You are on page 1of 60

Models of Growth and Inequality: An Evaluation

Hyeok Jeong and Robert M. Townsend January 2002 (In Progress)

Abstract This paper evaluates the diverse implications of models of growth and inequality for data by studying two benchmark models that have explicit micro underpinnings. Both models generate macro aspects of growth and inequality from the micro features of wealth-constrained self-selection. Using the Thai Socio-Economic Survey data, we estimate initial wealth distribution and key parameters that best t the micro foundations of the models and observe whether the model economies at those estimates can explain the macro aspects of growth and inequality. Both models capture the main aspects of growth and inequality dynamics but overdo their built-in heterogeneity while missing other diversity in the data. JEL Classication: C52, D31, O41 Keywords: Model Evaluation, Growth and Inequality, Wealth-Constrained Self-Selection

Introduction

Our purpose is to understand growth and inequality as implied by economic models that are explicit about micro underpinnings and impediments to trade. Our overall strategy is one of structural estimation and out-ofsample and post-sample testing. In the rst, estimation part, we choose the model parameters that maximize the likelihood in socio-economic survey data of the recurrent or stationary part of the decision problem of households in two distinct models. In the occupation choice model of Lloyd-Ellis and Bernhardt (2000), the LEB model, households face wealth constraints in nancing enterprise. In the nancial deepening model of Greenwood and Jovanovic (1990), the GJ model, households face wealth constraints in deciding upon costly entry to the formal nancial system itself. In the second, testing part of our strategy, we simulate each of these models at their estimated parameter values, and then compare the predicted shape and evolution of the income distributions of each of the two models to the actual shape and evolution of income distributions in socioeconomic survey data. We apply our methods to Thailand, a country that grew rapidly for the 1976-1996 period, but with increasing (and then decreasing) inequality. We know from Jeong (2000) and from his use of the Thai SocioEconomic Survey (SES) data that growth and inequality are explained in part, in the sense of an accounting decomposition, by occupational shifts from farmer to wage labor and non-farm self-employment and by increasing access to formal nancial intermediaries. We take these empirical ndings and these two relevant preexisting
Department of Economics, University of Southern California; Department of Economics, University of Chicago. We thank Xavier Gine and Kenichi Ueda for their help with the simulation programs used in this work. We also appreciate the helpful comments from Lars Hansen, Hugo Hopenhayn, Roger Moon, and the participants of the conference in Buenos Aires honoring Rolf Mantel and the Money and Banking workshop at the University of Chicago. Funding for the research was provided by the National Science Foundation and the National Institute of Health. Corresponding E-mails: hjeong@usc.edu and rtownsen@midway.uchicago.edu

models, LEB and GJ, as starting points and evaluate the ability of the two models to explain actual growth and inequality dynamics of Thailand. Much more generally, our goal is to explore and evaluate the diverse implications of theory for data. The underlying key parameters of the economic environment are estimated using micro data that best t the micro foundations of models, and then we observe whether the models at the estimated parameter values, when embedded in a theoretical dynamic context, can explain the macroeconomic aspects of growth and incomedistribution-dynamics. We are careful in our use of data in two senses. First, we use selected parts of the data, either by sub-sampling the population or by trimming the components of income, to match the data to the variables in the theory as closely as possible. Second, we separate the use of data in estimation and in testing. We use only the data of young households at initial dates in our estimation, as we match the stationary recurrent microeconomic part of the models. Then, we simulate the model economies at those parameter estimates over two decades generating the growth and inequality dynamics of the theory. We compare that prediction to subsequent data for the entire sample. We also formally test the models using the post-sample data, comparing in particular the forecasted income distribution with the actual one at the nal date.1 In general equilibrium models, the aggregate dynamics of growth and inequality are determined by various underlying factors interacting with each other. Both the LEB and GJ models have multiple sectors and the dierential pattern of capital accumulation across sectors aects the dierential levels in average sector income. These dierentials determine the incentives for shifting across sectors, and the sector composition of the economy thus changes over time. This, in turn, by changing the conditions of demand and supply, determines again the relative income gap across sectors. These underlying factors may counteract or reinforce each other, depending on the models structures and their parameter values, in determining the growth and income distribution dynamics at the aggregate levels. Thus we evaluate the models both in terms of aggregate as well as these decomposed features. Specically, we rst evaluate the models by observing whether they generate the aggregate dynamics as observed in the data. Then, we decompose the aggregate into the key driving or underlying factors in the theory, and observe whether these underlying factors have orders of magnitudes similar to the decomposed features of the actual data. In particular, compositional changes in population are the driving forces of the well-known Kuznets curve, an inverted-U shaped relationship between growth and income inequality dynamics, originally addressed in the context of industrialization in Kuznets (1955). The LEB and GJ models can deliver this Kuznets curve via compositional changes in occupation choice or in nancial participation. Indeed this is one of the principal
1 As Granger (1999) suggests, this separation of estimation from testing helps us to avoid the over-tting problem of model evaluation and is a well established procedure in many disciplines.

motivations given by the authors for their models. But each of the models provided explicit and distinct micro foundations for the so-called Kuznets dynamics observed at the macro level. From the above decomposition analysis, we evaluate whether the models generate the observed aggregate dynamics of growth and income inequality via Kuznets dynamics as they are actually observed in the data. Our intended contribution is to narrow the theoretical-empirical gap in the Kuznets curve and growthinequality literature. That is, we are deliberately between the two extremes of theory without measurement and measurement without theory. We uncover interesting features of the data while revealing the limitations of current theory. More generally, we hope that our model evaluation exercise might advance the mutual penetration of quantitative economic theory and statistical observation, as Frisch (1933) envisioned. We regard this paper as a natural rst step. Indeed we are surprised this step has not been taken before. In Section 2, we describe the two models of growth and inequality. In both models, growth and inequality dynamics at the aggregate level are linked through key self-selection features, of choice constrained by wealth at the micro level. Both models make explicit the micro underpinnings and economic environments that are supposed to generate growth and income-distribution-dynamics. This allows us to identify the sources of the growth and income-distribution-dynamics in each model economy. This guides us in turn in the choice of what variables to observe in the data, that is, according to the models. In Section 3, we take a preliminary look at these selected variables in the data to see if they behave as the models would imply. Section 4 describes how we implement more formal structural estimation of each models parameters, following the likelihood principle in Fisher (1925). We also make clear the orders of magnitudes of uncertainty around the estimated parameters, that is, standard errors, using bootstrap techniques. In Section 5, we simulate and compare the predicted paths of the models at these estimated parameter values to the actual Thai data. We compare by various means: trends of aggregate and decomposed dynamics of growth and inequality, the decomposition of accounting factors of aggregate growth and inequality change as suggested by Kuznets dynamics, and the predicted shape of the income distribution using Kolmogorov-Smirnov test statistics. These diverse and multiple ways of evaluating the models allows us to make clear the successes and failures of the models, and to oer suggestions for improved models and subsequent research.

2
2.1

Models of Constrained Self-Selection


LEB Model: Wealth-Constrained Occupation Choice

We consider as in Lloyd-Ellis and Bernhardt (2000), an extreme benchmark case: no credit market at all. The economy is populated by a continuum of family lineages of measure one evolving over discrete time t = 0, 1, 2.... Each agent is active for a period, then reproduces with one child. An agent with end-of-period wealth Wt at

date t maximizes individual preference over consumption ct , and bequest to children bt+1 as represented by the utility function u(ct , bt+1 ) subject to the budget constraint ct + bt+1 = Wt . The utility function is assumed to be homogeneous of degree one, strictly increasing, and strictly quasi-concave. There are two kinds of production technologies. One is available in a traditional sector, where everyone earns a safe subsistence return of a single consumption good. The other is available in a modern sector, where entrepreneurs hire capital kt and labor lt at each date t to produce the single consumption good according to a production function f (kt , lt ) that is strictly increasing and strictly concave. Each laborer is paid wage wt for its single unit of time at date t. The cost to capital is determined by its opportunity cost, a constant interest rate of unity tied to a backyard technology. Thus there are two occupation types in the modern sector, entrepreneurs and workers, while there is only one type in traditional sector, subsisters. There are two types of xed cost in the modern sector. The entrepreneur pays an initial setup cost xt to start up a business, and everyone in modern sector pays a xed cost of living . The setup cost represents the inverse of the innate entrepreneurial talent of each agent. It is independent of the inheritance level bt and randomly drawn from a time invariant distribution H . In the above model, an agent is distinguished by a pair of characteristics: initial inheritance bt and randomly drawn entrepreneurial talent xt , where we suppress the time subscript on these to emphasize the recurrent or stationary aspect. With a homogeneous-of-degree-one utility function, the optimal rules for consumption and bequest will be linear functions of wealth, and hence preference maximization is equivalent to end-of-period wealth maximization. Thus, given an equilibrium wage rate w, an agent of type (b, x) chooses his occupation to maximize his total wealth W : W = + b, f or subsisters f or wage earners f or entrepreneurs, (1)

= w + b ,

= (b, x, w) + b , where

(b, x, w) = max{f (k, l) wl (k + x)} s.t.


k,l

(2) (3)

0 k bx

Equation (1) suggests that there is a reservation wage level w = + below which every potential wageworker prefers to remain in subsistence sector. Likewise, if the wage rate exceeds that reservation wage, no one remains in subsistence sector. Therefore, the model implies that wage must be w = + when the subsistence sector coexists with the modern sector. Since households will at best be indierent between wage earning and susbsistence, we will focus here on the binary occupational choice, between being a wageworker or 4

being an entrepreneur. The higher is the initial wealth b, the more likely it is that an agent will be entrepreneur. A potentially ecient, low x agent may end up as a wageworker, constrained by low initial wealth b. Given inheritance b and market wage w, we can dene a marginal agent as one with entrepreneurial talent xm (b, w) who is indierent between being a worker and being an entrepreneur such that (b, xm , w) = w. If the setup cost is higher than this marginal setup cost, the household will be a worker for sure. However, with the inequality constraints 0 k b x, the setup cost x cannot exceed inheritance b, either. Therefore, given the wage w, the critical setup cost of a marginal agent with inheritance b who is willing to be an entrepreneur is given by: z (b, w) = min [b, xm (b, w)] . (4)

This is the key selection equation and we use it in later maximum likelihood estimation of the LEB model parameters. Note that choice of occupation is constrained by wealth. This LEB economy grows from two sources: capital accumulation of entrepreneurs in modern sector and occupational shifts from lower-income occupations to higher-income occupations. As wealth accumulates along the growth path, wealth constraints become less binding, i.e., the proportion of people whose wealth is less than the critical setup cost z (b, w) gets reduced and more households can set up their own businesses. This in turn changes the supply and demand conditions in the labor market and hence changes wage and also prots, and this in turn determines the income gap across occupations. These occupational shifts themselves change the aggregate shape of income distribution. All these factors aect the dynamics of income inequality and growth and are interwound each other. This accentuates the importance of the general equilibrium approach to understanding growth and income distribution dynamics.

2.2

GJ Model: Wealth-Constrained Financial Participation

Greenwood and Jovanovic (1990), hereafter denoted by GJ, model growth and the evolution of the income distribution as they relate to nancial deepening, that is, increasing participation in the nancial sector. There exists an entry cost that constrains that participation decision. Consider an economy with a continuum of agents with names on the unit interval [0, 1] . They live an innite discrete number of periods t = 0, 1, 2 . For every agent j , there are two technologies available that can convert the capital investment it at date t into a yield yt+1 at date t + 1. One technology oers a safe but relatively lower rate of return per unit capital and the other a risky rate of return t+1 + j,t+1 , where t+1 represents a common aggregate shock and j,t+1 an

idiosyncratic shock specic to agent j . The aggregate shock t+1 is governed by a time-invariant distribution function with support [ , ], and the idiosyncratic shock j,t+1 is governed by a time-invariant distribution function with support [, ] with E (j,t+1 ) = 0. Let j,t+1 = t+1 + j,t+1 be the composite technology shock, 5

and be its distribution function. GJ assume that the lower bound for the composite shock is positive, i.e., + > 0. Each agent can run one or both technologies, with his own portfolio share t for the risky one. Formally, each agent who invests it and has portfolio share t has beginning-of-period wealth kt+1 at t + 1 as kt+1 = t t+1 + (1 t ) it . and capital investment it , namely kt = ct + it . The lifetime objective is then to maximize ( ) X t E u(ct )
t=0

(5)

At the beginning of period t, an agent allocates his current disposable wealth kt into current consumption ct

subject to the sequence of resource constraints of kt = ct + it and law of motion in (5). Agents are heterogenous in their wealth level kt in each period t for two reasons: rst, the initial endowment k0 may be dierent across agents, distributed under a cumulative distribution function H0 . Second, the history of realizations of random shock {j,s }t s=0 diers across agents. There is another technology available other than physical production, namely nancial intermediation. An intermediary can run a countable number of trials for the risky technology and get a consistent test statistic of next periods return to the risky project. Then, the intermediary invests in the risky project only if this return exceeds the safe return . Furthermore, the intermediary can diversify the individual-specic idiosyncratic shocks by pooling participants resources. It can pay back a promised return r( t+1 ) per unit of capital invested at time t contingent on the realized aggregate shock t+1 . Therefore, every agent has an incentive to join the coalition of nancial intermediaries. Specically, to assure the benets of intermediation and the incentive to invest positive amount in production every period, GJ assume the following necessary conditions: E {r( )} > E { } > > 1/ . (6)

