Who are the Indian Middle Class? A Mixture Model of Class Membership Based on Durables Ownership1, 2

Sudeshna Maitra

Department of Economics, York University 1038 Vari Hall 4700 Keele Street Toronto, ON M3J 1P3 Email: smaitra@econ.yorku.ca Phone: 416-736-2100 Ext. 77052 Fax: 416-736-5987

July 2007

PRELIMINARY DRAFT, PLEASE DO NOT QUOTE!

I sincerely thank Barry Smith, whose insights have greatly benefited this research. All remaining errors in the paper are my own. 2 The research was undertaken while I was Research Fellow at The Conference Board, New York, and I am grateful to Bart van Ark, Ataman Ozyildirim and June Shelp for comments and support.
1

Who are the Indian Middle Class? A Mixture Model of Class Membership Based on Durables Ownership
Sudeshna Maitra York University July, 2007 ABSTRACT The size and consumption habits of the Indian middle class have evoked considerable interest in the media in the past two decades. Yet the definition of the middle class has been nebulous at best. I propose the use of a mixture model of class membership to identify and estimate the size of the lower, middle and upper classes in urban India, based on their distinct durables ownership patterns. Estimates using NSS data (55th Round, 1999-00) suggest that the urban middle class in India constitutes approximately 62% of urban households (which implies about 17% of all households) with mean ownership of 3 durable goods (out of 12). I also estimate the probability that each household in the sample belongs to a particular class and based on this information, back out some class-specific socioeconomic characteristics. The estimates suggest a larger urban middle class and lower class-defining income cutoffs than found (or used) in previous studies.

Keywords: middle class, durables ownership, EM algorithm JEL classifications: O15, I30, O10, O18

PRELIMINARY DRAFT, PLEASE DO NOT QUOTE!

1. Introduction
India’ growth achievements since the 1990s have put the living standards of Ins dians under global scrutiny. While the economic literature has primarily focussed on poverty and inequality (see Deaton and Kozel (2005) for a review), the fortunes of the ‘ new Indian middle class’have received substantial attention in the media and in business journals, as their earning potential and spending habits have important implications for the global economy. Yet there have been surprisingly few attempts to de…ne and identify the middle class in a rigorous manner. This paper seeks to address this gap in the literature by proposing a method to do so. Who are the Indian middle class? A broad de…nition – re‡ ected in most references to the middle class – places these households ‘ between’ the poor and the extremely rich. This potentially encompasses a very large and varied group of individuals, but the Indian middle class has been typically perceived to be an educated section of urban society employed in or seeking white collar jobs (Bardhan (1984), Sridharan (2004)). The size and characteristics of the Indian middle class deserve attention for several reasons. India possesses a sixth of the world’ population, and hence its s middle class constitutes a signi…cant portion of the global workforce as well as

a substantial market for …nal products. Second, the Indian middle class seems ideally placed to partake of the direct trickle-down bene…ts of high growth and to respond to economic incentives in a way that would make the growth sustainable (Sridharan (2004)). Finally, the growth and consumption habits of the middle class serve as a useful metric of how living standards in India are changing. Hence, it seems essential to develop a rigorous method for de…ning and identifying the Indian middle class. Prior studies that have attempted to analyze the middle class in India (Sridharan (2004), NCAER (2005), Ablett et al (2007), IBEF (2005)) have …rst imposed income cuto¤s for the di¤erent classes, and then proceeded to outline the characteristics (including consumption of durables) of the groups thus formed. Such an approach involves the use of several implicit assumptions –about who the di¤erent classes are and what their income levels must be – to which the results are extremely sensitive. In this paper, I propose the use of a mixture model to model the distribution of durables ownership in urban India. The mixture model yields a class structure and membership probabilities which can be used to determine who constitutes the middle class. The appeal of a mixture model lies in the fact that there are no external assumptions about who constitutes the classes, apart from the fact that 2

the classes are di¤erent. Since studies on the middle class seem to broadly agree that income or consumption is the most important basis for distinguishing between the classes, I de…ne the classes by an aspect of their consumption behaviour, viz. ownership of durable goods. The EM procedure then allows an estimation of the size and characteristics of the component classes in the population by identifying their distinct ownership patterns of durables. The unique solution generated by this approach provides an arguably more robust identi…cation of the classes than has been obtained thus far. The data comes from the 55th Round of the Indian NSS (1999-00). Durable ownership has featured prominently in discussions of living standards and the middle class (NCAER (2005), Ablett et al (2007), IBEF (2005)) in India. Hence I use data on durable ownership to de…ne and identify the classes. I focus on the total of 12 durable items –5 recreational goods (e.g. tape players), 4 household goods (e.g. refrigerators) and 3 transport goods (e.g. cars) – that a household may own at the time of interview. Since we are primarily interested in the middle class, which is largely perceived to be an urban phenomenon, I focus on the urban sub-sample of the NSS. However, the analysis may easily be extended to include the rural sub-sample as well. I …nd lower, middle and upper class households to constitute 20%, 62% and 3

