Inequality of Opportunity and Economic Growth in India

By taking the three circumstance variables of father’s education, caste identity, and sector of residence to define the inequality of opportunity share in the overall inequality, and [r]

Keywords: Inequality; Inequality of Opportunity; Growth; India

JEL Codes: I3; J62; N35; O15

+ I am thankful to my PhD supervisor, Prof Ashwini Deshpande, for helping me to think more

clearly and critically on the whole range of issues involved

* PhD Student, Department of Economics, Delhi School of Economics, University of Delhi, Delhi 110

007, India Email: mrinalini@econdse.org

2

1 Introduction

Economic inequality and its relationship with economic growth has interested researchers from at least

as early as the 1950s Arthur Lewis and Simon Kuznets are generally regarded as the pioneers in studying the relationship between economic growth and income inequality.1 Lewis (1954) and Kuznets (1955) suggested that income inequality is determined primarily by the level of economic growth These studies tested the causal effect of growth on inequality The research initiated by them has continued till date, with greater sophistication of method and new insights, but also with sharply varying results Studies took into account the plausible reverse causality in this relationship wherein growth determines inequality, but inequality too may be affecting growth Further, researchers also studied the relationship between inequality and growth, looking at the causal effect of inequality on growth Indeed, a common takeaway from a range of empirical and theoretical papers studying this relationship is that they lack consensus in the relationship they establish between inequality and growth According to some of these studies, inequality has a negative effect on growth (for instance, Alesina and Rodrik (1994); Persson and Tabellini (1994)); others demonstrate inequality’s positive impact on economic growth (Forbes (2000); Li and Zou (1998)), and a third category of papers establishes a non-linear, sign-changing relationship (for instance Banerjee and Duflo (2003)) Thus, while there is a great deal of research on the relationship between inequality and growth, the studies lack consensus

The World Development Report 2006 (WDR 2006) and Bourguignon et al (2007) reason that this

prevailing ambiguity over the effect of inequality on economic growth could be due to the failure to distinguish between different kinds of inequality It was hypothesized by Marrero and Rodriguez (2013) that the ambiguity may get resolved if one works with two constituents of economic inequality, namely, inequality of opportunity and inequality due to effort Roemer (1998) formalized the concept of inequality of opportunity in his detailed, path-breaking exposition Drawing on the familiar idea that outcomes or ‘advantages’ like income, consumption, wealth, health status, education, etc valued by individuals are dependent on a variety of factors, Roemer grouped these factors into two categories, namely, factors for which individuals cannot be held responsible, called ‘circumstances’, and those for which they can be held responsible, called ‘efforts’.2 Inequality in outcomes caused due to efforts is

1 World Development Report (2006) (henceforth, WDR (2006)), p 44

2 Roemer (1998) uses the term ‘advantage’ to refer to outcomes These terms – effort, circumstance, advantage, are now a part of common parlance in inequality of opportunity literature

3

considered to be normatively acceptable while inequalities caused due to circumstances are unanimously considered to be unfair and so ought to be eliminated as a matter of policy Equality of opportunity is defined as a situation where distribution of important outcomes is independent of circumstances Any society plagued with unequal opportunities has individuals’ economic success largely predictable in terms of family background such as caste, religion, etc as well as other exogenous factors like gender, region of residence, etc Inequality of opportunity refers to the inability of an individual in achieving an economic outcome due to the lack of opportunities And these opportunities are missing

because of characteristics of individuals beyond their control (circumstances)

The WDR 2006 primarily focused on the ‘instrumental relationship between equity and development’3, studying ‘the effects of unequal opportunities when markets are imperfect’4 Equity in WDR (2006) is defined in terms of greater equality of opportunity Drawing their motivation from WDR (2006),

Bourguignon et al (2007) argued that ‘inequality of opportunity might have more adverse consequences than the inequality which arises from differential rewards to effort’ on economic performance.5

The ‘total inequality’ in the aforementioned studies is made up of two components – one, inequality of opportunity, which is inequality caused due to circumstances and two, inequality due to effort or the residual inequality It is along these two distinct forms of inequality that we pursue our enquiry, by specifically looking at the relationship between growth and the share of inequality of opportunity (i.e., the proportion of inequality of opportunity in total inequality) This paper studies the effect of the interstate variation in share of inequality of opportunity on the interstate variation in growth for a panel

of seventeen major Indian states over a period of roughly two decades (1993-94 to 2011-12) This work happens to be the first of its kind in the Indian context, to the best of our knowledge This relationship has not yet been explored for India, and we intend to contribute to the literature by doing so In general too, i.e., not specific to India, this research topic (involving inequality of opportunity and growth) has a very small body of work with the question being asked in slightly modified form by Marrero and Rodriguez (2013) and Ferreira et al (2017)