These intermediary trading arrangements are costly, as in Townsend (1978, 1983), for example. There is an initial xed cost of of incorporating each participant j into the nancial coalition and a variable cost of (1 ) in proportion to the amount of funds each agent invests in the coalition. Thus the intermediary charges a lump sum entry fee q for each participant in exchange for the rights to operate an individuals project. In GJ, it is shown that the zero-prot condition for the intermediary implies r( ) = max { , } , and q = . (7)

Given this entry fee q = , not everyone joins the intermediation coalition. Only agents whose wealth levels exceed some critical level are willing to join that formal nancial system. This is the key selection decision of 6

the model and we use it in maximum likelihood estimation of the models parameters. Note again that choice of nancial sector participation is constrained by wealth. The decision making of an agent in period t who currently is not in the intermediation coalition can be summarized by the following functional equation: (P 1) : w(kt ) = max {u (kt it ) + E max [w(kt+1 ), v (kt+1 q )]}
it ,t

(8)

subject to (5), where E is the expectation with respect to the composite shock . Here, w() is the value function for the nonparticipants in the nancial sector, and v () the value function of participants. The latter is given by another functional equation: (P 2) subject to kt+1 : v (kt ) = max {u (kt it ) + E v (kt+1 )}
it

(9) (10)

= r( )it

where E is the expectation with respect to only the aggregate shock . In (P 2), the formulation assumes that once an agent enters the intermediated sector, he will never exit; this is proved in Greenwood and Jovanovic (1990), showing that v (k) > w(k) for every xed k. In this GJ model, nancial participation increases over time as the economy grows. With an income gap between participants and non-participants due to the intermediation technology, this nancial deepening is a source of growth. This change in the composition of the population of the economy aects the shape of aggregate income distribution over time. With dierent savings behavior as a function of wealth and dierent realizations of shocks, the income levels of those outside the nancial sector are dierent over time, and hence the average income gap between participants and non-participants changes over time. This in turn aects the growth and inequality dynamics. Thus the growth and income distribution dynamics are linked through various interacting forces.

Features from Data

We use the Thai Socio-Economic Survey (SES), a nationally representative micro survey conducted by the National Statistical Oce in Thailand. Over the recent two decades, between 1976 and 1996, eight rounds of repeated cross-sections were collected: 1976, 1981, 1986, 1988, 1990, 1992, 1994, and 1996. The sampling scheme of the SES is a clustered random sample stratied by geographic regions over the entire country. The sampling unit is the household, which is dened as a group of persons who make common provision for food and other living essentials. The general criteria for household membership are common housekeeping arrangements, sharing of principal meals, common nancial arrangements for supplying basic living essentials, and recognition of one member as head. The sample size varies from 10,897 households to 25,208 households, depending on survey year. Our analysis will focus on the economically active households in the SES data. We map the variables of the above models into this micro survey data and see how the associated variables in the data evolve over time in Thailand during this 1976-1996 period. Specically, we rst examine the wealth distribution and its association with the relevant household characteristics, i.e., occupation choice and nancial participation. Then, we describe the evolution of the aggregate growth and inequality of income and their components corresponding to each model.

3.1
3.1.1

Characteristics Choice Constrained by Wealth


Wealth Distribution

In both of the LEB and GJ models, selection, occupational choice in the LEB model and nancial participation in the GJ model, is constrained by wealth. Thus we need a variable for wealth to see if the Thai economy behaves as both models suggest. The SES does not record direct wealth estimates. However, it records the ownership of various household assets, the rental value of owned house, and the rental income from land, lodging etc. Using these pieces of information, we construct a proxy for wealth. We rst use the information on ownership of seventeen household assets A1 , , A17 .2 Principal component analysis provides us with an eigenvector (1 , , 17 ), where the variance of the linear combination of component asset variables weighted by this eigenvector, i.e., A = 1 A1 + + 17 A17 , is maximized among all possible linear combinations subject to the covariance structure of the component asset-ownership variables.3 That is, we nd a single variable, which best summarizes the variations of these seventeen asset variables, as a proxy for the latent wealth underlying them. This variable is not in monetary units though. To convert this physical asset index into monetary wealth,
2 We choose household assets recorded in common in intial and nal years of 1976, 1981 and 1996. The list includes phone, sofa, bed, stove, refrigerator, iron, electric pot, radio, tv, motor boat, car, motorcycle, sewing machine, having bedrooms in house, access to private water, equipment to use gasoline to cook, and access to electricity. 3 The values of the eigen vectors from the principal component analysis in both years of 1976 and 1996 are reported in the Appendix.

we need price data for all assets, which, unfortunately, we do not have. The only available price information related to household wealth is the monthly rental value wH of owned house.4 We thus assume that the rental price of the house can represent the average rental value of the above seventeen assets and multiply the above proxy A by the rental price of the house. This give us the ow value of physical assets other than the housing, denoted by wA , in 1990 Thai baht. Adding the rental value of owned house wH and other rental income wR , we then divide this sum by the monthly interest rate r to get an estimate of annual stock value of wealth5 : W = 1 A (w + w H + w R ). r (11)

This W is the variable that we use as a proxy for household wealth. We recognize that the levels of wealth may be poorly measured by this proxy variable and we do re-estimate the scale of wealth proxy in the estimation below. However, the distributional shape of the latent wealth is, we hope, well captured by this proxy. Figure 1.1 displays the resulting wealth distributions in 1976 and 1996, plotting the kernel density estimates of W in each year. Evidently the distribution of wealth in Thailand has shifted to the right, reecting the accumulation of wealth over two decades. However, the distribution has become more dispersed. Figure 1.2 displays the dierence in densities of the wealth distributions after normalizing about the mean levels. This gure clearly suggests that the distribution wealth has spread toward the tails and away from the mean, in particular toward the left tail, over the two decades. [F igures 1.1 1.2 Here] In the following subsections, we document observed trends in the key variables suggested by the two models, occupation choice and nancial participation, for the Thai economy during the 1976-1996 period. Specically, using the above wealth variable, we analyze whether these two self-selection features are plausibly related to wealth, as the LEB and GJ models suggest. 3.1.2 Occupation Choice

In the LEB model, selection is about the choice of occupation. The income unit of the SES is the household while the genuine unit of occupation choice in the model is the individual. Thus we need to dene occupation choice at household level. We have a variable named socioeconomic class of household in the SES, which is constructed from the employment status and employed sectors of the household head. According to the SES, the household head contributes to the total household income from eighty-three to ninety percent, depending on survey year. Thus the characteristics of the household head seem to be a reasonable proxy for the characteristics of the
the SES, the monthly average rental values of owned house are 320 in 1976 and 777 in 1996 both in 1990 Thai baht. calculate the representative interest rate, we use the annual interest rate of 14%, which is the time-series average of the lending rates of commercial banks, nance companies, and interbank loans between 1978 and 1996. (Data source: Economic Research Department at the Bank of Thailand.)
5 To 4 From

household as an individual and we use this SES variable to dene occupation at the household level. Among the economically active population in the data, there are three broad categories of occupations: worker, farmer, and non-farm business entrepreneur. The LEB model features a dynamic transformation among three occupation groups: worker, subsister, and entrepreneur. Workers and entrepreneurs are easily matched. We further identify as subsisters in the LEB model the farmers of the SES data.6 Figure 2 plots occupational composition over the period of 1976-1996. The main occupation transitions in Thailand were the switch from farmer to worker and an increasing fraction of entrepreneurs. The increasing fraction of entrepreneurs was nonlinear, stable at around 15% from 1976 to 1990, and then increasing after 1990 up to 19% by 1996. The occupational transition from farmers to workers was quite steady, and fast, during the entire period. The fraction of farmers decreased from 55% in 1976 to 31% in 1996 while that of workers increased from 29% in 1976 to 50% in 1996. These aspects of occupational transformation over time in Thailand are consistent with what the LEB model would suggest, i.e., wealth constraints in occupation choice are relaxed as wealth accumulates over time. In particular, Figures 3.1 and 3.2 illustrate observed occupational choice by wealth imputed classes in 1976 and 1996, respectively. In each given year, but more distinctly in 1996, the fraction of non-farm business entrepreneurs increases as wealth increases. [Figures 2- 3.2 Here] 3.1.3 Financial Participation

For the nancial participation variable, we classify the households into two groups, participants and nonparticipants, based on changes in nancial assets and liabilities. The SES provides us with data on changes of assets and liabilities with various nancial institutions due to nancial transactions by any member of the household. From these ow data, we identify the households that actually made nancial transactions. A household is categorized as participant if any member of the household used any of the formal nancial intermediaries commercial banks, a government savings banks, the Bank of Agriculture & Agricultural Cooperatives (BAAC), a government housing bank, nancial companies, or credit nanciers, and is otherwise as non-participant. There may be households who joined a nancial intermediary, but were inactive in nancial transactions for the entire month previous to the interview. Thus we may be under-reporting the participation rate. Figure 4 does show a signicant increase in the nancial participation rate among the economically active households, from 6% in 1976 to 27% in 1996. The trend of nancial deepening is also nonlinear, and quite dramatic after 1986. This nancial deepening accompanied by the wealth accumulation is consistent with the GJ model. Figures
6 Unfortunately, we cannot distinguish well in the data a subsistence farmer from an entrepreneurial farmer such as shrimp farmer, which might as well be categorized as an entrepreneur rather than a subsister.

10

5.1 and 5.2 illustrate nancial participation rate by wealth classes in 1976 and 1996, respectively. In each given year, we observe that the participation rate clearly increases with wealth. [Figure 4 - 5.2 Here]

3.2
3.2.1

Evolution of Growth and Inequality of Income


Aggregate Dynamics

Features of evolution of growth and inequality at aggregate levels in Thailand during 1976-1996 are displayed in Figures 6.1 and 6.2. Figure 6.1 overlays the levels of average income and income inequality, the latter measured by Theil-L entropy index. Evidently the Thai economy grew rapidly accompanied by a remarkable increase in inequality. The trends in levels are not monotone, moreover. Average income was more or less stable, and actually decreased slightly, for the rst half decade, between 1976 and 1986. Then, it doubled by 1996. Income inequality increased until 1992, with the exception of a downturn between 1986 and 1988, and decreased more substantially after 1992. Thus, we observe an inverted-U shaped relationship between development and inequality, as was hypothesized by Kuznets (1955). Figure 6.2 plots the growth rate of both variables. The annual growth rate of average income increased from 1986 until 1990, reaching 12%, and then tapered o to 6% by 1996. The growth rate of income inequality decreased until 1988, then increased shortly in 1990, and then decreased again, nally becoming negative after 1994. Interestingly, with the exception of the short period between 1986 and 1988, growth rates of average income and inequality co-move. [Figures 6.1 - 6.2 Here] In the following subsections, we describe the evolutions of growth and inequality of the key subgroups, emphasized in the theoretical models, namely occupation groups or nancial sector groups. 3.2.2 LEB Model Components

There are three types of income in the LEB model; subsistence income from the traditional sector, wage income, and prot income from the modern sector. Even though we cannot exactly categorize these income sources in the data, since the model abstracts from the detailed features of income sources of the real Thai economy, we do try to match as best we can the income variables in the model with the income sources in the data. Again, we identify the subsistence income in the LEB model with average farm income. We should keep two things in mind in this association: rst, the LEB model assumes that there is no risk in the subsistence sector, so the model misses a source of variation in farming prots; second, we drop all other sources of income of farmers such as part time earnings from non-farm activities, property income, and transfer income. Furthermore, we dene

11

income at the household level, not the individual level, and we do not adjust the family structure, because the model suggests that there is no variation in family structure. Thus we use total household income, not per capita income. We identify the wage of the LEB model with average wage earning of wageworkers, and prots of entrepreneurs of the LEB model with non-farm business prot. Similar qualications apply to these latter associations. Figure 7.1 plots the levels of average incomes of these three occupation groups. This illustrates the magnitudes of income gaps across occupations and the patterns of growth of the dierent factor incomes. Evidently entrepreneurs earn more than wage-earners who in turn dominate farmers. These incomes grew during the 1976-1996 period, at average annual growth rates of 1.4%, 3.2%, and 2.2%, for farm prot, wage, and non-farm prot, respectively. During the rst decade, 1976-1986, we do not observe much growth in factor incomes. Only the wage grew slightly, and prot incomes actually slightly declined. However, after 1986, growth was remarkable for every occupation; the average annual growth rates were 6.4%, 6.2%, and 6.1% for farm prot, wage, and non-farm prot, respectively. Figure 7.2 plots these growth rates. Evidently the growth rates of all three types of income co-moved, that is, increased and then decreased, until 1992. Then, those of the modern sector, i.e., wage and non-farm prot, decreased while farm prot rebounded. This catch-up growth by farmers contributed to a decrease in income gaps between farmers and non-farmers. Figure 7.3 plots the ratios of average incomes of entrepreneurs and wage earners relative to that of farmers. During 1976-1992, the income gap of wageworkers above farmers widened from 1.5 times to 2.6 times, and the income gap of entrepreneurs above farmers from 2.3 times to 3.4 times. Then, after 1992, the gaps narrowed to 2.2 times and to 2.8 times, respectively, owing to the above-described catch-up growth of farmers. According to the decomposition analysis in Jeong (2001), the decreased income gap across occupation groups associated with this catch-up growth accounts for the 85% of the inequality decrease in Thailand during 1992-1996 period. Figure 7.4 plots the levels of income inequality among households within each occupation group. This suggests that inequality levels increased within all three occupations and they co-moved, with the exception of the short period between 1992 and 1994. All these factors shape the growth and inequality dynamics at aggregate levels, as shown earlier in Figures 6.1 and 6.2. We shall return to this decomposition in Section 5 making more use of the LEB model itself. [Figures 7.1 - 7.4 Here] 3.2.3 GJ Model Components

