Changes in poverty and income distribution in Argentina: evidence via decompositions. very preliminary and incomplete

Transcription

1 Changes in poverty and income distribution in Argentina: evidence via decompositions. very preliminary and incomplete Florencia López Bóo Department of Economics, University of Oxford Manor Road, Oxford, OX1 3UQ, UK. November 5, 2007 Abstract Over the nineties, Argentina experienced a period characterized by increasing income inequality and poverty as per households surveys data, along with increasing growth as measured by the National Accounts. We analyze this issue along with the determinants of household incomes and inequality. We also introduce an axiomatically modified Datt-Ravallion decomposition, that separates changes in poverty rates into mean and inequality components. Earlier studies applying the Datt-Ravallion decomposition to other countries found the growth component to explain the largest part of observed changes in poverty. We show that this only holds for Argentina for the 2 years of crisis over the 11 year period we analyze. Previous decompositions results are extended in three key ways. First, the empirical density function is used to calculate the inequality component, without assuming a particular functional form for the Lorenz curve. Second, both components are recomputed without the vaguely defined Datt-Ravallion residual, which improves interpretability. Third, for the first time standard errors for the decomposition are provided. JEL Classification Codes: C16, D63, I30, I31, I32, O54 Keywords: decomposition of changes in poverty, poverty measures, incomes, inequality and growth. florencia.lopezboo@economics.ox.ac.uk. I am grateful for the early supervision of Prof. Sudhir Anand and substantive comments of Francis Teal. I also thank CSAE, CIDED and CEDLAS seminar participants, Nuffield Workshop in Economic and Social History, LACEA conference, and QEH seminar participants for many helpful suggestions. All errors are mine. 1

2 Introduction During the early 1990s Argentina had one of the lowest Gini coefficients of Latin America (i.e. Gini = 0.40). Despite this, total household income inequality during the 1980s and, in particular, the 1990s has risen more than in other Latin-American country (Li and Zou 1998, Szekely 2003). The Gini index actually ended up at 0.52 in 2001, after a three-year recession and a major macroeconomic crisis. 1 Szekely shows the increase in inequality across countries by estimating a regression for each country separately, being the dependent variable the Gini coefficient, and the independent variable, a year trend (t=1, 2,..n). We will perform similar regressions at the urban area level. The changes in poverty were also dramatic. Argentina had an annual per capita income officially calculated at (dollars) US$7418 per person in And, indeed, GDP per capita slumped from that value to US$2671 per capita in 2002, after the crisis. Yet, despite its relative wealth, it was a country with a high degree of poverty seemingly explained by the unequal distribution of incomes and the fact that growth of incomes from National Accounts was not being reflected in household income growth as per household surveys (Altimir and Beccaria 2001, Gasparini, Marchionni, and Sosa Escudero 2001). According to official figures, the overall poverty rate in 2001 was 35.4% of the individuals, and rates in the poorer Northeast and Northwest regions exceeded 50% (W.Bank 2000a). 2 The relationship between growth, inequality and poverty has been, and remains, one of the most controversial topics in development economics. From the 1950s to the early 1970s, the debate emphasized the possible trade offs between growth and income inequality. This derived from the Kuznets famous inverted U hypothesis, which posited that inequality rises during the initial phases of development and then declines after some crucial level is reached (Kuznets 1955). 1 Li, Squire and Zou found that the highest Ginis in Latin America for the period were: Brazil: 0.578, Mexico: 0.546, Honduras: 0.545, Panama : 0.52, Chile : and Colombia: The lowest one were Dominican Rep.: 0.469, Costa Rica: 0.46 and Venezuela: 0.44, which lie at the bottom of their not exhaustive list of Latin-American countries (Argentina was not in their sample). By taking the Argentine Gini coefficient of the total household income (without adjustment for underreporting, which is the same concept of income than the one considered by Li, Squire and Zou) in the period for the 16 urban municipalities would yield a Gini of 0,44 close to Venezuela, the lowest Gini in their Latin-American list. 2 Poverty rates estimations for Brazil were 30%, for Mexico 15% and for Chile: 15% (World Bank 2000) 2

