes

sign

ifica

nce

at1,

5,an

d10

perc

ent,

resp

ecti

vely

.

101

FIGURE 2.1: Constituencies in the merged dataset

102

TABLE 2.9: Impact of NOTA on vote shares by party

Change with NOTAN. of Percent of

Party candidates all voters Full model LogitBJP 506 32.90 -0.17 -0.63BSP 498 3.63 -0.02 -0.09BYS 102 0.10 0.00 -0.01CSM 54 0.22 0.00 -0.01GGP 44 0.20 0.00 -0.01INC 518 26.74 -0.14 -0.58Independents 468 4.96 -0.03 -0.11JGP 85 0.06 0.00 0.00MNF 10 0.06 0.00 0.00NPEP 133 1.30 -0.01 -0.02SP 194 0.43 0.00 -0.01ZNP 11 0.02 0.00 0.00Small Party 445 2.70 -0.01 -0.05Abstention 25.11 -1.20 -0.76Notes: The table shows, for each party, the total number of candidates and the correspond-ing share of all voters in the data (out of 101.168 million eligible voters). The last twocolumns are the simulated effects of introducing NOTA in the full random coefficientsspecification as well as in the more restrictive Logit model.

103

TABLE 2.10: The correlates of NOTA votes

(1) (2) (3) (4)Constituency characteristics:Reserved SC 0.005*** 0.005*** 0.002** 0.002**

(0.001) (0.001) (0.001) (0.001)Reserved ST 0.011*** 0.011*** 0.009*** 0.009***

(0.002) (0.001) (0.002) (0.002)Literacy -0.021*** -0.035*** -0.015** -0.026**

(0.006) (0.010) (0.006) (0.010)Size -0.006* -0.008** -0.003 -0.005

(0.003) (0.004) (0.003) (0.004)Percent male -0.215*** -0.230*** -0.133*** -0.190***

(0.031) (0.046) (0.033) (0.043)Percent SC 0.002 0.010 -0.006 0.006

(0.008) (0.009) (0.007) (0.008)Percent ST 0.010*** 0.008* 0.010** 0.007

(0.004) (0.004) (0.004) (0.004)No latrine 0.002 0.004

(0.004) (0.004)Water nearby 0.016** 0.014**

(0.006) (0.006)Water at home 0.011* 0.013**

(0.006) (0.005)Percent employed 0.015 -0.004

(0.011) (0.011)Rural workers -0.019*** -0.015***

(0.005) (0.005)Car ownership 0.023 0.020

(0.037) (0.034)Computer ownership -0.029 0.036

(0.057) (0.052)Phone ownership -0.009 -0.010*

(0.006) (0.005)TV ownership -0.003 -0.007

(0.006) (0.006)Candidate characteristics:Number of candidates -0.001*** -0.001***

(0.000) (0.000)No female -0.001 -0.001

(0.001) (0.001)<15% female -0.000 -0.000

(0.001) (0.001)Median age -0.000 -0.000

(0.000) (0.000)No SC 0.000 -0.000

(0.001) (0.001)<15% SC -0.000 0.000

(0.001) (0.001)No ST -0.000 -0.000

(0.001) (0.001)<10% ST 0.002 0.001

(0.001) (0.002)R2 0.57 0.60 0.63 0.66N 520 520 520 520Notes: The dependent variable is the share of NOTA votes among all votes. Regressions at the con-stituency level for the cross-section of constituencies affected by the NOTA policy in 2013. Mergeddataset: average demographic characteristics are from the census, average candidate characteris-tics are from the Election Commission. All regressions include state fixed effects. Robust standarderrors in parentheses. ***, **, and * indicates significance at 1, 5, and 10 percent, respectively.

104

TABLE 2.11: Effect of NOTA on turnout, excluding national election years

(1) (2) (3) (4)NOTA 0.033** 0.033** 0.030* 0.031*

(0.015) (0.016) (0.015) (0.015)Basic controls x xExtended controls x xExcluded years 2014 2014 2009, 2014 2009, 2014R2 0.18 0.20 0.19 0.21N 6139 6139 5680 5680States 25 25 22 22Notes: Estimates of the effect of the NOTA policy on turnout from Eqn. (2) using the repeatedcross section sample with specific years excluded. All regressions control for state and yearfixed effects, the log number of eligible voters in a constituency and its square, and the follow-ing state-level variables: labor force participation, real weekly household earnings, fractionof illiterates, fraction with primary school or less as highest education. The Extended controlsspecification also controls for reserved constituencies and the following state level variables:unemployment, sex ratio, fraction urban, and the growth rate of net domestic state product.Standard errors clustered by state in parentheses. ***, **, and * indicate significance at 1, 5,and 10 percent, respectively.