The GJ model features nancial participation as the relevant selection mechanism underlying growth and inequality dynamics. Here we observe the subgroup dynamics of growth and inequality, partitioning economically active households by the measured presence or absence of nancial sector participation. 12

Figure 8.1 plots the average income levels of the participating and non-participating groups. Evidently participants have higher average income levels. Income of each group grew during two decades, 2.2% per annum for participants and 2.0% for non-participants. The average incomes of both groups did not grow until 1986; they actually decreased slightly. In contrast, for the second decade 1986-1996, growth rates of both groups were remarkably high, at 5.1% for the participants and at 5.6% for the non-participants. These growth rates are plotted in Figure 8.2. The growth rates of both groups co-moved until 1992, then increased and then decreased, and nally diverged after 1992. Specically, the growth rate of the participants, which had peaked at 15% in 1990, decreased and tapered o to near zero by 1996. The growth rate of non-participants decreased only slightly, and then continued up after 1992. Thus we observe similar patterns of catch-up growth between the nancial participation groups as among the occupation groups. This catch-up growth reduced the income gap of the participant group over the non-participant group from 3 times to 2.4 times during 1992-1996, as is illustrated in Figure 8.3, nearly back to the starting point of 2.3 times in 1976. The income inequality levels within each group increased, as is shown in Figure 8.4. Again, all these factors contributed to growth with increasing inequality at aggregate levels in Thailand. We shall return to this decomposition in Section 5 making more use of the GJ model. [Figures 8.1 - 8.4 Here]

Model Parameter Estimation

One of the main goals of this paper is to evaluate if the two models, the LEB model and the GJ model, deliver the growth and inequality dynamics that we see in the Thai data featured in Section 3. Specically, we simulate each model using the wealth distribution estimated from Thailand and using the parameter values that best match the micro foundations of these models as measured in the SES data. For the purpose of parameter selection, the key parameters of the models are chosen from explicit structural equations, occupation choice for LEB and nancial participation for GJ. That is, we choose the parameter values of each model that maximize the likelihood of those choices as constrained by wealth, where the likelihood functions are formed and subject to the constraints exactly as the theories suggest. Here, in estimation we use the data at initial years only. We also use only the young households in our estimation, considering their wealth to be the previously-given, beginning-of-period wealth, as in the theories.7 This helps us to avoid an endogeneity problem. Later, in testing, we forecast the income distribution at the nal date (the year of 1996) by simulating the models at these parameter values, and compare the closeness of each model to the actual income distribution in Thailand. For this, we use the entire sample of households, not just the young households. In this sense, we evaluate the
7 The

young household is dened as the one, the head of which is 20 to 29 years old.

13

models by out-of sample and post-sample testing. In this section, we explain how the model parameters are chosen via formal estimation procedures. The model evaluation follows in Section 5.

4.1
4.1.1

LEB Model
Specication of Likelihood Function

Occupation choice is the key building block or micro foundation of the LEB model, and the mapping from wealth to occupation itself is stationary, conditional on wage w. Thus we may form a likelihood of occupation choice as is implied by the theory and then identify some of the key model parameters using micro data on wealth and occupation choice. In the LEB model, the utility of being a worker is equal to that of being a farmer as long as two occupations coexist and the demand for labor determines the fractions of workers and farmers. Thus the crucial occupation choice is binary, between being an entrepreneur or not, even though there are three occupation groups. Let yi denote the binary occupational choice of agent i such that yi = 0, 1, if agent i chooses to be a worker or a farmer if agent i chooses to be an entrepreneur (12)

Then, the probability of being entrepreneur for agent i is given by Pr{yi = 1} = Pr(xi z (bi , w)), where z is the marginal setup cost function for being entrepreneur as is characterized by (4) in Section 2. Equation (13) is an exact implication of the LEB model of occupational choice. Note that the critical value z (b, w) is determined by the production technology f (k, l). Presumed also in equation (13) is the parameterized distribution H of the talent variable x. We can thus apply maximum likelihood estimation, taking equation (13) to the data on occupational choice in the real Thai economy to estimate the parameters of technology and the distribution of the talent variable in the model. Specically, suppose we observe the occupational choices (y1 , . . . , yn ) and initial wealth (b1 , . . . , bn ) of n agents. Without loss of generality, we may order them such that the rst e agents are entrepreneurs and the rest are wageworkers. As xi is i.i.d., the likelihood function L is given by: L = =
e Q

(13)

i=1 n Q i=1

Pr(xi z (bi , w))

i=e+1

[Pr(xi z (bi , w))]yi [1 Pr(xi z (bi , w))]1yi .

n Q

[1 Pr(xi z (bi , w))]

(14)

14

Maximum likelihood estimation is done with the log likelihood function, which in this case is given by log L =
n X {yi ln[Pr(xi z (bi , w))] + (1 yi ) ln[1 Pr(xi z (bi , w))]} i=1

(15)

The parameters of technology and the talent distribution are chosen to maximize the log likelihood function in (15). Notice that this log likelihood function is dened conditional on market wage w, which is endogenously determined in the model. Thus the structural estimation using this log likelihood function faces a potential endogeneity problem. Fortunately, however, in the LEB model, during the initial periods when there exists surplus labor, the wage is determined by the reservation level at w = + , where the and are exogenously given. Thus, maximum likelihood estimation on the data at early dates in Thailand when there were surplus labor forces in rural areas avoids this endogeneity problem. We need to parameterize the technology and the distribution of x in order to specify the log likelihood in (15). In their original model, Lloyd-Ellis and Bernhardt used the following specications of preference u, technology f , and distribution function H for x:
$ $ u(ct , bt+1 ) = c1 bt+1 t

(16) (17) (18)

f (kt , lt ) = kt

2 2 + lt kt k + lt lt 2 t 2

H (x) = mx2 + (1 m)x

With the distribution function in (18), the support of x is unit interval [0, 1] with range of possible values for parameter m given by 1 m 1. (19)

This class of distributions subsumes the uniform distribution at m = 0, and more generally the parameter m determines the skewness of the distribution. When m > 0, x is skewed toward high setup costs, and ecient entrepreneurs becomes rarer still as m increases further. This skewness parameter plays a key role in the dynamic path of the economy as it determines the relative scarcity of entrepreneurs, the demand for labor, and hence the wage rate. The wage rate in turn determines occupational choice, and so on. With the Cobb-Douglas form of preference in (16), utility maximization is equivalent to the total income maximization at the saving rate of $, which is separate from occupation choice. Thus estimation of occupation choice does not identify the preference parameter $. We will discuss how $ is calibrated in later subsection. With the parameterized form in (18), the probability of being an entrepreneur for an agent i with initial wealth bi is Pr(x z (bi , w)) = mz (bi , w)2 + (1 m)z (bi , w) 15 (20)

The critical setup cost function z depends on technology parameters, so it is now specied as z (b, w; , , , , ). Given this, the log likelihood function in (15) can be written as log L(, , , , , m; w) =
n X {yi ln[mz (bi , w; , , , , )2 + (1 m)z (bi , w; , , , , )] i=1

(21)

+(1 yi ) ln[1 mz (bi , w; , , , , )2 (1 m)z (bi , w; , , , , )]} The likelihood function of occupation choice involves nonlinear combination of the technology parameters (, , , , ) and the talent distribution parameter m. As foreshadowed earlier, there is an additional pa-

rameter implicitly involved in this likelihood function, a scale parameter, which translates wealth in the Thai SES data to wealth in the model, in LEB-model units. That is, the likelihood function can be written only after we convert wealth in the Thai data into LEB-model units. The LEB model is not free from the choice of this scale parameter because the support of the random talent variable x as specied is bounded and additive. 4.1.2 Identication

The stationary nature of occupational choice in the LEB model allows us to estimate the parameters using only the data at initial years. However, if using only a single initial year, an identication problem occurs: only three out of ve production parameters (, , , , ) can be recovered. This is due to the modeling of technology with the above specic quadratic form (17). The prot maximization problem in (2) involves a wealth constraint on capital and not on labor, which gives us labor demand as a function of capital, and hence perfect correlation exists between capital and labor. Specically, labor demand l is linear in capital demand k, l= w + k (22)

so that the prot function can be written as a reduced function of capital only as a second-degree polynomial (b, x, w; , , , , ) = C0 + C1 k + C2 k2 x, where the coecients, which may depend on the wage, can be spelled out in detail: ( w)2 , 2 ( w ) C1 (w) = 1 + , 1 2 C2 = ( ). 2 C0 (w) = (24) (25) (26) (23)

Capital demand k thus depends on the technology parameters (, , , , ), wage w, and wealth b through the requirement that k b x, only by way of these three constants. Now note that the key determinant in occupation choice is the critical setup cost function z (b, w), which is derived by equating the above prot function in (23) with wage w. Let b (w) be the critical level of wealth 16

beyond which the wealth constraint does not bind in occupation choice and x (w) be the associated level of critical setup cost. Let b b(w) is the wealth level below which the wealth constraint binds exactly at the level of

setup cost and hence the capital employment hits the lower bound zero. These three parameters, b (w), x (w), and b b(w), conditional on wage w, fully characterize the occupational choice in the LEB model. Specically, the z (b, w) = x (w), if b b (w) q 2 C1 (w) + 1 (C1 (w) + 1) 4C2 (C0 (w) b w) , if b b(w) b < b (w) = b+ 2C2 = b, if b < b b(w), b b(w) = C0 (w) w, (27)

z function is given such that:

where

(28)
2

C1 (w) x (w) = b b(w) , 4C2 C1 (w) b (w) = x (w) . 2C2

(29) (30)

Thus, conditional on wage w, the three coecients C0 (w), C1 (w), and C2 , along with m, fully characterize the log likelihood function, which is reduced to the following form: log L(, , , , , m; w) = log L(C0 , C1 , C2 , m; w). Thus the production parameters are under-identied from the above maximum likelihood estimation with a single wage. Unlike models with a Cobb-Douglas form of production function, labor shares and capital shares in the LEB model are not stationary. Thus we cannot utilize the more conventional way of identifying production parameters, matching long-term time-series average values of factor shares with the production parameters. If the data on capital demand and labor demand at individual rm levels were available, we could solve this identication problem from the additional estimation of the linear relation between labor and capital given in (22). This would add two more identifying restrictions to the maximum likelihood estimation of occupation choice. Unfortunately, these data are not available. However, note that the z function and hence the likelihood function is dened conditional on wage. By adding variation in wage, we may add more identifying restrictions. Suppose we use wage variation over time, say between two initial years, 1976 and 1981. Then, we can uncover the full ve production parameters as follows. Given the wage in 1976, w76 , the critical setup cost function z and the log likelihood function are characterized by the three coecients C0 (w76 ),C1 (w76 ), and C2 , and similarly the C0 (w81 ),C1 (w81 ), and C2 , at the 17

wage in 1981, w81 . Note that the coecient C2 does not depend on wage, and hence should be the same over time, and this is imposed in the estimation. That is, we may identify the technology parameters by maximizing the following likelihood function:
76 76 81 81 log L(C0 , C1 , C2 , m; w76 ) + log L(C0 , C1 , C2 , m; w81 ),

(31)

76 76 81 81 where C0 = C0 (w76 ), C1 = C1 (w76 ), C0 = C0 (w81 ), and C1 = C1 (w81 ). By allowing exogenous growth

in the subsistence income , the reservation wage level will also exogenously vary over time. Thus we still can avoid the endogeneity problem in our estimation using the log likelihood function in (31) with two dierent wage levels. Now we can identify the full ve technology parameters from maximum likelihood estimation with the
76 76 81 81 two-year variation of wage. Specically, given the estimates C0 , C1 , C0 , C1 , and C2 , the ve production 76 81 parameters (, , , , ) can be identied as follows. First, from dividing C0 by C0 , we nd : p p 76 w 76 C 81 w81 C0 0 p p = . 76 81 C0 C0 76 81 Then, substituting this either into C0 or into C0 , we nd :

(32)

1 = 2

76 81 Substituting these , , and either into C1 or into C1 , can be found: p p 81 76 C 76 C 81 C0 C1 1 0 p p =1+ . 76 81 C0 C0

76 81 Using this and subtracting C1 from C1 , we get : 76 81 C1 1 w81 w76 C1 = . p 2 p 2 76 81 C0 C0

w81 w76 p p 76 81 C0 C0

!2

(33)

(34)

(35)

4.1.3

Finally, substituting , and into C2 , we get : !2 76 81 1 C1 C1 p p = 2C2 . 76 81 2 C0 C0 Estimation

(36)

We use only the sample of young households, whose heads age is between 20 and 29, in our maximum likelihood estimation, considering their wealth to be their initial wealth as the model suggests. This limited use of data helps us to avoid the following endogeneity issue. Suppose we observe a wealthy household whose occupational choice is an entrepreneur. Then, it can be either the case that the household is wealthy because of its previous occupational choice as entrepreneur, or the case that it chose to be an entrepreneur because of its high initial 18

wealth. The latter is what the LEB model suggests. Since the Thai SES is not a panel of households, we cannot distinguish between the two cases if we use the entire sample. Focusing on the young households, whose current wealth levels are assumed to approximate their initially inherited wealth levels, we seek to avoid this endogeneity problem. We also exclude the households who participate the nancial sector following the benchmark LEB model without credit, upon which the likelihood function in (14) is formed. We also use the SES data only at the initial two years of 1976 and 1981. For the consistent use of data, we use the annual average wage income of the young households from the SES at 33,796 for the year 1976 and 36,808 for the year 1981, both in 1990 baht. The above maximum likelihood estimation involves quite complicated non-linearities not only in the shape of the likelihood function itself but also in the constraints on the parameter space, which are imposed from the theory. Due to the specication of the talent distribution of x as in (18), with support being bounded in unit interval, the critical values b b and x in the z function should satisfy the following relations: 0 b b wt 1, 0 x wt 1,

t t ,C1 , and C2 should satisfy which implies that C0

t wt C0 t C0 wt t2 C1

0, 1.