18% of urban households, respectively. This implies an urban middle class of approximately 17% of the entire population, given that 28% of all Indian households were urban (2001 census, Indiastat). The mean number of goods owned by households in these classes are, respectively, 0:3, 3 and 6:3. Small standard errors of estimates support the existence of three classes with distinct ownership patterns of durables. The empirical approach involves maximum likelihood estimation. Maximum likelihood mixture models provide challenges in terms of parameter estimation and hypothesis testing. Here I use the Expectations Maximization (EM) algorithm for likelihood maximization (McLachlan and Krishnan (1996), Dempster et al (1977), Hastie et al (2001)). I provide a preview of the method in the next few paragraphs; Section 2 provides a detailed description of the model and methodology. I postulate the existence of three classes – lower, middle and upper – in a Three-Component Mixture Model framework, and focus on the total number of durable goods that a household owns at the time of interview. The objective is to estimate the population shares and durable-ownership density functions of the three component classes such that the likelihood of picking the sample is maximized. The likelihood is maximized using the EM (Expectations Maximization) al4

gorithm (McLachlan and Krishnan (1996), Dempster et al (1977), Hastie et al (2001)). The EM algorithm consists of 2 steps – the E step and the M step – which are iterated till convergence is obtained. Suppose that each household in the sample belongs to one of the three classes, represented by the dummy variables ( 1 ;
2)

(

ij

= 1 if household j belongs to class i, 0 otherwise). Since class
2)

memberships ( 1 ;

are unknown, I estimate, for each household, the expected

value of membership to each class conditional on the observed data on durable ownership. The conditional expectation is simply the probability that the household belongs to each class (since class membership can take values 0 or 1). This is the E (‘ Expectations’ step of the algorithm. To perform this step, I begin with ) initial guesses for the parameters of the class-speci…c densities. The conditional expectation of class membership is substituted for the latent class membership in the likelihood function which is then maximized to obtain estimates of class shares in the population and the density parameters. This is the M (‘ Maximization’ ) step of the EM. The E step is repeated with the values obtained in the M step and the EM iteration continues till convergence is obtained. The likelihood of a sample based upon a mixture model is very complex and traditional numerical optimization techniques such as Newton-Raphson break down. The EM optimum coincides with the likelihood optimum but is reached (somewhat slowly) using 5

iterated E and M steps. How do the mixture model estimates compare with existing estimates of the Indian middle class? The mixture estimates suggest larger middle and upper classes than are found by Sridharan (2004), Ablett et al (2007) and the NCAER and IBEF studies. Sridharan’ (2004) estimate of the middle class is between s 13% and 47% of urban households in 1998-99, depending on the breadth of his de…nition of middle class. Although these …gures are considerably less than the mixture estimate of 62% (of urban households), the numbers are hard to compare for two reasons. First, Sridharan has followed the NCAER approach and de…ned the classes by arbitrarily setting income cuto¤s. Second, each of his de…nitions of middle class includes the ‘ High’ income category1 and excludes the ‘ LowerMiddle’income category. Including the ‘ Lower-Middle’group and excluding the ‘ High’group in the de…nition of middle class, yields an urban-share estimate of 68:5% (using Sridharan’ estimates), which is much closer to 62%. This exercise s demonstrates the ambiguity that has traditionally dominated the identi…cation of the middle class, and recommends the new method presented here for its intuitive approach to the issue. Das (2001) makes a reference to the urban middle class as constituting 20% of
1

This is the highest income category in the analysis (Sridharan (2004).

6

the Indian population. While it is not clear how this …gure has been arrived at, it is nevertheless close to the EM estimate of 17% (of total households). The rest of the paper is organized as follows. The model is described in detail in Section 2. Section 3 presents results and Section 4 concludes the paper.