3 In place of the Kuznets curve, the recent literature points to different empirical regularities. As more time-series data becomes available, it appears that: first, aggregate inequality does not typically change dramatically from year to year. 3 Second, that results from cross-country studies are often not applicable to single countries given the existence of measurement errors and country-specific effects (Anand and Kanbur 1993, Ravallion 2001a, Dollar and Kraay 2002). Finally, that there is a sizeable heterogeneity across countries and time in the gains to the poor from a given rate of growth. The validity of the first point will be explored via both regression-based approach used in the recent literature along with a novel decomposition approach (Huppi and Ravallion 1991, Bruno, Ravallion, and Squire 1998, Li and Zou 1998). The effects of income inequality on poverty within a given country will be analyzed by using decompositions of changes in poverty rates. The specific question addressed in this paper is the following: how far the modest growth as measured per household surveys per se has (not) helped to reduce absolute poverty, given a significant change in inequality. To phrase this question slightly differently, how good was growth on average and what was the magnitude of the trickle down growth effect? To answer the first question one could investigate the change in the level of poverty arising from a change in economic growth, or due to a change in inequality. For instance, the magnitude of the inequality component may provide a useful measure of the degree of trickle down growth. I will also inspect whether changes in the inequality component have been associated to the way the mean was changing (growth spells vs. recession spells). Besides the empirical motivation, there is a methodological-policy concern as well. In order to understand the effectiveness with which growth and inequality changes has translated into poverty reduction through time; it may be useful to quantify its contribution in a simple and generalized way. The magnitude of the growth and inequality components may identify directly the effect of a particular phenomenon, which in turn may help to figure out the impact of a particular development policy. There is an extensive literature in Argentina on the determinants of inequality and poverty during the 1990s (FIEL 1999). However, so far none of them has explicitly explored and disentangled the impact of growth and inequality on poverty. In particular, there is no current literature (to the author s knowledge) that uses a precise, simple yet also general specification of changes in the distribution of incomes. Nor has any literature yet compared different decompositions empirically. These method will then be illustrated with microdata from Argentina. The data comes from the Permanent Household Survey EPH (Encuesta Permanente de Hogares) from 1992 to The EPH is 3 One study of panel data for 49 countries found that 91.8% of the variance in inequality was due to cross-country variance and only 0.85% was attributable to variance over time (Li, Squire and Zou 1998). When performed on a sub-sample of low and middle-income countries, the figures were 93.1% and 1.4% respectively. 3

4 a household survey conducted by INDEC (National Institute of Statistics and Census) which collects information on labor incomes, profits, interests, rents and pensions at the household level. The paper is structured as follows. Section 1 gives an overview of welfare changes in Argentina by using traditional methods. Section 2 explores the determinants of income and inequality trends by means of regressions using a new identification strategy for the inequality regressions. Subsequently, in section 3 a decomposition of changes in poverty into growth and distributional effects, by slightly changing the Datt-Ravallion method will be expounded, with the precise objective of looking at which component had contributed the most to increases (decreases) in poverty rates. The decomposition will provide a measure of the (infinitesimal) change in the poverty function with respect to inequality or growth, independently taken (ceteris paribus). This method will be contrasted to the regression based analysis in section 2 and to the Datt Ravallion and Shorocks decompositions. Section 4 reports results of the decompositions. Finally, Section 5 gives possible extensions and Section 6 concludes. 4

5 1 Overview of welfare changes in Argentina during the 1990s 1.1 The data I use the Argentine Permanent Household Survey (Encuesta Permanente de Hogares or EPH), which is undertaken semi-annually (May and October) in urban areas by the Instituto Nacional de Estadisticas y Censos (INDEC). The survey covers 31 urban areas at present, being representative of, approximately, 62% of the total urban population (INDEC 2000, W.Bank 2000b). 4 Thus, all the province s capitals and all the cities with more than 100,000 inhabitants according to the Population and Housing National Census of 1991, plus Alto Valle del Rio Negro (an urban-rural centre) are now covered by the survey. A small but non-negligible fraction of the population resides in rural areas and, unfortunately, relatively little is known about them as the EPH does not have systematic data collection on rural residents. The World Bank conducted a special survey after the 2002 crisis, and reported some information on small rural towns (Fiszbein, Giovagnoli, and Aduriz 2003) and the 1997 Encuesta de Desarrollo Social (EDS), conducted by the Information, Monitoring and Evaluation Programme System (SIEMPRO or Sistema de Informacion, Monitoreo y Evaluacion de Programas) and INDEC, is a valuable source of information on rural areas too. Since the rural population is excluded from these estimates, they may understate overall poverty levels, as what little evidence I have, suggests that poverty is higher in rural areas. 5 The structure of the survey is a stratified rotating panel (25% of the sample is replaced in each round), however this feature of the data will only be exploited in other chapters. It has to be noted that the replacement is always made keeping the characteristics from the population of previous samples. The population is sampled in two-steps. The units of the first phase are grouped in stratus defined based on the percentage of heads of household with complete primary education. These are called primary units (census ratio or blocks). Later a selection of primary units with a probability proportional to their size is performed (size is measured as quantity of private houses). In a second stage, houses are selected in a systematic way within the selected primary units, from a listing of houses in each of those units. Population and housing census provide the basic information in order 4 71% of Argentinas urban population lives in those 31 centers and Argentina urban areas represent 87.1% of the country, one of the largest shares in the world. While poverty estimates based on the EPH are usually called national, it should be stressed that the survey only covers large urban centres. The term urban refers to towns of 5,000 people or more. See accompanying Data Appendix for more details. 5 In 1998, rural poverty estimates for Salta and Misiones, two of the poorest provinces in the country, indicated poverty at approximately 75% for both provinces, while the urban poverty estimates were 40% and 45%, respectively (World Bank 2000). 5