TABLE 2.12: Effect of NOTA on turnout, controlling for redistricting

(1) (2) (3) (4) (5) (6)NOTA 0.033** 0.022* 0.031** 0.020* 0.030** 0.020

(0.016) (0.011) (0.015) (0.011) (0.014) (0.012)Basic controls x x xExtended controls x x xR2 0.21 0.23 0.21 0.23 0.21 0.23N 6084 6084 6084 6084 6084 6084States 22 22 22 22 22 22Notes: Estimates of the effect of the NOTA policy on turnout from Eqn. (2) using the repeated cross section sam-ple. Columns (1) and (2) are run on the states with available constituency boundary files. Columns (3) and (4)control for redistricting using the maximum overlap measure and columns (5) and (6) using the territorial frac-tionalization index. All regressions control for state and year fixed effects, the log number of eligible voters in aconstituency and its square, and the following state-level variables: labor force participation, real weekly house-hold earnings, fraction of illiterates, fraction with primary school or less as highest education. Even-numberedcolumns also control for reserved constituencies and the following state level variables: unemployment, sex ratio,fraction urban, and the growth rate of net domestic state product. Standard errors clustered by state in parenthe-ses. ***, **, and * indicate significance at 1, 5, and 10 percent, respectively.

105

TABLE 2.13: Effect of NOTA on turnout, robustness to state-specificevents

Excluded state Effect of NOTA NBasic controls Extended controls

Chhattisgarh 0.025* 0.035 6505(0.014) (0.021)

Maharashtra 0.031** 0.031* 6109(0.015) (0.015)

Delhi 0.029** 0.031* 6545(0.013) (0.016)

Jammu and Kashmir 0.030** 0.031* 6511(0.014) (0.016)

All four 0.028* 0.039* 5615(0.015) (0.019)

Notes: Estimates of the effect of the NOTA policy on turnout from Eqn. (2) using the repeated crosssection sample with specific states excluded. All regressions control for state and year fixed effects, thelog number of eligible voters in a constituency and its square, and the following state-level variables: laborforce participation, real weekly household earnings, fraction of illiterates, fraction with primary school orless as highest education. The Extended controls specification also controls for reserved constituenciesand the following state level variables: unemployment, sex ratio, fraction urban, and the growth rateof net domestic state product. Standard errors clustered by state in parentheses. ***, **, and * indicatesignificance at 1, 5, and 10 percent, respectively.

106

FIGURE 2.2: Distribution of NOTA vote shares across constituencies

Notes: NOTA vote share is measured as a fraction of total votes cast. N = 1176.

107

FIGURE 2.3: Impact of NOTA on turnout

Notes: Distribution of the changes in turnout across constituencies. Mean = 0.0119, median = 0.0100, N= 519.

108

FIGURE 2.4: Impact of NOTA on candidates’ vote shares

Notes: Distribution of the changes in vote shares across candidates. Mean = -0.0006, median = -0.0000,N = 3068.

109

110

Chapter 3

Firm Ownership and Wage Dispersion:

Evidence from Ghana Using Matched

Employer Employee Data

3.1 Introduction

The role of the firm in worker wage dispersion has been little studied empirically.

It is however considered as a very important determinant of wage inequality. The

environment in which a worker is working may have direct and indirect consequences

on his eventual earnings through measured or unmeasured channels. The idea of

unmeasured worker ability on the one hand and pure industry or firm effects on the

other hand are believed to be important factors affecting observed wage dispersion.

Existing empirical work has shown that neither of these explanations can be re-

jected but seperately identifying these effects using longitudinal data is difficult. As

Hamermesh (2008) points out, the effect would be overstated in favor of the group

on which more information is available. So, if information on workers is richer than

that on firms, the estimated effect of workers would look a lot bigger than that of

firms and vice-versa. This is more of a concern if we believe that the worker and

firm characteristics are correlated. Hamermesh notes that it is much better if we have

a matched employer-employee dataset. In this paper, I use matched data from the

Ghanaian manufacturing industry to analyze the role of firms in observed wage dis-

persion among workers. The empirical methodology of this paper closely follows

Barth et al (2014).

Barth et al (2014) is, to the best of my knowledge, the only paper that studies the

role of establishments in explaining wage inequality using US data. I apply their

methodology to study the same question in a developing country setting. Addition-

ally, I look at how patterns of inequality vary by firm ownership type. Specifically,

are private firms owned domestically different from foreign owned firms? Are the

firm effects explaining wage dispersion different based on ownership structure? I also

analyze how trends in earnings dispersion differ by worker skills.

In the Ghanaian manufacturing sector, we see majority of the manufacturing firms

in the private sector. 1This is largely due to the extensive economic reforms that have

been implemented since the mid to late 1980s. This paper also contributes to the litera-

ture in terms of analyzing earnings in the post-reform era in Ghana. I find an unusual

and very interesting pattern of wage inequality in the Ghanaian manufacturing sector

in this period. In the early period, as the reforms were perhaps sinking in, the vari-

ance in worker earnings increased but towards the end of the 1990s, by the time the

reforms had potentially taken full effect, the variance in earnings actually decreased.

However, while this pattern is true for domestic private firms in Ghana, firms with

1This is true in terms of the dataset I use. Census figures may be different nonetheless.

111

other types of ownership did not witness this decline in the later periods and they ex-

perience a secular increase in variance of earnings. Much of this paper is focussed on

analyzing this interesting pattern and estimating what role the firm component might

have played in this.