(37) (38)

4C2

Furthermore, since z is increasing and concave in b, the following inequalities should hold as well: b b(wt ) x (wt ),

x (wt ) b (wt ), which implies that C2


t C1

0, 0.

(39) (40)

Thus we solve the following constrained numerical optimization to do the maximum likelihood estimation for LEB: max
76 76 81 81 log L(C0 , C1 , C2 , m; w76 ) + log L(C0 , C1 , C2 .m; w81 )

76 ,C 76 ,C 81 ,C 81 ,C m,C0 2 1 0 1

subject to the constraints on parameter space: (19) and (37) to (40) for t = 76 and 81. This likelihood estimation is written at a given scale factor between the data unit and model unit. We vary this scale factor as well to 19

76 76 81 81 nd the scale factor that maximizes this likelihood at each set of estimates of C0 , C1 , C0 , C1 , C2 , and m. 76 76 81 81 Then, we uncover the ve production parameters from the maximizing values of C0 , C1 , C0 , C1 , and C2 at

a maximizing scale factor, using the identifying equations (32) through (36). Table 1 reports the estimated parameter values from the maximum likelihood estimation. The average value of log likelihood is -0.3368 at these estimates, which suggests the geometric mean value of the likelihood of the estimated occupation choice is 0.693. We include the standard errors of the estimates in parentheses. Since this estimation involves complicated non-linearities, we estimate the standard errors using bootstrap technique with 100 replications of random draws of the sample with replacement.

Table 1. Selected Parameters for LEB Model 0.5456 (0.0671) 0.3906 (0.0903) 0.2583 (0.0352) 0.0338 (0.0036) 0.1021 (0.0248) m -0.5933 (0.0580) scale 1.42E-6 (0.0088) log likelihood -0.3368

The occupation map of the LEB model partitions the type space (b, x) into three areas of occupation choice: the area of being workers or subsisters, that of being entrepreneurs subject to the wealth constraint in (3), and that of being unconstrained entrepreneurs. At the above estimates, this map is shown in Figure 9. Even though the information on being constrained is not available from the data, we may infer the magnitude of the wealth constraint in Thailand from this estimated occupation map together with the distributions of b and x at these estimates. [Figure 9 Here] First, the x s are estimated as 0.80 in 1976 and 0.74 in 1981, implying that the households whose realization of setup cost exceeds 0.80 under the 1976 wage and 0.74 under the 1981 wage will choose not to be entrepreneurs regardless of their wealth level. Figure 10 plots the estimated distribution function of the setup cost x at m = 0.5933. The cumulative distribution of x at x = 0.80 in 1976 is 0.896, which suggests that only 11.4% of the Thai households voluntarily choose not to be entrepreneurs regardless of their wealth level when the governing wage level is 1976 one. This fraction increases to 14.9% when the wage increases to its 1981 level. Therefore, the estimation suggests that most wageworkers or farmers in Thailand at least by 1981 chose to be so involuntarily either due to lack of luck, i.e., high realization of setup cost, or due to lack of wealth. The critical value x decreases in wage w and the order of magnitude of this involuntary occupation choice would be reduced as wage grows in the course of development. Second, k is the capital employment level of unconstrained entrepreneurs, and hence the minimum wealth level for being an unconstrained entrepreneur. The estimated occupation map in Figure 9 suggests that this 20

wealth level is 2.18 under the 1976 wage and 2.00 under the 1981 wage. Thus if the wealth level is lower than 2.18 in 1976 or 2.00 in 1981, the Thai household, even though being an entrepreneur, is subject to a wealth constraint. Figure 11 plots the wealth distributions of all households in Thailand in 1976 and 1981, the baht value being scaled by the above estimated scale parameter, with two vertical lines indicating the two critical wealth levels of 2.00 and 2.18, showing that the value of the cumulative distribution at each of these critical wealth levels is virtually one in both years. Thus the estimation implies that virtually all entrepreneurs in Thailand are constrained by their wealth level in running their businesses. [Figures 10 - 11 Here] 4.1.4 Calibration

The other parameters, saving rate $, subsistence income , and cost of living in modern sector are not related to the above occupational choice. We calibrate them separately from the above estimation but use the same SES data. Subsistence income is calibrated with annual average farm income of the SES in 1976 at 19,274 in 1990 baht. Converting this income into LEB units using the scale estimated from the above maximum likelihood estimation, we set at 0.0274.8 The Cobb-Douglas form of preferences over consumption and bequest yields simple optimal allocation rules; ct = (1 $) Wt and bt+1 = $Wt . Thus $ can be interpreted as a saving rate, which determines the speed of capital accumulation. The time series average estimate from the SES during 1976-1996 period for the saving rate is 0.25 with the standard error of 0.0204, . We calibrate $ at 0.23, onestandard-deviation below the mean, since the LEB economy at this value along with other chosen parameters displays a better t of observed aggregate inequality dynamics, by a mean-squared-dierence metric, than at 0.25. The cost of living parameter determines the wedge between being a wageworker in modern sector and being a subsister in traditional sector, and thus in turn determines the duration of unlimited labor supply before the upturn of the wage. Thus the parameter determines the wage dynamics and the relative income dynamics as well. We calibrate at 0.01, a value that best matches the aggregate inequality given the above estimates, again by a mean-squared-dierence metric.

4.2
4.2.1

GJ Model
Specication of Likelihood Function

Financial participation constrained by wealth is the key micro foundation of the GJ model. We form a likelihood function for the nancial participation decision using the pair of dynamic programming problems in (8) and (9) from Section 2.2. Let djt denote the participation decision of agent j at date t, which assigns 1 if agent j decides to participate in the nancial sector, and 0 otherwise. Then, from the two dynamic programming problems in
8 We set the exogenous growth at zero since otherwise we do not get the occupational transition from farmers to wageworkers at the above selected parameter values.

21

(8) and (9), the participation decision depends on wealth level kjt : djt = 1, = 0, if v (kjt q ) w(kjt ) if v (kjt q ) < w(kjt ). (41)

Townsend and Ueda (2001) follow Greenwood and Jovanovic and show that there exists a unique critical value k such that the participation decision formulated in (41) is equivalent to djt = 1, = 0, if kjt k if kjt < k (42)

when preferences are parameterized by CRRA utility function u(ct ) = c1 /1 .9 t

There is no closed form solution for k because there are no analytic solutions to the dynamic programming in (8). However, from the formulation of the dynamic programs in (8) and (9), it is clear that k is a function of the underlying parameters of the GJ model, GJ = (q, , , , , , ) so that k = k (GJ ), where is a set of parameters for the distribution function of the aggregate shock , and is a set of parameters for distribution function of the i.i.d. idiosyncratic shocks j . Recall that the law of motion is given by kjt = [j,t1 jt + (1 j,t1 ) ]ij,t1 for those outside the nancial system. Owing to the Markov structure of the dynamic programming problem in (8), the policy functions for portfolio and investment are time-invariant and depend on previous wealth and the underlying parameters GJ . That is, j,t1 = (kj,t1 , GJ ) and ij,t1 = i(kj,t1 , GJ ). Then, a non-participant at time t 1 will enter the nancial sector at time t only if kjt = [(kj,t1 , GJ ) jt + (1 (kj,t1 , GJ )) ]i(kj,t1 , GJ ) k = k (GJ ). Equivalently, jt kj,t1 , GJ i kj,t1 , GJ 1 k (GJ ) (1 kj,t1 , GJ ) .

That is, the participation decision of a previously non-participating agent j at date t with wealth kj,t1 can be rewritten as djt = 1, = 0, if jt kj,t1 , GJ if jt < kj,t1 , GJ , (43)

9 In the original GJ model, preferences are parameterized as a log utility function, which here is equivalent with the limit case as 1.

22

where kj,t1 , GJ kj,t1 , GJ i kj,t1 , GJ 1 k (GJ ) GJ (1 kj,t1 , ) . (44)

Given the level of previous wealth kj,t1 , the participation decision in (43) is stationary because the composite technology shock jt is drawn from the time-invariant distribution . Thus, once we solve for the time-invariant policy functions and i, we can form a likelihood function as follows, suppressing the time index t. Suppose we observe the participation decisions (d1 , , dn ) and previous wealth levels (k1 , , kn ) of n households, and that without loss of generality the rst p agents participate in the nancial sector and the remaining n p agents do not. Then, the likelihood function L can be written such as L = =
j =1 n Q j =1 p Q

[Pr( j kj , GJ )]1dj [1 Pr(j kj , GJ )]dj .

Q n Pr(j kj , GJ ) Pr(j < kj , GJ )


j =p+1

(45)

Then, the log likelihood is given by


n X log L GJ {(1 dj ) ln Pr( j kj , GJ ) + dj ln[1 Pr( j kj , GJ )]} = j =1 n X {(1 dj ) ln j ( kj , GJ ) + dj ln[1 j ( kj , GJ )]} = j =1

(46)

We choose the parameter vector GJ that maximize the log likelihood function in (46), satisfying the restrictions on parameter space given in (6). In contrast to the LEB model, here in the GJ model, we do not have any endogeneity problem for the structural estimation with this likelihood function. It remains to characterize the distribution function of the composite shock j . The GJ model assumes that aggregate shock follows a uniform distribution on [ , ], and the idiosyncratic shock j also follows a uniform distribution on [, ]. Then, the distribution function j of the composite shock j = + j is characterized as follows: j (s) = 1, = 1 = = f or s > + , f or < s + f or + < s , f or < s + f or s (47)

2 +s 4( )

, ( ) 2 s ( ) 4( )

(s )

= 0,

We substitute this distribution function j into the log likelihood function in (46). 23

4.2.2

Estimation

The critical level of capital k , and the policy functions and i need to be numerically calculated. This involves the computation of the pair of value functions v and w from the pair of the dynamic programming problems of (8) and (9) at each given parameter vector GJ = (q, , , , , , , , ). Due to the practical time constraints of this heavy computation, we search over the maximum likelihood value on a very coarsely partitioned parameter space around the benchmark values adopted in Townsend and Ueda (2001). The benchmark values in Townsend and Ueda (2001) were chosen as follows. The time discount parameter is set at = 0.96 following the typical business cycle literature, for example in Kydland and Prescott (1982). To expedite the computation, one period in the model is supposed to equal three years in the data. This is equivalent to lowering the time discount factor into 0.885 = (0.96)3 . The risk aversion parameter is set at = 1 as in the original GJ model. The technology parameters are calibrated from the actual Thai data. The xed entry cost parameter is calibrated at q = 5 to match the initial nancial participation rate of 6% at the initial wealth distribution estimated from the 1976 SES as above, and we presume no variable cost in intermediation, i.e., = 1. The safe return parameter is calibrated at = 1.132 = (1.042)3 to match the annual net return of 4% from the capital investment in agriculture estimated in Paulson and Townsend (1999) using the micro survey data on rural Thailand of Townsend et al. (1997), and also to satisfy the assumption > 1/ of the GJ model in (6). With these technology parameter values, we choose the aggregate shock parameter values of = 0.1 and = 2.9 to match the annual rate of return of 19% from non-agricultural business investment for the households with credit access, estimated again in Paulson and Townsend (1999), i.e., E {r( )} = 1.69 = (1.19)3 , and also to satisfy the restriction on parameter space E {r( )} > E { } > in (6). Here, we consider the agricultural projects as relatively safe investments compared with the non-agricultural business projects as we did for the subsistence sector versus modern entrepreneurial sector in the LEB model, again recognizing the potential problem of misalignment between theory and data from this classication of technology. Finally, the idiosyncratic shock parameters are taken as free parameters at = 0.05 and = 0.05 for three-year period. As mentioned above, the practical time constraints of heavy computation allowed to vary the parameter values in nine experiments around the neighborhood of this benchmark. We evaluate the likelihood values of the GJ model at these parameter sets and pick the parameter set with the highest likelihood value as maximum likelihood estimates. The experimented parameter sets and their associated likelihood values are reported in Table A.2 in Appendix.10 Here, again we use young households only to avoid the wealth endogeneity issue. Due to the stationary nature of the nancial participation of the GJ model, we also use the initial 1976 SES data
10 There are some observations that the MLE assigns zero probability in the GJ model while there were none in the LEB model. Even a single observation of the zero-probability event makes the likelihood zero, equivalently the log likelihood -. Thus we drop those observations in evaluating the log likelihood. The fractions of those observations are reported in Table A.2 as well.