6 to elaborate the sample framework. Approximately 40,000 households are surveyed each year. Following INDEC s methodology, for the basic poverty estimates and the analysis of the determinants of income poverty and inequality the data is considered as repeated cross-sections. Twelve May-rounds between 1992 and 2003 will be considered for this work. The urban areas we consider are: Gran La Plata, Gran Parana, Comodoro Rivadavia-Rada Tilly, Gran Cordoba, Neuquen-Plottier, Jujuy-Palpala, Rio Gallegos, Salta, San Luis- El Chorillo, Santa Rosa-Toay, Ciudad de Buenos Aires and Partidos del Conurbano. More details on the sampling, comparability and other issues are included in the Data Appendix. 1.2 Income measurement issues: sample, adjustment and misreporting Four components are needed for estimating absolute poverty: an income or expenditure estimate, a poverty line, a poverty measure and a definite population sample (Sen 1976). 6 In this section, the definition of the income aggregate, the selection of the sample, the adjustments made to the income measure and misreporting and non response issues are reviewed. In terms of sample selection and adjustments I follow FIEL and Gasparini (FIEL 1999, Gasparini and Sosa Escudero 1999). Appendix 1 gives more detail on adjustment for economies of scale and misreporting and Appendix 2 deals with poverty lines in detail while Appendix 3 and 4 define the poverty and inequality measures used in this paper. Definition of Income The definition of individual income taken by the EPH accounts for labor incomes (wage earnings and self-employed earnings), profits, interests, rents, scholarships, unemployment benefits and pensions. However, it ignores profits derived from capital wealth, proprietary ownership and durable goods such as vehicles. The EPH has various deficiencies in capturing capital income. Besides, modelling capital income and retirement benefits would be a difficult task, especially considering the scarce information provided by EPH. I will calculate different measures of household income (household adult-equivalent and household adult-equivalent adjusted by economies of scale) based on the household income measure, that is the sum of all individual income measures. 6 Consumption estimates based on expenditure data are a proxy for permanent income, and as such they are usually considered better metric of household welfare. The EPH, though, is mainly a labor markets survey and it does not collect this type of information. Expenditure surveys are only carried out every ten years in Argentina. 6

7 Valid observations The estimations were calculated on the basis that all household members answered in full the questions in the EPH survey referring to incomes. That is, partial incomes or failed interviews were all withdrawn. On average (for the 12 rounds), households with partial revenues or failed interviews were 9% of the total household sample; and zero total household incomes were around 1% of the same sample. 7 In part of the literature, observations with zero total household income considered valid by INDEC were used in the estimations of the equivalent household income statistics. However, in order to keep poverty and inequality numbers consistent with the official measures, households that did not declare revenues during the surveys period have been eliminated from the sample. The reasons are three: (i) obviously not receiving an income at all must be a short-run circumstance. To survive, a household must be receiving income from some source (i.e. charity or the State); (ii) the suspicion of underreporting is particularly severe in the cases where zero income is reported; and (iii) most of the members of families declaring zero incomes were inactive. Therefore, it was impossible to discern if they were poor people who give up to look for a job, elderly people without retirement benefits, or rich people who live from selling their assets. This selection could underestimate the effects of, for instance, unemployment on inequality and poverty. However, most of the unemployed are not eliminated from the sample because they are part of families in which at least one member receives a revenue. As usual, only workers with positive earnings are included in the individual labor income statistics. Price adjustments Given that the chosen poverty line is expressed in 1998 prices, all incomes were deflated by the 1998 Consumer Price Index. Therefore, all incomes are presented in real values by deflating nominal incomes by the consumer price index for the month when incomes reported in the survey where earned (April in our case for May waves). Prices are not uniform across regions either. INDEC provides information on regional prices that can be used to express incomes in real values of the Greater Buenos Aires (Ciudad de Buenos Aires plus Conurbano). We use incomes adjusted for regional prices throughout the analysis. Poverty and inequality are somewhat higher if we ignore that prices are lower outside the Greater Buenos Aires. Lastly and more difficult to address, given different consumption patterns across income ranges, the decision of using a unique price index for all deciles raises some concerns in terms of methodological soundness. Adjustments for demographic structure Many poverty reports are based on per-capita income disregarding the question of economies of scale. 7 The high rates of zero individual incomes are due to the existence of children, unemployed, inactive and retired people in many of the households 7