Analyzing the variance components, I find that this pattern is explained mainly

by fluctuations in the between-firm variation. The variation within-firms has been

steadily increasing. The initial rise in variance and the later fall seems to be due to

variation between firms rather than within firms. It is important to understand if this

pattern can be explained by gains and losses made by the workers at the top of the

earnings distribution. Interestingly, I find that while net gains to workers above the

90th percentile can explain about 61% of the initial surge in variance, it only explains

about 9% of the later fall. This suggests that workers below the 90th percentile must

have made significant gains in the later stages which resulted in closing down of the

earnings gap. This can possibly be attributed to economic reforms, though the empir-

ical analysis of this paper cannot identify if this is a pure causal effect or not.

I further analyze the interaction of worker and firm characteristics in a regression

setting and find that the variance in workers’ predicted wage from observable charac-

teristics have increased over time whereas the variance of the firms’ effect on wages

have declined over time. This pattern is observed for the average firm and on av-

erage for firms owned privately by domestic Ghanaians. For foreign owned firms

however, this is not replicated. The data suggests that the inverted u-shaped pattern

of variances of log real hourly earnings for private firms can mostly be explained by a

decline in the firm specific effect at the latter stages. The correlation between the con-

tribution of individual attributes to wages and the firm effect have declined over time.

112

This suggests that firms have gradually hired workers by observable characteristics,

independent of firm specific earnings, or in other words, their preference towards hir-

ing similar workers within a firm has decreased. I find that for high skilled workers,

inequality increases over time whereas it declines for low skilled workers.

The rest of the paper is organized as follows. Section 2 gives some background in-

formation. The empirical methodology is introduced in Section 3 and Data in Section

4. The main results are in Section 5. In Section 6, I briefly discuss how the estimated

effects are different over time periods and present conclusions in Section 7.

3.2 Background

3.2.1 The Ghanaian Manufacturing Sector

The Ghanaian economy went through economic reforms from 1983-1991. The manu-

facturing sector was believed to be the engine through which employment generation

and long term economic growth would be attained. (See Teal 1998). An immediate

consequence of providing impetus to the manufacturing sector is that demand for

skilled labor would rise. This was no different for Ghana particularly because the two

most flourishing manufacturing sectors in Ghana were wood products and textiles.

Detailed review of the impact of reforms on the Ghanaian economy can be found in

Baah-Nuakoh and Teal (1993). The manufacturing sector has contributed consistently

113

to around 8-10% of the GDP of Ghana ever since the reforms. Even though by stan-

dards expected by the reform, this may look like a small number but the sector em-

ploys over 250,000 people. 2 Apart from wood and textiles, food processing, smelting,

oil refining, cement and pharmaceuticals are other important sectors in manufactur-

ing in Ghana. A rising concern for the sector is the influx of cheap goods from other

economies like China which have massive comparative advantages in manufacturing.

This is not just a concern for the Ghanaian economy but most of the western African

nations. A recent study by Szabo (2014) points out the inefficiencies in the manufac-

turing sector of Ghana. Szabo estimates production functions for the manufacturing

firms and finds that firms mostly use more capital and less labor than the optimal

amount.

In the wake of this situation, it is worthwhile to look at the wage patterns in the

Ghanaian manufacturing sector which has been the center of attention for Ghanaian

policymakers. It remains indeterminate whether the sector has given the economy

the necessary fillip towards economic growth but we can see how much wages have

changed over time in this sector post the reforms and whether the pre dominance

of private owenrship has had significant effects on worker wages or not. Existing

research has pointed out that there is a link between economic growth and wage dis-

persion in African economies (see Agesa et al, 2011 for details in the context of Kenya).

2See Natcomm Report on Ghanaian Manufacturing Sector -http://www.natcomreport.com/ghana/livre/manufacturing.pdf

114

3.2.2 Related Literature and Conceptual Framework

The study of wage dispersion using matched employer employee datasets has gained

popularity in the last decade and a half. Abowd, Kramarz and Margolis (1999) in

their seminal paper introduced the use of a linked employer employee data set to run

a additively separable fixed effects framework using French data to disentangle the

true effect of a firm from that of the worker.

Fafchamps, Soderbom and Benhassine (2008) studied the African manufacturing

sector and analyzed the role of sorting in explaining wage gaps. Not much is known

however in terms of the role of firms in explaining dispersion in such poor countries.

This paper is an attempt to fill this void in the literature. To my knowledge, this is the

first paper to use a matched employer employee dataset to look at wage dispersion by

ownership pattern for a developing country. In particular the idea is to see if private

domestic ownership of firms leads to higher or lower wage dispersion as compared to

other types of ownership like state-owned, foreign or mixed.

The effect of unobserved worker ability and firm components on wage dispersion

maybe very different for firms owned and managed differently. Especially after eco-

nomic reforms, one might expect more FDI to flow in and more foreign firms being

setup. A body of literature analyzes the impact of FDI inflows on wage dispersion

(Saglam and Sayek, 2011; Heyman, Sjoholm and Tingvall, 2011; Almeida, 2007 and

Martins, 2004). The findings generally indicate that foreign firms lead to higher wage

dispersion and also pay higher actual wages. One explanation is that foreign owned

firms hire skilled workers and that leads to dispersion. Also, lower skilled workers

sort to other firms and hence inter- as well as intra-firm dispersion rises.