2. Methodology
2.1. Data The data used in the analysis comes from the urban sub-sample of the 55th Round of the Indian NSS (1999-00). The 48; 924 households in the sample are asked a battery of questions about their consumption habits and expenditures. For a list of 22 durable items, they are asked to report how many pieces of each good are in use at the time of the interview. I focus on 12 of these durable goods. These comprise 5 recreational goods (record player/gramophone, radio, television, VCR/VCP, tape/CD player), 4 household goods (electric fan, air conditioner, washing machine, refrigerator) and 3 transport goods (bicycle, motor bike/ scooter, motor car/ jeep)2 . For each of the 12 durables, I de…ne ‘ ownership’as an indicator that
The 10 items that have been left out are household furniture/ furnishings, sewing machine, stove and pressure cooker/ pan. These are omitted on account of being necessary items that may not be indicative of a- uence.
2

7

a household owns at least one piece of the durable at the time of interview. The variable of interest is the total number of durable goods (of the 12) –say, Y – that a household ‘ owns’ (by the de…nition above) at the time of interview. Figure A presents the distribution of Y in the sample. Table A presents summary statistics for the ownership variables. The bimodal distribution of Y in Figure A, along with positive skewness, suggest that a mixture model may be appropriate. A mixture model is one in which the observed density of Y is a weighted sum of densities of individual groups in the population. The goal is, therefore, to identify three distinct groups in the population such that their individual ownership densities or consumption patterns can, in combination, explain a distribution like that in Figure A3 . This is the idea that will be exploited in the Three-Component Mixture Model, estimated by an EM algorithm. The following subsections describe the model in detail.

2.2. The Three-Component Mixture Model Consider 12 durable goods and let Y represent the total number of these goods that a household owns at the time of interview, Y 2 f0; 1; 2 : : : 12g. Households can belong to one of three classes –1, 2 or 3 –which are de…ned by the pattern
The results of …tting two –instead of three –classes to the data are presented in Table B. A better …t is obtained with three classes (see Section 3).
3

8

of durable ownership of members. Assume that a household owns each good with a …xed probability (pi ), which depends on the class (i = 1; 2 or 3) to which it belongs. Assume also that each good is obtained independently. Hence the total number of goods owned by a class-i household follows a binomial distribution with parameters 12 and pi 4 . The probability of obtaining an observation y in the sample is given by:

P (y;

1;

2 ; p1 ; p2 ; p3 )

=

1 1 (y; p1 )

+

2 2 (y; p2 )

+ (1

1

2 ) 3 (y; p3 )

(1)

where i (y; pi )

i

represents the proportion of class i households in the population and

represents the (binomial) probability that the observation y comes from

a class-i household. This is a Three-Component Mixture Model. The likelihood function of the model described above can be written as
N Y L(y; ; p) = [ j=1

1 1 (yj ; p1 )

+

2 2 (yj ; p2 )

+ (1

1

2 ) 3 (yj ; p3 )]

where subscript j denotes the household, j = 1; 2; :::; N . The log likelihood func4 Allowing dependence in the ownership of di¤erent goods would necessitate several additional assumptions on the nature of dependence. Derivation of the density functions i in these cases becomes very complex.

9

tion is then:
N X j=1

log L(y; ; p) =

log [

1 1 (yj ; p1 )

+

2 2 (yj ; p2 )

+ (1

1

2 ) 3 (yj ; p3 )]

(2)

It is hard to obtain an analytical expression for the estimate of the parameters in (2). The EM algorithm is a tool used to simplify di¢ cult maximum likelihood problems such as the above (McLachlan and Krishnan (1996), Dempster et al (1977), Hastie et al (2001)) and is described in Section 2.3. The importance of the EM algorithm lies in its ability to …nd a path to the maximum likelihood point estimates where traditional numerical techniques typically fail.

2.3. Implementation of the Expectations Maximization (EM) Algorithm Suppose that each household belongs to a particular class and let the dummy variables ( 1 ;
2)

represent the class membership of households, i.e.

1j

= 1 if household j belongs to class 1 = 0; otherwise

2j

= 1 if household j belongs to class 2 = 0; otherwise

10

Then the likelihood and log-likelihood functions may be written as
N Y f j=1

LEM (y; ; p) =

1 1 (yj ; p1 )g

1j

f

2 2 (yj ; p2 )g

2j

f(1

1

(1 2 ) 3 (yj ; p3 )g

1j

2j )

log LEM (y; ; p) =

N X j=1

[

1j

log f
1j

1 1 (yj ; p1 )g 2j ) logf(1

+
1

2j

logf

2 2 (yj ; p2 )g

(3)

+(1

2 ) 3 (yj ; p3 )g]

It would be easy to …nd analytical expressions for parameter estimates from (3), if class memberships ( 1 ;
2)

were known. Since class memberships are unknown,
2)

the EM algorithm computes the expected values of ( 1 ;

conditional on the data,

plugs these into (3) and computes the maximands. The procedure is iterated till convergence is obtained. The steps involved are outlined below (McLachlan and Krishnan (1996), Dempster et al (1977), Hastie et al (2001)).