8 That means that what a family of two needs is exactly the half of what a family of four may need. With economies of scale a family of four with twice the income of the family of two is considered better off than the family of two. Therefore, per capita measures of income will understate the welfare of the family of four relative to the living standards of the family of two (Lanjouw and Ravallion 1995). Moreover, the needs of infants and children tend to be lower than those of adults. Measures of poverty based on per capita income rely on the estimation of the cost of basic needs for an average individual, but many large families do not consist of average individuals since they tend to have many children. Once again, not considering differences in need may lead to an over-estimation of the negative impact of family size, through the number of children, on poverty (Cowell and Mercader Prats 1999). INDEC has dealt with the issue of difference in needs, but not with economies of scale (INDEC 2002). We will then add to this discussion by comparing the adult equivalent measures with the ones adjusted by economies of scale. Unfortunately, there are no generally accepted methods for calculating equivalence scales. There are three main approaches for deriving equivalence scales: (i) the one relying on behavioural analysis to estimate equivalence scales, an approach that has generated a large literature, much of which is reviewed in Deaton (Deaton 1997) (ii) another approach using direct questions to obtain subjective estimates. 8 (iii) and, lastly, one approach that simply sets scales in some reasonable, but essentially arbitrary way. Deaton and Zaidi (2002) recommend, apart from the continuing use of per capita income, the arbitrary method (Council 1995, Deaton and Zaidi 2002, Browning 2004). 9 Adult equivalent Following INDEC s methodology to adjust incomes (which is an extension of the arbitrary method ), household composition is converted into male adult equivalents using standard conversion factors as the ones reported in Table 1 (see Apendix 1 for details and formulaes). The adult equivalent income aggregate for each household is constructed by first, summing up the income of all members living in the household, and then dividing the total household income by the total number of adult equivalents. Denoting by k i the number of members in household i, by ψ j the income of each one of them (j=1,..,k i ) and by q j the coefficient of equivalent adult corresponding to the age and gender of the member, total adult equivalent 8 One widely used subjective technique is the Leyden method pioneered by Van Praag (Van Praag and Warnaar 1997) 9 The most important critique to this approach is the one following the lead of Browning, Chiappori and Lewbel (2004). They argue that these equivalence scales assume that utility functions of the household and of a single individual can be comparable. They claim that individuals have utility, not households, so following the idea of the collective approach to household behaviour, they assume that each individual is characterized by his own utility function, so the only comparison they make is between the same person s welfare in different living arrangements. 8

9 income for household i, y i,q is defined as: ψj y i,q = (1) qj While not dealing straight away with economies of scale at the household level, this income aggregate is still better that the use of per-capita income for poverty measurement.following the same approach as the per capita income (which assigns to each individual of the household, the total income divided by the number of persons in that household), each adult equivalent income will be assigned to every individual of the household. The potential drawbacks of this approach are explained in Anand (1977). The most relevant drawback arises when household expenditures are mostly shared goods. In that case, total (instead of per capita) household income should be assigned to every individual (Anand 1977) Economies of scale Following Buhmann et al I incorporate economies of scale by raising (1) to 0.8 ( θ =0.8) which implies mild household economies of scale (Buhmann and et al. 1988, Gasparini and Sosa Escudero 1999, Deaton and Zaidi 2002). See Appendix 1 for more details. Non-response and misreporting One can suspect that the level of non-response is greater in the higher deciles of the distribution. If this is the case, and if one had considered those non-response families, there would have been a tendency to obtain higher inequality (and poverty) figures (Altimir 1986, Altimir and Beccaria 2001). Unfortunately, the Argentina s government has not published information of disposable income by sources. The only information available dates back to 1993 (Subsecretaria de Programacion Economica, 1998). With that information and data form ANSESS we estimated the structure of disposable income by sources in 1993 (wage earners, self-employed, entrepreneurs earnings, capital income, pensions and other transfers) and matched it with the EPH We apply the resulting coefficients to each income source (1.05 for pensions and transfers, 1.15 for wages, 1.3 for income of the self-employed, 4.1 for entrepreneurs income and 9.95 for capital income). Since I do not have information for other years, I apply the same coefficient to all EPHs. It has to be noted that for a given income source this procedure implies constant under-reporting among individuals and over time. However, Altimir et al (1986,2001) assume that the structure of the underreporting and the omissions did not change substantially from one year to the other in the sense of increasing or decreasing the relative difference between the measured inequality and the actual inequality. In 1992 I faced an additional problem, since there is no separate questions for entrepreneurs earnings. All these drawbacks 10 In the 1993 EPH is impossible to distinguish entrepreneur s earnings 9

10 imply that the adjustment wasquite weak. The change in poverty between 1992 and 2003 is smaller when considering the adjustment, but still quite substantial. Recorded inequality dramatically increases when applying the adjustment. However, recall that this simulation takes the source structure of disposable income fixed which is probably a strong assumption, given the indirect evidence on the increase in the share of capital income, profits and rents in national income during the 1990s Incomes trends National Accounts vs Household Surveys Real incomes are the arguments of all poverty, inequality and welfare measures. Thus, before indicators for these dimensions are computed, as it will be done in the next sections, some basic statistics on real incomes are shown. Table 2 shows real income for an aggregate of 12 urban areas for some selected years from 1992 to The real income reported by the EPH increase around 2% between 1992 and 1996 for my sample (not reported) 12 This change was in sharp contrast with the national accounts, according to which per capita GDP increased 8.9% in that period. The discrepancy may be due to increasing under-reporting in the EPH or overestimation of the GDP. It might also also be the consequence of an increase in the share of sources that are not well captured by the EPH-capital income, benefits and rents ( see Appendix 1). Between 1996 and 1998 as economy expanded, both adult equivalent income and per capita income grew by about 7% (6.9 for per-capita income and 6.6 for adult-equivalent income). A similar percentage is also reported by the national accounts (after taking into account 10% of inflation between 1993 and 1998 in Table 2). The crisis implied a 37.5% fall in the mean income reported by the EPH, which was higher than the 19% decline recorded by the national accounts Incomes by percentile It is important to note that incomes changes were not uniform across deciles. All income changes between 1992 and 2001 were clearly unequalizing as we can see with incomes of the 25th percentile going down all over the period for both income measures and p50 going down by 13% for the adultequivalent income, but with per-capita income slightly going up by 2.2% 11 For more details see summary of companion paper Underreporting and its impact on poverty and income distribution estimates in Argentina (Lopez Boo, 2004) in Appendix 1 12 Gasparini reports a fall of 7% when considering October waves for 16 urban areas instead of the 12 I use. The four urban areas I miss are San Juan, Santa Fe, Santiago del Estero and Tierra del Fuego, the firs three are relatively poorer than the others, and have experienced larger falls in mean income. 10