We do this exercise by first measuring inequality of opportunity in monthly per capita consumption expenditure (MPCE) for each state This gives us our main variable of interest – share of inequality of opportunity We supplement this with information on other explanatory variables to run our primary

3 WDR (2006), p 7

4 WDR (2006), p 7

5 Ferreira et al (2017)

4

regression6, and present results from several estimation methods including three specialized Generalized Method of Moments (GMM) techniques (difference GMM, system GMM, and collapsed system GMM); these methods have been used extensively in the existing literature on this topic

Our results show that the share of inequality of opportunity negatively affects economic growth Our research thus provides evidence for the hypothesis that inequality of opportunity is bad for growth, and

we use our result to build a case for promoting equal opportunities on grounds of achieving higher economic growth

The remaining paper is organized in the following manner In Section 2 we briefly discuss the papers establishing the concept of inequality of opportunity and its measurement in Indian context, followed by the few papers looking specifically at the effect of inequality of opportunity on growth The concept of inequality of opportunity and its measurement is discussed in detail in Section 3 In Section 4 we describe the dataset and the econometric specification used in the study We present the results from our analysis in Section 5, and offer concluding comments in Section 6

2 Review of Related Literature

Roemer (1998) is considered to be a seminal work in this field of research for formalizing the idea of equality of opportunity, even though the concept of equal opportunities was not new.7 The book provides a detailed analysis of the two constituents of total inequality (inequality of opportunity and inequality due to effort), beginning with a discussion on the two notions of equality of opportunity already prevalent when the book was written, namely, levelling the playing field, and the non-discrimination principle.8 Roemer (1998), while giving precise definitions of circumstances, effort, and opportunities, provided a setup for implementing an “equal opportunity policy” He provided separate discussions on achieving equal opportunities in indicators of health, education, etc Following the idea of equal opportunities put forth by Roemer (1998), researchers started empirically measuring the extent of inequality of opportunities in different economic indicators in several economies In the Indian context, Ashish Singh has extensively worked on inequality of opportunity, measuring unequal opportunities in a range of indicators like, earnings and consumption expenditure (Singh (2012b)); access to primary

6 Detailed description in Section 4

7 Dworkin (1981); Arneson (1989); Cohen (1989), while using different arguments, essentially build an argument for equality of opportunity

8 Roemer (1998), p 1

to have a bearing on growth The existence of such linkages is what motivates this line of research, and

we begin with a discussion of these linkages

Loury (1981) focuses on the distribution of earnings arising as a consequence of different family backgrounds of individuals He argues, it is commonly observed that earnings are positively related across generations9 and one explanation for this is that the skill sets of individual and their parental income are positively related Thus, in this study, family background, which qualifies as a circumstance variable, is shown to have an effect on human capital formation and therefore economic growth Another channel was studied by Galor and Zeira (1993), who show in their model that under imperfect credit markets, with borrowing rates higher than the rate at which funds are lent, individuals who are born in wealthy families will find it easier to invest in human capital (as they do not need to borrow funds at a high rate) So, the distribution of initial wealth affects the aggregate investment in human capital Furthermore, with human capital being an important determinant of final output and hence growth, they conclude that initial wealth distribution affects economic performance Individuals in the model are assumed to be identical in all aspects other than the wealth they inherit This implies that any kind of inequality observed in their final outcome is essentially inequality of opportunity, having a significant bearing on subsequent growth Banerjee and Newman (1993) also model the dependency of long-term prosperity and growth on initial distribution of wealth Economies starting with high inequality end up with low growth rates, low employment and low wages, in contrast to economies starting with low inequality, which have high growth rates, low unemployment and high wages These studies show that individuals born with unfavourable circumstances, for example with respect to gender, family background, caste affiliation, etc., irrespective of their innate abilities or exerted efforts, face hurdles while attempting to access credit, which curtail them from making desirable investment in human capital, which consequently negatively affects growth Chiu (1998) also demonstrates the effect

of circumstances on growth In a given economy, the socio-economic background an individual is born into plays a crucial role in determining how the innate abilities of these individuals are developed and

thus decisively affecting the final outcomes in their lives Those born in wealthier households get better