The EM Algorithm for a Three-Component Mixture Model 1. Start with initial guesses for the parameters, (
(0) 1 ; (0) (0) (0) (0) 2 ; p1 ; p2 ; p3 ).

2. Expectation (E) step: at the k th step, compute, as follows, the expected values (bi ) of class membership, conditional on the data (y1 ; y2 ; :::; yN ). 11
(k)

that a household belongs to class i, conditional on the data.

Since class memberships are binary, bi

(k)

is also the estimated probability

bij

(k)

= E( =

ij =(y1 ; y2 ; :::; yN ;

(k 1) ; 1

(k 1) (k 1) (k 1) (k 1) ; p1 ; p2 ; p3 ) 2

(4)
(k 1) (k 1) ) 3 (yj ; p3 ) 2

(k 1) (k 1) ) 1 (yj ; p1 1

+

(k 1) (k 1) ) i (yj ; pi i (k 1) (k 1) ) + (1 2 (yj ; p2 2

(k 1) 1

i = 1; 2; 3. 3. Maximization (M ) step: at the k th step, compute the parameters as follows. These are the maximands of the EM -log-likelihood function in (3), when ( 1;
2)

are replaced by their expected values conditional on the data. 1X = N j=1
N

(k) bi

(k) ij

(5)

pi b i = 1; 2; 3.

(k)

=

1 j=1 [ N 12 P

N P

(k) j yj

]
(k) j

j=1

4. Iterate steps 2 and 3 (the E and M steps) till convergence is obtained.

As output, the EM algorithm yields the following estimates: 12

1. bi : estimates of the proportion of class-i households in the population; i = 1; 2; 3 2. pi : estimates of the probability with which a class-i household owns a b durable good, i = 1; 2; 3 3. bij : the probability with each each household j belongs to class i, i = 1; 2; 3; j = 1; 2; :::; N The ownership probabilities pi and the corresponding class-speci…c densities b

identifying the distinct ownership patterns of the di¤erent classes. Moreover, the estimates of class shares bi tells us the size of the urban middle class in India.

b i (y; pi )

answer our motivating question –who are the Indian middle class? –by

pi , enable an assignment of each household into a particular class. This allows a b descriptive analysis of other class-speci…c household characteristics such as average per capita monthly expenditure, education of the household head, household type by employment and so on. The next section presents the results.

Finally, the estimated probabilities of class membership, bij , along with bi and

13

3. Results
3.1. EM Estimates The estimates produced by the EM algorithm are presented in Table 1 and Figures 1 to 3. The numbers in column (2) of Table 1 represent the population share of each class, bi . The middle class is estimated to constitute 62% of urban households.

This is roughly equivalent to 17% of the total population, since urban households accounted for about 28% of all Indian households in 2001 (2001 census, Indiastat). The lower and upper classes are found to constitute 20% and 18% of urban households, respectively. Asymptotic standard errors (obtained from the information matrix) are small, supporting the existence of three classes in the population. Column (3) reports estimates of the probability parameter pi for each class b

i = L; M; U . Lower class households are found to own a good with 3% probability while middle and upper class households own a good with probabilities of 25% and 52% respectively. Small standard errors support three distinct patterns of durable consumption behaviour5 . An alternative interpretation of the numbers in Column (3) is that 52%, 25%
5 The estimates (standard errors) of the di¤erences are as follows: pL pU = 0:5 (0:004); pL b b b pM = 0:23 (0:002) and pU pM = 0:27 (0:003) (L t Lower; M t M iddle; U t U pper). b b b

14

and 3% of households in the upper, middle and lower classes, respectively, own a representative durable good. This interpretation allows an extrapolation of the size of the urban market for a representative durable good, as it speci…es what proportion of the three classes will consume the good when it is introduced. The mean number of durable goods (out of 12) owned by class-i households is simply 12pi (the mean of the binomial distribution for class i). These estimates are reported in Column (4) of Table 1. The lower, middle and upper classes are found to own, on average, 0:3, 3 and 6:3 goods, respectively. Figure 1 plots the binomial density functions i at the estimated parameters

pi (i = 1; 2; 3). Classes 1, 2 and 3 are the lower, upper and middle classes, reb middle and upper classes peak at 3 and 6 durable goods, respectively.