11 between 1992 and 1998 and then going down by 10% from 1998 to This might show some changes in household composition over this period. On the other hand, the adult equivalent incomes (per capita income) of the 75th percentile went up by 8% (12%) from 1992 to 1998, while it went down by 8% (9%) from 1998 to 2001 for those respective measures. In contrast the impact of the last economic crisis (as measured from 2001/2003) was rather neutral, as all deciles lost almost 30% of their income. The exception was the poorest decile, where the loss was 14%, probably as a consequence of the implementation of the Heads of Household Program (PJH) in the first semester of this program transfers $150 a month (around US$50) to (mostly) poor households. Now we turn into poverty and inequality trends. 1.4 Poverty Trends Poverty measures While many social indicators improved during the 1990s (for example, infant mortality dropped from 25 deaths per 1000 births in 1991 to approximately 18 deaths per 1000 births in 1998, poverty levels have remained stubbornly high despite the rapid GDP growth experienced in the 1991/1998 period (W.Bank 2000b). Table 3 shows the evolution over time of the official estimates of poverty rates for the GBA and Table 4 my own estimates of a diverse set of poverty measures for the full country (see Appendix 3 for details). Since 1990, there has been a substantial drop in poverty, with the overall headcount ratio for all urban areas dropping from 33.7% of all individuals during the hyperinflation in October 1990 to 16.1% in May 1994 (the lowest poverty rate during the 1990s). However from 1994 on, poverty rates started to increase steadily, mainly influenced by the impact of the financial Mexican crisis in 1995 (which produced a decline in GDP and a sharp rise in unemployment) and the Brazilian Devaluation in 1999 (Argentina s largest trading partner). By the end of 2001, the failure to honour the foreign debt and the devaluation of the peso, the national currency, have added to the poor performance of welfare indicators during the last decade. In October 2002, 54.3% of the population was living in poverty. Other standard poverty measures, which include the poverty gap ratio (PGR) and the Sen Index (SI), showed a similar trend (Table 4). It is worth noting that while in 1992/2001 H nearly doubled its value, the PGR and SI multiplied its value by almost three. Roughly speaking, people with income around the poverty line became worse off (given the increase in H) over this period, while the poorest were even worse affected (given increases in the PGR and SI). The rate of extreme poverty fell from 6.6% in October 1990 to 3.5% in October 1994, increasing thereafter to reach 25.2% of all individuals in 11

12 May Plausible explanations of these trends, besides the appreciation of the exchange rate and the collapse in 2001, appear to be both the lower employment-growth elasticity and the larger wage differentials by skill (see Chapter 2 of this thesis). The movements in poverty and extreme poverty mirrored a dramatic fall in household income (Table 2), and a clear deterioration of labour market conditions Stochastic dominance However, in order to obtain results that are consistent with different poverty lines and as it is common in the literature, one needs to examine stochastic dominance to analyze changes over time. Given a particular poverty measure P (.) and a particular poverty line, z, P (.) is able to determine which distribution has more poverty than the other. Unlike inequality measurement in which dominance theory has a long tradition, dominance analysis in poverty measurement is still a relatively new field (Atkinson 1987, Foster and Shorrocks 1988, Jenkins and Lambert 1998). Using poverty dominance analysis, it is possible to say that all members of the class of a general additive poverty measures, including P α measures (see Appendix 3 for definitions) would rank one income distribution as having more poverty than another, given z. Also, given the controversy which usually goes with setting z, dominance criteria encompassing many z would be very desirable because a given poverty measure belonging to the class of general additive poverty measures would rank one income distribution as having more poverty than another, for a range of z. Along these lines, Figure 1 shows how the Kernel Density Function for the distribution of household adult-equivalent incomes for 2001 (also called the poverty incidence curve) is almost everywhere at least as high as the distribution of adult-equivalent incomes for However, the 1998 curve crosses the 1992 distribution at approximately $130 (expressed in 1998 pesos), which is at about the 25th percentile of the distribution. This implies that 2001 first order dominates 1998, for every z between 0 (zero) and the maximal income, and for all members of the class of general additive poverty measures. However, we do not know if 1998 in turn first order dominates 1992 for incomes higher than those at the 25t percentile. As another piece of evidence, Figure 2 shows second order welfare dominance by means of the Generalized Lorenz Curve (GLC) for the distribution of household adultequivalent incomes (Shorrocks 1983). The GLC for 1992 is everywhere at least as high as the distribution of adult-equivalent incomes for 1998 and the 1998 curve is everywhere at least as high as the 2001 distribution. 14 This 13 In October 2002, unemployment peaked to 17.8% and, in addition, 20% of the economically active population became underemployed or worked part-time due to the lack of good full-time jobs. 14 Generalized Lorenz Curve=Lorenz curve of x times the mean of x (where x is 12