It is therefore important to understand if firm behavior actually explains more of

115

the dispersion in wages when owned by foreign enterprises. The following empiri-

cal exercise is an attempt to answer this question. If foreign ownership indeed leads

to higher dispersion, then the firm component should explain more of the wage dis-

persion than worker components. Also, if domestic firms are selected by low skilled

workers, then worker components should explain dispersion more than firm effects

for such firms.

3.3 Empirical Methodology

3.3.1 Analyis of Variance

The empirical strategy I adopt is largely descriptive but provides valuable insight. I do

a variance decomposition of log real hourly earnings into within-firm and between-

firm elements. The matched component of my dataset facilitates this exercise. I follow

the method of Barth et al (2014) throughout this paper and analyze the variance com-

ponents as follows:

V ar(lnwif ) = V ar(Elnwif ) + V ar(lnwif − Elnwif ) (3.1)

where lnwif is the log real hourly earnings (before taxes) for individual ‘i’ working at

firm ‘f’. Elnwif represents the expected value of earnings of workers working at firm

‘f’. V ar(Elnwif ) is the between-firm variance and V ar(lnwif − Elnwif ) is the within

firm variance.

116

3.3.2 Percentile Analysis: Contribution of Earnings Gap to Variance

Barth et al (2014) also propose an arithmetic decomposition of the variance in log real

earnings to calculate the contribution of the difference in mean earnings for various

groups of individuals on the actual variance based on the percentile distribution of

earnings. Suppose we want to compare individuals above the p-th percentile to those

below the p-th percentile. Then if E(p) is the mean of log real hourly earnings of the

top p% and E(100−p) is the mean log real hourly earnings for the remaining workers,

then the variance can be decomposed as follows:

V ar(lnw) =p

100·100− p

100·[E(p)−E(100−p)]2+

p

100·V ar(p)+100− p

100·V ar(100−p) (3.2)

For this paper, I will focus on the earnings gap at the 90th percentile. So the above

reduces to:

V ar(lnw) = (0.1) · (0.9) · [E(10)− E(90)]2 + (0.1) · V ar(10) + (0.9) · V ar(90) (3.3)

The contribution of the earnings gap to variances is therefore given by the term

(0.1) · (0.9) · [E(10)− E(90)]2.

3.3.3 Worker Characteristics and Firm Fixed Effects

The third and final strategy of this paper is to run a modified Mincerian regression for

time periods ‘t’ as follows:

117

lnwif = β.Xif + δf + uif (3.4)

The vectorXif includes worker characteristics like hours worked, age, age squared,

experience, experience squared, education, education squared, parents’ experience

and parents’ experience squared for worker i in firm f . Using the methodology de-

scribed in Barth et al (2014), I decompose the variance of (4) as follows:

V ar(lnw) = V ar(κ) + V ar(δ) + 2cov(κ, δ) + V ar(u) (3.5)

In the above, κ = X · β and represents the predicted earnings from observable

worker characteristics. The term V ar(δ) represents the variance of earnings among

firms or the variance of the firm fixed effects. Another statistic of interest is the degree

of how similar workers are within a given firm, ρδ = cov(κ,δ)V ar(κ)

. Barth et al (2014) note

that if ρδ is close to zero, firms are independently hiring workers based on observed

characteristics, independent of firm earnings.

3.4 Data

The dataset for this study is a panel survey of the Ghanaian manufacturing sector

which took place from 1992 to 2003. The survey was the joint venture of Center

for Study of African Economies (CSAE), the University of Oxford, the University of

Ghana, Legon and the Ghana Statistical Office for the Regional Project for Enterprise

Development (RPED) and Ghana Manufacturing Enterprise Survey. The original sam-

ple of firms was drawn randomly from the 1987 census of manufacturing activities of

118

Ghana. From each of the firms a sample of workers were also interviewed in this pe-

riod. This allows matching the two databases to form a linked employer-employee

dataset. 3

Figure 3.1 shows the shares of firms by ownership categories from our sample. It

appears that the majority of firms are privately owned but some have foreign owner-

ship. For the eventual empirical analysis, I am going to break down the sample into

firms owned and managed by private Ghanaian agents, viz, Private Domestic Firms

and firms with some foreign ownership. The latter would include firms partially or

wholly owned by foreign owners. The remainder of the firms are state owned. Table

3.1 shows some descriptive statistics from plant level data. Column (1) reports means

for all firms whereas columns (2) and (3) report means for the other two categories.

All the means are averages across 12 years of data. Figure 3.2 gives the time series

plots of plant level averages of some of these observables by ownership types.

The average firm has about 75 workers whereas this number is much smaller for

private domestic firms (34) and much higher for foreign owned firms (194). The num-

ber of skilled workers as identified by the firms in their responses to the survey ques-

tion is 14 on average for the whole sample whereas for private domestic it is much

smaller at around 7 skilled workers per firm. The average education of workers is

about 10 years across firm types. Foreign owned firms seem to have workers with

higher average age and potential experience compared to the overall average. Most

foreign owned firms are located in the capital city of Accra. I also find that foreign

firms are more likely to export goods compared to private domestic firms. Value of

3The survey yields a dataset which is a balanced panel of 200 firms but a repeated cross section ofworkers. It was conducted over 7 rounds. The first three rounds were annual surveys, the next threecovered two years each and the last covered three years.