9 Jencks (1972)

6

access to education and other skill-developing programmes which give them an advantage over those born with probably better innate abilities but born in families where financial constraints result in these abilities not being equally well-developed and harnessed As a result the final outcomes of this latter group of individuals is often poorer than those of the first group of individuals Initial income distribution plays a role in determining the growth trajectory of the economy Chiu (1998) shows that as there is a redistribution of wealth from the rich to the poor, aggregate human capital in the economy will go up as the erstwhile rich who lose money (become poorer) will lower the spending of the human capital formation of their less talented (lower innate ability) children, while the erstwhile poor who have gained money (become richer) will spend more on their more talented children As the more talented children get more resources for their human capital formation, it builds a case for greater income equality Since human capital formation enhances growth, and there is a greater human capital formation if families have similar initial income, greater income equality leads to higher growth or better economic performance

The above models demonstrate the plausible channels through which circumstances may affect economic growth Motivated by such linkages, in the beginning of this decade, Marrero and Rodriguez (2013) approached the debate on the relationship between inequality and growth, empirically from a new direction Their paper tests the hypothesis that the two kinds of inequality – inequality of opportunity and inequality due to effort, affects growth in opposite ways They look at the relationship between inequality of opportunity, inequality due to efforts, and growth in the context of the United State of America They find a negative significant relationship between inequality of opportunity and growth, and a positive but insignificant relationship between inequality due to efforts and growth Ferreira et al (2017) also study the relationship in a cross-country analysis, but while they find a negative relationship between both kinds of inequality and growth, the relationship is not statistically significant

3 Understanding Inequality of Opportunity

We follow the ex-ante, parametric approach for estimation of inequality of opportunity as proposed by Ferreira and Gignoux (2011) This approach of parametric estimation follows Roemer’s (1998) definition

of equality of opportunity which requires distribution of outcomes to be independent of circumstances This in turn implies the following three conditions – (1) none of the circumstance variables should have a direct causal effect on y, (2) all the effort variables, if being considered explicitly in the estimation, should be distributed independently of circumstances, and (3) by assumption, random factors and luck

1 The measure is perfectly decomposable

2 The measure is additively separable

3 In the class of inequality indices using arithmetic mean as the reference income and satisfying Dalton transfer axiom, mean log deviation is the only path-independent decomposable inequality measure

Pigou-Using mean log deviation we thus get estimates for overall inequality in our outcome variable Now, to get to our measure of inequality of opportunity, we generate a parametrically standardized distribution

of our outcomes We give each and every individual the same circumstance variable, thereby eliminating any inequality that is associated with circumstances For generating this counterfactual distribution, we take outcomes to be a linear function of circumstances, effort and other factors (error term) Further,

we take effort itself to be endogenous to circumstances Error term is itself a function of circumstances

as well as some exogenous factors Following Ferreira and Gignoux (2011) we consider a log-linear functional form for the purpose of empirical estimation,

ln y = C + Eβ + u (1)

E = BC + v (2) where, y ∶ outcome variable (MPCE)

C ∶ vector of circumstance variables

E ∶ vector of effort variables

u, v ∶ other random factors (error term)

The reduced form of the above two equations, combining the direct effect of circumstance on the outcome variable and the indirect effect of circumstance through effort (E itself being a function of C) on the outcome variable, is given by,

10 Ferreira and Gignoux (2011)

8

ln y = C(α + βB) + vβ + u (3)

or,

ln y = Cθ + є (4)

where, C is the vector of circumstance variables

Ө is the coefficient given by α + βB

є is the error term given by vβ + uWe estimate the above equation using ordinary least squares (OLS) and get estimates for the predicted value of θ (θ�) and predicted value of error term for each individual (є̂i) Once we get the predicted error terms, we generate the counterfactual distribution for our outcome under the counterfactual of the same set of circumstances for the whole sample:

ŷi= exp ( C��θ�� + є̂i ) (5) where,C� = vector of sample average for circumstances

Ө� = estimated Ө from the above OLS regression

є�i= estimated єi for each individual from the above OLS regression (equation number (4))

We get this distribution of earnings ŷi by giving each and every individual the same circumstance variable, i.e., by replacing individual circumstance values with the sample average of each circumstance variable, thereby eliminating any inequality that is associated with difference in circumstances

Now we have everything required to estimate the measure of inequality of opportunity in our outcome

y The overall opportunity share of inequality (σi) of outcome y is given as the difference between the inequality in actual distribution of outcome and the inequality in the counterfactual distribution of outcome as a proportion of the inequality in actual distribution of outcome

σi = [ I({yi}) – I({y[ I({yi | Ci = C�}) ]

i)] (6)

or,

σi = [ I{yi} – I{y�i} ]

I{yi} (7) where,

yi = actual outcome distribution in our sample

y�i= counterfactual outcome distribution

9

I{yi} = mean log deviation= M1 Σ ln y̅ y

i , M being the number of observations

The numerator of equation number (7) gives the level of inequality of opportunity (difference between observed or factual inequality and inequality in counterfactual distribution) Equation (7) gives the measure of overall opportunity share in total observed inequality