spectively. The density of the lower class peaks at 0 durables, whereas that of the

Figure 2 plots the actual relative frequency of observations (Y ) in the data along with the predicted values. The …gure demonstrates a very good …t to the data6 . Figure 3 plots the probabilities bi that a household belongs to di¤erent classes

i (= 1; 2; 3). For example, households with low values of Y are most likely to
As an analytical exercise, a Two-Component (two classes) Mixture Model was …tted to the data by EM. The results are presented in Table B. The …t is clearly better in the ThreeComponent Model.
6

15

belong to the lower class (class 1) whereas those with the highest values of Y are certain to belong to the upper class (class 2).

3.2. Class Characteristics: A Descriptive Analysis Using the mixture estimates of bi and b i (y; pi ) it is possible to estimate the number

of observations of each value of Y that belongs to each class. Based on this computation, I randomly assign households to classes. As an example, suppose that there are 100 observations for Y = 0 and that the EM estimates predict that 60% of these belong to class 1, 10% to class 2 and 30% to class 3. I then randomly assign 60 of the 100 households with Y = 0 to class 1, 10 to class 2 and 30 to class 3. Likewise for each other value of Y . Assigning a class to each households allows a descriptive analysis of the average characteristics of each class. I focus on the durables ownership patterns for speci…c goods as well as a host of socioeconomic characteristics. The results are presented in Tables 2-3 and Figures 4-11 and discussed below. Tables 2(a) –(b) and Figures 4(a) –(b) demonstrate the durables consumption patterns of households belonging to the three classes (assigned by the procedure described above). Recreational and household goods appear to be more commonly

16

owned by all classes than are transport goods7 . Of these, electric fans and televisions are most popular among the top two classes, whereas fans and bicycles are most popular among the lower class. Table 3 reports the per capita monthly expenditures of households in each assigned class. These numbers suggest lower income cuto¤s for the di¤erent classes than has been used in prior studies. As an illustration, consider the following approximate calculation. At a household savings rate of 28% (Ablett et al (2007)) and the mean class-speci…c household sizes in the sample (see Table 3), median annual household incomes are Rs. 41354:16, 58420 and 104465 for the lower, middle and upper classes respectively. The NCAER study places the ‘ middle class’in the annual-household-income range of Rs. 200; 000 1; 000; 000 in 2001-

02. The class immediately below the middle class –viz. ‘ aspirers’–are also placed in an income range that appears too high, viz. Rs:90; 000 200; 000, annually8 .

Figure 5 plots the education levels of the household head, by class. The lower class has the highest component of illiterate heads (32%) whereas the upper class has the highest component of heads with a graduate degree (38%). Middle class
This could be partly attributable to the fact that, among the 12 goods considered, there are more recreational and household goods (5 and 4, respectively) than there are transport goods (3). 8 The NCAER study divides households into 4 classes: Deprived, Aspirers, Middle Class and Rich.
7

17

household heads are most likely to have secondary education (18%) although graduates comprise a comparable component as well (15%). A large proportion (18%) of middle class heads appear to be illiterate. Despite the mean proportion of literate middle-class-household members being 77% (see Table 3), this …nding is somewhat surprising given the perception of the middle class as white-collar workers. However, the phenomenon would be consistent with an environment of social mobility characterized by a large in‡ of lower class members into the ux middle class. Repeating the EM analysis for other rounds of the NSS could provide further insight into this phenomenon. Figure 6 presents a plot of household type by employment. Being urban residents, the proportion of households who are self-employed in agriculture is negligible. The largest component of households in each class are wage/salary earners. This fact is also mirrored in Figure 7 which plots sources of household income. Over 50% of households in each class have reported income in the past year from wages and salaries. Income from non-agricultural enterprises is reported by more than 30% of households in each class. A large proportion of households also report owning land. Income from interests and dividends is the third most highly reported source of income by the top two classes – 15% and 7% of upper and middle class households, respectively. For the lower class, income from ‘ other’ 18

sources is reported by considerably more households (12%) than is income from interests and dividends (2%). Figures 8 and 9 present a summary of the primary sources of energy used in cooking and lighting. LPG is most commonly used for cooking among the top two classes; …rewood and chips are most common among lower class households. For lighting, electricity is most common in all classes, although 25% of lower class households use kerosene as the primary source of energy. Finally, Figures 10 and 11 provide a summary of class composition by religion and social class. Hinduism is the religion of the majority in India, so it is not a surprise that Hindus constitute the largest component of all classes. However, Muslims and Christians form a larger component of the lower class (18% and 11% respectively) than the middle and upper classes (15% and 4% of the middle class while 10% and 4% of the upper class are Muslim and Christian, respectively). Likewise, Scheduled Castes and Tribes form a larger component of the lower than the middle and upper classes.