24

only in our estimation, and so we can evaluate the GJ model via post-sample testing as we do for the LEB model. Table 2 reports the chosen parameter values for the GJ model from the above estimation.11 It turns out that the nancial participation in the GJ model is the most likely at the benchmark parameter values, except for a higher risk aversion parameter at = 1.5.

Table 2. Selected Parameter Values for GJ model q 5 1.132 0.885 1.5 1.00 0.1 2.9 -0.05 0.05 log likelihood -0.1407

11 Here, due to the sparse grid points on the parameter space, the bootstrap standard errors do not deliver us a meaningful measure of uncertainty of the parameter estimates, and we do not attempt to estimate them for the GJ model.

25

Model Evaluation

To study the factors determining the trends of income inequality in the course of a countrys economic growth, Kuznets (1955) suggested that data be classied by average income levels for a suciently long span, say a generation, so that one can form long-term income-status groups. Then, long-term changes in the income distribution consist of the population shifts across income-status groups and changes in the income levels of these groups. He motivated this classied array of data by saying, If living members of society- as producers, consumers, savers, decision-makers on secular problems- react to long-term changes in income levels and shares, data on such an income structure are essential. Consistent with this idea, we classify the population according to the income-status groups that are suggested by the above models, by occupation groups as in the LEB model or by nancial participation groups as in the GJ model. Both models, LEB and GJ, are simulated here at the selected parameter values of Section 4. The simulations display how growth and distribution of income evolve over time according to the dynamic mechanisms of the models that are deliberately close to Thailand in their micro-foundation features. Then, we evaluate each model by matching each models simulated dynamic features to the actual data. Specically, we compare the implications of the models with the Thai data by three criteria: the dynamic paths of average income and inequality of income both at aggregate levels and subgroup levels, the decomposed Kuznets dynamics of growth and inequality change, and the forecasted shape of the aggregate income distribution at the nal 1996 date.

5.1

Numerical Simulation

The LEB model is simulated using the computer programs of Gine and Townsend (2001). At every date there is a distribution of beginning-of-period wealth, presumed to lie on some a priori grid. Guessing a wage, along with the parameters of technology, the regions of the occupation partition are pinned down. The distribution of talent then determines the fractions of the population choosing to be wageworker, subsister, or entrepreneur at each level of wealth. Adding up over all wealth levels, these population fractions should sum to one, and otherwise the labor market does not clear. This procedure is repeated to nd an equilibrium wage in a bisection algorithm. Thus end-of-period wealth is determined. Parameterizing preferences, a fraction $ of this wealth is saved, and this determines next periods distribution of beginning-of-period wealth. The distribution of setup cost for entrepreneurs adds additional diversity. The lower end point of the wealth distribution is the wealth of the household in the previous period who had least beginning-of-period wealth and the lowest talent (highest setup cost), and the upper end point is associated with the household in the previous period who had the highest beginning-of-period wealth and the highest talent (lowest setup cost). The initial condition of the model is the initial distribution of wealth. Here we will take the initial l976 wealth distribution as discussed earlier. One

26

period in the simulation is supposed to correspond to one year in the data. The LEB model is modied to include an exogenously embedded intermediated sector. Those in the intermediated sector can borrow and lend their wealth at an equilibrium interest rate, determined in a bisection algorithm. This sector is otherwise integrated with the rest of the economy that is, labor can migrate across sectors. Thus the wage and interest rate are determined simultaneously. At each period the number of households in the intermediated sector is specied exogenously, and made to increase at the observed trend rate of increase in participation as in the SES data, from 6% in 1976 to 26% in 1996. In a sense, then, we have made the model look a little more like the data, but in the comparison of LEB to the data below, we do not distinguish the data itself by nancial sector participation. The GJ model is simulated using the algorithm of Townsend and Ueda (2001). The burden here is nding the value functions w and v described earlier, which are not bounded, and the support of which is evolving over time. Still, the value function v has a closed form solution up to the risk aversion parameter , as does a value function for a ctitious sector, those never allowed to enter the nancial system. The value function w can be trapped between these two. The non-convex aspect of the problem, the one associated with the xed cost of entry q , disappears in the limit, as the horizon is driven to innity. Households choose some of the risky asset and when to enter the nancial system, and the value functions converge after iteration. The space of value functions is reasonably well approximated by Chebyshev polynomials. Policy functions, savings and portfolio share, are found by grid search with successive renements. In Townsend and Ueda (2001), to expedite the computation, one period in the model is mapped into three years in the data. This is equivalent to lowering the discount rate and inating the gross yields from the production technologies as was discussed in Section 4. The dynamic path of the GJ simulation depends on the realization of aggregate shocks. Thus we need to choose a specic path of realized outcomes of aggregate shocks to determine which GJ simulation is to be compared with the data. Here we pick the path that is closest, out of 10,000 simulated paths at the chosen parameter values, to the Thai data of growth and inequality using the mean-squared-error metric.

5.2

Comparison of Dynamic Trends

In the following comparison, we make three adjustments to the data. First, we rescale the Thai income data normalizing the initial average income levels so that we make the actual income distribution have the same center or mean as each model when the economy starts in 1976. This rescaling does not aect the Thai inequality levels, which are measured by the scale-free Theil-L entropy index. This initial normalization will help us to see the way in which average income and income inequality of the model economy diverge from the actual economy as they evolve over time, again starting from the initial distributions with same central tendency. Second, in comparing subgroup dynamics, we dichotomize the characteristics in the data as we did for estimation:

27

entrepreneur or non-entrepreneur for the LEB model and participant or non-participant in nancial sector for the GJ model. Third, we interpolate the annual values that are missing either in the simulations or in the data to make possible comparisons not only between a model and the data but also across models. 5.2.1 Aggregate Dynamics

The aggregate dynamics of the simulated economy are compared with the actual economy in Figure 12 for the LEB model and in Figure 13 for the GJ model. Each gure includes four panels that display average income levels, average income growth rates, income inequality, and proportion of the population in the higher-income group, i.e., entrepreneurs in LEB and nancial participants in GJ. Both models indeed deliver the overall features of growth with increasing inequality as observed in Thailand, even as they display their own discrepancies in levels and uctuations, as is now made explicit. The LEB model predicts less overall growth than the actual Thai economy, as is shown in Figure 12.1. Average income in Thailand grew by 87% over the two decades, while that of the LEB model by only 22%. This gap comes in part from the initial declining fraction of entrepreneurs, the higher-income group, as shown in Figure 12.4. Essentially, the initial distribution of wealth is too high and so for early periods that wealth is consumed and with lower wealth, xed costs are higher, and subsistence sector is more attractive.12 In any event, the LEB fraction of entrepreneurs at 13.7% at the initial date is quite close to that of Thailand, but the LEB fraction decreases to 6.6% until 1982, and then rises back to 13.9% in the end, by then well below the Thai fraction at 18.8%. However, the uctuations of the growth rate and the movement in the fraction of entrepreneurs after 1986 are quite well captured by the LEB model. Figure 12.3 suggests that the LEB model predicts a higher level of income inequality for the entire period, except at the nal date. The LEB model does, however, capture the pattern of increasing and then decreasing inequality after 1986, with a divergence in the timing and magnitude of the turning point. Initial changes in inequality may be attributable again to initial adjustments with an over-estimate of the initial wealth distribution and thus with movement back toward the subsistence sector. In short, these initial movements might be discounted somewhat. However, the increase in inequality from l985 to the mid 90s is robust across a variety of data sets and model parameters, and has to do with the increased rate of nancial expansion taking place in the background. On the other hand, the evidently large decrease in inequality at the end of the LEB simulation is due to a pick up in the equilibrium wage, and the timing and magnitude of that decrease is sensitive to alternative parameter values. This occurs because of the extreme specication of the LEB model, that marginal productivity in the subsistence sector is constant, not decreasing (see Gine and Townsend, 2001). The magnitude of the decrease in LEB inequality after the turning point is so large that the LEB inequality level at the nal date almost matches
12 This initial decline does not happen under alternative parameter values, as estimated in the Townsend-Thai data in Gine and Townsend (2001), for example.

28

that of Thailand. Thus, overall, with some discrepancies in levels, the LEB model generates the aggregate dynamic patterns of the Thai economy, particularly after 1986 (when signicant nancial deepening started in Thailand, as exogenously embedded in this LEB simulation.) The GJ model, in contrast, predicts more than double the overall growth of the Thai economy at 214%. Both the level and growth rate of average income of the GJ model are far more volatile than the actual Thai economy, as shown in Figures 13.1 and 13.2. Note, however, that the GJ model is able to track the pattern of change in Thai income, with its relatively at segment prior to l986 and its increasing trend afterward, though the GJ simulation does suer from wide uctuations which enter into the growth rate calculations. No doubt related, participation in the nancial sector in the GJ model, starting, by construction, at the same 6% rate as in the data, is relatively at until 1985, and only begins to increase at a higher rate at the very end of simulation (but only to 10% in the end, lower than the 26% of the Thai data). Striking, however, are the levels and overall magnitude of change of GJ inequality, as these are close to those in Thailand. However, GJ inequality does begin to decrease much earlier than actual Thai inequality and it rises back up when actual Thai inequality begins to decrease. This upright U-shaped, decreasing and then increasing, pattern of inequality comes as a surprise Greenwood and Jovanovic (1990) do not alert the reader that their model is capable of delivering this, though granted the simulation here is at alternative parameter values. Comparison of Figures 12 and 13 does not seem to suggest a uniform dominance of one model closer to the actual mechanism generating the Thai data. The LEB model performs poorly at rst but performs well after 1986, though its levels of inequality are too high and change too sharply in the end. Again this sharp decline of inequality in the end is due to the way LEB models wage movement. The LEB wage begins to grow, hence the gap between wage income and entrepreneurial income drops, suddenly when the surplus labor is exhausted. The GJ model appears to be consistent over the two decades with overall trends but has excessive volatility, making its overall performance of tting levels and growth of income poor. This excessive volatility of the GJ model is due to the large magnitude of the modeled aggregate shock at the chosen parameter values. In this respect, the LEB model matches the aggregate growth trends better than the GJ model. However, the prediction of the GJ model for inequality is more consistent in level, though period-by-period movements in inequality are contrary to what we see in the data.13 [Figures 12 - 13 Here] An overall summary assessment on the performance of models tting aggregate growth and inequality dynamics can be made by measuring the average distance of the simulated paths of the models from the actual paths of the data for each of the three aggregate dynamic features, i.e., income level path, growth rate path,
13 In Townsend and Ueda (2001), at alternative parameter values, there is a sustained increase in inequality projected for the future. Also, many of their simulations show an eventual rise in the participation rate, which is higher than the Thai data.

29

and inequality level path. As one metric of the distance between a model and the data, we use the root mean
T T squared dierence (RMSD) in time series between simulated path {x t }t=1 and actual path {xt }t=1 : v u T u1 X RM SD = t (x xt )2 . T t=1 t

Table 3 reports this distance of each model, measured by the RMSD metric, and it suggests that the LEB model ts aggregate dynamics in Thailand better than the GJ model for level and growth of income while the GJ model ts better than the LEB model for the level of inequality. Table 3. RMSD for Aggregate Dynamics Income Level 0.129 0.334 Growth Rate 0.047 0.367 Inequality Level 0.130 0.072