4 Data and Methodology

As we have elaborated till now, through our work we intend to understand the causal effect of the existing share of unequal opportunities on subsequent economic growth To get this, we estimate the following regression equation, number (8) –

gjt = β1 yj,t−k+ β2 IOp(y)j,t−k+ β3 LEj,t−k+ αj+ ηt+ µjt (8) where,

gjt = average annual growth rate of net per capita SDP during a k year interval to year t (k being the gap years between the two rounds of NSS data being used)

yj,t−k = lagged net SDP per capita in constant prices

IOp(y)j,t−k = lagged share of inequality of opportunity in MPCE

LEj,t−k = lagged life expectancy

αj = state fixed effects

ηt = time fixed effects

µjt = error term

We estimate this above equation in a panel dataset using four thick rounds of the National Sample Survey (NSS) Employment Unemployment Survey (EUS): Round 50 (1993-94), Round 55 (1999-2000), Round 61 (2004-05), and Round 68 (2011-12) Our dependent variable gjt is the average annual growth rate of per capita net state domestic product in state j, during a k year interval to year t We get this variable using the data released by the Central Statistics Office (CSO) on per capita state domestic product (PCSDP) at constant prices (2004-05 base) From the same data source (CSO) we also obtain our

of inequality of opportunity in MPCE among individuals in each state in the (t-k)th year period Inequality

of opportunity is estimated for males in the age group of 16-6511 We run the regression on seventeen major Indian states For computing inequality of opportunity, we use three circumstance variables, namely, father’s education, caste, and sector of residence (rural versus urban) For India, Singh (2012b) reports parental background and caste identity as influential circumstances viś-a-viś the final outcomes Further, our measure of inequality of opportunity is a lower-bound estimate of the true inequality of opportunity as data restrictions imply that all the relevant circumstance variables have not been included.12

LE or the life expectancy is measured as the percentage of population above the age of 65 This is an often used variable to capture health capital of an economy, considered to be an important contributor

to subsequent growth.13 This further translates into better human capital which leads to improved growth

We do not include male or female education as one of the controls in this regression, as father’s education is a variable already included in our estimation of inequality of opportunity We also refrain from including a control for urban versus rural living because our inequality of opportunity estimates already capture its effect on the resulting inequality

The above growth equation number (8) may be estimated using several econometric techniques First, the equation can be estimated using the OLS technique However, the OLS estimators may be biased due to the correlation between the state fixed effects (αj) and the lagged dependent variable ( yj,t−5),

11 While we would have liked to use gender as a circumstance variable too, the survey is not designed to capture some crucial information required, like father’s education level for adult women In India, where there in near universality of marriage, the likelihood of finding an adult unmarried woman in her parental home is very low in household surveys Post marriage, as women typically move into their marital homes, it becomes difficult to get their parental background.

12 Ferreira and Gignoux (2011)

13 Barro (2013)

11

irrespective of the way we treat(αj) Thus, OLS estimator of this dynamic panel data model will be inconsistent For the same reason, a random effects estimator too will give inconsistent estimates Doing a fixed effects (FE) estimation would get rid of the state fixed effects However, the demeaning process creates a correlation between the transformed lagged dependent variable ( yj,t−k− y̅j) and the transformed error term (µjt− µ̅j), leading to inconsistent fixed effects estimator This is termed as the Nickell bias

To solve for the above inconsistency problems, the model is transformed to get rid of the state fixed effects It involves taking first differences of the original model, i.e., our growth equation number (4),

gjt− gjt−1

= β1� yj,t−k− yj,t−(k−1)� + β2 �IOp(y)j,t−k− IOp(y)j,t−(k−1)�+ β3 �LEj,t−k− LEj,t−(k−1)� + �αj− αj� + (ηt− ηt−1) + (µjt− µjt−1) While this does get rid of the fixed effects(αj), correlation continues to exist between the first-differenced lagged dependent variable � yj,t−k− yj,t−(k−1)� and the error term(µjt− µjt−1) For example, for time period three, i.e., year 2004-05, the first-differenced lagged dependent variables is� yj,2− yj,1�, and the transformed error term is�µj,3− µj,2� However, now with the fixed effects washed out, we have the option of doing an instrumental variable estimation Difference GMM (Diff-GMM) and system GMM (Sys-GMM) methods were developed (Arellano and Bond (1991)) to cater to this problem, where internal instruments are employed In this technique, finding suitable external instruments for each variable is circumvented for datasets where longer lags of the regressors are available in the data to be used as instruments Under the Arellano-Bond estimation, the model is specified as a system of equations – one equation for each time period – and the instruments applicable

to each equation differ So, for later time periods, additional lagged values of regressors are available as instruments For our model, we use twice-lagged level regressors as instruments For time period three (NSS-EUS Round 61, 2004-05), there will be one instrument for each variable, for time period four (NSS-EUS Round 68, 2011-12), there will be two instruments for every variable Taking all these available instruments improves the efficiency of the Arellano-Bond estimator