4. Summary and Conclusion
I propose the use of a mixture model as a robust method for identifying and estimating the size of the urban middle class in India, when classes are de…ned by 19

their distinct patterns of durable ownership. Using a Three-Component Mixture Model and data on the total number of durables owned by households (NSS, 55th Round, 1999-00), I obtain estimates of the urban-population shares of the three classes (lower, middle and upper) as well as the probability that a household belonging to each class will own a durable good. The estimates are precisely estimated with small standard errors, supporting the existence of three distinct durables ownership patterns –hence, three distinct classes –in the Indian urban population in 1999-2000. The magnitudes of the share estimates indicate a larger urban middle and upper class (62% and 18%, respectively) than were found in previous studies (Sridharan (2004), NCAER (2005), Ablett et al (2007), IBEF (2005)). However, these previous studies have relied on several assumptions about who constitutes the classes, to which their results appear to be sensitive. The EM approach used here is free from such arbitrary assumptions and allows ‘ data to decide’who the constitutes the three classes based on their distinct durable ownership patterns. The solution obtained is unique. This recommends the usage of an EM algorithm to identify the classes and investigate the characteristics of component households.

20

References
[1] Ablett, Gold: sey Jonathan, The Global Rise Baijal, of Aadarsh et al. "The Market." Bird of

India’ s Report,

Consumer May 2007,

McKinat:

Institute

(available

http://www.mckinsey.com/mgi/publications/india_consumer_market/index.asp ) [2] Bardhan, Pranab. "The Political Economy of Development in India." Blackwell Publishing, 1984 [3] Das, Gurcharan. "India’ Growing Middle Class." The Globalist, November s 5, 2001 [4] Deaton, Angus and Kozel, Valerie. "Data and Dogma: The Great Indian Poverty Debate." RPDS Working Paper, Princeton University, January 2005 [5] Dempster, A. P., Laird, N. M. and Rubin, D. B. "Maximum Likelihood from Incomplete Data via the EM Algorithm." Journal of the Royal Statisticial Society. Series B (Methodological), 39 (1), 1977, pp. 1-38

21

[6] Hastie, Trevor, Tibshirani, Robert and Friedman, Jerome. "Elements of Statistical Learning: Data Mining, Inference and Prediction." Springer Series in Statistics, 2001 [7] India India’ s Brand Middle Equity Class Foundation Dream (IBEF). Shape." "Economic 2005 Indicators: at:

Takes

(available

http://www.ibef.org/artdisplay.aspx?cat_id=391&art_id=5788 ) [8] Indiastat (available at: http://www.indiastat.com) [9] McLachlan, Geo¤rey J. and Krishnan, Thriyambakam. "The EM Algorithm and Extensions." Wiley Series in Probability and Statistics, 1996 [10] NCAER Report. "The Great Indian Market." 2005 (Preview slides available at: http://www.ncaer.org/downloads/PPT/TheGreatIndianMarket.pdf ) [11] Sridharan, E. "The Growth and Sectoral Composition of India’ Middle Class: s Its Impact on the Politics of Economic Liberalization." India Review, 3 (4), pp. 405-428, 2004

22

Table A: Summary Statistics, Urban Sub-sample, NSS 1999-00, N = 48,924 households Variable Total number of goods 'owned' (Y ) If household 'owns': Record Player/ Gramophone Radio Television VCR/ VCP Tape/ CD Player Electric Fan Air Conditioner Washing Machine Refrigerator Bicycle Motor bike/ Scooter Motor car/ Jeep 'Owns' at least one durable good 'Owns' at least one recreational good 'Owns' at least one household good 'Owns' at least one transport good Total number of recreational goods 'owned' Total number of household goods 'owned' Total number of transport goods 'owned' 0.02 0.36 0.60 0.05 0.30 0.67 0.12 0.10 0.25 0.37 0.20 0.03 0.83 0.72 0.69 0.50 1.32 1.13 0.60 0.13 0.48 0.49 0.21 0.46 0.47 0.32 0.30 0.43 0.48 0.40 0.17 0.37 0.45 0.46 0.50 1.08 1.08 0.68 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 5 4 3 Mean 3.06 Std. Dev. Min. Max. 2.33 0 12 Notes Variable Used in EM Estimation '1' if household owns at least one piece of the item Recreational Goods

Household Goods

Transport Goods

Per Capita Monthly Household Expenditure 1018.73 1535.32
Fig. A: Distribution of Y
20

17 205987 48, 921 obs.