LEB GJ

5.2.2

Subgroup Dynamics

The aggregate dynamics of each model are now decomposed into the subgroup dynamics and compared to those in the data. We plot the levels and growth rates of subgroup average income, the income gap ratio between subgroups, and the inequality levels of each subgroup. We already observed at a more detailed level occupation group dynamics in Thailand, that is, with three categories of occupations in Section 3. Figure 14 presents these features for dichotomous occupation groups, namely entrepreneurs and non-entrepreneurs, to make the subgroup dynamics in the data more comparable with the LEB model. Figure 14.1 for the actual Thai economy suggests that average income grew in both occupation groups with similar trends. Indeed, the growth rates in Figure 14.2 for both occupation groups show co-movement except for the higher rate of entrepreneurs in the late 80s and then the catch-up-growth period for non-entrepreneurs during 1992-1996. The income gap ratio in Figure 14.3, measured as the ratio of average income of the entrepreneurs to that of non-entrepreneurs, declined from 2.3 times in 1976 to 1.9 times in 1996 but with a slight humped uctuation between 1988 and 1996 attributable again to an industrial expansion and increased protability at that time. Inequality levels also grew in both groups, again with co-movements evident from Figure 14.4, but with more erratic movements for entrepreneurs. Inequality levels are higher within the entrepreneurial group relative to the non-entrepreneurial group in most years - evidently there is either variation in protability of returns across entrepreneurs that is higher than the variation in income across farmers and wage earners or there is more measurement error. Yet subgroup inequality levels of both occupation groups converged over time. The occupation group dynamics of the best-t LEB economy are displayed in Figure 15. Average income grows with increasing inequality within each occupation group, accompanied by a declining income gap ratio 30

between entrepreneurs and non-entrepreneurs. Thus, qualitatively, the overall directions of subgroup dynamics seems to t the Thai economy. However, we do not see much co-movement in average income and income inequality between occupation groups in the LEB economy. In fact, we observe that growth rates of income across occupations move counter to each other, unlike the Thai economy. The inequality ordering between occupation groups is reversed: the subgroup inequality is much higher among non-entrepreneurs in the LEB model, due mainly to the merging of farmers and wage earners and the cost of living dierential between the two. Most extreme of all, the income gap ratio stands at seven to thirteen times, though still decreasing in trend as in the actual Thai data. [Figures 14 - 15 Here] Group dynamics by nancial sector participation in Thailand were already described in Section 3, and they are reproduced for convenience of comparison in Figure 16. Again, there is an income gap between participants and non-participants in the Thai data, a gap which eventually increases before declining in the end. Income levels tend to co-move, while inequality increases for both groups, though interwound. In Figure 17, we observe the subgroup dynamics of the best-t GJ model. The GJ model also captures the overall features of subgroup dynamics - subgroup average incomes grow (though more so for participants) with increasing inequality (though again more so for participants). A striking pattern of the GJ model is that the growth rates of both groups co-move closely. The income gap across participation groups increases as in the Thai data, but now in the model there are huge discrepancies in the order of magnitude of that gap: six to fourteen times in the GJ economy while only 2.3 to three times in Thailand. Inequality is higher among non-participants than among participants, even though the latter increases dramatically toward the level of the former. Note also that subgroup inequality levels are lower in GJ relative to the Thai economy, though that was also true for subgroup dynamics in the LEB model. We turn to a more explicit decomposition in the next subsection. [Figures 16 - 17 Here] In summary both models capture some aspects of subgroup dynamics in the actual Thai data, consistent with both capturing something of aggregate dynamics. But now discrepancies and anomalies are more apparent, and the t of each model to the Thai data begins to deteriorate. Both models produce income gaps across subgroups that are too large relative to their actual magnitude in the Thai data, and under-predict subgroup inequality. The LEB model is consistent with a declining income gap between entrepreneurs and nonentrepreneurs, and the GJ model is consistent with an increasing income gap between participants and nonparticipants. But the GJ model is able to produce co-movements across participants and non-participants, as in the Thai data, while the LEB model tends to produce the opposite, counter-movements across entrepreneurs 31

and non-entrepreneurs, unlike the Thai data. Likewise, the GJ model is able to produce higher inequality among non-participants than participants, which is true in the Thai data for a little less than the rst half periods, while the LEB model predicts more inequality among non-entrepreneurs, unlike the Thai data, which always display the opposite. 5.2.3 Decomposed Inequality Dynamics

Partitioning the population into subgroups by characteristics, there are two sources of the aggregate inequality in an accounting sense: across-group inequality and within-group inequality. The former inequality comes from the income gaps across subgroups, as if individual income levels were equal within each group. The latter is due to the variation in income levels among individuals within each group. Specically, using the Theil-L entropy index I , the inequality measure we use, these two sources of inequality can be decomposed as follows: I 1X log n i=1
n

yi

(48) (49)

= W I + AI, where W I =
K X

pk I k ,

and AI =

k=1

k=1

K X

pk log

(50)

Here, n denotes the population size, the average income at aggregate level, and yi the individual is income, k the subgroup index, pk the fraction of population in subgroup k, k the average income of subgroup k, and I k the income inequality within subgroup k. The W I denotes the within-group inequality, and AI the across-group inequality.14 Thus within-group inequality W I is the sum of the subgroup inequality levels weighted by their population shares, while the across-group inequality AI is the sum of log relative incomes of subgroups again weighted by their population shares. We can decompose the aggregate inequality at each given date, both for the models and the data, according to this decomposition formula. Figure 18 compares these decomposed inequality factors in Thailand with those of the LEB economy using the occupation-group partition. In Thailand, from Figure 18.1, within-occupationgroup inequality contributes 82% to 88% of total inequality. In contrast, within-occupation-group inequality in the LEB model, though increasing in importance over time, from 14% to 40%, is not the major factor. That is, LEB inequality is driven by across-occupation-group inequality much more than in the actual Thai data. We observe similar patterns for the GJ model in Figure 19. In the Thai economy, within-participationgroup inequality is the dominant contributing factor to the aggregate inequality while across-group inequality is small, even though its share increases quite fast from 7% to 14%. In the GJ model, within-participation-group inequality stays almost constant with a declining share while across-participation-group inequality dominates
(1967) originally suggested this index as an inequality measure. For a detailed discussion of its decomposability, see Bourguignon (1979) and Shorrocks (1980, 1984).
14 Theil

32

for all periods after 1980. Also the uctuations of GJ aggregate inequality are determined by the movements of across-group inequality. In summary, each model generates aggregate inequality dynamics by overstating across-group inequality. This may not be surprising. A model naturally concentrates attention on a few characteristics and so the eects of these characteristics would be exaggerated in the model economies relative to actual inequality in the data. Put dierently, given the presumed key elements, neither LEB nor GJ has probabilistic structures rich enough to capture within-group income variation as observed in the data. We return to this topic in the concluding remarks. However, we rst further decompose the changes of within-group inequality and across-group inequality and hence the change in aggregate inequality into their subgroup components. [Figures 18 - 19 Here]

33

5.3

Decomposition of Growth and Inequality Change

In this subsection, we suggest a simple method that decomposes total inequality change into the above various factors; compositional changes, diverging income-gap across subgroups, and subgroup inequality change. We also decompose income growth and thus identify the relative importances of the compositional eects on growth and inequality change, i.e., of Kuznets dynamics either via occupational choice or via nancial participation. 5.3.1 Decomposition Method

Consider an economy with dichotomous income-status groups, h and l, denoting high and low, respectively l the group h has higher mean income h t than the group l at t at date t for all periods. Then, the average l h l income level t at date t is the sum of h t and t weighted by the population shares pt and pt : h l l t = ph t t + pt t .

(51)

Thus, the growth of average income comes either from the growth within each subgroup, or from the population shifts from l to h, which can be formulated by the following discrete version of chain rule: = {ph h + (1 ph )l } + (h l )ph , (52)

where denotes the dierence operator over time and the upper bar the average over time. The rst two terms in (52) are the components of growth within subgroups, and the nal term the growth due to population shifts. Due to the additive structure of the Theil-L index, we can also decompose the change of aggregate inequality I , again applying the discrete chain rule to equations (49) and (50):

I = W I + AI,

(53)

W I = {ph I h + (1 ph )I l } + (I h I l )ph , AI = ph (h 1)( log h log l ) + (h l log h )ph , l

(54)

(55)

where W I denotes the change in within-group inequality, AI the change in across-group inequality, I k the inequality level of subgroup k and k
k

the relative income of subgroup k.15

As equation (53) suggests, the change of total inequality consists of that of within-group inequality W I and that of across-group inequality AI . Within-group inequality change W I in (54) is decomposed into two
15 The generalization of this decomposition to the case with more than two categories is quite straghtforward. See Mookherjee and Shorrocks (1982) for this.

34

types of terms exactly as in the growth decomposition (52). The rst two terms in (54) account for inequality change due to inequality change within subgroups, weighted by their population shares. The nal term in (54) accounts for within-group inequality change from population shift, the magnitude of which is governed by the dierence in subgroup inequality levels. If the higher-income group h has higher level of inequality than the lower-income group l on average, i.e., (I h I l ) > 0, then a population shift from l to h increases inequality through this term. The change of across-group inequality AI in (55) consists of two eects: a divergence eect, the rst term in (55), and a Kuznets eect, the second term in (55). The rst term in (55) can be interpreted as a divergence eect in the following sense. Note that ( log h log l ) approximates the dierence in growth rates between income-status groups h and l and that h > 1 since h > l , hence the coecient ph (h 1) is always positive. Thus, if higher-income group grows faster than the lower-income group, i.e., ( log h log l ) > 0, then, income levels between the two diverge and across-group inequality increases, and vice versa. The second term in (55) captures the eect of compositional change on across-group inequality. We call this term the Kuznets eect because the inverted-U-shaped inequality path along with growth is generated f h , and by this term. The coecient (h l log ) of ph in the second term in (55) is positive when ph p l
h

f f h , for a unique critical value of the population proportion of higher-income group p h .16 negative when ph > p

Thus, inequality rst increases when the population shifts from lower-income group to higher-income group,

and then eventually decreases after some critical level of development, here the fraction of higher-income group f h . This is the Kuznets curve that is verbally described and illustrated with numerical examples in the original p Kuznets 1955 paper. 5.3.2 Decomposition Results

We apply these decomposition formulae, (52), (54), and (55) to the models and the data to identify the relative importance of Kuznets dynamics as well as other subgroup factors. We thus evaluate whether the models deliver growth and inequality dynamics matching the magnitudes of Kuznets dynamics, and other subgroup factors, in the Thai data. Table 4 compares the growth decomposition results of the LEB model to the Thai data with respect to the dichotomous partition of occupation groups, entrepreneurs and non-entrepreneurs. The rst column corresponds to the subgroup income growth and the second column to the compositional growth. The nal column reports the two-decade total growth rate. The percentage contributions of each factor to total growth are included in parenthesis. As we already observed, the LEB model under-predicts income growth. The two-decade total growth rate is only 0.218 in LEB while it is 0.869 in Thailand. Table 4 further reveals that population shifts
16 The

f h is shown in Jeong (2001). existence and uniqueness of such critical value p

35

across occupation groups contributed 0.035 to actual Thai income growth but only 0.004 in the LEB model. This is one instance in which the model with its chosen categories still under-emphasizes the distinction across subgroups that is more apparent in the Thai data. Table 5 compares the decomposition results for inequality change between the LEB model and the Thai data by occupation groups. The rst and second columns report the contributions of within-group inequality change: the rst column corresponds to subgroup inequality eect and the second to the compositional eect on withingroup inequality change. The third and fourth columns report the contributions of across-group inequality changes: the third column corresponds to the divergence eect and the fourth column the composition eect on across-group inequality change, i.e., the Kuznets eect. The nal column reports the two-decade inequality growth rate. Total inequality change in the LEB model is far lower than in Thailand as was growth of income. The total inequality change in the LEB model is only 0.052 while it is 0.468 in the Thai data. Large discrepancies are observed in the magnitude of divergence eect between Thailand and LEB, though the signs are consistent. Specically, the income gap between entrepreneurs and non-entrepreneurs narrowed and contributed to a decrease in inequality by 0.051 in Thailand. In contrast, in the LEB model, convergence is associated with a decrease in inequality by 0.228. Furthermore, the Kuznets eect of occupational shifts, from non-entrepreneurs to entrepreneurs, increased inequality by 0.009 in Thailand while this eect is 0.004 in the LEB model. The composition eect on within-group inequality is small both in Thailand and the LEB model, at 0.000 and -0.001, respectively. A major force behind increasing total inequality is the increase in inequality within each occupation subgroup, at 0.510 in Thailand, and 0.277 in LEB. In summary, increase in subgroup inequality and the converging income gap between entrepreneurs and non-entrepreneurs are major factors of change in inequality in Thailand and also in the LEB model, but the model over-emphasizes the convergence eect and, though less dramatic, under-emphasizes the within-subgroup eect.

Table 4. Decomposition of Aggregate Income Growth by LEB Model Subgroup Composition Total Thailand 0.834 (96%) 0.035 (4%) 0.869 LEB 0.214 (98%) 0.004 (2%) 0.218 Table 5. Decomposition of Aggregate Inequality Change by LEB Model Within-Group Subgroup Composition 0.510 (109%) 0.000 (0%) 0.277 (533%) -0.001 (-1%) Across-Group Divergence Composition -0.051 (-11%) 0.009 (2%) -0.228 (-439%) 0.004 (7%) Total 0.468 0.052

Thailand LEB

36

Table 6 and 7 compare similarly the decomposition results of Thailand with those of the GJ model. Here the decompositions are done with respect to the partition by nancial sector participation. As noted earlier, the GJ model overstates the growth rate at 2.139, more than double that of actual Thai growth. Table 6 suggests that the contribution of nancial deepening to growth, i.e., the population shift from non-participants to participants, is quite signicant at 0.313 in Thailand, and in the GJ model as well, at 0.364, quite close to the actual Thai data. However, the subgroup growth in the GJ model is far larger at 1.775 than in Thailand at 0.556. The aggregate inequality change in the GJ model is also over-predicted, as was growth of income, at 0.725 than the actual rate, at 0.468. This is due to the large discrepancy in the divergence eect and subgroup inequality eect between the GJ model and Thailand. In the Thai data, the divergence eect contributes to inequality change only by 0.014 while the subgroup eect is 0.290. In contrast, in the GJ model, the divergence eect contributes 0.479, already exceeding the total inequality change in Thailand, while the subgroup eect is only 0.051. We observe a discrepancy in composition eect on within-group inequality change as well, but the magnitude of this discrepancy is relatively small. However, Kuznets eect is reasonably well captured by the GJ model. The composition eect of nancial deepening on inequality change in the GJ model, at 0.203, is relatively close to the Thai data, at 0.131. In particular, the percentage contribution of Kuznets eect on inequality at 28% exactly matches the actual Thai data. Thus the GJ model overstates the eect of diverging income levels between participants and non-participants on inequality increase while de-emphasizing the magnitude of changing inequality within subgroups, unlike the Thai data. Table 6. Decomposition of Aggregate Income Growth by GJ Model

Thailand GJ

Subgroup 0.556 (64%) 1.775 (83%)

Composition 0.313 (36%) 0.364 (17%)