In the context of our study (dynamic panel data model), if the series is highly autoregressive, i.e., time series are persistent with the coefficient for the lagged dependent variable close to one, and the number of time series observation is small, the Diff-GMM estimators have been found to have large

We do the standard tests for instrument validity and no autocorrelation, post our estimation The Hansen J-test of instrument validity has the null hypothesis of “the instruments as a group are exogenous” So, the higher the reported p-value, the better it is However, Roodman (2009a) strongly argues that this statistic should be treated with caution While a significant Hansen statistic implies that the set of instruments are invalid due to endogeneity, an implausibly high p-value (close to one) may be due to the effect of instrument proliferation That is, one should not read too much into very high p-values of the Hansen statistic because that suggests that due to the large number of instruments, the test has been weakened to the point that they are no longer informative However, in our model we get reasonable values for the Hansen statistic which do not indicate instrument proliferation The Arellano-Bond test for autocorrelation has a null hypothesis of no autocorrelation The test for AR(1) usually rejects the null hypothesis, but this is expected (Δµjt = µjt - µjt−1, and Δµjt−1 = µjt−1 – µjt−2 , both have

µjt−1)

5 Descriptive Statistics and Results

We begin by discussing some descriptive statistics of our dataset In Table 1, we present the average number of years of education for fathers in each state over the four rounds of data The average years

of father’s education has increased between the first round (1993-94) and the last round (2011-12) of

14 Blundell and Bond (1998), Blundell at al (2001), Bond et al (2001)

15 Ibid

Table 3 here

We report the fraction of individuals residing in urban areas in Table 4 This gives us a sense of urbanization of population over the four rounds This proportion too has gone up between Round 1 (1993-94) and (2011-12)

Table 4 here

In Table 5, we report the total inequality in MPCE over the four rounds Total inequality increases for all states between Round 1 (1993-94) and Round 2 (1999-2000) This result is also seen at the all-India level There is a fall in the total inequality observed post 1999-2000, i.e., Round 2

Table 5 here

Table 6 describes the share of inequality of opportunity in the four rounds of data The share of inequality of opportunity falls for all states except Punjab, Maharashtra, Madhya Pradesh, and Karnataka between 1993-94 (Round 1) and 1999-2000 (Round 2) Thereafter, all states witness an increase in the share of unequal opportunities in 2004-05 (Round 3), except Himachal Pradesh At the all-India level, 31% of the total inequality in MPCE was explained by circumstances in 2011-12 (Round 4), and as we have previously discussed, these figures represent the lower-bound for the actual share of inequality of opportunity that would have been observed had we taken into account the entire set of circumstances

For compatibility with other papers we present results from five specifications – OLS, FE, Diff-GMM, GMM, and collapsed sys-GMM For each GMM specification, we also present the Hansen J-test of instrument validity, and the Arellano-Bond autocorrelation tests

Sys-In tables 8 and 9, we have presented estimates from our two regressions Table 8 reports results from the model where we see the effect of total inequality on economic growth Table 9 is our primary model

of interest, whereby we look at the effect of inequality of opportunity on economic performance Our results from both the above models provide evidence for conditional convergence The sign of the coefficient on initial income is negative and significant at the 99% confidence level for two specifications, including that in diff-GMM, for both our models – with total inequality, and with the share of inequality of opportunity We do not get significant effect of total inequality on growth on any

of our specifications This is similar to the estimates in Ferreira et al (2017), where they get significant coefficients of total inequality only in their OLS and FE specification, which we know are very likely to give biased estimates All the GMM specifications give insignificant coefficients in our model as well as those in Ferreira et al (2017)

Table 8 here

The coefficient on the human capital variable (life expectancy) is not significant for any of our specifications This should not be too worrying because we have also included another variable on human capital – number of years of father’s education – while estimating our inequality of opportunity estimate, which is found to significantly affect inequality Our main variable of interest, share of inequality of opportunity, is defined as the share of inequality of opportunity in overall observed inequality (equation (3)) As can be seen in Table 9, FE and diff-GMM estimates give us the result that