119

total output is higher for foreign firms throughout the span of the data.

There is not much difference in the hours worked by the average worker in these

firms although it is slightly less for foreign owned firms. Foreign owned firms also

have higher wage bills. In terms of input usage and capital, the data suggests that on

average 22% of the inputs are imported. This number is somewhat smaller at under

20% for private domestic firms but significantly higher at 33% for foreign owned firms.

In terms of replacement value of plant and machinery, the mean for foreign firms is

above the average and that for private domestic is below the average for all firms.

3.5 Results

3.5.1 Within-Firm and Between-Firm Effects

Figure 3.3 is indicative of how log real hourly earnings varied in Ghana over the years

(1992-2003). A striking feature of this graph is that the variance of earnings seem to re-

semble an inverted u-shape for Ghanaian manufacturing firms. The variances seem to

have increased from 1992 to around 1998 after which the variances show a decline in

trend. This is intriguing because unlike secular increases in earnings inequality which

are common in the developed world, including the US, here I observe a decline after

an initial surge. . One possible reason for this pattern in the Ghanaian manufacturing

sector is the reforms that happened extensively in the late 80s and continued into the

early 90s. It is plausible that the effects of the reforms on the manufacturing sector,

especially for the domestic firms started accruing around the mid nineties. This might

be the reason that earnings inequality appears to have been checked since 1998 and

into the 2000s. A feature of the Ghanaian manufacturing sector was overstaffing. It is

120

believed that post-reform this has reduced a lot.4 This may have also contributed to

paying workers more efficiently and a resultant decline in earnings inequality. In Fig-

ure 3.3, I look at composition of this variance changes for different ownership patterns.

The Ghanaian manufacturing sector is dominated by private firms. So, I restrict my

analysis only to such firms and find that the inverted u-shaped pattern is replicated

only for domestic firms and not for the foreign owned firms.

The rest of this paper delves deeper into this phenomenon and attempts to analyze

the components underlying such patterns of dispersion. The preceeding discussion

first leads us into the following question: are these patterns due to within-firm or

between-firm variation in worker earnings and how do changes in these components

reconcile the overall trends in dispersion? To answer this question, I decompose the

variances following equation 3.1. The results are reported in Table 3.2. I report the total

variance of log real hourly earnings along with the within and between components

for 3 different years,viz, 1992, 1998 and 2003 to illustrate how the change in overall

variances compare to the change in the within and between components. I find that

for all firms, the within-firm variation has increased over time from 0.444 in 1992 to

0.514 in 1998 and 0.748 in 2003. However, the inverted u-shape of the total variance

curve can be explained mainly by the between-firm component of earnings. I find that

in the early period, the between firm component rises from 0.661 to 0.862 but it falls

to 0.319 in the latter half. This suggests that the between firm component seems to

explain the pattern that we observe in the dispersion of earnings.

Looking at changes over time, I find that the total variance increased in the early

period and declined in the latter by a larger magnitude. The change in within variation

4See Ackah, Adjasi and Turkson - http : //www.brookings.edu/ /media/Research/F iles/Papers/2014/11/learning−to− compete/L2CWP18Ackah−Adjasi− and− Turkson.pdf?la = en

121

increased. So what drives this decline is likely to be the change in between variation.

I find that the change in between-firm variance declined from 0.201 to -0.541.

In Table 3.3, I analyze the overall variances in earnings by firm type. There are

some interesting patterns to note. For private domestic firms, variance seems to be

initially flat but declines in the latter period whereas for private foreign firms it seems

to increase in the early period and remain stable in the latter period. If an influx of

foreign investment disproprotionately attracts skilled labor, this may lead to rising

inequality in the other sectior, viz, domestic in our case.5

3.5.2 Percentile Analysis

Given that we observe the inverted u-shaped pattern for variances in log real hourly

earnings, the question of gainers versus losers assumes importance. The purpose of

this exercise is to estimate whether gains to the workers at the top of the earnigns

distribution has contributed to the inequality? For our unique case, this question can

be reframed to accomodate the decline in inequality at the latter stages to see if top

earners have lost out on the gains they made or have the rest of the workers caught

up. I perform the empirical analysis described in the empirical methodology section

above and focus on the decomposition as per equation 3.3.

Table 3.4 presents the relevant results. I break up the analysis into two periods

of time. The surge phase (1992-1998) and the decline phase (1998-2003). In Panel A,

I analyze the surge phase. Column 2 reports statistics from 1992, column 3 reports

5 To investigate this further, I looked at the variances by worker skill and found that the inequalityfor low skilled workers has reduced over time consistently. This is indicative of the fact that as skilledworkers sort into foreign firms, the unskilled (or low skilled) workers end up in domestic firms anddomestic firms have a lower skilled pool to choose from. This may lead to a decline in dispersion forlow skilled workers.

122

statistics from 1998 and column 4 reports the changes from 1992-1998. I find that the

variance of log real hourly earnings increased by 0.271 log points from 1.105 in 1992

to 1.376 in 1998. I calculate the means of log real hourly earnings for two sets of

workers, viz, above the 90th percentile and below the 90th percentile. I find that in

1992, the difference in these means was 1.649 and it increased to 2.138 in 1998. Using

the expression in equation 3.3, I calculate the contribution of these mean differences

to the variance in earnings.