Total 0.869 2.139

Table 7. Decomposition of Aggregate Inequality Change by GJ Model Within-Group Subgroup Composition 0.290 (62%) 0.033 (7%) 0.051 (7%) -0.007 (-1%) Across-Group Divergence Composition 0.014 (3%) 0.131 (28%) 0.479 (66%) 0.203 (28%) Total 0.468 0.725

Thailand GJ

In sum, with respect to overall income growth, both models are o in explaining the magnitude, one way or the other, but both do reasonably well in explaining the decomposition of the data. The LEB model does attribute very little to the composition eect and most to within-group growth, just as in the Thai data, when wage earners and farmers are grouped together, as the model would dictate. The GJ model also does

37

well in accounting for compositional growth but it tends to over-emphasize the within-group growth eect. With respect to total inequality change, both models are again o in magnitudes, due to their magnication of divergence eect and dwindling of subgroup inequality eect, in particular for the GJ model. However, both are remarkably close to the data in terms of the contribution of population shifts to inequality. While the Kuznets eect is small for occupational shift and larger for nancial deepening, LEB and GJ are able to rationalize those magnitudes, respectively. Using the root-mean-squared-dierence metric on the decomposition results of Tables 4 to 7, Table 8 summarizes the tting performance of each model for the overall and decomposed changes, separately for growth and inequality change. This summary metric suggests that the LEB model ts the decomposition of income growth better than the GJ model and, though the metric is close, also for inequality change. Table 8. RMSD for Decomposition of Change Income Growth 0.520 1.017 Inequality Change 0.227 0.263

LEB GJ

5.4

End-of-Sample Test

As our nal criterion of model evaluation, we formally test each model, comparing the actual Thai income distribution in 1996 with the income distributions at that date, predicted by the models. This is an out-ofsample test in the sense that we estimated the models using the data only in initial years and do the testing on data in nal year. 5.4.1 Nonparametric Comparison of Distributions

The income distribution in the data is not necessarily close to some a priori parametric form. The income distributions from the models are endogenously determined, evolving over time with interacting features and it is also hard to expect to have parametric forms of these income distributions. Thus we compare distributional shapes between models and data in a nonparametric way. Let X1 , X2 , ..., Xn be an independent random sample from some xed identical underlying distribution F . The empirical distribution function Fn of this data is the random function dened as 1X Fn (x) = 1(Xi x), f or each x, n i=1
n

(56)

where 1() is an indicator function such that 1(Xi x) = 1, if Xi x, and 1(Xi x) = 0, otherwise. Fn is a cumulative distribution function constructed by assigning a mass of 1/n at each observation Xi . Fn (x) serves

38

as an unbiased and consistent estimator of the underlying F (x), with asymptotic normality (Darling 1957): n {Fn (x) F (x)} From the Glivenko-Cantelli lemma, we know that sup
<x<

N (0, F (x)[1 F (x)]).

|Fn (x) F (x)| 0, with probability 1 as n .

This suggests that the distance between the true distribution and the empirical distribution converges to zero almost surely as the sample size gets larger. These results allow us to test the goodness of t of the model to the data. Suppose a theory suggests a distribution function G as a data generating process for the underlying distribution F . We would like to test if this claim is statistically true or not by considering the goodness-of-t test with the following null hypothesis H0 : F = G, almost everywhere. (57)

Kolmogorov (1933) introduced the following sup-norm metric measuring distributional distance using the empirical distribution: Kn = n sup
<y<

|Fn (y ) G(y )|

Under the null hypothesis in (57), Kn must be small in large samples. The null is rejected if Kn is suciently large. Note that the distribution of Kn is independent of G(x) if the null is true, i.e., the test is distribution free. Indeed, Kolmogorov proved that when G is continuous, the asymptotic distribution of Kn has the following form:
n X

lim Pr {Kn < x} (1)j exp(2j 2 x2 ), at each x

j =

When we know the parametric form of G, we may compute Kn and apply this test statistic. Both the LEB and GJ models, however, do not have parametric forms for the aggregate income distribution. Fortunately, Smirnov (1939) extended the above results, comparing two nonparametric empirical distributions from two independent samples. Let Gm be the empirical distribution function constructed from the observed income prospect (Y1 , Y2 , ..., Ym ) generated from models. The test statistic is now given as: KSm,n = p mn/(m + n) sup
<y<

|Fn (y ) Gm (y )| .

(58)

Smirnov (1939) shows that the limiting distribution of KSm,n has the following form:
m,n X j =1

lim Pr {KSm,n < x} (1)j 1 exp(2j 2 x2 ), at each x 39

= 12

We use this Kolmogorov-Smirnov statistic as our goodness-of-t measure in comparing the distributional shapes of income from the theories and the data. 5.4.2 Test Results

The KS statistic in (58) for the univariate distribution can be easily extended to the multivariate distribution considering y as a random vector. Also the limiting distribution of the KS statistic, univariate or multivariate, is known to be a Brownian bridge, for which the critical value can be calculated from simulation and we can, in principle, test the closeness of joint distribution using the KS statistic.17 Thus, we could have tested the null for the joint distribution of the entire sequence of income distributions after initial periods. However, we do not pursue this joint test but instead compare the end-of-period income distributions for two reasons. First, as is clear in dening the KS statistic using sup norm, the rejection of the null for a marginal distribution implies the rejection of the null for the joint distribution. Considering the curse of dimensionality of nding the supremum over joint distribution of multiple-period income distributions, we rst attempt to test the closeness of a marginal distribution at a given period. Second, by focusing a distributional shape at a given period, we can better identify the underlying sources of discrepancies between the model and the data. In selecting the specic date to be xed, we choose the nal date to make the model economy relatively independent from the (estimated) initial distribution of wealth so that we can compare relatively genuine shape of model income distribution with the actual data. The second reason is related to the spirit of out-of sample test or post-sample forecasting. The Kolmogorov-Smirnov statistics are 0.493 for the LEB model and 0.163 for the GJ model, and the pvalues are virtually zero up to fourth decimal. Recall that the null hypothesis is that the distribution generated from the model is equal to the distribution observed in the data, and hence, to accept the null we would like a high p-value. Thus this null hypothesis for an end-of-period income distribution, hence for a joint distribution of entire sequence of income distributions as well, is rejected for both models by any standard statistical signicance level, though the GJ model does better in the sense that KS statistics is smaller for GJ than for LEB. A more detailed comparison of the shape of the income distribution functions reveals the underlying discrepancies. Figure 20 overlays the empirical distribution functions of LEB income onto that of the Thai data, both in 1996. The dierence is largest at the bottom left tail and also somewhat in the middle of the distribution. The concentration of wage earners who have the same wage income at the low end in the LEB economy is the source of the huge dierence. That is, the LEB distribution pops up when it reaches that low uniform wage while Thai income is more continuously distributed at the bottom half due to actual wage variation in the Thai data. Related, constancy in subsistence income in the traditional sector in the LEB model
17 We

thank Roger Moon for this suggestion.

40

does not capture the apparent variation in farm prot in the Thai data. The variation of LEB income at other positions is mainly due to variation in prot income in the business sector. The jump in the middle of the LEB distribution is due to the segmentation in income prospects between constrained entrepreneurs and unconstrained entrepreneurs. Unconstrained entrepreneurs with no internal heterogeneity all do the same thing in production and hence have virtually identical earnings, apart from interest income. Finally, the range of the support of the LEB income distribution is narrow relative to the actual Thai income distribution. In sum, the LEB model is too simple in its modeling of wage, farm and business income, and this misspecication seems serious in the sense that the above test rejects the LEB model. Figure 21 compares the empirical distribution functions of 1996 GJ income and 1996 Thai income. Evidently the matching performance of the GJ model is relatively better. Here we do not observe dramatic discrepancy. Income variation in the GJ model is, now in contrast, smoother overall. There may be in the model too much variation in income relative to the data, due perhaps to a relatively high variance of idiosyncratic shocks for non-participants in the GJ model. However, the GJ model still does not capture income at the extreme low end and the GJ income distribution is right-shifted relative to the Thai income distribution. More telling, the curvature in the middle part of the GJ income distribution is almost at. This is where we observe the largest discrepancy. The GJ model does not produce enough variation at the middle income levels. In fact, the GJ model tends to produce a twin peak in the income distribution, that is, a dip in the center of income distribution. Two factors may be at work. First and again missing in the GJ model is a specication of wage income. Second, there is a uniform critical level of wealth that determines entry in the GJ model, and those who have entered tend to move away from the center, leaving an indentation. [Figures 20 - 21 Here]

Conclusion

Both the LEB and GJ models deliver growth of income with an overall increase in inequality, a phenomenon observed in many countries and the motivation of these and other theoretical contributions. Quantitatively, the match at estimated parameter values with the Thai data is not terrible, indeed surprisingly good at times, and in both models as in the data, aggregate income and inequality dynamics can be understood as coming from choices that are constrained by wealth, producing income gaps across income-status groups and relatively slow but eventual population shifts toward high-income sectors. However, the LEB model does poorly in early periods and overall its levels of inequality are too high and changes too sharply in the end. It does perform well for the second decade. The GJ model appears to be more consistent with overall growth over the two decades in trend but has far too much volatility throughout,

41

due to the large magnitude of the modeled aggregate shock at the chosen parameter values. Thus the overall performance in tting income growth is better for LEB than for GJ. However, the prediction for the levels of inequality is better in GJ than in LEB. But the pattern of inequality movements is somewhat contrary, for both models, to what we see in the data. With respect to subgroup dynamics, within and across subgroups, both models produce income gaps across the target categories which are far too large relative to their actual magnitude in the Thai data, and both models dramatically under-predict the levels of subgroup inequality, drawing too much attention to the chosen heterogeneous categories in population. The LEB model is consistent with a declining income gap between entrepreneurs and non-entrepreneurs, and the GJ model is consistent with an in increasing income gap between participants and non-participants. In other respects, the GJ model seems to outperform the LEB model in generating the growth and inequality dynamics within and across subgroups. The GJ model is able to produce co-movements across participants and non-participants, as in the Thai data, while the LEB model tends to produce the opposite, across entrepreneurs and non-entrepreneurs, unlike the Thai data. Likewise, the GJ model is able to produce higher inequality among non-participants than participants, which is somewhat true in the Thai data, while the LEB model predicts more inequality among non-entrepreneurs, unlike the Thai data, which always displays the opposite. Both models are o in explaining the magnitudes of the overall income growth, but both do reasonably well in explaining the compositional eects. The LEB model attributes very little income growth to the composition eect and most to within-group growth, just as in the Thai data. The GJ model also does well in explaining the compositional growth, but it tends to over-emphasize the within-group growth eect. With respect to overall inequality change, both models are again o in magnitudes, but both are close to the data in terms of the contribution of population shifts. The eect of population shifts from low-income categories to highincome categories on inequality, that is, the Kuznets eect on inequality, is small for occupational shifts and large for nancial participation shifts as in the data, and LEB and GJ are able to rationalize those magnitudes, respectively. However, both models are radically o in the contributions to inequality of the income gap changes and of the inequality changes within subgroups, particularly so for GJ. Finally, in terms of the end-of-sample prediction for the shape of income distribution, both models suer from large discrepancies. LEB has too much homogeneity at the low end, stemming from a uniform wage and common income from subsistence agriculture, and too much homogeneity among rms unconstrained by wealth. GJ has too much heterogeneity at the low end of the income distribution, while failing to predict very poor segmentation of the Thai economy, and yet too little heterogeneity in the middle of the income distribution, due to homogeneity in the transactions costs, the presumed xed costs of entry to the formal nancial sector. However, the GJ model does better in its prediction of the shape of the Thai income distribution than the LEB 42

model. The formal Kolmogorov-Smirnov statistic suggests that the margin of probability accepting the null that the income distribution from a model is identical with the true data generating mechanism is higher for GJ than LEB, though this null is rejected for both models at any standard level of signicance. The better performance of the GJ model in tting the distributional shape than the LEB model is consistent with the better t of the aggregate inequality level path. A crude summary would be that the LEB model performs better than the GJ model in explaining the dynamic, aggregate and decomposed features in Thailand for the levels and growth of income. Both models are close in predicting the decomposed inequality changes, though the LEB model has a slight edge. However, the GJ model does better than the LEB model in explaining the levels of inequality as well as the shape of income distribution. We cannot conclude that one model strictly dominates the other. However, our goal is an increased understanding of growth and inequality as observed in actual economies and the construction of yet better models for these economies. Of the major lessons learned for Thailand, we would repeat the deciencies of each model, one at a time, as enumerated above and thus begin the process of introducing new elements. Some of them are obvious, such as greater heterogeneity in rm technologies, in labor and agricultural productivity, and in transactions costs and less sharing of idiosyncratic shock. Yet surely the goal is not heterogeneity per se but greater understanding achieved by stylized models. Further, it would be misleading to say that the models always overdo the chosen categorization, as with income dierentials, as both do remarkably well with Kuznets eects on inequality and they can under-estimate the contribution of population shifts, or over-estimate within group growth, in explaining income growth. Likewise, we would surely introduce human capital into each model to create more diversity. Indeed, Jeong (2001) estimates the empirical importance of educational expansion in accounting for the growth and inequality in Thailand. We could also try to blend the two models together, i.e., endogenous nancial sector participation with wealth-constrained occupation choice. But again this is no small task. We are to be able to estimate the models from micro data and simulate dynamic paths. Moreover such a respecication would not necessarily guarantee a closer t of such a hybrid extended model with the data. General equilibrium interactions, i.e., income dierentials and population shifts across income-status groups, are subtle and dicult to predict a priori. Even as they stand, we are surprised at the rich and complicated paths of income growth and inequality change that the two stylized models are able to generate.