0 0

Relative Frequency (%) 5 10 15

5 10 Total No. of Goods Owned (Y)

15

23

Table 1: Lower, Middle and Upper Classes in the Urban Sub-sample, Indian NSS, 55th Round (1999-00), N = 48,924 households

EM Estimates (Std. Error) (1) Category (Class) (2) Share of Urban Population 0.2034 (0.005) 0.6161 (0.005) 0.1804 (0.006) (3) (4) Probability of Mean No. of Owning a Goods Good (of 12)* 0.0257 (0.002) 0.251 (0.003) 0.5249 (0.004) 0.3084 (0.007) 3.012 (0.01) 6.2988 (0.014)

Lower (L ) Middle (M ) Upper (U )

* The 12 goods include 5 recreational goods (record player, radio, tv, vcr/vcp, tape/cd player), 4 household goods (electric fan, a/c, washer, fridge) and 3 trasnport goods (bicycle, motor bike/scooter, motor car/ jeep)

24

Table 2(a): Ownership by Durable Categories by Class in the Urban Sub-sample, NSS 1999-00, N = 48, 924 households Category Mean No. of Goods Owned by Households (Class) All (12 items) Lower (L) Middle (M) Upper (U) 0.31 3.01 6.30 Recreation Household Transport Goods Goods Goods (5 items) (4 items) (3 items) 0.12 1.37 2.51 0.11 1.06 2.52 0.07 0.58 1.27

Proportion of Households Owning At Least one Good in the Relevant Category, by Class All (12 items) 0.27 0.97 1.00 Recreation Goods (5 items) 0.12 0.85 1.00 Household Goods (4 items) 0.11 0.79 0.99 Transport Goods (3 items) 0.07 0.53 0.87

Table 2(b): Ownership of Individual Durable Goods by Class in the Urban Sub-sample, NSS 1999-00, N = 48, 924 households Proportion of Households Owning the Relevant Good, by Class Category (Class) Record Player Lower (L) Middle (M) Upper (U) 0.00 0.01 0.05 Recreational Goods Radio 0.07 0.39 0.58 TV 0.04 0.68 0.97 VCR/ VCP 0.00 0.02 0.19 Tape/ CD Player 0.01 0.27 0.71 Electric Fan 0.11 0.77 0.97 Household Goods Washing Air Cond. Machine 0.00 0.07 0.41 0.00 0.04 0.39 Fridge 0.00 0.18 0.75 Transport Goods Bicycle 0.07 0.43 0.53 Motor Bike/ Scooter 0.00 0.14 0.60 Motor Car/ Jeep 0.00 0.01 0.14

25

Table 3: Household Characteristics, by Class, in the Urban Sub-sample, NSS, 55th Round (1999-00) Category (Class) Mean Lower (L) Middle (M) Upper (U) Std. Dev.

Per Capita Monthly Household Expenditure Percentiles Min. Max. 25 423 532 842 50 625 762 1229 75 981 1140 1777 90 1421 1663 99 2791.43 3485

Other Household Characteristics Avg. No. of Proportion of Household Meals Per Day Literate Household Size Per Person Members (Mean) (Mean) (Mean) 2.34 2.38 2.41 0.64 0.77 0.88 3.97 4.65 5.12

791.26 859.109 961.785 1772.39 1469.57 1109.97

17 49 224

50528 205987 35612

2490.6 5390.08

Addendum: Percentiles of Per Capita Monthly Expenditure in the Entire Sample, N = 48, 921 Percentile Value 10 392 20 490 30 584 40 686 50 801 60 940 70 1120 80 1377 90 1815 99 3799.56

26

Fig. 1: EM-Estimated 'Density' Function of Y , by Class
0.8

0.7

0.6

0.5 phi1 (Lower) phi2 (Upper) phi3 (Middle) Rel. Freq of Obs. Probability

0.4

0.3

0.2

0.1

0 0 2 4 6 8 10 12 14 Total No. of Goods Owned (Y )