In 1992, the difference in mean log real hourly earnings for workers above and

below the 90th percentile contributed 0.245 log points of the total variation of 1.105.

This number increased to 0.411 in 1998. The increased contribution of these differences

was 0.166 over the surge period which accounts for 61.25% of the total increase in

variance during this time period. This suggests that the increased advantage of the

top earners over the rest had a major contribution to the surge in wage dispersion in

Ghana during 1992-1998.

I replicate the above analysis for the decline phase and report the statistics in Panel

B. Variance in log real hourly earnings declined from 1.376 in 1998 to 1.067 in 2003,

a change of 0.309 log points. The difference in means of log real hourly earnings for

workers above and below the 90th percentile decreased from 2.138 in 1998 to 2.062

in 2003. The contribution to the overall variance changed by 0.028 which is approx-

imately 9.06% of the total decline in variance. So the relative advantage of the top

earners was not entirely wiped out and was not the major reason why the surge in

inequality reversed. This suggests that workers in the rest of the distribtuion of earn-

ings made significant gains in the latter phases which could probably be attributed to

the sinking in of economic reforms.

123

3.5.3 Returns to Schooling and Earnings Dispersion

I run simple Mincerian regressions as described by equation 3.4 with X denoting the

observables like years of schooling, experience and experience squared only. I use

firm fixed effects and run these regressions for two periods of time, 1992-1998 and

1998-2003. OLS estimates of the coefficients are reported in Table 3.5. I find that the

returns to schooling increased over time whereas the marginal effect of experience

seems to have declined over time. Variance of the predicted earnings from education

and experience has increased from 0.109 to 0.133. This is not consistent with the fall in

variance of earnings over time suggesting that the decline in inequality is explained

more by firm factors and residuals and not necessarily by a decline in inequality of

returns to schooling and experience.

3.5.4 Does Predicted Wage Dispersion by Worker and Firm Charac-

teristics Vary by Ownership Structure?

This is the final section of the empirical analysis and I run the regression described

by equation 3.4 using all observable worker characterisitcs as controls. Since I have

identified two distinctive time periods in the data for studying variance in earnings,

I breakup the analysis in this segment by those time periods. So I run equation 3.4

for the surge period and the decline period seperately by pooling together data from

the years in those periods. Results are reported in Table ??. For the average firm,

variance in log real hourly earnings in the surge period was 1.460 and it decreased to

1.318 in the decline period. I decompose the variances following equation 3.5. I find

that the variance in predicted wage due to observed worker characteristics increased

124

from 0.235 log points to 0.331 log points whereas the variance from firm level earn-

ings decreased from 0.595 log points to 0.385 log points. This is consistent with the

idea that firm level characterstics seem to have contribued to the decline in wage in-

equality in Ghana from 1998-2003 rather than worker characteristics. Earlier, I found

that between firm variation seems to explain the inverted u-shaped pattern better than

the within-firm variation and the findings in this section corroborate those. I find that

the variance in the residuals have gone down over time. The sorting factor between

worker-firm, ρδ also decreased over time. This suggests that with time firms have

employed workers based on observed characteristics and independent of firm level

earnings. This also suggests the possibility that positive sorting was on the decline

where higher pay workers now joined lower average earnings firms and that helped

bridge the earnings gap and led to a reduction in inequality.

I am interested to know if the patterns for the average firm is universal among

ownership types or is it specifically driven by a particular organizational structure.

I split the above analysis to seperately study private domestic firms and private for-

eign firms. I find that the average effect is replicated for private domestic firms. The

variance of κ increased from 0.308 to 0.400 whereas the variance of the firm earnings

decreased from 0.387 to 0.232. The sorting factor also decreased. However, for firms

that are partially or wholly owned by foreign private enterprenerus, I find the oppo-

site. Recall that for these firms, we do not observe the inverted u-shaped pattern in

3.3 and there is a secular increase in inequality for these firms. The variance of κ in-

creased from 0.139 to 0.214 and variance of δ increased from 0.134 to 0.441 for these

firms. This suggests that the ownership structure seems to be an important element

that characterizes the unique wage variance structure in Ghana.

125

3.6 Conclusions

This paper uses a matched employer-employee dataset from the Ghanaian manufac-

turing sector to analyze earnings dispersion in Ghana from 1992-2003. The plot of

variances in log real hourly earnings over time indicate two clear periods of surge and

decline in inequality in Ghana. I analyze the variances in earnings from 1992-1998

when inequality increased and find that within and between firm variation increased

during this period. From 1998 to 2003, the between firm variation declined whereas

the within-variation still increased. This suggests that the fall in inequality in the latter

stages was driven by between-firm effects.

I find that the earnings gap between the workers above and below the 90th per-

centile explains about 61% of the rise in inequality from 1992-1998 but it only explains

about 9% of the fall in inequality from 1998-2003. This is suggestive evidence that

top earners did not necessarily lose out their advantage in the latter stages but other

workers made significant gains to wipe out this gap.