43

Appendix
Table A.1 reports the weight coecients, which comes from the eigenvector of the rst principal component in principal component analysis, that are used in constructing the physical wealth proxy variable A in each year of 1976, 1981 and 1996 in Section 3. This single proxy variable explains 26-36% of the total variation of seventeen assets.

Table A.1 Weight Coecients for Physical Wealth Proxy Assets Phone Sofa Bed Stove Refrigerator Iron Electric Pot Radio TV Motor Boat Car Motorcycle Sewing Machine Room in House Private Water Gasoline Cooking Equipment Electricity 1976 0.239 0.262 0.274 0.311 0.322 0.347 0.297 0.032 0.306 0.003 0.174 0.110 0.225 0.003 0.199 0.235 0.327 1981 0.197 0.269 0.279 0.314 0.325 0.318 0.317 0.107 0.323 0.014 0.168 0.126 0.168 0.000 0.277 0.278 0.251 1996 0.243 0.280 0.292 0.347 0.356 0.323 0.282 0.209 0.264 0.015 0.175 0.155 0.106 0.000 0.179 0.335 0.151

Table A.2 reports the nine parameter sets for GJ = (q, , , , , , , , ) in the GJ model and their associated average log likelihood values, which determines set 13 as the best parameter set in the likelihood sense. Table A.2 Log Likelihood Values of GJ model Experiments Experiment 11 12 13 21 22 23 31 32 33 q 5 5 5 5 5 5 5 5 5 1.132 1.132 1.132 1.132 1.132 1.132 1.132 1.132 1.132 0.885 0.885 0.885 0.885 0.885 0.885 0.885 0.885 0.885 1.0 1.2 1.5 1.0 1.2 1.5 1.0 1.2 1.5 1 1 1 1 1 1 0.92 0.92 0.92 0.1 0.1 0.1 0.3 0.3 0.3 0.1 0.1 0.1 2.9 2.9 2.9 2.804 2.804 2.804 3.234 3.234 3.234 -0.05 -0.05 -0.05 -0.25 -0.25 -0.25 -0.05 -0.05 -0.05 0.05 0.05 0.05 0.25 0.25 0.25 0.05 0.05 0.05 Log Likelihood -0.1938 -0.1659 -0.1407 -0.2373 -0.2310 -0.2033 -0.2308 -0.2188 -0.1589 Zero-Probability Event 3% 4% 6% 2% 2% 4% 2% 3% 5%

44

References
[1] Bourguignon, Francois (1979), Decomposable Income Inequality Measures, Econometrica V. 47, pp. 901-920 [2] Darling, D. A. (1957), The Kolmogorov-Smirnov, Cramer-von Mises Tests, Annals of Mathematical Statistics, V. 28, pp. 823-838 [3] Fisher, Ronald A. (1925), Theory of Statistical Estimation, Proceedings of Cambridge Philosophical Society, v. 22, pp. 700-725 [4] Frisch, Ragnar (1933), Editorial, Econometrica, v. 1, pp. 1-4 [5] Gine, Xavier, and Townsend, Robert M. (2001), Evaluation of Financial Liberalization: A General Equilibrium Model with Constrained Occupation Choice, mimeo, University of Chicago [6] Granger, Clive W. J. (1999), Empirical Modeling in Economics: Specication and Evaluation, Cambridge University Press [7] Greenwood, Jeremy and Jovanovic, Boyan (1990), Financial Development, Growth, and the Distribution of Income, Journal of Political Economy v. 98, pp. 1076-1107 [8] Jeong, Hyeok (2001), An Assessment of Relationship Between Growth and Inequality Using Micro Data from Thailand, mimeo, University of Southern California [9] Kolmogorov, A. N. (1933), Sulla determinazione empirica di una legge di distribuzione, Giornale dell Istituto Italiano delgi Attuari V.4: 83-91 [10] Kuznets, Simon (1955), Economic Growth and Income Inequality, American Economic Review, Papers and Proceedings v. 45, pp. 1-28 [11] Kydland, Finn E., and Prescott, Edward C. (1982), Time to Build and Aggregate Fluctuation, Econometrica, v. 50, pp. 1345-1370 [12] Lloyd-Ellis, Hew and Bernhardt, Dan (2000), Enterprise, Inequality, and Economic Development, Review of Economic Studies, V.67: 147-68 [13] Mookherjee, Dilip and Shorrocks, Anthony F. (1982), A Decomposition Analysis of the Trend in UK Income Inequality, Economic Journal v. 92, pp. 886-902 [14] Paulson, Anna and Townsend, Robert M. (1999), Entrepreneurship and Liquidity Constraints in Rural Thailand, mimeo, Northwestern University and University of Chicago 45

[15] Shorrocks, Anthony F. (1980), The Class of Additively Decomposable Inequality Measures, Econometrica v. 48 pp. 613-625 [16] Shorrocks, Anthony F. (1984), Inequality Decomposition by Population Subgroups, Econometrica v. 52, pp. 1369-1385 [17] Smirnov, N. (1939), On the Estimation of the Discrepancy between Empirical Curves of Distribution for Two Independent Samples, Bulletin Math ematique de lUniversit e de Moscou, v. 2: 3-16 [18] Townsend, Robert M. (1997) Micro Survey Data in Rural Thailand [19] Townsend, Robert M. and Ueda, Kenichi (2001), Transitional Growth with Increasing Inequality and Financial Deepening, mimeo, University of Chicago and IMF

46

1976 .4

1996

.3

.2

.1

0 0 50000 100000 wealth in 1990 baht 150000

Figure 1.1 Wealth Distributions

1996 kernel density - 1976 kernel density

.2

-.2

-.4 -50000 0 50000 normalized wealth 100000 150000

Figure 1.2 Spread of Wealth Over Time

farmer worker 56

entrepreneur

(%)

14 76 81 86 year 88 90 92 94 96

Figure 2. Trend of Occupational Composition

farmer worker 80

non-farm business

60

(%)

40

20

0 1 2 3 4 5 6 wealth deciles 7 8 9 10

Figure 3.1. 1976 Occupational Composition by Wealth Class


farmer worker 80 non-farm business

60

(%)

40

20

0 1 2 3 4 5 6 wealth deciles 7 8 9 10

Figure 3.2. 1996 Occupational Composition by Wealth Class

27

(%)

6 76 81 86 year 88 90 92 94 96

Figure 4. Trend of Financial Participation Rate

50

40

30

(%)

20

10

0 1 2 3 4 5 6 wealth decile 7 8 9 10

Figure 5.1. 1976 Financial Participation Rate by Wealth Class

50

40

30

(%)

20

10

0 1 2 3 4 5 6 wealth decile 7 8 9 10

Figure 5.2. 1996 Financial Participation Rate by Wealth Class

Average Income 6575

Theil-L .6548

Average Income

3224 76 81 86 year 88 90 92 94 96

.3930

Figure 6.1 Levels of Average Income and Income Inequality

Average Income .15

Theil-L

.1

.05

-.05 81 86 88 year 90 92 94 96

Figure 6.2 Growth Rates of Average Income and Income Inequality

Theil-L

entrepreneur farmer 10000 8000 6000 4000 2000 76 81

worker .15 .1 .05 0 -.05 86 88 year 90 92 94 96

entrepreneur farmer

worker

81

86

7.1 Average Income by Occupation


entrepreneur 3.5 3 2.5 2 1.5 76 81 worker

7.2 Average Income Growth by Occupation


entrepreneur farmer .7 .6 .5 .4 .3 worker

88 90 year

92

94

96

7.3 Income Ratio to Farm Income

86 88 year

90

92

94

96

76

81

7.4 Income Inequality by Occupation

86 88 year

90

92

94

96

Figure 7. Occupation Group Dynamics in Thailand

participant 15000

non-participant .15 .1 .05

participant

non-participant

10000

5000 0 0 76 81 -.05 86 88 year 90 92 94 96 81 86

8.1 Average Income by Participation

8.2 Average Income Growth by Participation


participant non-participant

88 90 year

92

94

96

participant to non-participant

3 2.8 2.6 2.4 2.2 76 81

.7 .6 .5 .4 .3

8.3 Participation Premium

86 88 year

90

92

94

96

76

81

8.4 Income Inequality by Participation

86 88 year

90

92

94

96

Figure 8. Participation Group Dynamics in Thailand

Figure 9. LEB Occupation Choice Map


1

0.9

Workers and Subsisters


0.8

1976 1981

0.74
0.7

0.6

x setup cost

0.5

Constrained Entrepreneurs
0.4

0.3

0.2

Unconstrained
0.1

Entrepreneurs

0 0 0.5 1 1.5 2

2.18

2.5

b inheritance

1 .895 .854 Cumulative Distribution

0 0 Setup Cost .74 .8 1

Figure 10. Estimated Distribution of Setup Cost in LEB

1976 1

1981

.5

0 Wealth in Estimated Scale 2 2.18

Figure 11. Estimated Initial Wealth Distributions in LEB

Thailand 1.5

LEB 1

Thailand

LEB

.5

.5

0 75 80

-.5

12.1 Average Income Level


Thailand .8 .7 .2 .6 .5 .4 75 80 85 year 90 95 .1 LEB .3

85 year

90

95

75

80

12.2 Average Income Growth Rate


Thailand LEB

85 year

90

95

0 75 80

12.3 Income Inequality

12.4 Fraction of Non-farm Entrepreneurs

85 year

90

95

Figure 12. Aggregate Dynamics Comparison in LEB

Thailand 1.5

GJ 1

Thailand

GJ

.5

.5

0 75 80

-.5

13.1 Average Income Level


Thailand .8 .7 .2 .6 .5 .4 75 80 85 year 90 95 .1 GJ .3

85 year

90

95

75

80

13.2 Average Income Growth Rate


Thailand GJ

85 year

90

95

0 75 80

13.3 Income Inequality Trend

13.4 Fraction of Participants

85 year

90

95

Figure 13. Aggregate Dynamics Comparison with GJ

Entrepreneur .8

Non-entrepreneurs .3

Entrepreneur

Non-entrepreneurs

.6

.2

.4

.1

.2

0 75 80

-.1

14.1 Average Income by Occupation


Entrepreneur to Non-entrepreneur
14

85 year

90

95

75

80

14.2 Average Income Growth by Occupation


Entrepreneur .8 Non-entrepreneurs

85 year

90

95

10

.6

.4 6 .2 2 75 80

14.3 Income Gap Ratio

85 year

90

95

75

80

14.4 Income Inequality by Occupation

85 year

90

95

Figure 14. Occupation Group Dynamics in Thailand

Entrepreneur .8

Non-entrepreneurs .3

Entrepreneur

Non-entrepreneurs

.6

.2

.4

.1

.2

0 75 80

-.1

15.1 Average Income by Occupation


Entrepreneur to Non-entrepreneur
14

85 year

90

95

75

80

15.2 Average Income Growth by Occupation


Entrepreneur .8 Non-entrepreneurs

85 year

90

95

10

.6

.4 6 .2

2 75 80

15.3 Income Gap Ratio

85 year

90

95

75

80

15.4 Income Inequality by Occupation

85 year

90

95

Figure 15. Occupation Group Dynamics in LEB

Participants 10

Non-Participants 1

Participants

Non-Participants

.5 5 0

0 75 80

-.5

16.1 Average Income by Participation


14

85 year

90

95

75

80

16.2 Average Income Growth by Participation


Participants .8 Non-Participants

85 year

90

95

Participants to Non-Participants

10

.6

.4 6 .2

2 75 80

16.3 Income Gap Ratio

85 year

90

95

75

80

16.4 Income Inequality by Participation

85 year

90

95

Figure 16. Participation Group Dynamics in Thailand

Participants 10

Non-Participants 1

Participants

Non-Participants

.5 5 0

0 75 80

-.5

17.1 Average Income by Participation

85 year

90

95

75

80

17.2 Average Income Growth by Participation


Participants .8 Non-Participants

85 year

90

95

Participants to Non-Participants

14

10

.6

.4 6 .2

2 75 80

17.3 Income Gap Ratio

85 year

90

95

75

80

17.4 Income Inequality by Participation

85 year

90

95

Figure 17. Participation Group Dynamics in GJ

Within Total .8 .6 .4 .2 0 75 80

Across .8 .6 .4 .2 0

Within Total

Across

18.1 Thailand

85 year

90

95

75

80

18.2 LEB

85 year

90

95

Figure 18. Decomposed Inequality Dynamics in LEB

Within Total .8 .6 .4 .2 0 75 80

Across .8 .6 .4 .2 0

Within Total

Across

19.1 Thailand

85 year

90

95

75

80

19.2 GJ

85 year

90

95

Figure 19. Decomposed Inequality Dynamics in GJ

Thailand 1

LEB

.5

0 0 5 10 Income 15 20

Figure 20. Comparison of 1996 Aggregate Income Distribution: LEB

Thailand 1

GJ

.5

0 0 5 10 Income 15 20

Figure 21. Comparison of 1996 Aggregate Income Distribution: GJ

You might also like