27

Fig. 2: Actual vs. EM-Predicted Distribution of Y
0.18

0.16

0.14

0.12

Probability

0.1 Rel. Freq of Obs. Predicted 0.08

0.06

0.04

0.02

0 0 2 4 6 8 10 12 14 Total No. of Goods Owned (Y )

28

Fig. 3: EM-Estimated Probability of Belonging to Each Class
1.2

1

0.8

Probability

0.6

Probability of Being Upper Probability of Being Middle Probability of Being Lower

0.4

0.2

0 0 2 4 6 8 10 12 14 Total Number of Goods Owned (Y)

29

Fig. 4(a): Ownership by Durable Categories by Class, Urban Sub-sample, NSS 1999-00

Transport Goods

Household Goods Upper (U) Middle (M) Lower (L) Recreational Goods

All

0.00

0.20

0.40

0.60

0.80

1.00

1.20

Proportion of Households Who Own the Good

Fig. 4(b): Ownership of Individual Goods by Class, Urban Sub-sample, NSS 1999-00

Motor Car/ Jeep Motor Bike/ Scooter Bicycle Fridge Washing Machine Air Cond. Electric Fan Tape/ CD Player VCR/ VCP TV Radio Record Player 0.00 0.20 0.40 0.60 0.80 1.00 1.20 Upper (U) Middle (M) Lower (L)

Proportion of Households Who Own the Good

30

Percentage of Class-i Households in Category (i = Lower, Middle, Upper) 10 15 20 25 30 35 0 Not literate Lit, NFEC/AEC Lit, TLC Lit, others Lit, below primary Primary Middle 5

Fig. 5: Level of Education, by Class

Category

31
Secondary Higher Secondary Grad. in agriculture Grad. In engineering Grad, in medicine Grad, in other

Lower Class Middle Class Upper Class

Percentage of Class-i Households saying 'Yes' (i = Lower, Middle, Upper) Percentage of Class-i Households in Category (i = Lower, Middle, Upper)
10 20 30 40 50 60 0 10 20 30 40 50 60 70 80 0

Owns land SelfEmployed

Received income from cultivation

Received income from fishing Regular Wg./Salary Earner

Received income from wage/salaried employment

Received income from nonagricultural enterprises Casual Labour
Category

Received income from pension

Fig. 6: Type of Employment, by Class

Fig. 7: Land Ownership & Source of Income, by Class

32
Upper Class Lower Class Middle Class

Received income from rent

SelfEmployed in Agriculture

Received income from remittances

Received income from interests and dividends

Others

Received income from other sources

Upper Class

Lower Class

Middle Class

Percentage of Class-i Households in Category (i = Lower, Middle, Upper) 100 120 100 20 40 60 80 10 20 30 40 50 60 70 80 90 0

Percentage of Class-i Households in Category (i = Lower, Middle, Upper)

0

Electricity Firewood and chips LPG Kerosene Gobar gas Dung cake Charcoal Electricity Others No cooking arrangement

Coke/ Coal

Kerosene

Other oil

Category Category

Gas

Fig. 9: Primary Source of Energy Used for Lighting, by Class

Fig. 8: Primary Source of Energy Used for Cooking, by Class

33
Upper Class Lower Class Middle Class

Candle

Others

No lighting arrangement

Upper Class

Lower Class

Middle Class

Fig. 10: Religion, by Class
90 80 70 60 50 40 30 20 10 0 Hindu Christian Jain Category Zoroastrian Lower Class Middle Class Upper Class

Percentage of Class-i Households in Category (i = Lower, Middle, Upper)

Fig. 11: Social Group, by Class
80

Percentage of Class-i Households in Category (i = Lower, Middle, Upper)

70

60

50 Lower Class 40 Middle Class Upper Class 30

20

10

0 Scheduled Tribe Scheduled Caste Other Backward Classes Others

Category

34

Table B: EM Results for a Two-Component Mixture Model

Density by Class
0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 0 5 Y 10 15

Probability

Rel. Freq. of Obs. phi1 (Lower) phi2 (Middle/Upper)

Actual vs. Predicted Distribution of Y
0.2 0.15 0.1 0.05 0 0 5 Y 10 15 T

Probability

Relative Freq. of Obs. Predicted

Two-Components Model: EM Estimates Prob. of Pop. Mean No. Class Owning a Share of Gds. Gd. Lower Middle/ Upper 0.43 0.57 0.09 0.38 1.08 4.52

35