Variance in the predicted wages from worker characteristics seem to have risen

consistently whereas variance from firm level earnings have decline in the later pe-

riod. This correlation between the predicted wage by worker characteristics and the

firm earnings have also declined. This leads us to believe that changing patterns of

sorting could be a possible reason for the decline in earnings inequality in Ghana.

Firms appear to have independently hired workers based on their observed traits and

even firms with low average earnings attracted hire pay workers. This meant that

inequality would be on the decline. These patterns are however only observed in pri-

vate domestic firms. Foreign owned firms do not seem to have such effects on earnings

variance.

126

3.7 References

1. Abowd, John M., Robert Creecy (2002). & Francis Kramarz. “Computing Person

and Firm Effects Using Linked Longitudinal Employer-Employee Data.” Cornell

University Working Paper

2. Abowd, John M., Francis Kramarz. & David Margolis (1999). “High Wage Work-

ers and High Wage Firms” Econometrica 67

3. Agesa, Richard U., Jacqueline Agesa (2011). & Andrew Dabalen. “Changes in

Wages, Wage Inequality and the Return to Human Capital Skills in Kenya: 1977-

2005.” Review of Development Economics 15

4. Ahiakpor, Ferdinand. & Raymond Swaray (2015). “Parental expectations and

school enrolment decisions: Evidence from rural Ghana.” Review of Development

Economics 19

5. Almeida, Rita (2007). “The labor market effects of foreign owned firms.” Journal

of International Economics 72

6. Baah-Nuakoh, Amoah. & Francis Teal (1993). “Economic Reform and the Man-

ufacturing Sector in Ghana.” RPED Country Studies Series

7. Barth, Erling., Alex Bryson., James C Davis., & Richard Freeman (2016). “It’s

Where You Work: Increases in Earnings Dispersion across Establishments and

Individuals in the United States.” Journal of Labor Economics

8. Card, David., Ana Rute Cardoso. & Patrick Kline (2013). “Bargaining and the

Gender Wage Gap: A Direct Assessment.” IZA DP 7592

127

9. Card, David., Jorg Heining. & Patrick Kline (2013). “Workplace Heterogeneity

and the Rise of West German Wage Inequality.” Quarterly Journal of Economics

128

10. Data Source Acknowledgement: Regional Project on Enterprise Development &

Ghana Manufacturing Enterprise Survey, Rounds I to VII (12 years: 1992-2003)

11. Fafchamps, Marcel., Mans Soderbom. & Najy Benhassine (2009). “Wage Gaps

and Job Sorting in African Manufacturing.” Journal of African Economics 18

12. Hamermesh, Daniel (2008). “Fun with matched firm-employee Data: Progress

and Road maps.” Labour Economics 15

13. Heyman, Fredrik., Fredrik Sjoholm. & Patrik Gustavsson Tingval (2011). “Multi-

nationals, cross-border acquisitions and wage dispersion.” Canadian Journal of

Economics

14. Martins, Pedro S. “Do Foreign Firms Really Pay Higher Wages? Evidence from

Different Estimators.” IZA DP No. 1388 (2004).

15. Saglam, Bahar Bayraktar & Selin Sayek (2011). “MNEs and wages: The role of

productivity spillovers and imperfect labor markets.” Economic Modelling 28

16. Szabo, Andrea (2014). “Measuring Firm-level Inefficiencies in the Ghanaian

Manufacturing Sector.” Working Paper, University of Houston

17. Teal, Francis (1998). “The Ghanaian Manufacturing Sector 1991-95: Firm Growth,

Productivity and Convergence.” Working Paper, CSAE, University of Oxford

128

FIGURE 3.1: Share of Firms by Ownership Type

129

TABLE 3.1: Descriptive Statistics

All Firms Private Domestic Any Foreign(1) (2) (3)

Number of Workers 74.50 34.53 194.41

Number of Skilled Workers 13.69 6.85 35.01

Years of Education of Average Worker 10.02 9.65 10.99

Potential Experience of Average Worker 15.26 6.85 35.01

Age of Average Worker (in years) 32.60 30.11 39.32

Weekly Hours Worked by Average Worker 45.90 46.87 43.56

Located in Accra (=1 if Yes) 0.59 0.53 0.70

Log Annual Wage Bills 16.58 15.77 18.68

Log Real Output 17.36 16.69 19.41

Exports(=1 if Yes) 0.33 0.29 0.46

Log Replacement Value of Plants and Machinery 17.67 16.65 20.80

Percentage of Inputs Imported 22.03 19.13 32.56

Observations (Firm X Years) 2400 1464 804

Notes: Data comes from the CSAE-RPED database as discussed in the data section. Any Foreign cor-responds to firms partially or wholly owned by foreigners. Private domestic refers to firms owned andmanaged by private Ghanaian residents entrepreneurs. Total observations are 2400, for private domes-tic it is 1464 and for foreign owned it is 804. Remaining 132 are observations for state firmsAll figuresare means computed from plant level database of 200 firms overall for 12 years. .

130

TABLE 3.2: Analysis of Variances of Log Real Hourly Earnings of Workers

1992 1998 2003(1) (2) (3)

All Firms

Total Variance 1.105 1.376 1.067

Within 0.444 0.514 0.748Between 0.661 0.862 0.319

Decomposing the Change over time 1992-1998 1998-2003

Total Change 0.271 -0.305

Within Change 0.070 0.234Between Change 0.201 -0.541

Data from CSAE-RPED on the Ghanaian Manufacturing Sector from 1992-2003. I matcched the workerand firm files to create the linked dataset based on which I perform the analysis. The arithmetic calcu-lations are based on the method described in Section 3.1 in the text.