13 implies that 1992 second-order dominates 1998, and 1998 in turn secondorder dominates This will hold if and only if welfare in 1992 was higher than welfare in 1998 and welfare in 1998 was higher than welfare in 2001 (Fields 2001b). 15 This situation was clearly compounded by rising income inequality as we could observe in Section We analyze this matter further in the next section. a given distribution of incomes). 15 To endorse these results Jenkins and Lambert provide a new device, the three I s of poverty curve, which play a similar role to the Generalized Lorenz Curves. They describe income distribution and implement tests of whether one distribution dominates another. Those curves have been plotted as well. Figure 4 shows the curves with the poverty line set at $160 (1998 pesos, per adult-equivalent, per month) which illustrates the same trend depicted with the kernel density functions and the Generalized Lorenz Curves. 13

14 1.5 Inequality Trends There was a sharp increase in inequality between the 1980s and the end of 1990s. In Figures 2 and 3 and Tables 5 to 7 we see figures for the period we will concentrate on ( ). The only exception to this increasing trend was from 1990 to 1992 when the distribution became more egalitarian, due to the equalizing effect of the removal of the inflationary tax which tends to affect the poor most (Canavese, Sosa Escudero, and Gonzalez Alvaredo 1999). 16 Nevertheless, inequality began to increase again in 1993, jumping sizably in 1994, after the Mexican crisis. It must be noted that the change is quite modest in 1994 for the Gini index when considering both per capita income and adult-equivalent income, however that change is pretty big for the adult equivalent income adjusted by economies of scale. Finally, the second big jump happens in 2002, when the Gini index for adult equivalent incomes reaches 0.54, the highest value since reliable data is available. 17 Table 5 shows the Gini for our three measures of income (household adult equivalent income with and without economies of scale and household per capita income; while Table 6 shows the Gini for Labor income (from profits, self-employed and wage employees) and for Retirement benefits income. Regardless of definition of income, the Gini has gone up until 2002 and it went slightly down in The only exceptions occurs in 1995 (when the Gini goes down for the adult equivalent income with economies of scale, aescy) and in 1997 (when the Gini does not change for household per capita income). The same trends are observed regardless of the income inequality measure (see Table 7). For the nine inequality measures we consider here (Coefficient of Variation, Standard deviation of the logs, Mehran, Piesch and Kakwani measures, Theil index, Entropy, Mean log deviation and the Coefficient of Variation Square) the trend is exactly the same regardless of the income measures used, besides the exception noted above for the Gini in particular years. The myriad of reasons which may have caused these changes in income, poverty and inequality will be analyzed in the next two sections. Firstly, we look at regression based analyses in section 2 and secondly, we decompose changes in poverty between growth and inequality components, analyzing the residual term in a novel way. 16 Inflation fell from a peak of over 1300 percent in 1990, to less than one percent during 1996/98, jumping again to 41% in Real economic growth, on the other hand, increased from an average rate of -1.1% during the 1980s, to 5.8% during the 1991/98 period, then dropping to -10.9% in 2002, after the crisis. 17 The INDEC releases household data since 1974, but periodically only from 1980/82 on. 14

15 2 Determinants of income and inequality trends 2.1 Income regressions Tables 8 and 9 show the regressions of the log of incomes on a year trend, a set of household level, region-level, sector level and macro variables. As expected we observe a negative trend with the year variable, even when we augment the specification with region, sector and macro variables (columns 2 to 4). The coefficient on the trend is more pronounced for the adultequivalent income and the per capita income than the adult equivalent with economies of scale measure (results for the two latter are available under request). This may be evidence for households size changing over time. The number of males in the household is significant and positively related with household per capital income, however once we control for economies of scale, the effect is significant and negative as we see in table 8. This seems to be supporting evidence to use adjusted income measures rather than the raw per capita income. Other household size/dependency measures have the expected sign (number of child in the household and number of other family members in the household) as the education of the household head has. However, it is interesting to note that the magnitude of the coefficient is bigger for adult equivalent measure than for the rest of the measures. Also, the size of the coefficients goes down when we control for urban areas, sector and macro variables. Both the rate of unemployment (or number of unemployed over total size of the household) and the college rate in in the household (number of college graduates over total size of the household) have the expected signs and are very robust to any specification. However, it striking the magnitude of the unemployment rate that more than offsets the effect of the college rate. Turning into the more aggregate variables, the importance of import penetration is significant and positive. Lastly, when we introduce the formality variable in table 9 for the year for which is available we see that is almost as important the proportion of college graduates in the household as the proportion of formal employees in the household. The differences between regions is also striking. For instance, take Ciudad de Buenos Aires (the richest city) and Jujuy or Salta (two of the poorest cities) and the signs and magnitudes show very clearly the regional differences in incomes in the country. For instance, Comodoro, Rio Gallegos, Ciudad de Buenos Aires and Partidos are the only urban areas with positive coefficients in the income regressions. 15

16 2.2 Inequality regression For the inequality regressions (table 10) we use variation across and within provinces. We calculate two inequality indices (Gini and Theil indices) by urban area over 11 years (11*12=132 observations) to see how time, urban area, unemployment rates at the urban area level, trade and rates of return to college over secondary completion variables have affected inequality over time. In Table 10 we only report results for the Gini index as the results for the Theil index give similar outcomes. We see all the urban areas dummies are significant, except for San Luis; and that the year trend has a positive and significant coefficient, even when augmenting the regression with region and macro variables. Average unemployment time series at the urban area level seem to be very significant and positively correlated with increases in inequality; and this effect becomes even more important when we control for trade, and also when using the fixed effects estimation in column (4). Import penetration (as measured by imports over value added) is negatively correlated with the log of the Gini index, but this relationship becomes insignificant as soon as we include the rates of return to college in the equation and also with the fixed effects estimation. Rates of return to college completion (over secondary completion) are not significant, but are still positively related to the log of the Gini index for household adult-equivalent income. This makes sense as rates of return should be associated with labor income, but the relationship with household income is less direct. consider only 1995 onwards, we see how the average of formal workers in the economy is very significantly and negatively related to the Gini index and the Theil Index. 16

17 3 Modified Datt Ravallion decomposition of changes in poverty Recent theoretical advances have brought the interplay between growth, poverty and inequality back into a prominent position in development theories. Benabou provides an excellent survey of early contributions in this rapidly growing literature (Benabou 1996). However, in concerning the existing methods for quantifying this relationship, the literature is still rather new and scarce despite recently attracting increased attention (Datt and Ravallion 1991, Kakwani and Subbarao 1992, Kakwani 1993, Shorrocks 1999, Kolenikov and Shorrocks 2005) Unfortunately, the various existing inequality measures are not particularly useful if we want to investigate whether shifts in income distribution helped or hurt the poor during a period of overall economic contraction (expansion). It is certainly not possible to conclude that a reduction in inequality will reduce poverty using any measure which satisfies the Pigou- Dalton criterion. Moreover, even when a specific reduction (increase) in inequality does imply a reduction (increase) in poverty, the change in the inequality measure can be quite uninformative because it does not quantify the impact. One of the objectives of this paper is to provide a full picture of poverty changes and understand those changes in terms of inequality and mean income changes. Decomposition o poverty changes will then help to quantify the contribution of each component of interest. Some of the earlier decompositions procedures were proven to violate some intuitively natural axioms. Here, in order to make the decomposition more tractable and intuitive, we apply the axiomatic approach proposed by Kakwani (1997) 18 and compare it with earlier decompositions (i.e Datt Ravallion and the SOS or Shapley-Owen-Shorrocks decomposition procedure (Datt and Ravallion 1991, Shorrocks 1999)). Shorrocks, for instance, proposes a solution that considers the marginal effect on poverty of eliminating each of m contributory factors in sequence, and then assigns to each factor the average of its marginal contribution in all possible elimination sequences. This procedure yields an exact additive decomposition of P (the poverty function) into m contributions (see Appendix 5 for a full explanation of this procedure) without any requirement of linearity of the change in poverty function. Moreover, this is a symmetric decomposition and, in particular, the contributions will not depend on the order in which factors are considered. As stated in Datt and Ravallion ( An adding up decomposition procedure that introduces the idea of average growth and average inequality effects, the sum of which is equal to the total change in poverty. Datt and Ravallion rejected this solution as being arbitrary (Datt and Ravallion 1991, p.278 third footnote). Far from being arbitrary, the analysis suggests that this is exactly the outcome that results from applying a systematic decomposition procedure to the growth-inequality issue. 17

18 p )...if the mean income or the Lorenz curve remains unchanged over the decomposition period, then the residual (R) vanishes. The residual exists whenever the poverty measure is not additively decomposable between and L, i.e. whenever the marginal effects on the poverty rate of changes in the mean (Lorenz curve) depend on the precise Lorenz curve (mean). Separability of the poverty measure between the mean and Lorenz parameters is also required for the decomposition to be independent of the choice of the reference. The methodology explained below will then impose some requirements in the poverty change function, which in turn will become an additively decomposable function between any m components (such as the growth and the inequality component here). The only (though strong) assumption required will be that the poverty function is homogeneous of degree one. This is an artifact used to make the decomposition manageable empirically. We will discuss the shortcomings of this assumption by comparing the size of the residual and the way it could be attributed to each component (IN PROGRESS). In terms of previous empirical results, using data from Brazil and India, Datt and Ravallion found the growth component to explain the largest part of observed changes in poverty. Applying the same methodology to African, Latin-American, East and Southeast Asian countries, Demery et al (1995) came to the conclusion that first, poverty changes are largely determined by economic growth and second, that changes in inequality are of secondary importance to changes in poverty in the great majority of cases (Demery, Sen, and Vishwanath 1995). This paper will show that for Argentina this does not seem to hold, neither across groups nor by sub-periods under the period under study. It is still contentious whether this 20% increase in inequality (as measured by the Gini index) was due to a domestic Argentine phenomenon or a global economic trend that affected Argentina after 1990, when her major move away from a high-inflation economy with heavy state intervention and protection to a more open, private-sector oriented economy, took place. It will be verified that the inequality component dominates the growth component for the whole period. Accordingly, it may be worth analyzing first, the circumstances under which each component arises as the crucial one and, second, the type of decomposition which might be the more accurate in order to isolate the impacts of growth/inequality on poverty. Bourguignon, Ferreira et al (2001) found two chief limitations to this decomposition method. First, the analysis relies on summary measures of the distribution rather than the full distribution. Second, the decomposition of changes in poverty measures often leaves an unexplained residual of a nontrivial magnitude. Here, an attempt will be made to somehow work out both problems. The first problem will be tackled somehow by using the cumulative empirical distribution instead of a parameterized distribution function (i.e. instead of using a lognormal or Singh Maddala that bound the num- 18