TABLE 3.3: Variances of Log Real Hourly Earnings of Workers by Own-ership Type

1992 1998 2003(1) (2) (3)

Total Variance 1.105 1.376 1.067

Firms with Full Domestic Ownership 1.321 1.371 1.151

Firms with Some Foreign Ownership 0.311 0.671 0.834

Data from CSAE-RPED on the Ghanaian Manufacturing Sector from 1992-2003. I matcched the workerand firm files to create the linked dataset based on which I perform the analysis. The arithmetic calcu-lations are based on the method described in Section 3.1 in the text.

131

TABLE 3.4: Contribution of Earnings Gap to Variance in Earnings

(1) (2) (3)

Panel A 1992 1998 Change

Variance (Log Real Hourly Earnings) 1.105 1.376 0.271

Mean (Log Real Hourly Earnings) Above 90th percentile 5.988 6.488Mean (Log Real Hourly Earnings) Below 90th percentile 4.339 4.350Difference in Means (Above - Below 90th Percentile) 1.649 2.138

Contribution of Difference in Means to Variance 0.245 0.411 0.166Percentage of Total Change in Variance 61.25

Panel B 1998 2003 Change

Variance (Log Real Hourly Earnings) 1.376 1.067 -0.309

Mean (Log Real Hourly Earnings) Above 90th percentile 6.488 7.036Mean (Log Real Hourly Earnings) Below 90th percentile 4.350 4.974Difference in Means (Above - Below 90th Percentile) 2.138 2.062

Contribution of Difference in Means to Variance 0.411 0.383 -0.028Percentage of Total Change in Variance 9.06

Data from CSAE-RPED on the Ghanaian Manufacturing Sector from 1992-2003. The contribution of thedifference in means has been calculated as per the description in Section 3.2 in the text. The variance inlog earnings can be arithmetically decomposed following Barth el al (2014) as in equations (2) and (3)to calculate the contribution to variance. The percentage share has been calculated as the change in thecontribution as a share of the total change in the variance.

132

TABLE 3.5: Returns to Schooling and Experience

Dep Var: Log Real Hourly Earnings

1992-1998 1998-2003(1) (2)

Schooling 0.042*** 0.063***(0.008) (0.002)

Experience 0.068*** 0.046***(0.012) (0.004)

Experience2 -0.0013*** -0.0005***(0.0003) (0.0001)

Mean of Dep Var 4.568 4.654

Variance of Dep Var 1.460 1.318

Variance (X · β) 0.109 0.133

Data from CSAE-RPED on the Ghanaian Manufacturing Sector from 1992-2003. I matcched the workerand firm files to create the linked dataset based on which I perform the analysis. Log real hourlyearnings are measured in Ghanaian cedis. All regressions include firm fixed effects. Robust standarderrors reported in parentheses. ***,** and * represent significance at 1%, 5% and 10% respectively.

133

TABLE 3.6: Variance Decomposition: Estimated Firm Effects and Pre-dicted Wage from Worker Characteristics

Surge in Inequality Decline in Inequality

1992-1998 1998-2003

(1) (2)

All Firms

Earnings Dispersion: V ar(lnw) 1.460 1.318Individual Characteristics: V ar(κ) 0.235 0.331Firm Fixed Effect: V ar(δ) 0.595 0.385Cov(κ, δ) 0.175 0.164Sorting: Worker-Firm ρδ 0.745 0.495Residuals: V ar(u) 0.498 0.385

Firms with Private Domestic Ownership

Earnings Dispersion: V ar(lnw) 1.551 1.369Individual Characteristics: V ar(κ) 0.308 0.400Firm Fixed Effect: V ar(δ) 0.387 0.232Cov(κ, δ) 0.124 0.081Sorting: Worker-Firm ρδ 0.403 0.202Residuals: V ar(u) 0.580 0.535

Firms with Any Foreign Ownership

Earnings Dispersion: V ar(lnw) 0.803 0.825Individual Characteristics: V ar(κ) 0.139 0.214Firm Fixed Effect: V ar(δ) 0.134 0.441Cov(κ, δ) 0.036 -0.075Sorting: Worker-Firm ρδ 0.259 -0.350Residuals: V ar(u) 0.808 1.413

Data from CSAE-RPED on the Ghanaian Manufacturing Sector from 1992-2003. Regressions are ex-tensions of mincerian equations and include the following independent variables: age, age squared,education, education squared, experience and its squared and firm fixed effects. Dependent variable islog real hourly earnings, κ is predicted wage from observed worker characteristics, δ is firm effect onearnings and ρδ is the correlation between the worker characteristics and firm effects.

134

FIGURE 3.2: Time Series Plots of Firm Averages by Ownership Types:Firm Dataset of 200 Firms over 12 Years

135

FIGURE 3.3: Variances of Log Real Hourly Earnings by Ownership Type

136

© copyright by somdeep chatterjee may 2016

Documents