19 ber of parameters of the function). Since the parameterization of a Lorenz curve will no longer be required to calculate the inequality component, this will make the decomposition much simpler and more general. The second point will be approached by applying to the Argentine data an axiomatic technique through which the Datt-Ravallion residual is not required to maintain the decomposition as in Shorrocks (1999) or Kolenikov and Shorrocks (2005) 3.1 The Datt Ravallion decomposition Datt and Ravallion treat the subject as follows. The poverty measure P at date (or region/country) t is written as P t (z, F t ). In fact the residual that may eventually appear is also a part of the inequality component that failed to be explained by the special form of the Lorenz curve chosen or because of the reference data chosen. This is explained by the fact that inequality in a distribution can change in infinite ways (and this depends partly on the reference date chosen). We now define the growth and inequality components: he growth component of a change in the poverty measure is defined as the change in poverty due to a change in the mean income while holding the Lorenz curve L t constant at some reference level L r : G(t, t + n; r) = P (z, µ t+n, L r ) P (z, µ t, L r ) (2) The inequality component is the change in poverty due to a change in the Lorenz curve L t while keeping the mean income constant at the reference level µ r : I(t, t + n; r) = P (z, µ r, L t+n ) P (z, µ r, L t ) (3) Therefore, a change in poverty between t and t + n may change due to a change in the mean income, µ t, or due to a change in the Lorenz curve, L t, and can be decomposed as follows 19 : P t,t+n = P t+n P t = P (G t,t+n (.), I t,t+n (.), R t,t+n (.)) (4) where this function may also be written: P t,t+n = P t+n P t = G(t, t + n, r) + I(t, t + n, r) + R(t, t + n, r) (5) In (7) the first two arguments in the parentheses refer to the initial and terminal dates of decomposition and the last argument makes explicit the 19 It would also be possible to explain the decomposition in an alternative way, using an elasticity-based presentation of this approach. For instance, the elasticity of the headcount ratio with respect to the growth component, G, would be G (µ/h), where µ is the mean income and H is the headcount ratio. 19

20 reference date, r, with respect to which the observed change in poverty is decomposed. In the next bus-sections we see how the inequality component can be measured in a more generalized way, and how to deal with the functional form of (7) in order to get rid of the residual. 3.2 The empirical cumulative density function approach and the inequality component The most commonly used measures of poverty are completely defined by a cumulative distribution function and a poverty line. Let y denote household adult-equivalent income and F t (y) denote the cumulative distribution function over the whole population which represents the proportion of persons with an income less than or equal to y at time t. The mean household adultequivalent income, µ t, will be used to measure the growth component and let z denote the poverty line (which will be the same for all individuals and will be deflated at 1998 pesos for all years). It will now be illustrated how the cumulative density function and the mean of a given distribution can be used to measure the inequality component. Bearing in mind that the Lorenz curve is a mean-normalized integral of the inverse of a distribution function and if L (p) denotes the slope of the Lorenz curve, and y is the income level which cuts off the bottom p percent, we can write (Kakwani 1980): L(p) = 1 µ F 1 (y)dy (6) And by deriving (2) we obtain: y = F 1 (p) = µl (p) (7) where L(p) is the equation of the Lorenz curve, giving the fraction of total income that the pth fraction of the population possess (for a population ordered in ascending order of income). Figure 5 shows the cumulative distribution function which may be used to estimate the poverty rate associated with any poverty line and illustrates a situation in which, given a fixed poverty line, growth raises all incomes by the same factor, leading to a simple rightward shift of the initial year distribution function from F t to F t. For instance, F t are the average incomes of periods t and t + n, respectively. By construction, F t and F t+n have the same mean. If one focuses, for instance, on the headcount ratio, then this shift reduces the headcount ratio from H t to H t. 20 If economic growth is instead accompanied by inequality-reducing redistribution, the eventual distribution F t+n will exhibit less spread than F t, producing a further reduction in the 20 ThusH t is the estimated poverty rate that would have been observed in the final period if the inequality had remained at the initial level, t. 20